Volumen 02

Seminario de Matem atica Financiera MEFF-UAM
Volumen 2. A nos 1998 y 1999
Directores: Santiago Carrillo Men endez ndez P Jos e Luis Ferna erez
c Santiago Carrillo Men endez, Jos e Luis Fern andez P erez y otros
Preparaci on de la edici on: Pablo Fern andez Gallardo Maquetaci on: Aula Documental de Investigaci on. Mart n de los Heros, 66. 28008. Madrid Edita: Instituto MEFF Imprime: JUMA
ISBN: 84-699-6766-5 Dep osito Legal M-52027-2001 Printed in Spain - Impreso en Espa na Este libro no podr a ser reproducido total o parcialmente, ni transmitirse por procedimientos electr onicos, mec anicos, magn eticos o por sistemas de almacenamiento y recuperaci on inform aticos o cualquier otro m etodo, ni su pr estamo, alquiler o cualquier otra forma de cesi on de uso del ejemplar, sin el permiso previo, por escrito, del titular o titulares del copyright.
Pr ologo
Este segundo volumen del Seminario de Matem atica Financiera MEFF-UAM es una recopilaci on de las conferencias que, durante los a nos 1998 y 1999, han dictado un selecto grupo de profesionales en la cita mensual que convocan los profesores Santiago Carrillo y Jos e Luis Fern andez. Pese a que el acto tiene que competir en el valioso tiempo de los asistentes con otras actividades m as l udicas que se celebran simult aneamente, el Seminario mantiene un numeroso p ublico. El prestigio de los ponentes y el inter es de los temas que plantean, fruto del trabajo cotidiano y de investigaci on, explican el exito de audiencia y tambi en el seguimiento de muchos eles adeptos en toda la geograf a espa nola quienes, gracias al uso de internet, tienen acceso inmediato a la ponencia escrita. Esta colaboraci on con la Universidad Aut onoma de Madrid se enmarca dentro de la pol tica de MEFF de promover la innovaci on en el sector nanciero de nuestro pa s y de impulsar el intercambio continuo de ideas entre los acad emicos y los profesionales de los mercados. Sin duda tanto unos como otros se enfrentan ahora m as que nunca a un mundo que obliga a poner en cuesti on la relevancia emp rica de los modelos te oricos y de las herramientas de gesti on predominantes en cada momento. Y creemos que este tipo de encuentros puede ser una ayuda para interpretar y abordar la compleja realidad. La publicaci on de un libro siempre es el resultado del esfuerzo que realiza un equipo de personas. Y esto es evidente en este ejemplar en el que han participado, adem as de los directores, los autores, el transcriptor y mis compa neros de Instituto MEFF. A m me corresponde la labor graticante de prologar el libro y de expresar en nombre de MEFF nuestro agradecimiento a todos los que lo han hecho posible y muy especialmente a los directores, Santiago Carrillo y Jos e Luis Fern andez, a los que profeso una gran admiraci on y afecto.
Remedios Romeo Directora General Instituto MEFF
Indice
A modo de introducci on . . . . . . . . . . . . . . . . . . . . . . . . . . . Cobertura en tiempo discreto de una cartera de opciones Rafael Salinas Mart nez de Lecea . . . . . ix 1 9
On the relevance of Modeling Volatility for Pricing Purposes . . . Manuel Moreno El uso de las mixturas de distribuciones gaussianas en nanzas: aplicaci on en valoraci on y cobertura de opciones nancieras . . . . Susana Corcuera y Juan Carlos Garc a C espedes La problem atica de medici on de la performance . . . . . . . . . . . . ndez Prosper Lamothe Ferna Equivalencia de los distintos m etodos para valorar empresas por descuento de ujos. Distintas alternativas para valorar el ahorro de impuestos debido al apalancamiento y sus implicaciones sobre la valoraci on . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ndez Pablo Ferna
41 81
92
Redes neuronales para aplicaciones en negocios . . . . . . . . . . . . 129 An bal R. Figueiras Vidal Volatilidad estoc astica, volatilidad incierta . . . . . . . . . . . . . . . 139 Antonio Sanchez Calle Short-Term Options with Stochastic Volatility: Estimation and Empirical Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 n y Gonzalo Rubio Gabriele Fiorentini, Angel Leo May static pricing models be useful when pricing catastrophelinked derivatives? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181 s, In aki R. Longarela y Julio Lucia Alejandro Balba Derivatives as Tradeable Assets . . . . . . . . . . . . . . . . . . . . . . 213 Terry J. Lyons Information Transmission around Block Trades on the Spanish Stock Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 Miguel Angel Mart nez, Mikel Tapia y J. Yzaguirre
vii
Market integration measures . . . . . . . . . . . . . . . . . . . . . . . . 255 s y Vicente Meneu Angel Pardo, Alejandro Balba Crash Prediction: Science or Alchemy? Common Points between Earthquakes, Sand Piles and Markets . . . . . . . . . . . . . . . . . . 281 Gabriele Susinno Una modelizaci on GARCH bivariante del futuro sobre el Ibex-35. Implicaciones para la cobertura del contado . . . . . . . . . . . . . . 291 Juan Angel Lafuente Risk Assessment and Extreme Value Theory in Finance . . . . . . 315 rez Alberto Sua Valoraci on y medici on de riesgo de pasivos bancarios sin vencimiento denido . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331 s M. Tarriba Unger Jesu An alisis cr tico del mercado el ectrico espa nol . . . . . . . . . . . . . 349 Rafael de Benito Uso de las martingalas en la valoraci on de activos derivados . . . . 375 Eloy Fontecha Modelos multivariantes de valoraci on de bonos convertibles con riesgo de cr edito . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 s P Jesu erez Colino
viii
A modo de introducci on
Este segundo volumen del Seminario MEFF-UAM de Matem atica Financiera saldr aa la luz coincidiendo con el inicio del sexto a no de actividad mensual. En el se recogen las diecinueve ponencias realizadas en dicho seminario a lo largo de los a nos 1998 y 1999. Una mirada al ndice de este volumen permite hacerse una idea de la diversidad de temas abordados en estas sesiones. El problema de la modelizaci on de la volatilidad ha sido quiz as el tema estrella si medimos su inter es por el n umero de conferencias que lo han abordado desde diversos a ngulos: Antonio S anchez, por una parte y Gabriele Fiorentini, Angel Le on y Gonzalo Rubio por otra, se han aproximado a los modelos de volatilidad estoc astica (tambi en a los modelos de volatilidad incierta en el caso del primer ponente). Manuel Moreno ha abordado el problema de la modelizaci on de la volatilidad en los modelos de evoluci on de tipos de inter es y su impacto en la valoraci on, mientras que Juan Angel Lafuente nos aproxima a los modelos GARCH bivariantes y sus aplicaciones. Se ha estudiado el uso de metodolog as de implantaci on reciente al entorno nanciero en varias de las ponencias: redes neuronales (An bal Figueiras), teor a de valores extremos (Alberto Su arez) y mixturas de normales (Susana Corcuera y Juan Carlos Garc a C espedes). Ha habido ocasi on de abordar temas tan diversos como la medici on de la performance de una cartera (Prosper Lamothe), la valoraci on de empresas (Pablo Fern andez), los derivados sobre cat astrofes (Alejandro Balb as, I naki Longarela y Julio Luc a), la situaci on del mercado el ectrico espa nol (Rafael de Benito), la cobertura en tiempo discreto de las carteras de opciones (Rafael Salinas) o el uso de modelos multivariantes en riesgo de cr edito (Jes us P erez Colino). Eloy Fontecha abord o el uso de los m etodos de martingalas en la valoraci on de activos derivados. Terry Lyons se interes o por la din amica conjunta del subyacente y de la call cuando esta es un valor negociado y de las consecuencias de este hecho para la cobertura. Gabriele Susinno nos aproxim o a la econof sica con una conferencia dedicada a la predicci on de crisis burs atiles. Tambi en han encontrado acomodo en las sesiones del seminario los interesantes trabajos de M.A. Mart nez, M. Tapia y J. Yzaguirre (transmisi on de la informaci on en las transacciones por bloques), de Angel Pardo, Alejandro Balb as y Vicente Meneu (medidas de integraci on de mercados) y de Jes us Tarriba (valoraci on y medici on de riesgos de pasivos bancarios sin vencimiento denido).
ix
Todo ello hace que este volumen sea una muestra patente de la interesante convergencia entre la investigaci on acad emica sobre nanzas y la pr actica nanciera. En total doce de estos trabajos tienen autores acad emicos y los autores de los siete restantes son practitioners, aunque resulte dif cil, para quien no los conozca personalmente, decidir qui en es qui en a la vista de los t tulos de las ponencias. Esta convergencia, se na de identidad del Seminario MEFFUAM, es consecuencia del talante abierto y de la dedicaci on de los ponentes y, tambi en, de los asistentes, a quienes dedicamos, con nuestro afecto y sincera gratitud, este volumen.
Santiago Carrillo Men endez Jos e Luis Fern andez P erez
Cobertura en tiempo discreto de una cartera de opciones

Rafael Salinas Mart nez de Lecea1
Resumen Bajo las hip otesis b asicas de Black & Scholes, se obtienen expresiones anal ticas del ratio de cobertura eciente, para una cartera de opciones europeas, que minimiza la varianza de los rendimientos de la cartera de cobertura cuando el ajuste de la misma se realiza de forma discreta cada t. Como ilustraci on se realizan diferentes simulaciones mediante un arbol binomial que nos permiten evaluar como se diferencian el ratio de cobertura tradicional (delta) del ratio de cobertura de m nima varianza para distintas sendas del activo subyacente.
Introducci on
Una de las primeras dudas que plantea la implementaci on pr actica de la cobertura en delta de un libro de opciones es la frecuencia optima de ajuste del libro. Los modelos te oricos de valoraci on de opciones tipo Black, Scholes & Merton est an basados en un ajuste continuo de la cartera de cobertura; sin embargo los costes de transacci on, los horarios de mercado y la liquidez de los mismos no permiten realizar dicho ajuste te orico. En este contexto el dilema del gestor se divide entre cubrir el libro con la mayor frecuencia posible, lo cual implica incorporar toda la volatilidad intrad a e interd a al coste de nuestra gesti on, o realizar una cobertura discreta que intente evitar los costes que introduce el ruido a corto plazo de la serie de precios, pero que obliga a incorporar una hip otesis sobre cu al va a ser la tasa de variaci on del activo subyacente durante el tiempo entre fechas de ajuste. Otra situaci on equivalente a la anterior es la que se plantea en las opciones a corto plazo de mercado organizado, durante la u ltima semana a vencimiento. En este contexto, y especialmente para opciones ATM y OTM, el simple paso de un d a afecta signicativamente los ratios de cobertura, lo cual obliga a tener que realizar continuos ajustes manuales en los programas de gesti on. Siguiendo a Robins & Schachter (1994) y Robins, Sanders & Schachter (1996), y bajo las hip otesis b asicas de Black & Scholes, a continuaci on se obtienen expresiones
1 Rafael Salinas Mart nez de Lecea es Director de Gesti on de Recursos Propios del Banco Bilbao Vizcaya Argentaria. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de enero de 1998.
Rafael Salinas
anal ticas del ratio de cobertura eciente, para una cartera de opciones europeas, que minimiza la varianza de los rendimientos de la cartera de cobertura cuando el ajuste de la misma se realiza de forma discreta cada t. Como ilustraci on se realizan diferentes simulaciones mediante un arbol binomial que nos permiten evaluar como se diferencian el ratio de cobertura tradicional (delta) del ratio de cobertura de m nima varianza para distintas sendas del activo subyacente.
El ratio de cobertura de m nima varianza
Consideremos una cartera de opciones vendidas con cobertura din amica con activo subyacente en el momento inicial t: Ht = jt w1t St Dt , (1)
donde Ht es el valor de la cartera en el momento t, St es el valor del activo subyacente, jt w1t es el ratio de cobertura (siendo w1t el delta de la opci on y jt un factor de ajuste respecto a dicho ratio de cobertura tradicional) y Dt es el valor de la opci on call con precio de ejercicio K y tiempo al vencimiento (T t). El rendimiento de esa cartera de cobertura durante un periodo de tiempo t ser a Ht jt w1t St Dt = , Ht jt w1t St Dt siendo la varianza de dichos rendimientos var Ht Ht =
2 2 w1t var(St ) + var(Dt ) 2 jt w1t cov (St , Dt ) jt . (jt w1t St Dt )2
(2)
(3)
Diferenciando con respecto a jt e igualando a cero obtenemos el ratio de cobertura que nos minimiza la varianza de los rendimientos: jt w1t = St var(Dt ) Dt cov (St , Dt ) . St cov (St , Dt ) Dt var(St ) (4)
Bajo los supuestos del entorno Black & Scholes, podemos obtener una expresi on anal tica para la ecuaci on (4), desarrollando expresiones para la cov (St , Dt ), la var(St ) y la var(Dt ). En primer lugar, la expresi on de la varianza de una variable aleatoria lognormal es
2 2 t var(St ) = St e e
2
1 ,
(5)
donde es la volatilidad del activo subyacente y su tasa de rendimiento esperado. En segundo lugar, sabemos que cov (St , Dt ) = E (St+t Dt+t ) E (St+t ) E (Dt+t ) . (6)
Aplicando un resultado de Rubinstein (1984), seg un el cual el valor esperado de una opci on europea al nal de un periodo t es igual al valor que dicha opci on tendr a hoy si el nivel del activo subyacente fuera S e t y el precio de ejercicio K er t (donde r es el tipo de inter es sin riesgo), esto es E (Dt+t ) = Dt (S e t , K er t , T t, r, ) , y conociendo los momentos de una variable aleatoria lognormal, se obtiene que E (St+t ) E (Dt+t ) = St et St et N (x) Ker(T tt) N (x T t) , (8) donde 1 S t e t 2 x= (T t) . ln + 2 K er (T tt) T t Por otra parte, se puede demostrar que
+
2 1 N (A + B z ) ez /2 dz = N 2
(7)
A 1 B2
(9)
donde N () es la funci on de distribuci on de la normal y A y B son constantes. Aplicando este resultado al primer sumando de la parte derecha de la ecuaci on (6) se obtiene E (St+t Dt+t ) = S t e t St e t+
2
K er (T tt) N De igual forma, y en tercer lugar,
t (x + T t t (x T t + T t N
2 2 var(Dt ) = E (Dt +t ) (E (Dt+t )) .
(10)
La expresi on anal tica de E (Dt+t ) ya ha sido indicada en (8). De forma similar a (9) y aplicando el siguiente resultado: N (A1 + B1 z1 ) N (A2 + B2 z2 ) f (z1 , z2 , 12 ) dz1 dz2 = N2 A1 1+
2 B1
A2
2 1 + B2
, 12
on donde N ( ) es la funci on de distribuci on normal univariante, f (z1 , z2 , ) es la funci de densidad normal bivariante, N2 ( ) es la funci on de distribuci on normal bivariante
Rafael Salinas
y A1 , A2 , B1 y B2 son constantes, se obtiene que

2 E (Dt +t ) =
S t e(
/2+) t
N2 y ; y ;
t T t
2 St e(
/2+) t
K er (T tt) N2 y
2
+ K er (T tt) donde g y = = 1 ln T t t g+ . T t
N2
t T t t g T t;g T t; T t
T t;g;
S t e t K er (T tt)
2 (T t) 2
Para calcular las expresiones anteriores solo se requiere conocer los par ametros que entran en la formula de Black & Scholes y la tasa de rendimiento esperado del activo subyacente .
Resultados emp ricos
Para una opci on call ATM, con un subyacente que cotiza a 50, un tasa anual de crecimiento del activo del 20.2%, volatilidad anualizada del 36.105% y tipo de inter es sin riesgo de 9.8%, Robins & Schachter obtienen los siguientes resultados: (a) El ratio de cobertura de m nima varianza es mayor que el ratio delta. En concreto para distintos vencimientos obtienen: Tiempo a vto. 63 d as 42 d as 21 d as Delta 0.5745 0.5609 0.5432 Ratio Cobertura M nima-Varianza con t 1 d a 3 d as 5 d as 0.5785 0.5905 0.5946 0.5660 0.5778 0.5890 0.5535 0.5738 0.5937
(b) Empleando una simulaci on de Monte Carlo, con mil series de precios, calculan el rendimiento total de la cobertura para distintos vencimientos y obtienen una medida del riesgo de la gesti on mediante la desviaci on est andar de dichos rendimientos. Los resultados que obtienen son los siguientes: Para un reajuste diario de la cartera la volatilidad anualizada de los rendimientos de la cartera de cobertura son:
Cobertura en tiempo discreto de una cartera de opciones Vencimiento 21 d as 42 d as 63 d as Cobertura Delta 9.45% 7.95% 9.84% Cobertura Minima-Varianza 8.52% 7.12% 8.80%
Si el reajuste se realiza cada 5 d as los resultados son Vencimiento 21 d as 42 d as 63 d as Cobertura Delta 15.04% 12.10% 13.78% Cobertura Minima-Varianza 11.38% 9.39% 11.43%
En t erminos de esta medida, ellos interpretan que la cobertura de m nima varianza tiene menos riesgo que una cobertura delta tradicional. Con datos reales de opciones americanas sobre el ndice S&P100, Robins, Sanders & Schachter realizan un an alisis similar al anterior. Con horquillas de precios de opciones y datos simult aneos del S&P100 para los primeros 42 d as h abiles de 1986, realizan tantas simulaciones como sean posibles vendiendo opciones en la mejor oferta, gestionando la cobertura en delta y en ratio de m nima varianza (calculado de forma num erica) para distintos t y recomprando posteriormente las opciones al precio de demanda. Los resultados obtenidos en dicho an alisis emp rico son similares a los de la simulaci on anterior.
Ejemplo num erico
Con objeto de comparar el ratio de cobertura de m nima varianza con el tradicional delta, se han generado, en un contexto de neutralidad ante el riesgo, a rboles binomiales para un activo subyacente cuya rentabilidad es del 2.5%, volatilidad anualizada del 25% y siendo los tipos de inter es del 4.5%. Para vencimientos de 1, 3, 6 meses y 1 a no se comparan el delta tradicional con el ratio de cobertura de m nima varianza, con distintos plazos de ajuste t a lo largo de las distintas ramas del a rbol binomial (ver cuadros anexos). Los resultados que se obtienen en este ejemplo coinciden con los resultados ya mostrados de Robins & Schachter. (a) Para un determinado vencimiento los ratios de cobertura de m nima varianza son mayores que el delta tradicional cuanto menor es la frecuencia de ajuste de la cartera. (b) Para un mismo vencimiento la diferencia entre ambos ratios es mucho m as sensible a la frecuencia del ajuste en opciones at y out of the money.
Rafael Salinas
Cuadro 1. Diferencia ratio m nima varianzadelta B&S (valor absoluto en puntos porcentuales). Vencimiento anual, tipo 4.5%, dividendo 2.5%, volatilidad 25%, ajuste cada 3 semanas.
Cuadro 2. Diferencia ratio m nima varianzadelta B&S (valor absoluto en puntos porcentuales). Vencimiento anual, tipo 4.5%, dividendo 2.5%, volatilidad 25%, ajuste mensual.
Resumen y conclusiones
En un contexto de cobertura de opciones en tiempo discreto, la cartera que incorpora opciones y su cobertura din amica con activo subyacente deja de ser una cartera sin riesgo, a diferencia de lo que ocurr a con la cartera de gesti on din amica en tiempo continuo.
Dado que el rendimiento esperado de dicha cartera es la rentabilidad del activo sin riesgo, parece intuitivo obtener un ratio de cobertura que minimice la varianza de los rendimientos de la cartera de cobertura. Robins & Schachter, dentro de los supuestos est andar de teor a de opciones, obtienen una expresi on anal tica para dicho ratio de cobertura de m nima varianza en el caso de una call europea y comprueban emp ricamente que dicha cobertura presenta una menor varianza, en los rendimientos, que la tradicional gesti on en delta. Por contra dicho ratio de cobertura requiere estimar cu al es la tasa esperada de rendimiento del activo subyacente entre fechas de ajuste de cartera. Un simple ejemplo generado a partir de un a rbol binomial nos permite comprobar las diferencias entre el ratio de cobertura delta y el de m nima varianza. Futuras l neas de an alisis a desarrollar en este campo pueden incluir: obtener resultados equivalentes para opciones put y comparar resultados en un libro completo de opciones, comprobar cu al de los dos m etodos de cobertura es m as robusto a estimaciones err oneas de la volatilidad y/o rentabilidad por dividendos; an alisis de las implicaciones que para la valoraci on del riesgo puede generar tener un libro de opciones cubierto con ratios de m nima varianza pero cuyo riesgo se mide en t erminos de delta, estimar y comparar la gamma de ambas estrategias de cobertura, etc.
Referencias
[1] Robins, R.P. & Schachter, B: An analysis of the risk in discretely rebalanced option hedges and delta-based techniques, Management Science 40 (1994), no. 6, pp. 798808. [2] Robins, R.P., Sanders, R.W. & Schachter, B: An empirical investigation of variance reduction throught non-delta-neutral hedging, The Journal of Derivatives 4 (1996), no. 2, pp. 5969. [3] Rubinstein, M: A simple formula for the expected rate of return of an option over a nite holding period, Journal of Finance 39 (1984), no. 5, pp. 15031509.
Rafael Salinas Mart nez de Lecea Gesti on de Recursos Propios Banco Bilbao Vizcaya Argentaria Paseo Castellana, 81, pl. 3 28046-Madrid, Espa na e-mail : salinas.martinez@grupobbva.com
On the Relevance of Modeling Volatility for Pricing Purposes

Manuel Moreno1
Abstract This paper presents a two-factor (Vasicek-CIR) model of the term structure of interest rates. We assume that default-free discount bond prices are determined by the time to maturity and two factors, the long-term interest rate and the spread. Under no-arbitrage conditions, a general bond pricing equation is derived and a closed-form expression for bond prices is obtained. Empirical evidence of the models performance in comparison with a double Vasicek model is presented. We conclude that the modeling of the volatility in the long rate process improves the tting and prediction of medium- and long-term rates. However, for the shortest maturities, it is not so clear which is the best model to use.
Introduction
The evolution over time of interest rates for default-free zero-coupon bonds is a topic that has been extensively analyzed in the nancial literature. Initially, the analysis of this evolution was performed by means of one-factor models which assume that movements in interest rates are driven by changes in the short-term (instantaneous) riskless interest rate (see, among others, Vasicek (1977), Cox et al (1985) or Chan et al (1992)). However, it is now widely accepted that interest rates are aected by more than one state variable. In this direction, several papers as Richard (1978), Brennan and Schwartz (1979), Schaefer and Schwartz (1984), Cox et al (1985), Longsta and Schwartz (1992), Due and Kan (1996), Chen (1996), Boudoukh et al (1999) and Dai and Singleton (2000), use multiple factors to explain the future movements that interest rates may show. There is substantial empirical evidence2 that shows that movements in interest rates can be decomposed in three types of basic changes related to the level of interest rates, the slope, and the curvature of the yield curve. As the curvature is usually the less important explanatory variable when dealing with spot interest rates,
1 Manuel Moreno Fuentes es Profesor Titular del Departament dEconom a i Empresa de la Universitat Pompeu Fabra y miembro del Centre de Recerca en Econom a Financera de dicha Universidad. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de febrero de 1998. 2 See, for instance, Jones (1991), Litterman and Scheinkman (1991), Zhang (1993) and Knez et al (1994).
10
Manuel Moreno
we can think that movements in spot interest rates may be reasonably well explained by the two rst factors. In fact, this is the motivation for the model previously presented and developed by Moreno (1996) which uses the long-term interest rate and the spread of interest rates as state variables (that is, the dierence between the short-term interest rate and the long-term interest rate is used as a rough measure of the slope of the yield curve). In that paper, both factors are assumed to follow a Vasicek process and, therefore, both variables (1) show mean reversion to a certain long term value and (2) their diusions reect a constant variance term. Under these assumptions, a general bond pricing equation was derived and a closed-form expression for zero-coupon bond and for interest rate derivatives prices was computed, This paper also presented the empirical performance of this model in relation to an alternative one-factor model. It can be argued that one of the assumptions made by Moreno (1996), namely, the constant variance in the diusion of the processes followed by both factors is too restrictive from an empirical point of view.3 This restrictive feature leads to the present paper whose main objective is to analyze if modeling the volatility improves the empirical performance of the Moreno (1996) model. Thus, the same state variables will be used although we will assume that the long-term interest rate does not follow a Vasicek process but a root-square (CIR-type) process. This alternative model will be denoted hereafter as the Vasicek-CIR model and it can be considered, from a theoretical point of view, as an special case of the Schaefer and Schwartz (1984) model. The schedule of this paper is as follows. Section 2 presents the main assumptions of the Vasicek-CIR model and provides the basic pricing equation that any derivative asset must satisfy. This equation, with the appropriate terminal condition, allows us to obtain the price of any asset that, at maturity, pays a certain payo as indicated in such terminal condition. In this section we (a) compute the analytical expression that indicate the price of any discount bond under the assumptions given by this two-factor model and (b) recall the analogous formula that was obtained by Moreno (1996) (Vasicek-Vasicek model hereafter). Section 3 analyzes the empirical behavior of both models by comparing the usefulness of these alternative formulas to t and forecast bond prices, that is, the in- and out-of-sample performance of such expressions. The data analyzed correspond to Spanish interest rates and bond prices for dierent maturities during the period 1991-1995. Finally, Section 4 summarizes and concludes.
3 Interest rate volatility is usually increasing in interest rate level although there is no consensus about the exact relationship between volatility and level. See Chan et al (1992), A t-Sahalia (1996), Conley et al (1997) and Stanton (1997) for this issue.
11
The Bond Pricing Equation
In this section we present the two-factor (Vasicek-CIR) model that we will use to price default-free discount bonds by deriving (and solving) the pricing equation which must be veried by the prices of these bonds. The main assumption of this model is that the price, at time t, of a default-free discount bond that pays $1 at maturity T depends only on the current values of two state variables and time to maturity, = T t. The main motivation for the factors to be used is the empirical evidence (see footnote 2) that changes in interest rates are a combination of movements in (a) the level of interest rates, (b) the slope and (c) the curvature of the yield curve, whose eect is usually negligible. Therefore, we can use the long-term rate and the spread as the variables that help us to explain the movements in the general level of interest rates and changes in the relationship between the short and the long end of the yield curve. With both variables, we can also try to explain the intermediate movements of the yield curve.4 Although most previous studies use the short-term interest rate as one of the state variable, we redene these variables and, analogously to Schaefer and Schwartz (1984), we use the long-term rate, denoted by L, and the spread, denoted by s, the dierence between the short-term rate, denoted by r, and the long-term rate. This selection of state variables allows us to use the assumption of orthogonality between them.5 Once chosen these variables, we assume that their evolution over time is given by the following stochastic dierential equations6 : ds dL = 1 (s, L) dt + 1 (s, L) dw1 , = 2 (s, L) dt + 2 (s, L) dw2 , (1)
where t denotes calendar time, and dw1 and dw2 are standard Brownian processes 2 2 where E [dw1 ] = E [dw2 ] = 0, dw1 = dw2 = dt, and (by the orthogonality assumption) it is veried that E [dw1 dw2 ] = 0. 1 () and 2 () are the expected instantaneous rates 2 2 of change in the state variables and 1 () and 2 () are the instantaneous variances of changes in these factors. Let P (s, L, t, T ) P (s, L, ) be the price, at time t, of a default-free discount bond that pays $1 at maturity T = t + . We can express the instantaneous percentage
4 Two alternative couples of factors to be used may be: (a) the long-term interest rate and the short-term interest rate and (b) the short-term interest rate and the spread. However, the above two variables are chosen because of a better analytical tractability. 5 This assumption simplies the analytical tractability of the model. Empirical evidence that supports this assumption can been seen in Ayres and Barry (1980), Schaefer (1980), Nelson and Schaefer (1983) and, for the Spanish case, in Moreno (1996). 6 After presenting this generic model and deriving the general pricing equation, we will particularize it to obtain the Vasicek-CIR model
12
Manuel Moreno
change in the price of this bond as the sum of its expected rate of return and the unexpected variations in return due to the random changes in the factors dP (s, L, t, T ) = (s, L, t, T ) dt + s1 (s, L, t, T ) dw1 + s2 (s, L, t, T ) dw2 . P (s, L, t, T ) (2)
The steps to be given to obtain the bond pricing equation are very standard7 and can be summarized as follows: 1. Application of It os Lemma. 2. Setting up of a (hedging) portfolio, composed of three bonds with dierent maturities, that is instantaneously riskless. 3. Under no-arbitrage conditions, the expected rate of return of this portfolio must equate the instantaneous riskless rate of interest. These three steps jointly with a little algebra lead us to the following partial dierential equation: 1 2 2 [ () Pss + 2 () PLL ] + [1 () 1 () 1 ()] Ps 2 1 [2 () 2 () 2 ()] PL + Pt r P + = 0, (3)
where subscripts denote partial derivatives. The coecients 1 () and 2 () can be interpreted as the market prices of the spread and long-term rate risk, respectively. Therefore, given the stochastic process (1) we have assumed for both variables, (3) is the fundamental equation for the pricing of default-free discount bonds of dierent maturities which depend solely on the spread, the long-term interest rate, and its time to maturity. In this equation we deal with the market prices of risk, i (), because the only way to tie down the bond prices in our (partial equilibrium) model is by means of these (exogenous) parameters. The solution of the equation (3), subject to the terminal condition given by the nal payment of the bond, P (s, L, 0) = 1, s, L, is the price of the discount bond we are looking for. The coecients of the bond pricing equation (3) are the parameters of the stochastic process (1) which was assumed for the two factors and the market prices of the risk related to both state variables. As this equation is too general to be solved analytically, we will make the following assumptions about these coecients: Assumption 1 The market price of the spread risk is linear in this variable, that is 1 () = a + bs .
7 For
more details, see Moreno (1996).
13
Assumption 2 The market price of the long-term rate risk is proportional to the square root of this variable, that is 2 () = d L . Assumption 3 Each of the state variables follow a diusion process ds dL = k1 (1 s) dt + 1 dw1 , = k2 (2 L) dt + 2 L dw2 .
(4)
The motivation for the rst two assumptions is that a constant market price of risk is too restrictive and quite unrealistic. The rst assumption generalizes of the one presented in Vasicek (1977) while the second one is similar to the one obtained in Cox et al (1985). Regarding the third assumption, the rst process, known as OrnsteinUhlenbeck process, has been used previously by Vasicek (1977) while the second one was proposed by Cox et al (1985). Both processes show mean reversion, an important stylized fact that interest rates usually show. In the process assumed for the spread, we nd a constant variance in the diusion term while the variance of the long-term rate is proportional to its level. For each state variable, ki > 0 is the coecient of mean reversion which reects the speed of adjustment of the variable towards its long-run mean value, i , and dwi are standard Brownian motions. Under these three assumptions, we can rewrite the equation (3) as 1 2 1 2 1 Pss + q1 ( 1 s) Ps + 2 L PLL + q2 ( 2 L) PL + Pt (L + s) P = 0 , (5) 2 2 subject to the terminal condition P (s, L, 0) = 1, where q1 q2 = k1 + b 1 , = k2 + d 2 , 1 = (k1 1 a 1 )/q1 , 2 = k2 2 /q2 . s, L , (6)
Solving the partial dierential equation (5) we obtain the following proposition: Proposition 1 The value at time t of a discount bond that pays $1 at time T , P (s, L, t, T ) P (s, L, ), is given by P (s, L, t, T ) = A( ) eB ( )sC ( )L , (7)
14 where = T t and A( ) = A1 ( ) = A1 ( )A2 ( ) , 2 exp 1 B 2 ( ) + s (B ( ) ) 4q1 2 exp (q2 + ) 2 (q2 + ) exp{ } + ( q2 ) 1 eq1 , q1 2(exp{ } 1) , (q2 + ) exp{ } + ( q2 )
Manuel Moreno
, , (8)
2 2k2 2 /2
A2 ( ) = B ( ) = C ( ) = with q1 = k1 + b1 , q2 = k2 + d2 , Proof: See Appendix.
2 2 1 = (k1 1 a1 )/q1 , s = 1 1 /(2q1 ),
2 = k2 2 /q2 ,
2 + 2 2 . q2 2
(9)
The terms in equation (8) verify 0 < Ai ( ) < 1, > 0, 0 < B ( ) < , > 0, 0 < C ( ) < , > 0, Ai (0) = 1, Ai () = 0, i = 1, 2 B (0) = 0, B () = 1/q1 , C (0) = 0, C () = 2/(q2 + ) .
(10)
Substituting t = T into (7), it is shown that the terminal condition for the price bond, P (s, L, 0) = 1, s, L, is satised. Moreover, it is also derived that P (0, 0, ) = A( ) = A1 ( )A2 ( ) < 1, > 0 . It can be checked that the following realistic features are veried
s
lim P (s, L, ) = lim P (s, L, ) = lim P (s, L, ) = 0 ,

L
that is, when any of the arguments included in the bond price formula tends to innity, the price converges to zero. It is also easily shown that the bond price function is decreasing and convex in both factors and decreasing with the time to maturity. Once we have obtained the expression (and properties) for the bond price formula under the Vasicek-CIR model, we will recall the assumptions made by Moreno (1996) and the corresponding pricing formula that was derived in that paper: Assumption 1 (equal to Assumption 1) The market price of the spread risk is linear in this variable, that is 1 () = a + b s .
15
Assumption 2 The market price of the long-term rate risk is linear in this variable, that is 3 () = c + d L . Assumption 3 Each of the state variables follow a diusion process of Vasicek type ds dL = k1 (1 s)dt + 1 dw1 = k3 (3 L)dt + 3 dw3 (11)
Under these assumptions, we can rewrite the equation (3) as 1 2 1 2 1 Pss + q1 ( 1 s)Ps + 3 PLL + q3 ( 3 L)PL + Pt (L + s)P = 0 , 2 2 where q1 q3 = k1 + b1 , = k3 + d3 , 1 = (k1 1 a1 )/q1 , 3 = (k3 3 c3 )/q3 . (12)
The solution of the dierential equation (12), subject to the terminal condition given by the payo of the bond at maturity (see equation (6)), was established in the following proposition: Proposition 2 (Proposition 1 in Moreno (1996)) The value at time t of a discount bond that pays $1 at time T , P (s, L, t, T ) P (s, L, ), is given by P (s, L, ) = D( ) eE ( )sF ( )L , where = T t and = D1 ( )D3 ( ) , 2 D1 ( ) = exp 1 E 2 ( ) + s (E ( ) ) , 4q1 2 D3 ( ) = exp 3 F 2 ( ) + L (F ( ) ) , 4q3 q1 1e E ( ) = , q1 1 eq3 F ( ) = , q3 with q1 = k1 + b1 , q3 = k3 + d3 ,
2 2 1 = (k1 1 a1 )/q1 , s = 1 1 /(2q1 ),
(13)
D( )
(14)
3 = (k3 3 c3 )/q3 ,
2 2 L = 3 3 /(2q3 ).
(15)
Proof: It is similar to the proof of Proposition 1 and it is omitted for the sake of brevity.
16
Manuel Moreno
Empirical Application
In this section, we describe the empirical application in which we compare the tting and forecasting behavior of the Vasicek-CIR and the Vasicek-Vasicek models. This comparison is performed analyzing the in- and out-of-sample properties of both models. The dataset consists of daily Spanish interest rates and zero-coupon bond prices and cover the period 1991-1995.8 For each day of this period, we have interest rates (in annualized form) and bond prices for ten dierent maturities: 1, 7, and 15 days, 1, 3, and 6 months, and 1, 3, 5, and 10 years. The interest rates corresponding to the shortest and longest maturity (1 day/10 years) are used as proxies of the short- and long-term interest rate, respectively. The main descriptive characteristics of the state variables used in both two-factor models are: 1. For both interest rate series, the unconditional average is larger than 10%. Short-term interest rates are larger than this mean value until October 1993 while the long-term interest rates exceed this level in the whole period except from June 1993 through June 1994. On the other hand, the spread has a mean value very close to zero and ranges between 4% and 8%. 2. The short-term rate is more volatile and moves into a wider interval than longterm rate does. 3. Both state variables show an uniformly high degree of serial correlation. 4. Most of the changes in short-term interest rates are smaller than 100 basis points while changes in long-term rates are much smoother. As a consequence, changes in the spread are quite similar to changes in short-term interest rates. 5. It is seen a small decrease - in mean - in interest rates through the sample period. 6. Evidence of mean reversion in spread and interest rates is derived. 7. The theoretical assumption about the orthogonality between the state variables is empirically corroborated. Next we present the empirical performance of both models. We recall that both models use the same factors and the dierences between them derive from the alternative processes assumed for the long-term rate, L. Each state variable of the two competing models, s and L, follows a diusion process (see equations (4) and (11)). The diusion parameters of these processes
8 For more details on these data, see N un ez (1995) for technical details on the estimation procedure used by the Bank of Spain, and Moreno (1996) for a descriptive and graphical analysis.
17
(ki , i , i , i = 1, 2, 3) are estimated by the Generalized Method of Moments presented in Hansen (1982)9 . The econometric specication in discrete time is st st1 = a1 + b1 st1 + s t, Lt Lt1 = a2 + b2 Lt1 + L t , Lt Lt1 = a3 + b3 Lt1 + L t , so that ki = bi , i = ai , bi
2 s t IID (0, 1 ) 2 L t IID (0, 2 Lt1 ) 2 L t IID (0, 3 )
i = 1, 2 , 3 .
Table I includes the estimation results obtained for the sample period 1991-1995 and shows that the parameters bi of the discrete time specication (and hence, the diusion parameters ki ) are signicantly dierent from zero. So, there is evidence of mean reversion in all the state variables. In both models, the long-term interest rate tends to a mean value close to 10% while the spread tends to a mean value close to zero. Comparing the two processes assumed for the long-term rate, it may be interesting to recognize that, under the CIR model, the long rate reverts faster to its long-term value than when considering the Vasicek assumption. After estimating the parameters of the diusion processes followed by the factors in both models, these values are used to obtain the remaining parameters of equations (7) and (13). Thus, similarly to Moreno (1996), we use the specications P = P (q1 , q2 , s |k1 , k2 , 1 , 2 , 1 , 2 ; s, L, ) + , P = P (q1 , q3 , s , L |k1 , k3 , 1 , 3 , 1 , 3 ; s, L, ) + , (16)
where P is the observed price of the discount bonds available at time t, P () is the closed-form pricing equation for each model (see equations (7) and (13)) and is an error term. The parameters of the equations (16) (qi , i = 1, 2, 3, s , L ) are estimated on a daily basis for the period 1991-1995 by means of a panel of data where we have daily yield curves containing a cross-section of discount bond prices. Therefore, we have a matrix with 1230 rows and 10 columns where each row includes the (ten) zerocoupon bond prices available at each day and each column contains the bond prices for a certain maturity. We estimate the non-linear equations (16) for each day of the period 1991-1995. The estimation of the rst equation provides the parameters of the Vasicek-CIR model (that is, q1 , q2 , s ) while the estimation procedure, when applied to the second equation, provides the parameters of the Vasicek-Vasicek model, that is, q1 , q3 , s , L .
9 For details on this technique and its applications in the estimation of continuous-time models, see Moreno and Pe na (1996).
18
Manuel Moreno
Estimation results for the daily parameters of the Vasicek-CIR model are included in Table II.10 This table shows the average of the estimated parameters obtained for (a) the full sample period and (b) the sample period divided year by year and reects that all the parameters are positive and highly signicant. The evolution over time of these parameters can be seen graphically in Figure 1. This gure shows that the highest values are attained in 1991 while the lowest (and more stable) parameters correspond to the period 1994-1995. In the next step, we can compute the values, day by day, of the market prices of risk related to each state variable using the estimated parameters obtained from equation (16) jointly with the expressions (9) and (15) and the Assumptions 1 and 2. A graphical representation of these values, for the Vasicek-CIR model, can be seen in Figure 2. The average values of these prices - under both models - for the whole period and for every year, are included in Table III. Analyzing the two factors of the Vasicek-CIR model, both market prices of risk are highly signicant and have a similar behavior across the period 1991-1995: each price has always the same sign during all the period 1991-1995 (for the long-term rate, the market price of risk is always positive and the risk related to the spread has always a negative price) and both series of market prices are specially low in 1992 and the second half of the period 1991-1995. In absolute terms, the price of the spread risk is, at every moment, much higher than the price of risk of the long-term rate. On the other hand, for the Vasicek-Vasicek model, we can observe a very dierent behavior between both factors and with respect to the alternative model. Thus, the two basic features for the prices of risk in this model are: 1. For the full period, the market prices of risk for both state variables are positive and signicantly dierent from zero. 2. Analyzing this period on a yearly basis, the parameters are also signicantly dierent from zero but they show a changing sign. Thus, the mean market price of risk of the spread is negative in the last two years of the sample period while the average of the market price of risk related to the long-term rate is negative in 1991-1992 and 1994. Finally, we will use the values of the diusion parameters jointly with the parameters estimated in the equation (16) to analyze the tting and forecasting power of both two-factor models. The within- and out-of-sample periods are 1991-1994 and 1995, respectively. First, the in-sample estimated data, for each day of the period 1991-1994 and for both models, are provided by the inclusion of the (daily) estimated parameters and the estimated parameters of the diusion processes in the non-linear equation (16).
10 We do not show the results for the Vasicek-Vasicek model that can be seen in Table VIII in Moreno (1996). That paper provided all the results for the Vasicek-Vasicek model that are included in the following tables in this paper.
19
Next, we will compare the out-of-sample properties of both models by using the k -step-ahead forecasts that are generated for the bond prices. These t + k -time forecast values are built using the coecients estimated from time t. This procedure is repeated for each day of 1995. After obtaining the in- and out-of-sample forecasts, the (within and out-of-sample) pricing errors of both models are computed to compare each other. Thus, we dene, for time t, the error, et , and the percentage error, P Et , as t , et = P t P P Et = t Pt P 100 , Pt
t are, respectively, the observed and the estimated (tted or forecasted) where Pt and P price, for time t, of the zero-coupon bond with a certain maturity. For both models, the within-sample (absolute and percentage) pricing errors are shown in Figures 3 and 4. For all the maturities, it can be seen that the VasicekVasicek model provides a very large pricing error in May, 1993. This error coincides with a sharp change in the short-term rate and in the spread. For both models, neither gure suggests a systematic pattern in these pricing errors. Denoting by N the number of days of the period to be analyzed, we use the pricing errors to compute several accuracy measures that help us to compare the empirical performance of both models: 1. Mean Error (ME). This measure weights equally the daily errors. Therefore, positive values can be oset with negative values and, thus, this measure may be small even with large errors. Its expression is ME = 1 N
N
et =
t=1
1 N
t ) . (Pt P
t=1
2. Mean Absolute Error (MAE). As the mean error, this measure gives an equal weight to the daily errors but positive and negative errors do not cancel out. It is dened as M AE = 1 N
N
|et | =
t=1
1 N
t | . |P t P
t=1
3. Root Mean Squared Error (RMSE). It is usually the most common measure of accuracy and its denition is 1 N
N
RM SE =
(et
t=1
)2
1 N
t )2 . (Pt P
t=1
20
Manuel Moreno
4. Mean Percentage Absolute Error (MAPE) Similarly to the mean absolute error, the absolute value of the error is used but each error is weighted by the current value of the bond price. Its expression is M AP E = 1 N
N
|P E t | .
t=1
5. Root Mean Squared Percentage Error (RMSPE). This measure is similar to the root mean squared error but, similarly to the MAPE, the daily errors are weighted by the actual bond prices. It is given by 1 N
N
RM SP E =
(P Et )2 .
t=1
These ve descriptive measures, for both models, are computed for 1991-1994 (within sample period), for 1995 (out-of-sample period) and for dierent subperiods. The within and out-of-sample results are reported in Tables IV-VI and Tables VII-X, respectively. The performance of both models, for the within-sample period, is included in Table IV. For this period, the Vasicek-Vasicek model underprices the short- and mediumterm bonds and overprices the bonds whose maturity is beyond one year. On the other hand, the Vasicek-CIR model overprices the bonds with maturities up to three months as well as the 5-year bonds. All the statistics included in Table IV reect that both models provide a very big accuracy to the observed bond prices for all the maturities. It can be seen that the pricing errors are increasing with the maturities (the longer the maturity, the larger the error price) but the MAPE for the Vasicek-Vasicek and the Vasicek-CIR models is smaller than 0.26% and 0.02%, respectively. Moreover, in the Vasicek-CIR (Vasicek-Vasicek) model, this statistic is always smaller than 0.01% (0.08%) except for the 5-year bond price. Comparing both models, the Vasicek-CIR model outperforms the Vasicek-Vasicek model for all maturities and for all the statistics. In short-term bonds, with maturities smaller than one month, both models provide negligible errors and the relative improvement obtained with the Vasicek-CIR model is not very large. On the other hand, focusing on the maturities beyond one month, the VasicekCIR model provides a huge improvement: the errors from the Vasicek-Vasicek model are decreased in more than 88% for all these maturities. The biggest improvement in accuracy is achieved in the medium-term maturities (six month and one year) and in 5-year bonds, maturity in which the error measures from the Vasicek-CIR model are about 6% of the error measures provided by the Vasicek-Vasicek model. This conclusion is obtained for all the statistics.
21
Table V includes the results obtained for the year 1992. Similarly to the period 1991-1994, the Vasicek-Vasicek model underprices the shortest maturities (up to six months). In contrast to what happened in the whole within-sample period, this underpricing can also be seen in the Vasicek-CIR model which, on the other hand, overprices the bonds that mature beyond one year. In this year, both models t the observed data specially well. The error measures for both models are decreased in more than half for most of the maturities with respect to the whole period. Comparing both models, the error measures of the Vasicek-CIR model, as in the within-sample period, are about 10% of the statistics provided by the Vasicek-Vasicek model for all the maturities longer than 15 days. Therefore, the main conclusion for this year is the same than the obtained for the period 1991-1994: for the shortest maturities, both models t specially well to the data but, for most of the remaining maturities, the Vasicek-CIR model provides a remarkable large improvement in accuracy. Several subperiods have been analyzed and, basically, the same conclusions are reached. For illustrative purposes, Table VI includes the within-sample results obtained for 1-year bonds for each semester of the period 1991-1994. Looking at every statistic, it is shown that the Vasicek-CIR model ts better than its competing model in all the semesters and it works specially well in 1991 and 1994 while the Vasicek-Vasicek model obtains its best performance in the rst semester of 1992 and in the second one of 1994. Based on a mean absolute error (MAE or MAPE) criterion, the superiority of the Vasicek-CIR model implies an improvement of about 90% in all the semesters but the last one in which the errors from the Vasicek-Vasicek model are decreased in just 77%. For all the semesters, the mean absolute percentage error of the Vasicek-CIR model is around 0.003% that is about fteen times smaller than the obtained for the VasicekVasicek model. This superiority is specially remarkable in the second semester of 1991 and in the rst one of 1994 when the absolute value of the errors are decreased in more than 96%. The forecasting power of both models is analyzed by computing one- and ve-step ahead forecasts11 of bond prices for every maturity and for every day of the year 1995. Measures of the forecasting pricing errors are included in Tables VII-X. Tables VII-VIII include the measures for one-step-ahead forecasts. Table VII provides the one-step-ahead measures for the whole out-of-sample period. It can be seen that (1) both models forecast quite well, (2) the forecasting power decreases with the time to maturity, and (3) for both models, the MAPE (RMSPE) is always smaller than 0.36% (0.48%). It can also be seen that both models perform similarly for the
11 Ten-step-ahead
forecasts were also computed. Results are available upon request.
22
Manuel Moreno
shortest maturities and, in fact, the Vasicek-Vasicek model performs slightly better than the Vasicek-CIR model. For maturities beyond one month, the Vasicek-CIR model forecasts better than the Vasicek-Vasicek model showing that the modeling of the volatility in the long-term rate process helps to predict the movements in the medium- and long-term interest rates. This superior forecasting performance is not monotonic in the time to maturity: the largest improvement is obtained in the 1-year bond prices when the error measures from the Vasicek-Vasicek model are decreased in more than 21% (15%) when working on a (root) mean absolute criterion. In the remaining bonds, the improvement in the relative forecasting power is much smaller (2 7%) and never exceeds 12%, value that is obtained when forecasting the 5-year bonds. Table VIII provides the error measures obtained, for every month of 1995, from 1-year bonds in which the Vasicek-CIR model achieves its best relative performance. Both models perform better in the second semester (with a MAPE (RMSPE) smaller than 0.06% (0.08%)) than in the rst one, when the MAPE (RMSPE) reaches 0.1% (0.12%). It can also be seen a similar behavior in both models from January to April (with a small superiority of the Vasicek-Vasicek model in this period) and in the two last months of 1995. In the period May-October, the Vasicek-CIR model leads to an improvement of the out-of-sample performance that ranges between 20% (in May) and 65% (in July). Finally, Tables IX-X show the results obtained when the forecasting horizon is ve days. Table IX includes the measures obtained with the out-of-sample errors for all the maturities in 1995. Although these measures are bigger than in the previous forecasts, they are reasonably small as reected in the MAPE or the RMSPE that are, for both models, smaller than 0.8% and 1%, respectively. Results are analogous to the obtained with the previous predictions: (1) the forecasting power decreases with time to maturity, (2) the Vasicek-Vasicek model outperforms its rival model in the maturities smaller than six months, and (3) the VasicekCIR forecasts better than the Vasicek-Vasicek model in the maturities beyond six months. However, this improvement is usually quite small (between 1% and 3%) and only increases until 5% when forecasting 1-year bond prices. Finally, Table X provides the quarterly results obtained with ve-step-ahead forecasts for 1- and 5-year bonds. In both maturities, it can be seen a better forecasting behavior in the second half of 1995 than in the rst one. For 1-year bonds, both models show a MAPE smaller than 0.2% in all the quarters and the Vasicek-CIR model outperforms the Vasicek-Vasicek model whose error measures are decreased between 5% and 15% in the period April-September. Focusing on the forecasting errors for 5-year bond prices, the MAPE statistic ranges between 0.6% and 1%. In this case, the Vasicek-CIR model improves the fore-
23
casts from its competing model in 10% from July to September and, in the remaining subperiods, its improvement is much smaller (1 2%).
Conclusions
This paper has presented a two-factor (Vasicek-CIR) model in continuous time for the analysis of the terms structure of interest rates and its empirical behavior has been analyzed with respect to a second alternative model. The main common characteristic of these two models is that both employ the same factors (state variables) to explain the unexpected changes that interest rates may show in the future. These factors are the long-term interest rate and the spread, the dierence between the short- and the long-term interest rate. The Vasicek-CIR model assumes that the spread follows a Vasicek process while the long-rate is modeled as a CIR-type process. On the other hand, the VasicekVasicek model has assumed that both variables follow a Vasicek process. This second model was previously presented by Moreno (1996) which also developed its pricing properties, the implications on the term structure of interest rates and analyzed its empirical properties with a sample of daily interest rates and bond prices during the period 1991-1995. The main objective of this paper is analyze and compare the empirical performance of both models in the period 1991-1995. Therefore, we can determine if modeling the volatility of the long-term interest rate may help us to explain the future movements of interest rates. As a starting point, we have derived a bond pricing equation whose solution indicates the price of a zero-coupon bond under certain assumptions, namely, this price depends solely on the current values of two state variables (mentioned above) and the time to maturity of the bond. After this solution is obtained, it is used to analyze the tting and forecasting properties of this model. These properties are, in a posterior stage, compared with the ones derived from the Vasicek-Vasicek model. The parameters of our competing models have been estimated in two steps. In the rst one, the parameters of the diusion processes have been estimated by the Generalized Method of Moments by Hansen (1982). Once these values are obtained, the remaining parameters are estimated by using a cross-section technique. As a result of this combination of estimation methods, we have been able to obtain the daily market prices of risk corresponding to the state variables for both models. It has been shown that, for the Vasicek-CIR model, these market prices of risk are highly signicant and have a constant sign during all the period 1991-1995. On the other hand, under the Vasicek-Vasicek model, it can be seen that these prices are signicantly dierent from zero and positive for 1991-1995 although they show a changing sign when analyzing on a yearly basis this period.
24
Manuel Moreno
Finally, we have analyzed the tting and forecasting power of both models. The within- and out-of-sample periods are 1991-1994 and 1995, respectively. After computing the within- and out-of-sample forecasts, the pricing errors of both models (and several accuracy measures) for dierent subperiods have been obtained to compare one each other. All these statistics show the following facts: (1) both models provide a very big accuracy to the observed bond prices for all the maturities, (2) the pricing errors are increasing with the maturities, (3) the MAPE for the Vasicek-Vasicek and the Vasicek-CIR models is smaller than 0.26% and 0.02%, and (4) in the Vasicek-CIR (Vasicek-Vasicek) model, this statistic is always smaller than 0.01% (0.08%) except for the 5-year bond price. Comparing both models, it can be seen that (1) the Vasicek-CIR model outperforms the Vasicek-Vasicek model for all maturities and for all the statistics, (2) in maturities smaller than one month, both models provide negligible errors and the relative improvement obtained with the Vasicek-CIR model is not very large, (3) dealing with maturities greater than one month, the errors from the Vasicek-Vasicek model are decreased in more than 88% for all these maturities, (4) the biggest improvement (about 90 94%) in accuracy is achieved in the medium-term maturities and in 5-year bonds. Several subperiods have been analyzed and the same conclusions are reached. The forecasting power of both models has been analyzed by computing one- and ve-step ahead forecasts for every maturity and for every day of the year 1995. Looking at one-step-ahead forecasts, it has been shown that - similarly to the within-sample period - both models perform quite well, the forecasting power decreases with the time to maturity and, for the shortest maturities, both models perform in a similar way. However, for maturities beyond one month, the Vasicek-CIR model forecasts better than the Vasicek-Vasicek model although this superior forecasting behavior is not monotonic in the time to maturity. Best relative performance is obtained in the 1-year bond prices when the error measures from the Vasicek-Vasicek model are decreased in more than 21%. In the remaining bonds, the improvement is much smaller ranging between 2% and 12%. Finally, dealing with ve-step-ahead forecasts, all the statistics reect a worse performance than in the previous (shorter) forecasts although similar results are shown: (1) the forecasting power decreases with time to maturity, (2) the Vasicek-Vasicek model outperforms its rival model in the shortest maturities and (3) the Vasicek-CIR forecasts better than the Vasicek-Vasicek model in the maturities beyond six months. In this case, this improvement is usually quite small, between 1% and 5%. Therefore, the main conclusion is that the modeling of the volatility in the longterm rate process can help (in a large amount) to t the observed data and can improve - in a reasonable quantity - the prediction of future movements in the medium- and long-term interest rates. However, for shorter maturities, it has been shown that the pricing errors are, basically, negligible and it is not so clear which is the best model to be used.
25
Appendix: Proofs
Proof of Proposition 1 The method of the separation of variables allows us to write the solution of the equation (5) subject to (6) as P (s, L, t, T ) = X (s, t, T ) Z (L, t, T ) , where X (s, t, T ) solves the equation 1 2 Xss + q1 ( 1 s)Xs + Xt sX = 0 2 1 subject to the terminal condition X (s, T, T ) = 1, s , and Z (L, t, T ) is the solution of the equation 1 2 LZLL + q2 ( 2 L)ZL + Zt LZ = 0 , 2 2 with terminal condition Z (L, T, T ) = 1, L . To solve equation (18), we posit a solution of the type X (s, t, T ) = X (s, ) = A1 ( )eB ( )s . Hence, the equation (18) becomes A1 ( ) 1 2 2 B ( ) q1 ( B ( )s s = 0 , 1 s)B ( ) 2 1 A1 ( ) where, from (19), the terminal conditions are given by A1 (0) = 1, B (0) = 0 . (24) (23) (22) (21) (20) (19) (18) (17)
Equation (23) is linear in the variable s and, therefore, it becomes null when the corresponding coecients are equal to zero. Hence, this equation is equivalent to the following system of rst-order dierential equations q1 B ( ) + B ( ) 1 = 0 A ( ) 1 2 2 1 B ( ) q1 = 0, 1 B ( ) 1 2 A1 ( ) subject to the terminal conditions (24). (25) (26)
26
Manuel Moreno
We rst solve (25) with terminal condition B (0) = 0. Including this solution in (26), integrating this equation, and using A1 (0) = 1, we obtain B ( ) A1 ( ) where 1 eq1 , q1 2 = exp 1 B 2 ( ) + s (B ( ) ) 4q1 =
2 2 s = 1 1 /(2q1 ).
(27)
Replacing (27) into (22), we obtain the nal expression for X (s, t, T ). In a similar way, to solve equation (20), we posit a solution of the type Z (L, t, T ) = Z (L, ) = A2 ( )eC ( )L . Hence, the equation (20) becomes 1 2 A2 ( ) 2 LC 2 ( ) q2 ( C ( )L L = 0 , 2 L)C ( ) 2 A2 ( ) where, from (21), the terminal conditions are given by A2 (0) = 1, C (0) = 0 . (30) (29) (28)
As equation (29) is linear in the variable L, this equation is equivalent to the following system of rst-order dierential equations 1 2 2 C ( ) + q2 C ( ) + C ( ) 1 = 0 , 2 2 A ( ) q2 = 0, 2 C ( ) + 2 A2 ( ) subject to the terminal conditions (30). We rst solve (31). It is straightforward to show that this equation can be rewritten as where c1 = 2 dC ( ) 2 (C ( ) c )(C ( ) c ) = d , 2 1 2 c2 = q2 < 0, 2 2 =
2 + 2 2 . q2 2
(31) (32)
q2 + > 0, 2 2
Integrating this equation and using the terminal condition C (0) = 0, it is obtained that C ( ) c2 c2 1 1 ln = + ln , C ( ) c1 c1
27
and a little algebra leads to C ( ) = 2 (exp{ } 1) . (q2 + ) exp{ } + ( q2 ) A2 ( ) = 0. A2 ( ) (33)
Once we know C ( ), we can solve the equation (32) or, equivalently k2 2 C ( ) + Integrating, we have ln[A2 ( )] = k2 2 Let y = exp{ }. Then, more algebra gives C ( )d = 2 2 ln((q2 + ) exp{ } + ( q2 )) (q2 + ) 2 . 2 C ( )d + kA . (34)
Replacing this expression in (34) and applying the condition A2 (0) = 1, the nal expression for A2 ( ) is given by 2 exp (q2 + ) 2 A2 ( ) = (q2 + ) exp{ } + ( q2 )
2 2k2 2 /2
(35)
Including (33) and (35) into (28), we obtain the nal expression for Z (L, t, T ). This expression, jointly with (22), gives the closed-form formula for the default-free discount bond prices for all maturities. 2
Tables and Figures

Table I. Estimates of the Diusion Parameters
This table provides the parameter estimates (with t-values in parentheses) of the processes followed by each state variable. The sample period is from January 1991 to December 1995. The parameters are estimated by means of the Generalized Method of Moments applied to the following equations
st st1 = a1 + b1 st1 + s t, Lt Lt1 = a2 + b2 Lt1 + L t , Lt Lt1 = a3 + b3 Lt1 + L t ,

Variable Spread Long-Term Rate (CIR process) Long-Term Rate (Vasicek process) a -2.08 106 (-0.0210) 1.079 103 (3.5991) 0.732 103 (2.2968) b -0.01544 (-3.0756) -0.01042 (-3.6285) -0.00728 (-2.3988)
2 s t IID (0, 1 ) 2 L t IID (0, 2 Lt1 ) 2 L t IID (0, 3 )
k 0.01544 (3.0756) 0.01042 (3.6285) 0.00728 (2.3988)
-0.1347 103 (-0.021) 0.103532 (33.2787) 0.100574 (20.881)
0.003467 0.070703 0.001159
28
Manuel Moreno
Table II. Averages of Pure Cross-Sectional Regressions

This table contains the estimation results, for each day of the period 1991-1995, of the parameters (qi , i = 1, 2, s ) in the closed-form pricing equation for the Vasicek-CIR model, P (s, L, t, T ) = P (s, L, ) = A( )eB ( )sC ( )L , where A( ) A1 ( ) B ( ) with q1 = k1 + b1 , q2 = k2 + d2 , = = = A1 ( ) A2 ( ) exp
2 2 exp (q2 + ) 1 2 B 2 ( ) + s (B ( ) ) , A2 ( ) = 4q1 (q2 + ) exp{ } + ( q2 )
2 2k2 2 /2
1 eq1 2(exp{ } 1) , C ( ) = q1 (q2 + ) exp{ } + ( q2 ) 1 = (k1 1 a1 )/q1 , 2 = k2 2 /q2 ,

2 /(2q 2 ) s = 1 1 1
2 + 2 2 q2 2
Numbers in parentheses represent the average of the t-statistics of cross-sectional regressions. Numbers in square brackets [.] represent the standard deviation of the time series of parameter estimates. 1991-1995 0.6302 (120.66) [0.5376] 0.6461 (87.88) [0.6430] 0.0939 (784.10) [0.0170] 1991 0.9068 (99.92) [0.7612] 0.8175 (92.79) [0.7786] 0.1059 (1559.51) [0.0103] 1992 0.4618 (116.21) [0.2526] 0.3490 (77.04) [0.1967] 0.0919 (572.70) [0.0150] 1993 0.8731 (70.97) [0.7065] 1.2430 (117.71) [0.7156] 0.1010 (945.38) [0.0081] 1994 0.5440 (154.12) [0.1749] 0.5823 (104.31) [0.4736] 0.0828 (580.79) [0.0174] 1995 0.3577 (162.88) [0.1515] 0.2227 (46.68) [0.1431] 0.0876 (254.74) [0.0199]
q1
q2 s
Table III. Averages of Market Prices of Risk

This table contains the estimation results, for each day of the period 1991-1995, of the market prices of risk (i , i = 1, 2, 3) related to each state variable in both two-factor models. Numbers in parentheses represent the average of the t-statistics of these estimates. Numbers in square brackets [.] represent the standard deviation of the time series of market prices of risk estimates. 1991-1995 -16.6370 (-89.24) [13.0414] 2.8754 (85.99) [2.9495] Panel A: Vasicek-CIR 1991 1992 -25.4189 -11.1207 (-89.14) (-76.07) [21.2431] [7.3795] 3.8128 1.5872 (91.24) (74.87) [3.6390] [0.9426] Model 1993 -19.0044 (-86.83) [11.5403] 5.5101 (116.41) [3.4294] 1994 -15.8773 (-118.56) [6.1945] 2.4110 (102.51) [1.9340] 1995 -11.6185 (-75.26) [5.9674] 0.9865 (44.00) [0.6534]
1 (Spread) 2 (Long rate)
1 (Spread) 3 (Long rate)
Panel B: Vasicek-Vasicek Model 1991-1995 1991 1992 1993 0.2386 3.1513 1.8632 2.6534 (5.09) (46.88) (34.10) (1.19) [12.2793] [10.1866] [5.6779] [13.3915] 4.8419 -8.4789 -6.9682 37.8514 (15.22) (-49.31) (-30.41) (30.75) [41.8299] [32.8720] [15.2471] [59.6728]
1994 -0.7678 (-2.07) [18.0933] -1.7900 (-29.76) [44.0095]
1995 -5.7401 (-54.31) [7.5040] 2.8856 (1.75) [20.7386]
29
Table IV. Within-Sample Pricing Error Measures. 1991-1994

This table contains the within-sample pricing error measures of both two-factor models for the period 1991-1994. We consider zero-coupon bonds with face value of $1 and with maturities ranging from 7 days to 5 years. We have computed ve dierent error measures: the mean error (ME), the mean absolute error (MAE), the root mean squared error (RMSE), the mean absolute percentage error (MAPE) and the root mean squared percentage error (RMSPE). Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year ME -0.000000 -0.000001 -0.000002 -0.000001 0.000006 0.000020 0.000003 -0.000077 Vasicek-CIR Model MAE RMSE 0.000001 0.000004 0.000002 0.000012 0.000004 0.000034 0.000008 0.000035 0.000013 0.000060 0.000032 0.000115 0.000061 0.000133 0.000119 0.000187 MAPE 0.000083 0.000175 0.000382 0.000803 0.001373 0.003515 0.008256 0.019644 RMSPE 0.000366 0.001214 0.003471 0.003651 0.006402 0.012915 0.018230 0.030982
Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year
Vasicek-Vasicek Model ME MAE RMSE -0.000000 0.000001 0.000004 0.000000 0.000003 0.000015 0.000003 0.000012 0.000053 0.000034 0.000074 0.000273 0.000109 0.000200 0.000712 0.000241 0.000480 0.001457 -0.000081 0.000517 0.001156 -0.000481 0.001547 0.002887
MAPE 0.000106 0.000321 0.001239 0.007631 0.021114 0.053265 0.070541 0.256526
RMSPE 0.000421 0.001519 0.005413 0.028254 0.075686 0.162536 0.159114 0.467754
Table V. Within-Sample Pricing Error Measures. 1992

This table contains the within-sample pricing error measures of both two-factor models for the year 1992. Zero-coupon bonds with face value of $1 and with maturities ranging from 7 days to 5 years. Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year ME 0.000000 0.000000 0.000001 0.000004 0.000011 0.000023 -0.000012 -0.000069 Vasicek-CIR Model MAE RMSE 0.000000 0.000000 0.000000 0.000001 0.000001 0.000001 0.000004 0.000005 0.000012 0.000015 0.000033 0.000041 0.000034 0.000048 0.000119 0.000146 MAPE 0.000021 0.000038 0.000084 0.000428 0.001304 0.003671 0.004963 0.020964 RMSPE 0.000033 0.000064 0.000140 0.000551 0.001630 0.004636 0.007023 0.025712
Vasicek-Vasicek Model ME MAE RMSE 0.000000 0.000000 0.000001 0.000000 0.000001 0.000002 0.000001 0.000005 0.000009 0.000005 0.000037 0.000062 0.000002 0.000109 0.000188 -0.000042 0.000305 0.000510 -0.000081 0.000349 0.000500 0.000208 0.001013 0.001542
MAPE 0.000035 0.000124 0.000533 0.003838 0.011595 0.034493 0.050313 0.181938
RMSPE 0.000068 0.000230 0.000940 0.006357 0.019969 0.057827 0.072872 0.281741
30
Manuel Moreno
Table VI. Within-Sample Pricing Error Measures for 1-year Bonds

This table contains the within-sample pricing error measures of both two-factor models for each semester of the period 1991-1994. We consider zero-coupon bonds with face value of $1 and with maturity of 1 year. Period 1991:I 1991:II 1992:I 1992:II 1993:I 1993:II 1994:I 1994:II ME 0.000003 0.000007 0.000031 0.000015 0.000040 0.000039 0.000019 0.000008 Vasicek-CIR Model MAE RMSE 0.000023 0.000030 0.000023 0.000033 0.000037 0.000045 0.000028 0.000037 0.000052 0.000308 0.000039 0.000045 0.000024 0.000029 0.000026 0.000047 MAPE 0.002556 0.002587 0.004192 0.003132 0.005884 0.004286 0.002616 0.002821 RMSPE 0.003426 0.003700 0.005008 0.004217 0.034621 0.004886 0.003080 0.005079
Period 1991:I 1991:II 1992:I 1992:II 1993:I 1993:II 1994:I 1994:II
ME -0.000138 0.000386 -0.000062 -0.000022 0.000586 0.000480 0.000753 -0.000096
Vasicek-Vasicek Model MAE RMSE MAPE 0.000365 0.000708 0.041316 0.000663 0.001082 0.074359 0.000269 0.000428 0.030164 0.000342 0.000584 0.038968 0.000774 0.003241 0.086797 0.000493 0.001365 0.053847 0.000794 0.001435 0.085701 0.000113 0.000444 0.012277
RMSPE 0.079741 0.121425 0.048048 0.066439 0.364635 0.149408 0.154830 0.048524
Table VII. Comparison of One-Step-Ahead Forecasts. 1995

This table contains the out-of-sample pricing error measures of both two-factor models for the year 1995. We compute one-step-ahead forecasts for prices of zero-coupon bonds with face value of $1 and with maturities ranging from 7 days to 5 years. Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year ME 0.000000 0.000001 0.000002 0.000006 0.000015 0.000036 0.000168 0.000310 Vasicek-CIR Model MAE RMSE 0.000016 0.000024 0.000032 0.000047 0.000067 0.000098 0.000176 0.000254 0.000292 0.000417 0.000457 0.000650 0.001273 0.001703 0.001847 0.002397 MAPE 0.001649 0.003247 0.006782 0.017962 0.030523 0.050227 0.173836 0.315120 RMSPE 0.002412 0.004745 0.009899 0.025969 0.043680 0.071588 0.233587 0.411416
Vasicek-Vasicek Model ME MAE RMSE 0.000000 0.000016 0.000024 0.000000 0.000032 0.000047 -0.000001 0.000067 0.000098 -0.000016 0.000180 0.000255 -0.000065 0.000321 0.000436 -0.000209 0.000580 0.000767 -0.000239 0.001320 0.001790 0.000961 0.002096 0.002760
MAPE 0.001648 0.003242 0.006781 0.018363 0.033589 0.063822 0.180321 0.357493
RMSPE 0.002410 0.004738 0.009874 0.026064 0.045608 0.084462 0.245460 0.473140
31
Table VIII. Comparison of One-Step-Ahead Forecasts. 1-year Bonds

Out-of-sample pricing error measures of both two-factor models for each month of 1995. One-stepahead forecasts for prices of zero-coupon bonds with face value of $1 and with maturity of 1 year. Period 1995:I 1995:II 1995:III 1995:IV 1995:V 1995:VI 1995:VII 1995:VIII 1995:IX 1995:X 1995:XI 1995:XII ME -0.000102 0.000037 -0.000149 0.000044 -0.000086 0.000056 0.000081 0.000180 -0.000013 0.000047 0.000147 0.000217 Vasicek-CIR Model MAE RMSE 0.000830 0.001048 0.000592 0.000726 0.000758 0.000971 0.000679 0.000821 0.000650 0.000798 0.000299 0.000367 0.000171 0.000284 0.000285 0.000396 0.000339 0.000485 0.000187 0.000264 0.000274 0.000377 0.000462 0.000650 MAPE 0.091516 0.065100 0.083832 0.074892 0.071588 0.032940 0.018823 0.031256 0.037148 0.020504 0.029984 0.050265 RMSPE 0.115583 0.079775 0.107340 0.090439 0.087900 0.040464 0.031279 0.043447 0.053150 0.028900 0.041194 0.070714
Period 1995:I 1995:II 1995:III 1995:IV 1995:V 1995:VI 1995:VII 1995:VIII 1995:IX 1995:X 1995:XI 1995:XII
Vasicek-Vasicek Model ME MAE RMSE -0.000128 0.000806 0.001028 0.000042 0.000598 0.000736 -0.000157 0.000751 0.000966 0.000007 0.000642 0.000755 -0.000564 0.000861 0.000993 -0.000553 0.000726 0.000900 -0.000336 0.000522 0.000727 -0.000311 0.000480 0.000628 -0.000476 0.000550 0.000713 -0.000183 0.000272 0.000346 0.000055 0.000278 0.000374 0.000233 0.000458 0.000651
MAPE 0.088825 0.065687 0.082970 0.070767 0.094891 0.080044 0.057492 0.052822 0.060301 0.029797 0.030352 0.049808
RMSPE 0.113421 0.080922 0.106741 0.083297 0.109457 0.099249 0.080025 0.069009 0.078096 0.037896 0.040852 0.070898
Table IX. Comparison of Five-Step-Ahead Forecasts. 1995

Out-of-sample pricing error measures of both two-factor models for1995. Five-step-ahead forecasts for prices of zero-coupon bonds with face value of $1; maturities ranging from 7 days to 5 years. Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year ME 0.000000 0.000001 0.000003 0.000019 0.000062 0.000194 0.000868 0.001312 Vasicek-CIR Model MAE RMSE 0.000029 0.000041 0.000057 0.000081 0.000121 0.000171 0.000327 0.000471 0.000581 0.000855 0.001072 0.001558 0.003318 0.004242 0.004669 0.005904 MAPE 0.002910 0.005754 0.012148 0.033447 0.060846 0.117955 0.452690 0.795712 RMSPE 0.004115 0.008149 0.017260 0.048175 0.089560 0.171596 0.581056 1.011933
Vasicek-Vasicek Model ME MAE RMSE 0.000000 0.000029 0.000041 0.000001 0.000057 0.000081 0.000001 0.000120 0.000171 -0.000003 0.000327 0.000469 -0.000017 0.000590 0.000857 -0.000052 0.001140 0.001600 0.000457 0.003296 0.004247 0.001967 0.004852 0.006077
MAPE 0.002909 0.005751 0.012137 0.033434 0.061792 0.125428 0.449916 0.826338
RMSPE 0.004112 0.008135 0.017204 0.047936 0.089773 0.176266 0.582409 1.039909
32
Manuel Moreno
Table X. Comparison of Five-Step-Ahead Forecasts. 1- and 5-year Bonds

This table contains the out-of-sample pricing error measures of both two-factor models for each quarter of the year 1995. We compute ve-step-ahead forecasts for prices of zero-coupon bonds with face value of $1 and with maturity of 1 and 5 years. We report ve dierent error measures: the mean error (ME), the mean absolute error (MAE), the root mean squared error (RMSE), the mean absolute percentage error (MAPE) and the root mean squared percentage error (RMSPE). Panel A: 1-year Bonds Period 1995:I 1995:II 1995:III 1995:IV Period 1995:I 1995:II 1995:III 1995:IV Vasicek-CIR Model MAE RMSE MAPE 0.001652 0.002308 0.182288 0.001082 0.001431 0.119313 0.000698 0.000985 0.076596 0.000860 0.001161 0.093834 Vasicek-Vasicek Model ME MAE RMSE MAPE -0.000513 0.001642 0.002300 0.181183 -0.000186 0.001326 0.001612 0.146189 -0.000005 0.000761 0.001005 0.083563 0.000520 0.000836 0.001153 0.091139 ME -0.000505 0.000180 0.000447 0.000667 RMSPE 0.255000 0.157693 0.108032 0.126462 RMSPE 0.254149 0.177747 0.110399 0.125664
Panel B: 5-year Bonds Period 1995:I 1995:II 1995:III 1995:IV Period 1995:I 1995:II 1995:III 1995:IV Vasicek-CIR Model MAE RMSE MAPE 0.004647 0.006103 0.822578 0.005908 0.007249 1.021432 0.004360 0.005489 0.729830 0.003765 0.004417 0.609400 Vasicek-Vasicek Model ME MAE RMSE MAPE -0.000945 0.004740 0.006169 0.838764 0.002615 0.005902 0.007305 1.019201 0.003061 0.004897 0.005972 0.820306 0.003184 0.003853 0.004517 0.623691 ME -0.001216 0.001559 0.001996 0.002977 RMSPE 1.086229 1.251052 0.918156 0.711505 RMSPE 1.097282 1.258717 1.000055 0.727507
33
Parameter q1 (Vasicek-CIR Model) 1991-1995 Parameter q1 10 5 0 0
200
400 600 800 1000 Time Parameter q2 (Vasicek-CIR Model) 1991-1995
1200
1400
Parameter q2
10 5 0 0
200
800 1000 Time Parameter s* (Vasicek-CIR Model) 1991-1995
400
600
1200
1400
Parameter s*
0.2 0.1 0 0
200
400
600 Time
800
1000
1200
1400
Figure 1: Plot of pure-cross parameters in the Vasicek-CIR model.

Market Price of the Spread (Vasicek-CIR Model) 1991-1995 0 -50 -100 -150 -200 0
Market Price
200
400
600 Time
800
1000
1200
1400
Market Price of the Long-term Rate (Vasicek-CIR Model) 1991-1995 40 30 20 10 0 0
Market Price
200
400
600 Time
800
1000
1200
1400
Figure 2: Plot of market prices of risk in the Vasicek-CIR model.
34
Manuel Moreno
1 0.5 Error 0 -0.5 -1
x 10
-4
Error for 7-day Bonds
-1.5 0 x 10
-4
100
200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Error for 15-day Bonds
900
1000
4 2 Error 0 -2
-4 0
100
-3
200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model
Error for 1-month Bonds
900
1000
2 1 Error 0 -1
x 10
-2 0 x 10
-3
100
200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Error for 3-month Bonds
900
1000
8 6 Error 4 2 0
-2 0
100
900
1000
35
20
x 10
-3
Error for 6-month Bonds
Error
10
0 0 100 200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Error for 1-year Bonds 0.04 900 1000
Error
0.02
0 0 100 200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model 900 1000
Error for 3-year Bonds 0.03 0.02 Error 0.01 0 -0.01 0
100
200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Error for 5-year Bonds
900
1000
0.01 0 Error -0.01 -0.02 -0.03 0
100
900
1000
Figure 3: Within-Sample Errors of the Vasicek-Vasicek and Vasicek-CIR Models.
36
Manuel Moreno
Percentage Error for 7-day Bonds 0.01 Percentage Error
-0.01 0 100 200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Percentage Error for 15-day Bonds 0.04 900 1000
Percentage Error
0.02 0 -0.02 -0.04 0
100
900
1000
Percentage Error for 1-month Bonds 0.2 Percentage Error 0.1 0 -0.1 -0.2 0
100
200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Percentage Error for 3-month Bonds
900
1000
0.8 Percentage Error 0.6 0.4 0.2 0 -0.2 0 100 200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model 900 1000
37
Percentage Error for 6-month Bonds 2 Percentage Error 1.5 1 0.5 0 -0.5 0 100 200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Percentage Error for 1-year Bonds 4 Percentage Error 3 2 1 0 -1 0 100 200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model 900 1000 900 1000
Percentage Error for 3-year Bonds 4 Percentage Error 3 2 1 0 -1 0 100 200 300 400 500 600 700 800 --- Vasicek-Vasicek Model, ... Vasicek-CIR Model Percentage Error for 5-year Bonds 2 Percentage Error 900 1000
-2
-4 0
100
900
1000
Figure 4: Within-Sample Percentage Errors (Vasicek-Vasicek and Vasicek-CIR).
38
Manuel Moreno
References
[1] A t-Sahalia, Y. (1996). Testing Continuous-Time Models of the Spot Interest Rate. Review of Financial Studies 9, 2, 385426. [2] Ayres, H.R. and J.Y. Barry (1980). A Theory of the U.S. Treasury Market Equilibrium. Management Science 26, 6, 539569. [3] Boudoukh, J., M. Richardson, R. Stanton and R.F. Whitelaw (1999). A Multifactor, Nonlinear, Continuous-Time Model of Interest Rate Volatility, mimeo, New York University. [4] Brennan, M.J. and E.S. Schwartz (1979). A Continuous Time Approach to the Pricing of Bonds. Journal of Banking and Finance 3, 133155. [5] Chan, K. C., G.A. Karolyi, F.A. Longsta and A.B. Sanders (1992). An Empirical Comparison of Alternative Models of the Short-Term Interest Rate. Journal of Finance 47, 3, 12091227. [6] Chen, L. (1996). Interest Rate Dynamics, Derivatives Pricing, and Risk Management. Springer-Verlag, Berlin. [7] Conley, T.G., L.P. Hansen, E.G. Luttmer, and J. Scheinkman (1997). ShortTerm Interest Rates as Subordinated Diusions. Review of Financial Studies 10, 3, 525577. [8] Cox, J.C., J.E. Ingersoll and S.A. Ross (1985). A Theory of the Term Structure of Interest Rates. Econometrica 53, 385408. [9] Dai, Q. and K. Singleton (2000). Specication Analysis of Ane Term Structure Models. Journal of Finance 55, 5, 19431978. [10] Due, D. and R. Kan (1996). A Yield-Factor Model of Interest Rates. Mathematical Finance 6, 4, 379406. [11] Jones, F.J. (1991). Yield Curve Strategies. Journal of Fixed Income, 3341. [12] Knez, P.J., R. Litterman, R., and J. Scheinkman (1994). Explorations Into Factor Explaining Money Market Returns. Journal of Finance 49, 5, 1861 1882. [13] Litterman, R. and J. Scheinkman (1991). Common Factors Aecting Bond Returns. Journal of Fixed Income 1, 1, 5461. [14] Longsta, F.A. and E.S. Schwartz (1992). Interest Rate Volatility and The Term Structure: A Two-Factor General Equilibrium Model. Journal of Finance 47, 4, 12591282.
39
[15] Moreno, M. (1996). A Two-Mean Reverting-Factor Model of the Term Structure of Interest Rates. Economics and Business Working Paper, num. 193. Universitat Pompeu Fabra de Barcelona. Also as Finance and Banking Discussion Papers Series, num. 23. Universitat Pompeu Fabra de Barcelona. [16] Moreno, M. and J.I. Pe na (1996). On the Term Structure of Interbank Interest Rates: Jump-Diusion Processes and Option Pricing. In Forecasting Financial Markets, edited by Christian Dunis, John Wiley & Sons Ltd., Sussex, England. [17] Nelson, J. and S.M. Schaefer (1983). The Dynamics of the Term Structure and Alternative Portfolio Immunization Strategies. In Innovations in Bond Portfolio Management: Duration Analysis and Immunization, eds. G.O. Bierwag, G.G. Kaufman, and A. Toevs, Greenwich, CT: JAI Press. [18] Nu nez, S. (1995). Estimaci on de la Estructura Temporal de los Tipos de Inter es en Espa na: Elecci on entre M etodos Alternativos. Documento de Trabajo 9522. Banco de Espa na. [19] Richard, S.F. (1978). An Arbitrage Model of the Term Structure of Interest Rates. Journal of Financial Economics 6, 3357. [20] Schaefer, S. and E.S. Schwartz (1984). A Two-Factor Model of the Term Structure: An Approximate Analytical Solution. Journal of Financial and Quantitative Analysis 19, 4, 413424. [21] Stanton, R. (1997). A Nonparametric Model of Term Structure Dynamics and the Market Price of Interest Rate Risk. Journal of Finance 52, 5, 19732002. [22] Vasicek, O. (1977). An Equilibrium Characterization of the Term Structure. Journal of Financial Economics 5, 177188. [23] Zhang, H. (1993). Treasury Yield Curves and Cointegration. Applied Economics 25, 3, 361367.
Manuel Moreno Departament dEconom a i Empresa Universitat Pompeu Fabra Carrer Ram on Trias Fargas, 25-27 08005 Barcelona, Spain e-mail: manuel.moreno@econ.upf.es
El uso de las mixturas de distribuciones gaussianas en nanzas: aplicaci on en valoraci on y cobertura de opciones nancieras
Susana Corcuera y Juan Carlos Garc a C espedes1
Resumen Pocos a nos despu es de que el modelo de Black-Scholes fuera universalmente aceptado por la comunidad nanciera, empiezan a ser maniestas sus principales deciencias: una de ellas, el smile de volatilidad, nos empuja claramente a considerar modelos con un decrecimiento menor que la normal en las colas (colas m as gruesas que la normal) y muchas veces con probabilidad desigual a grandes descensos que a grandes ascensos. Por otro lado, BS cuenta entre sus grandes ventajas con un equilibrio razonable entre la bondad del modelo y su sencillez. Una mixtura de dos normales es, desde esta perspectiva, un modelo que incorpora de forma natural fen omenos como el smile de volatilidad y la mayor densidad de probabilidad en las colas, debido a su mayor exibilidad y su mejor explicaci on de la formaci on de los precios, sin, por otro lado, alejarse demasiado del mundo de la normalidad, lo que permite un uso sencillo y directo del modelo. Este art culo desarrolla las principales ideas acerca del uso de una mixtura de normales para la valoraci on de productos nancieros: qu e es y qu e representa una mixtura como distribuci on de probabilidad, por qu e es una buena idea para explicar la formaci on de precios, etc. Se desarrolla una f ormula de valoraci on y se dan expresiones expl citas para opciones plain-vanilla, las cuales resultan ser combinaciones lineales (con dos t erminos) de las f ormulas an alogas en BS. Se presentan tambi en estudios pr acticos sobre opciones s/IBEX 35 y S&P500 y, nalmente, se analiza el funcionamiento de las griegas (derivadas) en el modelo, conjuntamente con ejemplos pr acticos basados en la cobertura de una cartera.
1 Susana Corcuera es analista t ecnico de Riesgo en el BBVA. Juan Carlos Garc a C espedes es Director de Metodolog as de Riesgos de BBVA. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de marzo de 1998.
42
Susana Corcuera y Juan Carlos Garc a C espedes
Introducci on
Una mixtura de distribuciones es, en general, una distribuci on de probabilidad cuya funci on de densidad es una cierta combinaci on ponderada de otras dos (o de otras muchas). Son representativas de experimentos en los que con una cierta probabilidad p1 se extrae un n umero con la densidad f1 , con probabilidad p2 de la f2 ,. . . y con probabilidad pn de la densidad fn . El modelo encaja bien en un contexto nanciero en el que con una cierta probabilidad (alta) los precios oscilan en una banda corriente de poca volatilidad, y con una probabilidad mucho m as baja se produce un sobresalto que conduce a fuertes subidas o a fuertes bajadas (noticias, reacciones en cadena, stop-losses,. . . ). Posiblemente hoy una de las mayores deciencias de los modelos normales (BS) radique en que los mayores errores de predicci on se cometen en las colas de la distribuci on (sucesos extremos). Este hecho nos lleva a la aparici on del fen omeno de smile, para el que no hay explicaci on ni posible cobertura (precisamente porque queda fuera del marco del modelo); tambi en nos lleva a la aparici on de una nueva problem atica, no tanto en el a rea de valoraci on como en el de la medici on de los riesgos, disciplina que centra precisamente su inter es en la medici on de la cola de la distribuci on, donde, por desgracia, el modelo normal infravalora. Sin embargo, muchas caracter sticas de la normal, sobre todo en la zona central de la distribuci on, permanecen: simetr a, m axima frecuencia en torno a la media, etc. El modelo de mixtura es por tanto capaz de mantener las caracter sticas importantes de normalidad de las series (porque en el fondo est a formado por dos normales), pero por otro lado introduce una mayor riqueza en la modelizaci on permitiendo distribuciones con colas m as gruesas y asim etricas. Entre las aplicaciones inmediatas de las mixturas de normales se pueden mencionar: La medici on de riesgos. La estimaci on y valoraci on de la sonrisa/mueca de volatilidad. La cobertura del riesgo de sonrisa/mueca. La valoraci on de opciones. En el campo de las opciones nancieras, la modelizaci on mediante mixturas de gaussianas permite el uso casi directo de la conocida teor a de valoraci on de opciones en el contexto Black-Scholes, con realmente muy pocas modicaciones, a la vez que incorpora de forma natural la existencia del smile.
El uso de las mixturas de distribuciones gaussianas en finanzas
43
Mixtura de distribuciones gaussianas
Sea la variable aleatoria X tal que X = I X1 + (1 I ) X2 , donde X1 y X2 son dos variables aleatorias con distribuciones X1 N (1 , 1 ) y X2 N (2 , 2 ) respectivamente2 e I es una variable aleatoria con distribuci on I B (p), siendo B (p) una distribuci on de Bernoulli con probabilidad de exito p. Se dice entonces que la variable aleatoria X es una mixtura de dos normales. Se dice que X es una mixtura de n normales cuando X = I1 X1 + I2 X2 + + In Xn , donde nuevamente X1 , X2 , . . . , Xn son variables aleatorias con distribuciones X1 N (1 , 1 ), X2 N (2 , 2 ), . . . , Xn N (n , n ), y la distribuci on de Ii viene dada por la siguiente expresi on: Ii = 1, 0, si B = i, si B = i,
donde la variable aleatoria B es B : {1, 2, . . . , n}, con probabilidades3 P (B = k ) = pk , para cada k = 1, . . . , n. Aunque como vemos el concepto de mixtura sirve para un n umero arbitrario de distribuciones, el modelo de valoraci on se ha desarrollado para una mixtura de dos, y as se entender a en lo sucesivo4 . Un ejemplo simple de una mixtura de 2 normales, con probabilidades iguales del 50% , ser a el siguiente: se lanza una moneda al aire y si sale cara se extrae una realizaci on de una variable aleatoria N (0, 1); si sale cruz se extrae una realizaci on de una variable aleatoria N (1, 2). Un poco m as generalmente, una mixtura de dos normales cuyos par ametros son (1 , 2 , 1 , 2 ) es una distribuci on de probabilidad que reeja un experimento consistente en - con probabilidad p se toma un numero aleatorio con distribuci on N (1 , 1 ), - con probabilidad 1 p se toma un n umero aleatorio con distribuci on N (2 , 2 ).
2 de la mixtura se expresan como funci on de La media mixtura y la varianza mixtura los par ametros de sus dos componentes gaussianas:
Media de la mixtura: mixtura = p 1 + (1 p) 2 .

denota una normal de media y desviaci on t pica . probabilidad de que Ii = 1 es pi que, a su vez, es la probabilidad de extraer un n umero de la normal i- esima. 4 Es posible trabajar con mixturas de m as de dos normales; de hecho es posible dejar el n umero de normales como un par ametro m as que se puede estimar. Sin embargo, en tal caso el n umero de par ametros empieza a crecer y el modelo deja de ser u til (demasiados grados de libertad).
3 La 2 N (, )
44
Varianza de la mixtura: Caso particular 1 = 2 :

2 2 2 mixtura = p 1 + (1 p) 2 .
Caso general:
2 2 2 nixtura = p 1 + (1 p) 2 + p (1 p) (1 2 )2 .
Representando gr acamente la desviaci on t pica de la mixtura en funci on de la probabilidad p (tomando 1 = 1, 2 = 2 jos) para varias diferencias de medias (1 2 ) se obtiene el gr aco siguiente:
Para varianzas de las normales componentes jas, la varianza de la mixtura depende por tanto de dos factores: la diferencia entre las medias (diferencias mayores en las medias generan obviamente varianzas mayores5 ), y nalmente depende de la probabilidad p, llegando a alcanzar un m aximo para pmax = Con pmax obtenemos una varianza
2 max = 2 2 1 2 1 1 1 2 2 (1 + 2 ) + (1 2 )2 + . 2 4 4 (1 2 )2 2 2 1 1 1 2 + . 2 2 (1 2 )2
2.1
La mixtura frente a la normal de igual media y volatilidad
En cada uno de los cuatro gr acos siguientes se representa una mixtura con una cierta media y varianza frente a la distribuci on gaussiana que tiene iguales par ametros. Se pone as de maniesto que las mixturas permiten generar las distribuciones leptoc urticas y asim etricas caracter sticas de los activos nancieros.
5 Es
intuitivamente obvio si pensamos en lo que signica el experimento.
45
Par ametros de la mixtura: p = 0.5, 1 = 0, 1 = 1, 2 = 0, 2 = 2
Par ametros de la mixtura: p = 0.5, 1 = 1, 1 = 0.5, 2 = 1, 2 = 1
Cual es la utilidad de la mixturas de gaussianas? Por qu e se plantea su uso, frente a otras posibles distribuciones? B asicamente por cuatro razones: 1. Una mixtura de dos distribuciones normales es una distribuci on de probabilidad con relativamente pocos par ametros6 , pero que a su vez, permite la exibilidad suciente para ser un buen modelo para un n umero importante de experimentos probabil sticos. 2. Su interpretaci on en muchas situaciones econ omicas es natural; por ejemplo, ante la posibilidad de que el banco central modique sus tipos de intervenci on : (a) Con una determinada probabilidad alta, p, el banco central bajar a tipos, lo que dar a lugar a un escenario de baja volatilidad. Esto es lo que descuenta la mayor a del mercado. (b) Sin embargo existe una peque na probabilidad de que el banco central mantenga tipos, esper andose en tal caso que la volatilidad en el mercado sea mayor, con tensiones sobre el nivel de tipos. Esta situaci on se puede modelizar de forma sencilla mediante una mixtura, suponiendo que la distribuci on de tipos de inter es es, con probabilidad p, N (1 , 1 )
6 Por ejemplo, una mixtura de 2 normales s olo tiene 5 par ametros, y puede representar adecuadamente distribuciones asim etricas, con dos m aximos, leptoc urticas, etc.
46
y con probabilidad 1 p, N (2 , 2 ), donde 1 < 2 < 0, p
1 p y 1
2 .7
3. Los modelos nancieros de saltos (jump-diussion processes ) pueden en cierta forma asimilarse a mixturas de normales (con, por ejemplo, grandes diferencias entre las las medias o la volatilidad de las componentes normales). En este sentido, la mixtura es un modelo m as sencillo que el de salto y que puede captar muchas de sus caracter sticas. 4. La admisi on de la mixtura como la distribuci on base de explicaci on de la formaci on de los precios en el mercado tiene dos consecuencias positivas de gran relevancia: Por su estrecha relaci on con la distribuci on normal, la aplicaci on de las mixturas de normales en la valoraci on de derivados nancieros es muy sencilla puesto que, fruto de esta estrecha relaci on las f ormulas que se derivan son ligeras transformaciones de la conocida f ormula de Black-Scholes. Su mayor riqueza le permite captar mejor la formaci on de los precios, una prueba de ello es que el smile de volatilidad no s olo queda explicada por la mixtura sino que nace como consecuencia inmediata del modelo.
El modelo Black-Scholes falla: la sonrisa y la mueca
En los mercados de opciones aparece lo que se denomina sonrisa-mueca de volatilidad, esta es una expresi on para subrayar que la volatilidad impl cita es una funci on del strike con un aspecto gr aco de una sonrisa, a lo que debe su nombre.
Este fen omeno est a indicando que la distribuci on de rendimientos al plazo de la opci on no es normal (lo que es lo mismo, que la distribuci on de precios no es lognormal). Para argumentar esta armaci on sup ongase el caso de la valoraci on de una opci on de tipo call asumiendo, para una mayor simplicidad, que el tipo de inter es libre de
7 La Uni on monetaria es otra situaci on econ omica en la que el modelo de mixtura encaja perfectamente. Con una cierta probabilidad un pa s determinado entrar a en la UME, dando lugar a un escenario de tipos de inter es bajos y baja volatilidad; con otra probabilidad no entrar a, dando lugar a un escenario de tipos de inter es altos y sobre todo de alta volatilidad.
47
riesgo es cero8 . El perl de pagos de una opci on call, con la distribuci on del subyacente superpuesta se representa en el gr aco siguiente:
Como el tipo de inter es es cero, el precio de la opci on viene dado por por
Preciocall =
(x K ) f (x) dx .
En el caso de que f (x) fuese una distribuci on lognormal entonces se tendr a el modelo Black-Scholes. Qu e pasar a si f (x) no fuese lognormal? En cualquier caso siempre se puede armar que existe una distribuci on lognormal y(k) (x) tal que se cumple:
K
(x K ) f (x) dx =
(x K ) y(k) (x) dx .
A la desviaci on est andar (k ) de la distribuci on lognormal y(k) (x) que cumple la relaci on anterior se le llama volatilidad impl cita9 . A continuaci on se dibujan las dos distribuciones, f (x) e y(k) (x), para un determinado strike K1 . T picamente, las distribuciones reales de mercado presentan leptocurtosis, es decir, son de cuerpo m as estrecho y colas m as gruesas; las gr acas se presentan siguiendo esta losof a:
8 Aunque los argumentos y los gr acos que se adjuntan se han particularizado al caso de una opci on de tipo call, el desarrollo te orico es general y sirve para cualquier tipo de perl de pago. 9 N otese que (K ) es en general una funci on de K : para cada strike K , la volatilidad impl cita es diferente.
48
Las distribuciones anteriores, por denici on, cumplen que

K1
(x K1 ) f (x) dx =
K1
(x K1 ) y(K1 ) (x) dx ,
donde, como ya se ha dicho, (K1 ) es la volatilidad impl cita correspondiente al strike K1 . Sup ongase que se desea calcular la volatilidad impl cita (K2 ) para otro strike diferente, K2 (con K2 > K1 ). Sin necesidad de m as demostraci on, dado que la distribuci on f (x) tiene las colas m as gruesas (por ser leptoc urtica) que y(K1 ) (x), es evidente que la distribuci on y(K2 ) (x) que hace que
K2
(x K2 ) f (x) dx =
K2
(x K2 ) y(K2 ) (x) dx
debe cumplir que (K2 ) > (K1 ). De ello se deduce que colas gruesas (y simetr a) en la distribuci on implican smile (sim etrico) en las volatilidades impl citas. Si la distribuci on f (x), adem as de colas gordas, es asim etrica, da lugar a que aparezca la mueca de volatilidad en lugar de la sonrisa. Se deja al lector el ejercicio de comprobar dicha armaci on.
49
En resumen, se tiene que: Distribuci on Asim etrica Mueca alegre Distribuci on Sim etrica Sonrisa
Colas gruesas
Colas delgadas
Mueca triste
Tristeza
Como se ha visto, la raz on de la mueca/sonrisa se explica porque la distribuci on del subyacente no es lognormal. Una manera de tener esto en cuenta es utilizar la mixtura de normales para aproximar/modelizar la distribuci on impl cita de los subyacentes. En el siguiente apartado se desarrollar a con detalle el modelo de valoraci on de opciones en base a una distribuci on subyacente que coincide con la mixtura de dos gaussianas.
Estimaci on de densidades impl citas a partir de precios de opciones
Sea St la variable aleatoria que representa los precios de un activo y sup ongase que el rendimiento instant aneo, Rt , de este activo tiene una distribuci on de probabilidad que es una mixtura de dos normales. Sean X1 y X2 dos variables aleatorias con distribuciones X1 N (1 , 1 ), X2 N (2 , 2 ) respectivamente, donde N (, ) denota una normal de media y desviaci on t pica , y sea I otra variable con distribuci on I B (p), siendo B (p) una distribuci on de Bernoulli con probabilidad de exito p. Se dice entonces que la variable aleatoria Y = eI X1 +(1I ) X2 es una logmixtura de normales. V ease que la variable aleatoria Y es tambi en la siguiente: Y = eX1 , eX2 , si I = 1, = I eX1 + (1 I ) eX2 . si I = 0.
Es decir, una logmixtura de normales es lo mismo que una mixtura de lognormales. Si, como se ha dicho, Rt = log(St /St1 ) es una variable aleatoria que toma valores seg un una mixtura de normales, entonces se tiene que la distribuci on de St /St1 = eRt es, seg un lo anterior, una logmixtura de dos normales y por tanto tambi en una mixtura de lognormales.
50
Con lo anterior, queda claro que si los rendimientos de un activo son una mixtura de normales, el activo tiene como distribuci on una mixtura de lognormales. En tal contexto, cu anto vale una opci on sobre dicho subyacente? Si se aplica el argumento de neutralidad al riesgo10 , el precio C de una opci on o derivado, viene dado por C = er t E (f (St )) , donde r es el tipo de inter es libre de riesgo, f () es el pago nal del derivado, St es el precio del subyacente y E es la esperanza bajo la probabilidad riesgo-neutra. Esto equivale a C = er t f (St ) g (St ) dSt ,
donde g () es la distribuci on riesgo neutral de St . Supongase que g (St ) no es lognormal, sino que es una mixtura de lognormales; es decir, g (St ) = p g1 (St ) + (1 p) g2 (St ) , en donde g1 (St ) es la densidad lognormal (1 , 1 ) y g2 (St ) es la densidad lognormal (2 , 2 ). Aplicar el argumento de neutralidad al riesgo supone que la esperanza del activo descontado es igual al precio actual. Aplicar esto mismo al precio del propio subyacente implica que E (St ) = S0 = er t S0 p e(1 +1 /2) t + (1 p) e(2 +2 /2) t , er t = p e(1 +1 /2) t + (1 p) e(2 +2 /2) t .
2 2
2 2 2 2
Si llamamos m1 = 1 + 21 y m2 = 2 + 22 (donde m1 y m2 son las dos medias de las dos distribuciones lognormales y 1 y 2 son las medias de las dos normales), se tiene que er t = p em1 t + (1 p) em2 t . Esta igualdad dice que las medias de las dos distribuciones lognormales est an relacionadas con el tipo de inter es libre de riesgo11 . Si se asume que r, m1 y m2 son valores cercanos a 0, entonces la igualdad anterior se puede aproximar, haciendo una reducci on al primer orden, por (1 + r t) = r p (1 + m1 t) + (1 p) (1 + m2 t) p m1 + (1 p) m2 . =
10 El argumento de neutralidad al riesgo: el precio del activo derivado es el valor presente de los ujos esperados bajo la probabilidad riesgo-neutra. 11 En el entorno Black-Scholes, como s olo existe una normal, la media de la misma queda inmediatamente determinada por el tipo de inter es libre de riesgo.
51
Es decir, el tipo de inter es libre de riesgo es la media ponderada de las medias de las dos lognormales de la mixtura. Volviendo a la expresi on inicial, el precio del activo derivado es C = er t f (St ) g (St ) dSt = er t f (St ) p g1 (St ) + (1 p) g2 (St ) dSt ,
que podemos reescribir como C = = er t p f (St ) g1 (St ) dSt + (1 p) f (St ) g2 (St ) dSt
p em1 t er t em1 t
f (St ) g1 (St ) dSt f (St ) g2 (St ) dSt .
+ (1 p) em2 t er t em2 t
Si recordamos que g1 () es una distribuci on lognormal (1 , 1 ) o, en t erminos de 2 m1 , una distribuci on lognormal (m1 1 /2, 1 ) y que, an alogamente, g2 () es una distribuci on lognormal (2 , 2 ) o, en t erminos de m2 , una distribuci on lognormal 2 (m2 2 /2, 2 ), observamos que la expresi on anterior equivale a que C = er t p em1 t C1 + (1 p) em2 t C2 , donde C1 es el precio de la opci on bajo Black-Scholes con m1 como tipo de inter es libre de riesgo y 1 como volatilidad, C2 es el precio de la opci on bajo Black-Scholes con m2 como tipo de inter es libre de riesgo y 2 como volatilidad. Operando la expresi on anterior: C= O bien C=p e m 1 t e m 2 t C + (1 p ) C2 . 1 p em1 t + (1 p) em2 t p em1 t + (1 p) em2 t e m 1 t , p em1 t + (1 p) em2 t e m 2 t , p em1 t + (1 p) em2 t p em1 t C1 + (1 p) em2 t C2 . p em1 t + (1 p) em2 t
Si se denen 1 y 2 seg un 1 = 2 =
la expresi on anterior queda C = p 1 C1 + (1 p) 2 C2 .
52
La condici on sobre el tipo de inter es queda como sigue: 1 = p 1 + (1 p) 2 . El precio del derivado tambi en se puede expresar como: C = p e(m1 r) t C1 + (1 p) e(m2 r) t C2 . Si se denen m1 r = l1 y m2 r = l2 , entonces se tiene que C = p el1 t C1 + (1 p) el2 t C2 . Adem as se debe cumplir que: 1 = p el1 t + (1 p) el2 t . Luego es decir, C = p C1 + (1 p ) C2 , donde p = p e
l1 t
C = p el1 t C1 + (1 p el1 t ;
El precio de la opci on bajo la mixtura es una combinaci on lineal de los precios de la opci on bajo las dos distribuciones gaussianas (B-S) ponderadas por la probabilidad transformada p . Esta probabilidad transformada p es una correcci on de la probabilidad original en base a la distancia de cada media al tipo de inter es libre de riesgo.
4.1
La paridad put -call
Antes de proceder a la estimaci on de densidades a partir de los precios de las opciones, es conveniente analizar una cuesti on previa que luego aparecer a con insistencia y que est a relacionada con los tipos de inter es y la paridad put -call. Existe una igualdad que deben cumplir las opciones, llamada paridad put-call , seg un la cual, suponiendo que no existen dividendos (la existencia de dividendos hace aparecer un factor m as pero b asicamente debe seguir cumpli endose una expresi on del mismo tipo), debe vericarse que C + X er t = P + S . En un entorno real de mercado la existencia de una horquilla de precios convierte la igualdad anterior en dos desigualdades, dos condiciones para que no se pueda arbitrar: C alto + X er1 t P bajo + S , C bajo + X er2 t P alto + S ,
53
donde C alto es el precio que hay que pagar para comprar una call, C bajo es el precio que se recibe al vender una call. P alto es el precio que hay que pagar para comprar una put. P bajo es el precio que se recibe al vender una put. es que se recibe al invertir. r1 es le tipo de inter es que se paga por los pr estamos. r2 es el tipo de inter Resolviendo las desigualdades anteriores se llega a que, para evitar que exista arbitraje, debe ocurrir que r1 r2 1 log t 1 log t P bajo + S C alto X P alto + S C bajo X = rmax , = rmin .
Resulta que una vez estimados los l mites rmax y rmin , estos son funci on del strike, es decir, que lo que se debe cumplir es que r1 r2 rmax (X ) rmim (X ) X , X .
Gr acamente, para el caso de opciones sobre el ndice S&P (fecha 21/5/98, vencimiento 20/6/98), rmin y rmax tienen la siguiente forma:
54
Para el IBEX, el subyacente no es el ndice sino el futuro sobre el IBEX 35, as que la paridad put -call cambia ligeramente. El gr aco para los tipos de inter es de no arbitraje es el que se adjunta (fecha de c alculo, 20/5/98 para opciones de vencimiento 19/6/98)12 :
4.2
La estimaci on de distribuciones impl citas de subyacentes
Consideremos ahora precios reales de opciones sobre un subyacente determinado. Se plantea el problema de tratar de encontrar dos normales cuya mixtura valore lo m as ajustadamente posible los precios de dichas opciones en el entorno de mercado en el que estos se producen. Sean c1 , c2 , . . . , cn precios reales de opciones de compra (calls ) y p1 , p2 , . . . , pm precios de opciones de venta (puts )13 . Entonces debe cumplirse que14 C (K ) = donde es C1 es el precio de la call de strike K bajo Black-Scholes con m1 como tipo de inter libre de riesgo y 1 como volatilidad.
12 Como se ver a posteriormente, estas bandas de tipos de inter es cuadran perfectamente con el tipo de inter es que estiman las mixtura. El tipo de inter es estimado por la mixtura es, en general, un tipo constante, e intermedio entre ambas bandas; es el tipo de inter es constante que mejor ajusta a la vez con los precios bid y ask. 13 La elecci on de puts y calls para la estimaci on s olo es debida a que estos son los contratos de opciones m as l quidos y a que sus precios aparecen en las principales fuentes suministradoras de informaci on de precios. 14 A lo largo del art culo se han usado X y K indistintamente para denotar el strike.
p em1 t C1 (K ) + (1 p) em2 t C2 (K ) , p em1 t + (1 p) em2 t
55
C2 es el precio de la call de strike K bajo Black-Scholes con m2 como tipo de inter es libre de riesgo y 2 como volatilidad. Por otro lado, igualmente debe cumplirse que P (K ) = siendo es P1 es el precio de la put de strike K bajo Black-Scholes con m1 como tipo de inter libre de riesgo y 1 como volatilidad. P2 es el precio de la put de strike K bajo Black-Scholes con m2 como tipo de inter es libre de riesgo y 2 como volatilidad. Consid erense las siguientes funciones:
n m
p em1 t P1 (K ) + (1 p) em2 t P2 (K ) , p em1 t + (1 p) em2 t
Ea (m1 , m2 , 1 , 2 , p)
=
i=1 n
(C (Ki ) ci ) +
i=1
(P (Ki ) pi )
=
i=1 m
p em1 t C1 (Ki ) + (1 p) em2 t C2 (Ki ) ci p em1 t + (1 p) em2 t p em1 t P1 (Ki ) + (1 p) em2 t P2 (Ki ) pi p em1 t + (1 p) em2 t
n
2 1/ 2
+
i=1
y Er (m1 , m2 , 1 , 2 , p) =
i=1
C (Ki ) ci ci
+
i=1
P (Ki ) pi pi
Estas funciones son los errores absolutos y relativos, respectivamente, para cada elecci on de los par ametros de la mixtura: = (m1 , m2 , 1 , 2 , p); trataremos de obtener los par ametros que minimizan las funciones anteriores. Antes de continuar, convendr a hacer alguna matizaci on adicional. Existen b asicamente dos formas de proceder en relaci on al tipo de inter es libre de riesgo (r): El valor de r no afecta directamente a las estimaciones (ninguna de las funciones de error son funciones de r)15 . En este sentido, el m etodo seguido consiste en estimar los par ametros de la mixtura y, bajo condiciones riesgo-neutras, calcular el tipo de inter es como () r= 1 log p em1 t + (1 p) em2 t . t
15 En realidad s , porque en la f ormula de valoraci on se sustituy o en su momento r por su relaci on con m1 , m2 y p; en lo que se quiere hacer enfasis es en que r no aparece expl citamente. Esto no ocurre por ejemplo en el caso en el que el subyacente es el futuro. V eanse las estimaciones para las opciones sobre el futuro del IBEX.
56
Otra posibilidad es minimizar imponiendo la restricci on anterior, es decir, imponiendo que m1 , y m2 tienen que ser tales que cumplan (*), con un tipo r prejado de antemano16 . Hay que recordar que r no est a completamente determinado, dados los precios bid y ask como se ha visto, para evitar el arbitraje, para cada strike, los tipos r1 y r2 deben cumplir las siguientes relaciones: r1 rmax , r2 rmin .
En este sentido, por tanto, no deber a existir ninguna preferencia a la hora de imponer un tipo de inter es dentro de dicha banda. Finalmente, una vez estimados los par ametros de la mixtura a partir de precios reales de opciones, se obtiene la distribuci on mixtura que mejor ajusta a dichos precios. La distribuci on te orica as obtenida se podr a utilizar, entre otras cosas, para valorar cualquier otro tipo de opci on diferente de calls o puts, sin que la valoraci on dependa de circunstancias como la necesidad de obtener una volatilidad impl cita para cada strike (esto podr a ser arbitrario para strikes que no cotizaran en mercado), o mucho m as importante, resuelve el problema de valorar otros derivados nancieros sin conocer la volatilidad impl cita que como sabemos, depende no s olo del strike sino adem as de caracter sticas del instrumento (v ease el apartado que trata sobre la valoraci on de digitales).
4.3
Resultados de las estimaciones para opciones sobre el ndice S&P 500

D as: 18/05/98, 20/05/98, 21/05/98 (Fuente: Bloomberg) Resultados de minimizaciones de la funci on de errores absolutos mediante m etodos de Newton y gradiente conjugado.
D a 18/05/98 Resultados sobre 60 precios de mercado; 30 calls (15 bid y 15 ask ), 30 puts (15 bid y 15 ask ), vencimiento 20/6/98. Nivel del subyacente: 1111.8. Se estima una u nica mixtura con todos los precios, con strikes desde el 1075 hasta el 1150. 1.- Los par ametros obtenidos sin restricci on en las medias de la mixtura (dejando r libre) sobre la condici on riesgo-neutro son: Par ametros Anualizados En t p 0.24 0.24 M1 -71.6% -6.27% 1 19.89% 5.89% 1p 0.76 0.76 m2 26.3% 2.31% 2 10.98% 3.25%
16 En este caso se est a imponiendo la condici on de no arbitraje para la media de la mixtura en la estimaci on, y es como si en lugar de tener 5 grados de libertad (5 par ametros que estimar), se tuvieran s olo 4 grados (4 par ametros que estimar, al imponer el tipo de inter es, se estima m1 o m2 , pero nunca ambas).
57
Estos par ametros implican que el tipo de inter es r es 3.48%, que adem as es el tipo de inter es constante que hace que la paridad put -call se cumpla para todos los strikes. Ea (m1 , m2 , 1 , 2 , p) Er (m1 , m2 , 1 , 2 , p) = = 5.63 (la suma total de precios es 1330.75), 24.44% .
La m axima diferencia absoluta en primas (estimadas frente a reales) se produce para el strike 1095 y vale 1.38: precio real 34.38, precio mixtura 35.7.
La m axima diferencia relativa en primas se produce para el strike 1140 y es un 6.1% precio real 11.50, precio mixtura 10.80.
Como antes se ha detallado, el tipo de inter es r estimado es el que hace que se cumpla la paridad put -call, para todos los strikes. Hay que recordar que se est a estimando una u nica distribuci on (mid ) a partir de precios bid y ask. El precio de las opciones bajo esta distribuci on est a casi en todos los casos dentro de la banda bid -ask, y da lugar a una u nica volatilidad impl cita para put y call del mismo strike ; por ello la linea de smile bajo dicha distribuci on es u nica. Una vez estimados a partir de precios de opciones los par ametros de la mixtura que mejor ajustan dichos precios, se puede obtener una representaci on de la distribuci on impl cita (como mixtura), de las dos normales que la componen y de la normal de igual media y varianza que la mixtura:
58
2.- Si se preja el tipo de inter es de mercado (para este ejemplo se ha tomado r = 5.25%), y se obliga a la mixtura a cumplir la condici on de arbitraje sobre el tipo de inter es impuesto: Par ametros Anualizados En t p 0.24 0.24 M1 -68.1% -5.97% 1 19.05% 5.64% 1p 0.76 0.77 m2 27.0% 2.37% 2 11.49% 3.40%
Error absoluto Ea = 8.48, error relativo Er = 43.76%. El cambio en los par ametros estimados es muy peque no; sin embargo, el error absoluto aumenta de 5.63 a 8.48 para ajustar las medias y los pesos a un tipo de inter es del 5.25%. Dentro de la banda de tipos de inter es permitida es l cito cualquier tipo que se escoja; de hecho nos vale 3.48%, que tambi en est a dentro de la banda. En este sentido es preferible no imponer restricciones sobre r de este estilo y s olo comprobar si el tipo de inter es estimado cumple estar o no dentro de la banda. D a 20/05/98 Resultados sobre 60 precios de mercado; 30 calls (15 bid, y 15 ask ), 30 puts (15 bid y 15 ask ), vencimiento 20/6/98. Subyacente en el nivel 1119.10. Par ametros Anualizados En t p 0.21 0.21 M1 -76.3% -6.48% 1 18.90% 5.51% 1p 0.79 0.76 m2 26.0% 2.20% 2 10.64% 3.10%
Tipo de inter es estimado: 5.23 %. Errores absolutos: 5.29. Errores relativos: 23.75%. La m axima diferencia absoluta en precio se produce para el strike 1150 de 1.15: precio real 33.88, precio mixtura 35.03.
La m axima diferencia relativa en precio se produce para el strike 1140 de 7.02%: precio real 14.13 , precio mixtura 13.13.
59
D a 21/05/98 Resultados sobre 54 precios de mercado; 28 calls (14 bid, y 14 ask ), 28 puts (14 bid y 14 ask ), vencimiento 20/6/98. Subyacente en el nivel 1114.6 Par ametros Anualizados En t p 0.35 0.35 M1 -49.3% -4.06% 1 17.48% 5.01% 1p 0.65 0.65 m2 31.7% 2.60% 2 8.99% 2.58%
Tipo de inter es estimado: 3.71 %. Errores absolutos: 6.07. Errores relativos: 30.92%. La m axima diferencia absoluta se produce para el strike 1115 con -1.34: precio real 23.88, precio mixtura 22.54.
La m axima diferencia relativa se produce para el strike 1150 con 10.57%: precio real 52.50, precio mixtura 51.32.
Las gr acas de ajuste en precios y smile son similares a las de los d as anteriores. Estimados los par ametros para tres d as, se puede analizar el cambio en la distribuci on, y el cambio en el smile.
4.4
Resultados para estimaciones en opciones sobre el futuro de Ibex-35

D as: 18/05/98, 19/05/98, 20/05/98, 21/05/98 Fuente: Datos de cierre de MEFF de opciones call y put sobre el futuro Ibex-35 Vto 19/06/98
Se han probado tambi en para el IBEX los diferentes m etodos de minimizaci on: cuadrados de errores absolutos, relativos, variaciones logar tmicas. . . No obstante, el hecho de que el spread (de horquilla) en t erminos absolutos permanezca constante en algunos activos, hace que las diferencias relativas bid /mid, ask /mid en los precios out of the money sean muy altas (frente a los precios in y at the money ), y en este caso ponderan mucho m as en la funci on de error, lo que proporciona un mejor ajuste
60
en la zona muy fuera del dinero y un ajuste muy malo en el resto de las zonas de precios cuando se utilizan criterios de minimizaci on relativos. Este ha sido el caso del IBEX, por lo que se ha preferido como criterio de minimizaci on la suma de los errores absolutos cuadr aticos. En el caso en el que el subyacente es un futuro, la f ormula de valoraci on BlackScholes es ligeramente diferente de la habitual. Asumiendo que el comportamiento del futuro sobre un subyacente es id entico al del mismo subyacente pagando dividendos a una tasa igual al tipo libre de riesgo, la f ormula de valoraci on (B-S) queda en C P = er (T t) F N (d1 ) K N (d2 ) , = er (T t) K N (d2 ) F N (d1 ) ,
para la call y la put respectivamente, donde d1 = log(F/K ) + ( 2 /2) (T t) , T t d2 = log(F/K ) ( 2 /2) (T t) . T t
En el caso de la valoraci on en entorno mixtura, se utiliza la f ormula habitual: C (K ) = pero donde ahora es C1 es el precio de la call de strike K bajo Black-Scholes con m1 como tipo de inter libre de riesgo y 1 como volatilidad. C2 es el precio de la call de strike K bajo Black-Scholes con m2 como tipo de inter es libre de riesgo y 2 como volatilidad. F es el precio del futuro a fecha de valoraci on. K es el strike. Por tanto C1 tiene la forma17 :
1 C1 = er (T t) F N (d1 1 ) K N (d2 ) ,
p m1 Deltat C1 (K ) + (1 p) em2 t C2 (K ) , p em1 t + (1 p) em2 t
donde d1 1 d1 2
17 La
= =
2 log(F/K ) + (m1 r 1 /2) (T t) , T t 2 log(F/K ) (m1 r 1 /2) (T t) . T t
expresi on para C2 es an aloga a la de C1 .
61
De forma an aloga se obtiene la f ormula para la put 18 . Como se ve, se sigue b asicamente en la l nea de la teor a Black-Scholes, con muy pocas modicaciones. Se aplica la formula de valoraci on obtenida a los precios de opciones con subyacente el contrato de futuros de MEFF Ibex-35-PLUS. Como se ha dicho, el m etodo de estimaci on de los par ametros de la mixtura es minimizar la funci on de errores absolutos Ea :
n
Ea (m1 , m2 , 1 , 2 , p) =
i=1
C (Ki ) ci
+
i=1
P (Ki ) pi
donde ahora el precio te orico se obtiene con la f ormula anterior (v alida para el futuro). Sin embargo, ahora r aparece de forma expl cita en la expresi on que se desea minimizar, en Ea hay un par ametro m as que antes. El problema se trata de dos formas: 1. Se estima r junto con el resto de los par ametros de la mixtura, sin ninguna restricci on; esto en general conduce a un valor de r muy cercano a cumplir la condici on de riesgo neutral. 2. Se estiman los par ametros imponiendo r tal que cumpla la condici on riesgo neutral (se ja r a un nivel). Si se analiza c omo depende la funci on de objetivo del tipo de inter es r, se observa que - Habitualmente se obtiene una estimaci on de r en el intervalo (1.5%, 4%). - Por lo general la suma de errores cuadr aticos es muy poco sensible a variaciones de r en dicho intervalo. - Cuando se deja r libre, se suele estimar un tipo de inter es ligeramente por debajo de la condici on riesgo-neutro, aunque, en t erminos de errores absolutos, hay muy poca diferencia en la estimaci on entre imponer el tipo de inter es r que cumple la condici on riesgo neutral o dejarlo libre. En general, se ha tomado la serie entera de precios publicados teniendo en cuenta que, en primer lugar, son precios de cierre19 y, en segundo lugar, el m etodo de c alculo de precios de cierre en MEFF es u ltimo precio cruzado salvo en precios muy ITM y OTM donde no suele haber liquidez, y entonces los precios publicados son solamente indicativos, por lo que la mayor a de las veces el mercado no est a en esos niveles; esto puede crear algunas dicultades en las estimaciones.
18 El hecho de considerar que la tasa r de dividendos es igual en cada una de las dos lognormales que constituyen la mixtura es claramente arbitrario; se puede pensar que los dividendos son diferentes para ambas normales, manteniendo que la tasa esperada de dividendos sea r. La ventaja de suponer que es igual es que s olo se a nade un par ametro m as al modelo, y adem as, como veremos, es suciente para valorar bien las opciones. 19 Con esto se quiere hacer especial hincapi e que no son precios s ncronos.
62
D a 18/05/98 Estimaciones nales sobre 225 precios de cierre; 102 call s y 123 puts. Sin condici on Riesgo-neutro (SinRN) Hay varias soluciones muy cercanas seg un, sobre todo, el dato inicial de r con el que comienzan las iteraciones. Se ha barrido una banda grande de par ametros iniciales, obteni endose los siguientes m nimos relativos para los par ametros que a continuaci on se detallan (Eabs es la suma de errores absolutos cuadr aticos y Erel es la suma de errores relativos cuadr aticos):
p 0.276 0.281 0.281 m1 -110.56% -110.83% -111.07% 1 39.19% 38.72% 38.67% 1p 0.724 0.719 0.719 m2 44.59% 44.14% 44.13% 2 24.21% 24.07% 24.06% Tipo 3.84% 2.68% 2.64% RN 3.86% 2.70% 2.66% Eabs 116.20 115.98 115.98 Erel 0.961 0.920 0.916
Suma total de precios mercado: 107,853. La relaci on entre el error absoluto y la suma de precios es: 115.98/107, 853 = 0.00108. Con condici on Riesgo-neutro (ConRN) Se impone en la minimizaci on la condici on riesgo-neutral: R= log(p em1 t + (1 p) em2 t ) . t
As se obtiene una u nica soluci on para cualquier valor inicial de los par ametros:
p 0.282 m1 -110.70% 1 38.71% 1p 0.718 m2 44.16% 2 24.04% Tipo 2.63% RN 2.63% Eabs 116.02 Erel 0.913
El ratio entre el error absoluto y la suma de precios es: 116.02/107, 853 = 0.0010753. Con los par ametros obtenidos, se puede dibujar la distribuci on de la mixtura y compararla con la de una normal de igual media y varianza, o bien dibujar la mixtura junto con sus gaussianas constituyentes (gura de la derecha):
63
Se puede as mismo ver el smile de mercado frente al smile te orico dado por la mixtura.
D a 19/05/98 Estimaciones sobre 176 datos, 119 call s y 57 puts. Se han obtenido u nicamente dos soluciones para las dos estimaciones (sin condici on RN, con condici on RN) que se detallan a continuaci on: Sin restricci on p 0.276 Tipo 1.61% m1 -108.56% RN 3.00% 1 38.25% Eabs 78.734 1p 0.724 Erel 1.6106 m2 42.20% Eabs /total 0.001040 2 23.93%
Con restricci on p 0.268 Tipo 3.00% m1 -109.04% RN 3.00% 1 38.62% Eabs 79.08 1p 0.732 Erel 1.5916 m2 42.73% Eabs /total 0.001045 2 24.16%
La suma total de precios reales de mercado es 75.699. Con cualquiera de las dos soluciones, el ajuste tanto en precio como en el smile es pr acticamente igual. Hay que recordar que se est a utilizando el mismo tipo de inter es, tanto para valorar bid como el ask. En este sentido, la estimaci on est a proporcionando un cierto tipo
64
medio, el tipo medio que aplicado tanto a la compra como a la venta de opciones ajusta mejor, minimiza al m aximo el error respecto de los precios bid y ask. Es por esto por lo se han utilizado todos los precios, tanto bid como ask, para hallar la distribuci on impl cita, aunque evidentemente se cometa siempre el error de la horquilla, porque en realidad el mercado deber a tender a disminuir las horquillas conforme aumente la liquidez, y as la distribuci on bid y la distribuci on ask deber an tender hacia la estimada con todos los precios; pero todo esto, con unas horquillas tan grandes en determinados precios, (sobre todo en niveles relativos) puede ser causa de inestabilidades en los tipos de inter es estimados. D a 20/05/98 242 precios de opciones (112 call s, 130 puts ). Se obtienen las siguientes estimaciones: Sin restricci on p 0.343 0.330 0.327 0.325 Tipo -1.75% 1.00% 1.50% 2.00% m1 -94.58% -92.74% -92.48% -92.15% RN -2.19% 1.18% 1.79% 2.40% 1 33.90% 34.07% 34.10% 34.14% Eabs 93.313 93.636 93.763 93.911 1p 0.657 0.670 0.673 0.675 Erel 4.9585 5.0519 5.0680 5.0854 m2 43.30% 44.89% 45.17% 45.45% Eabs /total 0.000825 0.000828 0.000829 0.000830 2 20.89% 21.07% 21.11% 21.14%
Con restricci on p 0.336 0.328 0.325 0.321 Tipo -0.14% 1.00% 1.50% 2.00% m1 -93.25% -93.56% -93.74% -94.05% RN -0.14% 1.00% 1.50% 2.00% 1 33.89% 34.31% 34.50% 34.68% Eabs 93.45 93.76 94.10 94.55 1p 0.664 0.672 0.675 0.679 Erel 5.0129 5.0491 5.0676 5.0852 m2 44.33% 44.65% 44.78% 44.91% Eabs /total 0.000826 0.000829 0.000832 0.000836 2 20.99% 21.10% 21.14% 21.19%
Como se ve, para cualquiera de las soluciones anteriores la variaci on de los par ametros es muy peque na. Se puede dibujar la distribuci on de la mixtura frente sus gaussianas constituyentes.
65
D a 21/05/98 Estimaciones realizadas sobre 230 precios (110 call s, 120 puts ). Se obtienen dos u nicas soluciones, una sobre minimizaci on libre, y otra con restricci on RN. p 0.305 Sin restricci on Tipo 0.76% p 0.307 Con restricci on Tipo 0.40% m1 -87.09% RN 0.86% m1 -87.25% RN 0.40% 1 36.03% Eabs 110.93 1 35.98% Eabs 111.04 1p 0.695 Erel 0.2174 1p 0.693 Erel 0.2027 m2 37.54% Eabs /total 0.000987 m2 37.38% Eabs /total 0.000988 2 21.25%
2 21.16%
66
4.5
Comentarios nales respecto a la estimaci on de las distribuciones impl citas
Es conveniente, antes de terminar con el cap tulo de estimaci on de densidades, mostrar en m as detalle los problemas ya mencionados que en la estimaci on introduce el diferencial de precios bid -ask. Recordemos que se han probado b asicamente tres tipos de estimaciones, dependiendo de cu al vaya a ser la funci on objetivo que se minimice: a saber, la suma de los cuadrados de (1) las diferencias absolutas de los precios, (2) las diferencias porcentuales de los precios o (3) las diferencias de los logaritmos de los precios. En general, el primer criterio ajusta mejor el precio para las opciones los precios in the money, los otros dos ajustan mejor los precios out of the money. En general se ha utilizado el criterio 1, minimizaci on de la suma de cuadrados de diferencias de precios, ya que se ha observado que este criterio es m as robusto que los otros dos. Trabajar con criterios relativos como son las diferencias porcentuales o las diferencias de logaritmos tiene el problema de los spreads entre bid y ask. Es habitual que los spreads entre bid y ask para las opciones sean m as o menos constantes independientemente del strike ; esto hace que para las opciones out of the money la relaci on entre el spread (ask bid ) y el precio medio de la opci on (bid +ask )/2 sea incluso superior al 100%. En tal caso, intentar minimizar, por ejemplo, la suma de cuadrados de las diferencias porcentuales ocasionar a que el ajuste de los precios at o in the money sea malo. A continuaci on se presenta la relaci on entre el spread de las opciones y el precio medio para diferentes strikes.
Se ve que el ratio de spread -precio en las opciones sobre el S&P es muy estable (compar andolo al IBEX 35); ello principalmente es debido a que en la muestra de S&P no hay opciones muy dentro o fuera del dinero.
67
De hecho, la estimaci on de la densidad impl cita del subyacente para las opciones sobre el S&P por cualquiera de los tres m etodos proporciona resultados muy similares, como se puede ver en la gr aca adjunta (densidad impl cita del S&P el 21/5/98 obtenida a partir de precios de opciones con vcto. 20/6/98), cosa que desde luego no sucede trabajando con el IBEX35, donde la minimizaci on bien de variaciones porcentuales, bien de variaciones logar tmicas, ocasiona estimaciones completamente diferentes y que adem as ajustan p esimamente salvo para el tramo muy fuera del dinero20 .
Valoraci on de opciones con la mixtura
Sup ongase que se desea valorar una opci on digital seg un se presenta en el gr aco adjunto. Como dato, es conocido el precio de la call para el mismo strike que la opci on digital. Qu e precio hay que dar a la opci on digital? La respuesta inmediata ser a calcular mediante un modelo de tipo B-S el precio de la opci on digital, para lo cual, qu e volatilidad impl cita usar? La correspondiente a la call de igual strike ? Se ver a que en general la respuesta es no.
20 Existe otro criterio de minimizaci on m as complejo: se trata de la suma de los cuadrados de las diferencias de precios entre el precio estimado y la banda de spread . Si el precio estimado cae dentro de la banda la diferencia es cero; si el precio cae fuera, la diferencia se toma respecto al extremo de la banda mas cercano al precio. Se han hecho algunas pruebas con este criterio sobre opciones S&P y las diferencias con respecto al criterio normal de m nimos cuadrados han sido muy peque nas. El problema de este cuarto criterio es que en general proporciona m ultiples soluciones.
68
Recordando la expresi on que determina la volatilidad impl cita para una opci on call de strike K , c(K ) , es tal que se cumple:
Precio
call
=
K
(x K ) f (x) dx =
(x K ) y(K ) (x) dx .
De la misma manera se puede denir la volatilidad impl cita de una opci on digital de strike K : D(K ) es tal que se cumple

Precio
digital
=
K
D f (x) dx =
K
D y(K ) (x) dx .
Es evidente que, en general, c(K ) = D(K ) . Es decir, la volatilidad impl cita que se debe usar para valorar una opci on digital es diferente de la volatilidad impl cita derivada a partir del precio de una opci on call, y esto es as porque la volatilidad impl cita es un concepto que no s olo depende del strike, sino del tipo de opci on. En cierto sentido es un concepto como el de TIR, que depende de caracter sticas del instrumento, del bono, de la opci on21 . Mediante el uso de las mixturas es posible solventar este problema, se puede estimar directamente la densidad del subyacente a partir del precio de las opciones call y put para diferentes strikes, utilizando como aproximaci on la mixtura de gaussianas. Una vez estimada la mixtura que mejor aproxima a la distribuci on real f (x) del subyacente se puede calcular el precio de cualquier otro tipo de opci on, por ejemplo una opci on digital (binaria, cash or nothing ), integrando directamente sobre la mixtura. No obstante, en relaci on con lo anterior, la valoraci on en mercado de cualquier derivado utilizando el smile producido por las calls y las puts es una pr actica habitual. A continuaci on se puede ver un ejemplo de valoraci on de una opci on digital sobre el ndice S&P 500 y el IBEX 35, y los errores que produce utilizar la volatilidad impl cita de las calls y puts de igual strike. En los ejemplos comparativos que se presentan a continuaci on, se ha valorado una opci on digital por B-S utilizando como volatilidad impl cita la correspondiente a los calls y puts, de igual strike 22 , y por otro lado se ha valorado esta opci on con la mixtura (usando la densidad estimada a partir de calls y puts, v ease en el anexo B la derivaci on del precio de la digital en el entorno mixtura).
21 Aunque en general es cierto lo antedicho, existen algunos casos particulares que parecen saltarse dicha regla; as , por ejemplo, las opciones call y put s que cumplen que para igual strike la volatilidad impl cita es igual para ambas. Esto es debido a que debe cumplirse la paridad call -put, tanto en el entorno B-S como en el de mixturas, por lo que debe suceder que C + K er t = P + S , de donde se deduce que C es igual a P . 22 El tipo de inter es que se ha introducido en la f ormula de B-S ha sido el que surge de la condici on riesgo-neutro tras la optimizaci on.
69
Datos del S& P d a 18/5/98 Se obtienen los par ametros para la mixtura con los que se valorar a la opci on digital.
Par ametros p 0.24 m1 -71.6% 1 19.89% 1p 0.76 m2 26.3% 2 10.98%
Por otro lado, el smile de mercado del d a 18/5/98 para el S&P que se va a utilizar para valorar la opci on por Black-Scholes se presenta en la siguiente tabla:
strike volat. strike volat. 1075 19.96% 1115 16.98% 1085 19.21% 1120 16.65% 1090 18.82% 1125 16.33% 1095 18.44% 1130 16.03% 1100 18.06% 1135 15.75% 1105 17.69% 1140 15.48% 1110 17.33% 1150 15.01%
Las diferencias obtenidas en precio se representan en el gr aco siguiente: el eje de abcisas representa el nivel de strike de una opci on digital que paga 1 cuando el subyacente queda por debajo del strike, y el eje de ordenadas el precio.
El siguiente gr aco presenta las diferencias porcentuales entre ambos precios:
Cuando m as cerca se est a del vencimiento, obviamente las dos valoraciones convergen, pero seg un aumenta el plazo hasta el vencimiento se producen las mayores diferencias entre un modelo y el otro. Como ejercicio te orico, si se realiza esta valoraci on con exactamente la misma situaci on en cuanto a volatilidades impl citas, pero para un plazo residual de tres meses para vencimiento, se comprueba que se producen discrepancias de hasta 25$ sobre 100 en los strikes m as ATM.
70
Datos del S& P d a 19/5/98 Los par ametros de la mixtura estimados para ese d a y el smile de mercado son:
strike volat. strike volat. 1075 20.17% 1115 16.99% 1085 19.40% 1120 16.62% 1090 19.00% 1125 16.26% 1095 18.59% 1130 15.93% 1100 18.18% 1135 15.61% 1105 17.77% 1140 15.32% 1110 17.37% 1150 14.79%
Se obtiene de nuevo la valoraci on por las dos v as:
Datos del IBEX 35 d a 18/5/98 Los par ametros de la mixtura para el IBEX 35 son:
p 0.282 m1 -110.70% 1 38.71% 1p 0.718 m2 44.16% 2 24.04%
El smile utilizado para valorar con B-S se representa en la siguiente tabla:

strike 8150 8200 8250 8300 8400 8500 8550 8700 8800 8900 9000 9100 9150 9200 9250 9300 volatilidad 43.01% 42.83% 42.64% 42.44% 42.01% 41.54% 41.30% 40.50% 39.92% 39.32% 38.71% 38.08% 37.76% 37.44% 37.12% 36.81% strike 9350 9400 9450 9500 9600 9650 9700 9750 9800 9850 9900 9950 10000 10050 10100 10150 volatilidad 36.50% 36.19% 35.88% 35.58% 34.99% 34.71% 34.43% 34.16% 33.90% 33.64% 33.39% 33.15% 32.92% 32.69% 32.48% 32.27% strike 10200 10250 10300 10350 10400 10500 10600 10700 10800 10850 10900 11000 11200 11350 volatilidad 32.07% 31.87% 31.68% 31.50% 31.33% 31.00% 30.70% 30.42% 30.17% 30.05% 29.93% 29.72% 29.34% 29.09%
71
Finalmente, se obtienen las diferencias en precio que se presentan a continuaci on:
Cobertura de la sonrisa y la mueca de volatilidad
En el modelo de B-S, uno de los supuestos b asicos es que la volatilidad es constante (no depende del precio de ejercicio); sin embargo, como ya hemos visto, en el mercado se observa que esto no es as , que existen la sonrisa y la mueca de volatilidad. A pesar de eso, se suele utilizar el modelo de B-S, tomando diferentes volatilidades impl citas para diferentes precios de ejercicio, calculando diferentes precios y griegas para diferentes precios de ejercicio. En particular se considera en el entorno B-S la vega como una aproximaci on o indicador del riesgo que se corre ante cambios en la volatilidad impl cita. Sin embargo, como hemos visto, no existe una volatilidad impl cita, sino una curva de volatilidades impl citas (m as correctamente, una supercie de volatilidades impl citas si tenemos en cuenta la dimensi on temporal el plazo de las opciones) que depende del strike. En este contexto, qu e sentido tiene la vega de una cartera de opciones con igual subyacente y diferentes precios de ejercicio? En tal caso, la vega mide el riesgo ante movimientos paralelos en la curva de volatilidades impl citas, es decir, incrementos o decrementos iguales en las volatilidades impl citas para todos los precios de ejercicio. Pero, qu e pasa si el movimiento no es paralelo? Si el cambio es en la forma de la curva de volatilidades impl citas, la cobertura vega que dicta el entorno B-S es inadecuada; es el caso, por ejemplo, de rotaciones de la curva de smiles .
72
En el modelo de opciones para mixturas, el smile no es m as que una consecuencia de que la distribuci on del subyacente no es lognormal, y es en general leptoc urtica y asim etrica. Ahora, con la mixtura, en lugar de los dos par ametros (media y volatilidad) del modelo B-S, se tienen 5, dos medias, dos volatilidades y una probabilidad23 . La contrapartida es que con esos cinco par ametros se modeliza toda la curva de volatilidades. En este contexto las griegas del entorno B-S sufren ligeras transformaciones. Ahora, adem as de Delta, Gamma y Theta aparecen dos Vegas, dos Rhos y Psi (en el anexo A aparecen todas las f ormulas de valoraci on y las derivadas en el entorno mixtura as como su relaci on con las f ormulas en el entorno B-S).
6.1
Las derivadas con mixturas de opciones
En el entorno mixtura, las derivadas son ocho (tres m as que en el entorno B-S): Derivada del precio de la opci on respecto del subyacente (Delta). Derivada de la Delta de la opci on respecto del subyacente (Gamma). Derivada del precio de la opci on respecto de la volatilidad de una de las dos normales que com ponen la mixtura (Vega 1). Derivada del precio de la opci on respecto de la volatilidad de la otra de las normales que componen la mixtura (Vega 2). Derivada del precio de la opci on respecto de la media de una de las normales que componen la mixtura (Rho 1). Derivada del precio de la opci on respecto de la media de la otra de las dos normales que componen la mixtura (Rho 2). Derivada del precio de la opci on respecto del tiempo (Theta)24 . Derivada del precio de la opci on respecto de la probabilidad que interviene en la mixtura (Psi). Las tres nuevas derivadas que aparecen ahora tienen que ver precisamente con el smile y su cobertura; de hecho, los par ametros que caracterizan el smile en el entorno mixtura son las dos volatilidades (1 y 2 ), la distancia entre las dos medias (m1 m2 )
realidad el modelo de B-S, tal y como se utiliza en el mercado, requiere muchos m as de dos par ametros: requiere tantas volatilidades como strikes se coticen; si hay n strikes cotizados, el modelo requiere n + 1 par ametros. 24 En el art culo se presenta la mixtura como una aproximaci on de la distribuci on del subyacente para un plazo determinado, sin embargo no se ha determinado un proceso estoc astico para modelizar la evoluci on en el tiempo del subyacente. En este sentido, el modelo de valoraci on es est atico, proporciona valoraciones para un plazo y nada dice de cu anto vale la opci on para plazos diferentes. En tanto en cuanto no se d e dicho proceso, no tiene sentido hablar de evoluci on del precio de las opciones en el tiempo ni tiene sentido por tanto hablar de Theta.
23 En
73
y la probabilidad (p). Al igual que en el entorno B-S, es posible cubrir los riesgos minimizando al m aximo las distintas derivadas de las opciones. Como se ha visto, el modelo de mixtura contempla perfectamente la existencia del smile ; sin embargo, el modelo descrito es est atico, describe las opciones para un vencimiento dado pero no sirve para, conociendo el valor de las opciones para un determinado vencimiento, valorar otras opciones de diferente vencimiento. En todo momento se ha supuesto que existe un determinado proceso estoc astico tal que la distribuci on del subyacente que sigue dicho proceso es una mixtura de gaussianas en un instante t del tiempo. Pero, qu e pasa antes y despu es de t? Con el modelo que aqu se presenta no es posible armar nada25 y por tanto nada se puede decir acerca de c omo es Theta (derivada de la opci on respecto del tiempo) en este contexto. A pesar de lo anterior, en el anexo A se presenta una Theta que corresponde a una pura extrapolaci on de la Theta del modelo B-S al entorno mixtura, que a pesar de la forma tan sui generis como se ha calculado parece funcionar bien. Para comprobar hasta qu e punto sirven las derivadas en el entorno mixtura se han realizado dos ejercicios. El primero ha consistido en comprobar c omo funciona una aproximaci on de primer orden para el precio de las opciones. Con datos de tres d as de cotizaciones de opciones sobre el ndice S&P se han estimado todos los par ametros de las mixturas que mejor ajustan y se han calculado las derivadas (en entorno mixtura). A continuaci on, utilizando el precio del primer d a como base, se han calculado los precios para los otros d as usando la siguientes aproximaci on de primer orden (salvo en el subyacente): Preciot+1 Preciot + Precio = Preciot + Delta S + 1 Gamma (S )2 2 +Vega2 2 + Rho1 m1 + Rho2 m2 + Psi p2 + Theta t .
As , se ha comparado el precio real con el precio obtenido a partir de la aproximaci on; los resultados se pueden ver en las dos gr acas siguientes: la primera gr aca presenta el precio de las opciones call sobre el S&P el d a 18/05/98 (vcto. 19/06/98).
25 Los autores est an trabajando en esta l nea, intentando caracterizar el proceso estoc astico que da lugar a que la distribuci on del subyacente sea una mixtura para cualquier instante del tiempo. Todo parece indicar que existe m as de un proceso con dichas caracter sticas y lo ideal ser a identicar, de todos ellos, aqu el que no s olo explica el smile sino que adem as se ajusta a la estructura temporal de las volatilidades impl citas.
74
La segunda presenta los smiles de los tres d as: el smile del 18/05/98 es la l nea discontinua, el d a 20 sufri o una ligera rotaci on y el d a 21 casi una traslaci on paralela.
En el u ltimo gr aco se presenta, en l nea discontinua, las diferencias reales de precios de las opciones call entre el 18/05/98 y los d as 20 y 21. La l nea continua representa las diferencias de precios calculadas a partir de la expansi on de Taylor cuasilineal (salvo en el subyacente) antes comentada. Como se ve, las diferencias entre ambas curvas, real y aproximada a trav es de las derivadas, son m nimas; esto indica que las derivadas de la mixtura funcionan bien.
El segundo ejercicio ha consistido en construir, en el momento t, la cobertura a una posici on en opciones. Concretamente, se deseaba cubrir una posici on de 1.000.000 de opciones at the money tomando posiciones en otras cuatro opciones de diferente strike. En el cuadro adjunto se ve el resultado del ejercicio.
strike 1090 1100 1105 1110 1115 1120 1125 1130 1135 1140 1150 Posici on en 980,869 -1,608.442 Valor 18/5/98 35,067.628 -51,888.619 Importe en ptas. Variaci on 20/5/98 Variaci on 21/5/98 5,230.504 250.929 -8,121.848 -277.499
1,000.000
22,742.391
4,084.510
-202.959
-916.806 539.575 Importe total
-11,700.632 5,828.180 22,742.391
-2,448.559 1,257.602 2.209
724.767 -484.673 20.565
75
La variaci on de la opci on at the money (strike 1115) hubiera sido, sin cobertura, de +4.084.510 y 202.959; con la cobertura tan s olo de 2209 y 10565. Como se ve, la cobertura ha funcionado perfectamente26 .
Anexo A: precio y derivadas en el entorno mixtura

Aqu se presenta en detalle el precio y las derivadas de una opci on bajo el supuesto de mixtura. El valor de la opci on, de tipo call por ejemplo, en Black-Scholes (B-S) ser a: C = S N (d1 ) X er t N (d2 ) , en donde d1 y d2 vienen dados por d1 d2 = = log(S/X ) + r + 2 /2 t , t log(S/X ) + r 2 /2 t = d1 t . t
En el caso de usar una mixtura de normales el precio de la opci on, C , vendr a dado por: p em1 t C1 + (1 p) em2 t C2 C= , p em1 t + (1 p) em2 t en donde, para el caso de nuevo de la opci on call por ejemplo: C1 C2 m1 m2 d1 1 d1 2 d2 1 d2 2
m1 t = S N (d1 N (d1 1Xe 2) , 2 m2 t = S N (d1 X e N (d2 2) , 2 1 , = 1 + 2 2 = 2 + 2 , 2 2 log(S/X ) + (m1 + 1 /2) t = , 1 t 2 log(S/X ) + (m1 1 /2) t = t , = d1 1 1 1 t 2 log(S/X ) + (m2 2 /2) t = , 2 t 2 log(S/X ) + (m2 + 2 /2) t = t . = d2 1 2 2 t
26 La cobertura se ha calculado mediante la opci on solver de EXCEL 5.0, buscando la posici on en las otras cuatro opciones que minimiza la variaci on esperada de la cartera. Por otra parte, ser a muy interesante realizar el mismo ejercicio para varios d as seguidos (al menos durante un mes); en tal caso se ver a mucho mejor la efectividad de la cobertura.
76
Tambi en se puede expresar el precio de la opci on como Cm = p C1 + (1 p ) C2 , en donde p = p el1 t , m1 r = l1 .
Como se ve, conociendo la f ormula de valoraci on de la opci on en entorno B-S, es posible conocer la f ormula de valoraci on en entorno mixtura; no es m as que una combinaci on lineal en donde los pesos de la combinaci on son una transformaci on sencilla de la probabilidad de la mixtura.
A.1 Las griegas con mixturas

Veamos el caso particular de opciones tipo call bajo entorno B-S o bajo entorno mixtura (se deja al lector el ejercicio de calculo de las derivadas para opciones tipo put, digitales,. . . ). Delta: La delta de una opci on es la derivada del precio respecto del subyacente. = C . S
En el entorno B-S se ten a que = N (d1 ). En el entorno Mixtura, la delta de la opci on es, como el precio, la combinaci on lineal de dos deltas: m = p 1 + (1 p ) 2 ,
2 en donde p = p el1 t , m1 r = l1 , 1 = N (d1 1 ) y 2 = N (d1 ).
Gamma: La gamma de una opci on es la derivada de la delta respecto del precio: = . S
En el entorno B-S, la gamma viene dada por = N (d1 ) . S t
En el entorno mixtura, la gamma es una combinaci on lineal de dos gammas: m = p 1 + (1 p ) 2 , donde p = p el1 t , m1 r = l1 , 1 = N (d1 1) , S 1 t 2 = N (d2 1) . S 2 t
77
Vega: La vega de una opci on es la derivada del precio respecto de la volatilidad = C .
En el entorno B-S la vega viene dada por: = S t N (d1 ). En el entorno mixtura, no existe una u nica vega, existen dos, una por cada normal constituyente de la mixtura: 1 2 m = p 1 , m = (1 p ) 2 , donde p = p el1 t , m1 r = l1 , 1 = S t N (d1 1) , 2 = S t N (d2 1) .
Rho: La Rho de una opci on es la derivada del precio respecto del tipo de inter es Rho = C . r
En el entorno B-S la rho viene dada por Rho = X t er t N (d2 ). En el entorno mixtura no existe una rho, existen dos: realmente existen las derivadas del precio de las opciones respecto de las dos medias de las normales constituyentes de las mixturas. Rho1 m = p (Rho1 + C1 t) , en donde p = p el1 t , m1 r = l1 y Rho1 = C1 = X t em1 t N (d1 2) , m1 Rho2 = C2 = X t em2 t N (d2 2) . m2 Rho2 m = (1 p ) (Rho2 + C2 t) ,
Psi: La Psi de una opci on es la derivada del precio respecto de la probabilidad p: = C . p
En el entorno B-S, no existe esta derivada. En el entorno mixtura, Psi viene dada por m = el1 t (C1 C2 ) , en donde m1 r = l1 . Theta: La theta de una opci on es la derivada del precio respecto del tiempo: = En el entorno B-S, theta viene dada por: = S N (d1 ) r X er t N (d2 ) . 2 t C . t
78
En el entorno mixtura, theta no es exactamente una combinaci on lineal, existe un t ermino adicional: m = p 1 + (1 p ) 2 + l1 p , (C2 C1 ) , en donde m1 r = l1 y 1 2 S N (d1 ) 1 1 m1 X em1 t N (d1 2) , 2 t S N (d2 ) 1 2 m2 X em2 t N (d2 = 2) . 2 t =
Anexo B: el precio de las digitales en el entorno mixtura

El valor de la opci on digital en Black-Scholes (B-S) es: C = X er t N (d2 ) , en donde d2 viene dado por: d2 = log(S/X ) + (r 2 /2) t . t
En el caso de usar una mixtura de normales, el precio de la opci on, C , viene dado por: p em1 t C1 + (1 p) em2 t C2 Cm = , p em1 t + (1 p) em2 t en donde, para el caso de nuevo de la opci on call, m1 m2 d1 2 d2 2 = = = =
2 1 , 2 2 2 + 2 , 2 2 log(S/X ) + (m1 1 /2) t , 1 t 2 log(S/X ) + (m2 2 /2) t . 2 t
1 +
Tambi en se puede expresar el precio de la opci on como Cm = p C1 + (1 p ) C2 , en donde p = p el1 t y m1 r = l1 .
79
Referencias
[1] Alexander, Carol y P. Williams, Modelling the term structure of kurtosis: A comparision of neural network and GARCH methods, 1.997. [2] Apuntes del seminario, La gesti on de la volatilidad: Valoraci on, Predicci on y Trading, organizado por Instituto MEFF y la Escuela Europea de Finanzas. Enero 1.998 [3] Baxter, Martin y Andrew Rennie, Financial Calculus, an introduction to derivative pricing. Cambridge University Press, 1.996. [4] Diebolt, Jean y Christian Robert, Bayesian estimation of nite mixture distributions. Part I: Theoretical aspects. L.S.T.A. Universit e Paris VI. [5] Diebolt, Jean y Christian Robert, Bayesian estimation of nite mixture distributions. Part II: Sampling implementation. L.S.T.A. Universit e Paris VI. [6] Due, Darrell. Security Markets, Stochastic Models, Academic Press. [7] Due, Darrell, Dynamic Asset Pricing Theory. Second edition, Princeton University Press, 1.996. [8] Dupire, Bruno. Pricing and hedging with smiles. Mathematics of derivative securities. Cambridge University Press, 1.997. [9] Feelders, A.J., Learning from biased data using mixture models. Data Distilleries Ltd., Amsterdam, The Nederlands. [10] Green, Peter J. y Sylvia Richardson, On Bayesian analysis of mixtures with an unknown number of components, presented at the Workshop on Model Interpretation and Model Robutness in Highly Structured Stochastic Systems. Revised version, October 1.996. [11] Hamilton, J. A quasi-Bayesian approach to estimating parameters for mixtures of normal distributions. Journal of Business and Economic Statistics 9 (1991), no. 1, pp. 2739. [12] Haykin, Simon. Neural Networks: A comprehensive foundation. Macmillan Publishing Company, Inc. 1.994. [13] Hendricks, Darryll, Evaluation of Value at Risk models using historical data. Economic Policy Review, Federal Reserve Bank of New York 2, April 1.996, pp. 3969. [14] Hull, John y Alan White, Value at Risk when daily changes in market variables are not normally distributed forthcoming Journal of Derivatives, Noviembre 1.997.
80
[15] Hull, John y Alan White, Evalutating the impact of Kurtosis and skewness on derivative prices forthcoming Net Exposure, Noviembre 1.997. [16] Hull, John, Options, futures and other derivative securities, Third Edition, Prentice Hall. [17] J.P. Morgan/Reuters, RiskMetricsTM , Technical Document, Fourth Edition, New York, 1.996. [18] Lamberton D. y B. Lapeyre, Introduction to Stochastic Calculus applied to Finance. Chapman & Hall, 1.996. [19] Robert, Christian P, Mixtures of distributions: Inference and estimation. Dept. de Math., Universit e de Rouen et Crest, Insee, Paris. [20] Roeder, Kathryn y Larry Wasserman, Practical Bayesian density estimation using mixtures of normals. Carnegie Mellon University, 1.995 [21] Stephens, Matthew, Bayesian methods for mixtures of normal distributions. PhD Thesis, University of Oxford, 1.997. [22] Subu Venkataraman, Value at Risk for a mixture of normal distributions: The use of quasi-bayesian estimation techniques Economic Perspectives, Federal Reserve Bank of Chicago. [23] Zangari, P. An improved methodology for measuring VaR RiskMetrics Monitor, Reuters/JP Morgan, 1.996. [24] Zangari, P. What risk managers should Know about mean reversion and jumps in prices RiskMetrics Monitor, Reuters/JP Morgan, 1.997.
Susana Corcuera Gesti on Global de Riesgo de Mercado BBVA Paseo de Recoletos, 12, planta 2 28001-Madrid, Espa na e-mail: scorcuera@grupobbva.com
Juan Carlos Garc a C espedes Gesti on Global de Riesgo de Mercado BBVA Paseo de Recoletos, 12, planta 2 28001-Madrid, Espa na e-mail: jcgarcia@grupobbva.com
La problem atica de medici on de la performance

ndez1 Prosper Lamothe Ferna
Resumen En este trabajo se realiza un an alisis comparativo de los diferentes m etodos propuestos en la literatura nanciera para analizar la denominada performance de las carteras. A partir de an alisis de los ratios cl asicos, el autor introduce las nuevas medidas de performance que aparecen en la investigaci on y en la industria a partir de los a nos 90. Adicionalmente, se incide en los aspectos a tener en cuenta con carteras internacionales y en las carteras mixtas, donde la selecci on del ndice de referencia o benchmark no es evidente. Por otro lado, se comenta la problem atica que supone para la medida de la performance la existencia de diferentes estilos de inversi on y algunas soluciones al sesgo que produce este factor.
Introducci on
Uno de los aspectos m as interesantes de la gesti on de carteras es la medida de la calidad del gestor ponderando la rentabilidad alcanzada y el riesgo asumido. La calidad de la gesti on se concept ua a trav es del t ermino performance que no tiene f acil traducci on al castellano. En general los requisitos para una buena gesti on de carteras son dos: 1. Habilidad para obtener una rentabilidad mayor que un ndice de mercado equivalente a la cartera en t erminos de riesgo. 2. Capacidad para eliminar el riesgo no sistem atico de la cartera. El impresionante desarrollo de la industria de gesti on de activos ha fomentado la tecnolog a de medici on de la performance. Para medir la performance lo primero que tenemos que hacer es medir la rentabilidad obtenida en la realidad por la cartera. De forma peri odica, los inversores deben medir la performance de la cartera con un sistema que permita responder a las siguientes cuestiones:
1 Prosper Lamothe Fern andez es Catedr atico de Econom a Financiera de la Universidad Aut onoma de Madrid y presidente de Delta Investigaci on Financiera. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de abril de 1998.
82 Cu al es la rentabilidad global de la cartera?
Prosper Lamothe
C omo se descompone esta rentabilidad en t erminos de plusval as, resultados de cambio y dividendos? Cu ales son las decisiones que han contribuido m as a la rentabilidad? Qu e diferencial se ha obtenido frente a otros ndices de referencia? Se a nade valor de forma signicativa en la selecci on de acciones en alg un mercado espec co? Se est a logrando el objetivo de diversicaci on?, etc. Una primera estimaci on de la performance de la cartera se logra a trav es del c alculo de los denominados ndices de performance.
2
2.1
Los ndices cl asicos de performance

Ratio de Sharpe
Rc Rf , c
El ndice o ratio de Sharpe (1966), se calcula por la siguiente expresi on: S=
siendo Rc la rentabilidad de la cartera en el periodo de an alisis, Rf la rentabilidad del activo libre de riesgo en el periodo de an alisis y c la desviaci on t pica de la rentabilidad de la cartera durante el periodo de an alisis. Con el ratio de Sharpe estimamos la remuneraci on al riesgo que obtiene cada gestor en t erminos del diferencial de rentabilidad sobre la tasa libre de riesgo por cada punto porcentual de desviaci on t pica del rendimiento de la cartera. En el caso de carteras internacionales, un aspecto a analizar en la aplicaci on de este ndice es la denici on de la tasa libre de riesgo. Para una cartera global, esta puede ser la rentabilidad de las letras del tesoro norteamericanas. En el caso de carteras regionales, la mejor opci on es utilizar la rentabilidad del activo libre de riesgo en la econom a m as representativa de la zona. En u ltimo caso, podemos utilizar la rentabilidad libre de riesgo en la moneda base del inversor. La medida de Sharpe no presupone la vericaci on de ning un modelo, (como por ejemplo, el CAPM), mide el riesgo de forma global, pudiendo aplicarse para un amplio espectro de fondos (renta variable, mixtos, fondos internacionales, etc.). Esto hace que el ndice de Sharpe siempre se estime para cualquier cartera en el an alisis de su perl rentabilidad-riesgo. Por ejemplo en el caso de los denominados fondos de gesti on alternativa (hedge funds ), el ndice de Sharpe es uno de los escasos ratios de performance que se pueden aplicar.
tica de medicio n de la performance La problema
83
2.2
Indice de Treynor
Rc Rf , c
El ndice de Treynor (1966), se calcula por la siguiente expresi on: T =
ndice anterior y donde c es la beta de la signicando Rc y Rf lo mismo que en el cartera. El ratio de Treynor supone que la beta es un buen indicador del riesgo sistem atico de la cartera y en cierta medida asume la vericaci on del CAPM. En el caso de carteras internacionales, la beta la debemos estimar en base al ndice m as representativo del universo de acciones de posible inversi on. En algunos casos, por ejemplo, carteras globales, lo m as apropiado es estimar el ndice por subcarteras (por ejemplo, Am erica, Eurolandia, etc.). La asunci on de un determinado modelo de equilibrio del mercado de capitales limita su aplicaci on frente al ndice de Sharpe y otros ratios. Tal como aparece en la gura 1, el ndice de Treynor nos estima el angulo de la l nea de equilibrio rentabilidad-riesgo sistem atico que alcanzan los gestores de los fondos. En el ejemplo que manejamos, hemos partido de la base que los gestores han superado a la l nea de equilibrio del mercado de capitales representada por LEMC, obteniendo un mayor premio al riesgo sistem atico de sus carteras. Rentabilidad 18% A .. ...B .......... . . . . . . . . . . .......... 6 LEMC .......... 6 . . . . . . . . . . 12% .......... .......... . . . . . . . . . ...... Rf = 8% 15%
Beta
Figura 1: Indice de Treynor: LEMC es la l nea de equilibrio del mercado de capitales. Supuesto de un 4% de prima de rentabilidad Rm Rf .
84
Prosper Lamothe
2.3
Indice de Jensen
Como sabemos, seg un el CAPM, Ri Rf : i (Rm Rf ) + ui , siendo ui una variable aleatoria. Si un gestor es mejor que el resto del mercado, debe superar de forma consistente las primas de rentabilidad por riesgo que se obtienen en dicho mercado. Si se denomina a la rentabilidad obtenida en la cartera i por una gesti on mejor que la del resto del mercado, Ri Rf : i + i (Rm Rf ) + ui . Y en base a los valores de i , sabemos que si i > 0 se ha desarrollado una buena gesti on, y para i < 0 la gesti on ha sido negativa. sticamente signicativa para que el gestor Realmente i debe ser positiva y estad haya obtenido una adecuada performance. Si = 0, los gestores adec uan su cartera a la l nea de equilibrio del mercado de capitales. Si > 0, los gestores est an incorporando valor a nadido en su gesti on y si < 0, los gestores realizan una mala labor ya que no son capaces de alcanzar la rentabilidad correspondiente a una estrategia pasiva en el mercado. En el caso de una cartera internacional, se debe tener muy en cuenta la composici on de la cartera estrat egica a la hora de estimar el de Jensen. As , si la composici on estrat egica es 50% en USA, 50% en Jap on, tenemos dos soluciones: a) Estimar las betas y la medida de Jensen de la cartera en relaci on a un ndice sint etico formado en un 50% por un ndice USA y en el otro 50% por un ndice japon es. b) Calcular el ndice de Jensen a nivel de subcarteras. Esta es la opci on m as aconsejable, ya que nos permite evaluar la calidad de nuestra gesti on a nivel individual en los diferentes mercados en que invertimos. El ndice de Jensen tiene problemas similares al ratio de Treynor. Para su correcta aplicaci on se deber a vericar el CAPM, lo cual es una cuesti on muy debatida, no aceptada en t erminos generales y que se sale del a mbito de este trabajo. Adicionalmente, tanto el ndice de Treynor como el ndice de Sharpe son inadecuados en la medici on de la performance de carteras de renta ja y/o activos distintos a las acciones. De los tres ndices cl asicos, el ndice de Sharpe es el m as utilizado en los mercados. Adicionalmente, se han desarrollado otros ratios e ndices que comentaremos en los siguientes apartados.
85
3
3.1
Los nuevos ndices de performance

El ratio de informaci on
En la pr actica de la gesti on de carteras se ha extendido el uso del denominado ratio de informaci on que mide la performance de la cartera en relaci on a la evoluci on del ndice o ndices de referencia (benchmarks ). El ratio se calcula por la siguiente expresi on: RI = Rc Rb , cd
siendo Rc la rentabilidad de la cartera en el periodo de an alisis, Rb la rentabilidad del ndice de referencia o benchmark durante el periodo de an alisis, y cb la desviaci on t pica de los diferenciales de rentabilidad entre la cartera y el benchmark durante el periodo de an alisis. Este ratio presenta varias ventajas con respecto a las medidas anteriores: No asume el cumplimiento de ning un modelo espec co de equilibrio del mercado de capitales como los ratios de Treynor y Jensen. Mide adecuadamente el valor a nadido por el gestor al estimar la diferencia de rentabilidad entre la cartera y el benchmark. Esta medida es m as correcta que la del ndice de Sharpe. Por ejemplo, el ndice de Sharpe puede dar valores negativos para fondos de renta variable en un periodo de ca das de los precios burs atiles. Signica eso una mala performance de todos los gestores? No, depende de la rentabilidad alcanzada en relaci on a los ndices de mercado. En general, para los fondos y carteras gestionados en base a un ndice o benchmark, constituye la medida ideal de la calidad de su gesti on al relacionar el diferencial de rentabilidad alcanzado con el riesgo relativo asumido por el gestor al separarse en mayor o menor grado de su ndice de referencia o benchmark (riesgo activo). En nuestra opini on, el ratio de informaci on, junto con otros ndices similares de medida de la performance, se ir an imponiendo progresivamente en la pr actica y en los ambientes acad emicos.
3.2
La medida M 2 de Modigliani-Modigliani
Franco y Leah Modigliani (1997) han propuesto otras medidas de performance que por simplicidad han denominado ndices M 2 . En la metodolog a propuesta, el primer paso consiste en igualar las carteras en t erminos de riesgo en relaci on al benchmark o ndice de referencia. As , si el riesgo de nuestra cartera, estimado por la desviaci on t pica de los rendimientos anuales es i y la del benchmark B , sabemos que podemos
86
Prosper Lamothe
igualar el riesgo de la cartera al ndice de referencia apalanc andonos (mediante endeudamiento) o desapalanc andonos en una proporci on di , de modo que (1 + di ) i = B , y por lo tanto, di = B 1. i
A partir de aqu , M 2 se calcula como la rentabilidad de la cartera objeto de estudio apalancada (o desapalancada) para replicar al benchmark : M 2 = (1 + di ) ri di rf . De forma an aloga, podemos calcular una medida M 2 en base a betas, que denomina2 remos Mbeta , con razonamientos an alogos a los anteriores, por la expresi on
2 Mbeta =
BB Bi
(ri rf ) + rf ,
(1)
y adicionalmente podemos estimar un M 2 por la igualdad

2 rB , M 2 = Mbeta 2 por su valor en (1) y siendo rB la rentabilidad del benchmark. Sustituyendo Mbeta agrupando t erminos, obtenemos que
M2 =
Jensen . Bi
un valor a nadido. En t erminos de jerarquizaci on, las medidas M 2 no aportan ning Ahora bien, nos permiten enriquecer el an alisis de la performance con la obtenci on de rentabilidades de carteras equivalentes en t erminos de riesgo.
La medici on de la sincronizaci on
Los ndices que hemos estudiado en ep grafes anteriores nos permiten evaluar globalmente la performance de la cartera. Ahora bien, no distinguen las causas por las que los gestores han podido obtener una buena performance. En un primer an alisis, sin considerar todav a los aspectos espec cos de las carteras internacionales, podemos decir que los gestores pueden obtener rentabilidades superiores al benchmark por dos razones: a) Porque seleccionan bien los t tulos que integran la cartera. Es decir, por una buena selecci on de valores. b) Porque anticipan correctamente los movimientos del mercado, lo que les permite obtener benecios modulando su exposici on a la renta variable en base a sus precisiones. Esto se denomina sincronizaci on.
87
La cuesti on es c omo medir lo aportado por los gestores realizando actividades de market timing, es decir a trav es de su sincronizaci on con el mercado. Tal como indica Freixas (1996) y Matall n (2000), existen dos m etodos para estimar la sincronizaci on: el m etodo cuadr atico y el m etodo basado en la opci on de reestructuraci on de la cartera. Seg un el m etodo cuadr atico, la beta de la cartera variar a progresivamente con las variaciones del mercado. Aumentar a cuando el mercado se halle en fase alcista, por lo que la rentabilidad de la cartera sincronizada ser a mayor que la de la cartera de beta constante (no sincronizadas) y en fases de correcci on del mercado la rentabilidad tambi en ser a mayor al reducir la beta con una gesti on de menor riesgo en la cartera. Una forma de medir este efecto es identicar la convexidad obtenida con la sincronizaci on a nadiendo un t ermino cuadr atico en la modelizaci on de Jensen. As , rct rft = c + c (rmt rft ) + (rmt rft )2 + ucy . Si c es positivo y estad sticamente signicativo, existen indicios de que los gestores est an logrando una adecuada sincronizaci on. El segundo m etodo asume que el gestor consigue prever s olo dos tipos de periodos distintos, los de mercado alcista y bajista, sin poder modular de forma progresiva la beta de su cartera. En este caso2 , la regresi on es la siguiente: rct rft = c + c (rmt rft ) + c max(0, rf rmt ) + uct . Bajo este supuesto, el nivel de sincronizaci on se corresponde a una opci on put con precio de ejercicio igual al rendimiento del activo libre de riesgo. Por ambos m etodos, el valor del par ametro no dar a lo a nadido a trav es de la selecci on de valores y el par ametro reejar a el rendimiento aportado por sincronizaci on. Ambos modelos son u tiles para medir la sincronizaci on. La elecci on de uno y otro se debe realizar en funci on de las caracter sticas propias del proceso de gesti on. En cualquier caso, es muy f acil estimar ambas regresiones por lo que se puede calcular los y s por los dos m etodos.
La atribuci on global de la performance
Las carteras formadas por varias categor as de activos como las carteras de acciones internacionales exigen un tratamiento espec co a la hora de medir la calidad de la gesti on que realizamos. Planteada una asignaci on estrat egica de la cartera en t erminos de bandas de posible inversi on por cada categor a de activos junto con los niveles de
2 Ver
Freixas (1996).
88
Prosper Lamothe
medios de inversi on a efectos de medici on de la performance debemos estimar la contribuci on del gestor en t erminos de: - Asignaci on de activos: es decir la contribuci on positiva o negativa por el peso asignado a las diferentes categor as de activos dentro de las bandas preestablecidas en las que se permite hacer market timing. - Selecci on de valores: consistente en la medici on de los aportado a trav es de la selecci on de t tulos, segmentos de la curva, etc. dentro de cada categor a de activos. Para una correcta medici on es fundamental una adecuada selecci on de ndices de referencia o benchmarks para cada categor a de activos. A partir de la cartera estrat egica, los ndices o benchmarks y los pesos reales mantenidos, obtenemos las siguientes expresiones: Asignaci on de valores: siendo AAi la contribuci on de la Asignaci on de Activos de la categor a i,
i el peso real asignado a la categor a i; Wr i el peso asignado a la categor a i; Wb i la rentabilidad geom etrica del mercado i, medida por su benchmark ; rb i i i AAi = (wr wb ) (rb Rb ) ,
etrica de la cartera benchmark : Rb la rentabilidad geom Rb =

i=1 i i wb rb ;
La Asignaci on de Activos conjunta de la cartera se obtiene por agrupaci on (n es el n umero de categor as b asicas):
n
AAc =
i=1
AAi .
i i i SVi = wr (rr rb ),
Selecci on de valores: siendo
SVi la contribuci on de la selecci on de valores de la categor a i,

i el peso real asignado a la categor a i; Wr

i Rr la rentabilidad geom etrica real de la cartera i; i la rentabilidad geom etrica del mercado i, medida por su benchmark. Rb
89
Para a nadir valor la rentabilidad de la cartera ha de ser superior a la de su ndice de referencia, y la contribuci on a la rentabilidad de la cartera ser a mayor cuanto mayor sea el peso de la categor a. La Selecci on de Valores conjunta de la cartera se obtiene por agrupaci on:
n
SVc =
i=1
SVi .
Se puede observar claramente que la medida de la contribuci on de la asignaci on de activos y de la sincronizaci on se orientan a la estimaci on de la misma variable: la capacidad del gestor (gestores), de a nadir valor modulando su exposici on a los diferentes activos. La medida de sincronizaci on es u til para el investigador externo que quiere medir de forma ex ogena, la capacidad de los gestores de a nadir valor modicando el riesgo sistem atico de las cartera en funci on de las expectativas sobre la evoluci on del mercado. Los consultores y gestores que tienen datos internos y m as completos de las carteras aplicar an las metodolog a comentada en este ep grafe para evaluar la capacidad de sincronizaci on con el mercado de un determinado proceso de inversi on.
Otros factores que afectan a la performance
Otra cuesti on importante en la medida de la performance es el estilo de inversi on. Con los ratios cl asicos podemos tener un gran sesgo derivado de la no identicaci on del estilo de gesti on de la cartera. Por ejemplo, emp ricamente se ha demostrado que las carteras value baten a medio y largo plazo al ndice general. La soluci on a este problema puede basarse en dos enfoques: Estimar alfas (al estilo de Jensen) en base a modelos multifactoriales de explicaci on del rendimiento de la cartera. En la medida en que el modelo multifactorial identique el riesgo espec co del factor estilo, los alfas obtenidos me permiten una mejor evaluaci on de la performance. Utilizar el m etodo propuesto por Sharpe (1992), que parte de la identicaci on inicial de los estilos de gesti on desarrollados en las carteras, utilizando tambi en modelos multifactoriales. Obviamente la mejor soluci on te orica ser a disponer de un benchmark espec co para cada estilo, lo que no es siempre posible. Dada la importancia que est a alcanzando la correcta medida de la performance para todos los instrumentos de inversi on colectiva, no dudamos que en los pr oximos a nos se desarrollaran t ecnicas m as depuradas que las actuales para evitar el riesgo del estilo de gesti on.
90
Prosper Lamothe
Referencias
[1] R. Dembo (1997), Optimal Portfolio Replication, Algo Research, Working Paper Series. [2] F. Fabozzi (1995), Investment Management, Prentice-Hall, New Jersey. [3] F.Galan (1996), Riesgo y Rentabilidad de los Fondos de Inversi on Mobiliaria en Espa na, Tesis Doctoral. Universidad de C ordoba. [4] P. Lamothe (1999), Gesti on de carteras de acciones internacionales, Editorial Pir amide, Madrid. [5] M.C.Jensen (1969), Risk, the Pricing of Capital Assets, and the Evaluation of Investment Portfolios, Journal of Business, Abril. [6] J.C. Matall n Saez (2000), Gesti on de carteras: evaluaci on de los fondos de inversi on mobiliaria en Espa na, Tesis Doctoral, Universidad Jaime I, Castell on. [7] F. Modigliani (1997), Risk-Adjusted Performance, The Journal of Portfolio Management, Invierno, p ags. 4554. [8] L. Modigliani (1997), Risk-adjusted Performance, en Morgan Stanley Dean Witter(ed.), The Risk Book, Londres, p ags. 46-57. [9] G. Rubio (1993), La evaluaci on de los fondos de inversi on el an alisis de la composici on mensual de las carteras, An alisis Financiero 59, p ags. 6483. [10] G. Rubio (1995), Further Evidence on Performance Evaluation: Portfolio Holdings, Recommendations and Turnover Costs, Review of Quantitative Finance and Accounting 5, p ags. 127153. [11] W.F. Sharpe (1996), Mutual Fund Performance, Journal of Business 39, Enero. [12] W. Sharpe (1992), Asset Allocation: Management Style and Performance Measurement, Journal of Portfolio Management, Invierno, p ags. 719. [13] J.L. Treynor (1966), How to Rate Management Investment Funds, Harvard Business Review 43, Enero-Febrero. [14] J.L. Treynor, K.K. Mazuy (1992), Pueden los fondos de inversi on anticipar el mercado?, An alisis Financiero 58, p ags. 2631.
Prosper Lamothe Fern andez Departamento de Financiaci on e Investigaci on Comercial Universidad Aut onoma de Madrid Campus de Cantoblanco, s/n 28049-Madrid, Espa na e-mail: lamothe@delta-finance.com
Equivalencia de los distintos m etodos para valorar empresas por descuento de ujos. Distintas alternativas para valorar el ahorro de impuestos debido al apalancamiento y sus implicaciones sobre la valoraci on
ndez1 Pablo Ferna
Resumen Este art culo se centra en la valoraci on de empresas por descuento de ujos. En la primera parte se demuestra que los cuatro m etodos m as habituales de valoraci on por descuento de ujos (free cash ow descontado al WACC; ujos para las acciones descontados a la rentabilidad exigida a las acciones; Capital cash ow descontados al WACC antes de impuestos; y Adjusted Present Value) proporcionan siempre (un periodo, multiperiodo, ujos perpetuos, cualquier estructura temporal de los ujos, ratio de endeudamiento constante o variable) el mismo valor. Este resultado es l ogico, porque todos los m etodos analizan la misma realidad bajo las mismas hip otesis; s olo dieren en los ujos que toman como punto de partida para la valoraci on. Las discrepancias de las diversas teor as sobre la valoraci on de una empresa provienen del c alculo del ahorro de impuestos debido al apalancamiento (DVTS). El art culo muestra y analiza 7 teor as distintas sobre el c alculo del DVTS: Modigliani y Miller (1963), Myers (1974), Miller (1977), Miles y Ezzell (1980), Harris y Pringle (1985), Ruback (1995), Damodaran (1994), y el m etodo de los pr acticos. Se demuestra que el m etodo de Myers (1974) proporciona resultados inconsistentes. Este trabajo tambi en introduce una nueva interpretaci on del resto de las teor as: que incorporan unos costes de quiebra en la valoraci on, entendiendo como tales la diferencia entre el valor de la empresa que proporcionan Modigliani-Miller (costes de quiebra cero) y el valor de la empresa que proporcionan dichas teor as. Una manera sencilla de analizar esta diferencia es a trav es del free cash ow ajustado al riesgo del negocio. Para analizar los resultados de las distintas teor as conviene tener en cuenta que el DVTS no es propiamente el valor actual del ahorro de impuestos debido al pago de intereses
1 Pablo Fern andez es profesor de Finanzas en el IESE. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de mayo de 1998.
92
ndez Pablo Ferna
descontado a una tasa, sino la diferencia de dos valores actuales: el valor actual de los impuestos que paga la empresa sin deuda menos el valor actual de los impuestos que paga la empresa con deuda. El riesgo de de los impuestos que paga la empresa sin deuda es inferior al riesgo de los impuestos que paga la empresa con deuda. Tambi en se muestran los cambios que se producen en las f ormulas de valoraci on cuando el valor de la deuda no coincide con su valor nominal.
Introducci on
Este art culo se centra en la valoraci on de empresas por descuento de ujos. En primer lugar se demuestra que los cuatro m etodos m as habituales de valoraci on por descuento de ujos (WACC, ujos para las acciones, Capital cash ow y APV) proporcionan el mismo valor. Las discrepancias en la valoraci on de una empresa provienen del c alculo del ahorro de impuestos debido al apalancamiento (DVTS). El art culo muestra y analiza 7 teor as distintas sobre el c alculo del DVTS. Una de las conclusiones es que el m etodo de Myers (1974) proporciona resultados inconsistentes. Una interpretaci on del resto de las teor as es que incorporan unos costes de quiebra en la valoraci on, entendiendo como tales la diferencia entre el valor que proporcionan Modigliani-Miller (costes de quiebra cero) y el valor que proporcionan dichas teor as. Una manera sencilla de analizar esta diferencia es a trav es del free cash ow ajustado al riesgo del negocio. La secci on 2 del art culo muestra los seis m etodos m as habituales de valorar empresas por descuento de ujos: free cash ow descontado al WACC; ujos para las acciones descontados a la rentabilidad exigida a las acciones; Capital cash ow descontados al WACC antes de impuestos; APV; free cash ows ajustados al riesgo del negocio descontados a la rentabilidad exigida a los activos; y cash ows disponibles para las acciones ajustados al riesgo del negocio descontados a la rentabilidad exigida a los activos. La secci on 3 es una breve rese na de los art culos m as relevantes sobre valoraci on de empresas por descuento de ujos. La secci on 4 muestra las principales f ormulas de valoraci on que se derivan de los art culos m as relevantes: Modigliani y Miller (1963), Myers (1974), Miller (1977), Miles y Ezzell (1980), Harris y Pringle (1985), Ruback (1995), Damodaran (1994), y el m etodo de los pr acticos. La secci on 5 muestra con un ejemplo las diferencias en la valoraci on seg un las alternativas presentadas en la secci on 4. La secci on 6 analiza con mayor detalle el origen de las diferencias en la valoraci on seg un los diversos autores: del c alculo del ahorro de impuestos debido al apalancamiento (DVTS). Tambi en proporciona las diferencias en la valoraci on seg un las distintas teor as. En el anexo 1 se muestran las abreviaturas utilizadas en este art culo. En el anexo 2 se muestran los cambios que se producen en las f ormulas de valoraci on cuando el valor de la deuda no coincide con su valor nominal.
n de empresas por descuento de flujos Valoracio
93
M etodos de valoraci on de empresas por descuento de ujos

A. A partir del free cash ow y del WACC (coste ponderado de los recursos) La f ormula (1) indica que el valor de la deuda (D) m as el de los recursos propios (E) es el valor actual de los free cash ows (FCF) esperados que generar a la empresa, descontados al coste ponderado de la deuda y los recursos propios despu es de impuestos (WACC): E0 + D0 = VA0 [WACCt ; FCFt ] . (1)
Hay cuatro m etodos fundamentales para valorar empresas por descuento de ujos2 :
La denici on de WACC o coste ponderado de capital (en ingl es, weighted average cost of capital ), viene dada por (2): WACCt = Et1 Ket + Dt1 Kdt (1 T) . Et1 + Dt1 (2)
Ke es la rentabilidad exigida a las acciones, Kd es el coste de la deuda y T es la tasa efectiva del impuesto sobre los benecios. Et1 + Dt1 son valores de mercado3 . B. A partir del cash ow esperado disponible para las acciones (CFac) y de la rentabilidad exigida a los recursos propios de la empresa (Ke) La f ormula (3) indica que el valor de las acciones (E) es el valor actual neto de los cash ows disponibles para las acciones esperados (CFac) descontados a la rentabilidad exigida a los recursos propios de la empresa4 , (Ke): E0 = VA0 [Ket ; CFact ] . (3)
La f ormula (4) indica que el valor de la deuda (D) es el valor actual neto de los cash ows disponibles para la deuda esperados (CFd) descontados a la rentabilidad exigida a la deuda (Kd). D0 = VA0 [Kdt ; CFdt ] . (4)
2 Las f ormulas que aparecen a continuaci on son v alidas si el tipo de inter es de la deuda coincide con la rentabilidad exigida a la misma (Ke), o dicho de otro modo, que el valor de mercado de la deuda es id entico a su valor contable. Las f ormulas para el caso en que esto no sucede aparecen en el anexo 2 y en Fern andez (1999), pp.389-391. 3 Realmente, los valores de mercado son los valores que se obtienen en la valoraci on con la f ormula (1). Por esto, la valoraci on es un proceso iterativo: se descuentan los free cash ows al WACC para calcular el valor de la empresa (D + E), pero para obtener el WACC se necesita el valor de la empresa (D + E). 4 Con frecuencia se denomina coste de los recursos propios o coste de las acciones.
94 La expresi on que relaciona el FCF con el CFac es5 : CFact = FCFt + Dt It (1 T) ,
ndez Pablo Ferna
(5)
donde Dt es el aumento de deuda e It son los intereses pagados por la empresa. Es obvio que CFd = It Dt . La suma de los valores que proporcionan las f ormulas (3) y (4) es id entica al valor proporcionado por (1):6 E0 + D0 = VA0 [WACCt ; FCFt ] = VA0 [Ket ; CFact ] + VA0 [Kdt ; CFdt ] . C. A partir del capital cash ow (CCF) y del WACCBT (coste ponderado de los recursos, antes de impuestos) Los capital cash ows son los cash ow disponibles para todos los poseedores de t tulos de la empresa, sean estos deuda o acciones, y equivalen al cash ow disponible para las acciones (CFac) m as el cash ow que corresponde a los tenedores de deuda (CFd). La f ormula (6) indica que el valor de la deuda hoy (D) m as el de los recursos propios (E), es igual al capital cash ow (CCF) descontado al coste ponderado de la deuda y los recursos propios antes de impuestos7 (WACCBT ). E0 + D0 = VA[WACCBT ; CCFt ] . La denici on de WACCBT es WACCBT = Et1 Ket + Dt1 Kdt . Et1 + Dt1 (7) (6)
La expresi on (7) se obtiene de igualar (1) con (6). WACCBT representa la tasa de descuento que asegura que el valor de la empresa obtenido con ambas expresiones es el mismo8 : E0 + D0 = VA[WACCBT ; CCFt ] = VA[WACCt ; FCFt ] . La expresi on que relaciona el CCF con el CFac y con el FCF es CCFt = CFact + CFdt = CFact Dt + It = FCFt + It T , donde Dt = Dt Dt1 e It = Dt1 Kdt .
5 Obviamente, el free cash ow es el hipot etico cash ow disponible para las acciones cuando la empresa no tiene deuda. 6 De hecho, una manera de denir el WACC es: el WACC es la tasa a la que se debe descontar el FCF para obtener el resultado que proporcionan (3) y (4). 7 BT viene de before taxes (antes de impuestos). 8 De hecho, una manera de denir el WACC BT es: el WACCBT es la tasa a la que se debe descontar el CCF para obtener el resultado que proporcionan (3) y (4).
(8)
n de empresas por descuento de flujos Valoracio D. Valor actual ajustado (APV)
95
La f ormula del valor actual ajustado, APV (adjusted present value ) (9) indica que el valor de la deuda (D) m as el de los recursos propios (E) de la empresa apalancada, es igual al valor de los recursos propios de la empresa sin apalancar Vu m as el valor actual neto del ahorro de impuestos debido al pago de intereses (DVTS): E0 + D0 = Vu0 + DVTS0 . (9) Existen varias teor as para el c alculo del DVTS, que analizaremos en la secci on 4 de este trabajo9 . Si Ku es la rentabilidad exigida a las acciones en la empresa sin deuda (tambi en llamada rentabilidad exigida a los activos), Vu viene dado por Vu0 = VA0 [Kut ; FCFt ] . Por consiguiente, DVTS0 = E0 + D0 Vu0 = VA0 [WACCt ; FCFt ] VA0 [Kut ; FCFt ] . Los cuatro procedimientos descritos proporcionan siempre (un periodo, multiperiodo, ujos perpetuos, cualquier estructura temporal de los ujos, ratio de endeudamiento constante o variable) el mismo valor de la empresa si se utilizan correctamente. Hay discrepancias para calcular el APV: existen diversas teor as sobre la magnitud del DVTS, que analizaremos en este trabajo. La magnitud del DVTS tiene implicaciones en la valoraci on y afecta a: - el valor de las acciones (E) y el de la empresa (E + D), - la relaci on entre la rentabilidad exigida a los activos (Ku) y la rentabilidad exigida a las acciones en la empresa apalancada (Ke); - la relaci on entre el WACC y la rentabilidad exigida a los activos (Ku). Podemos hablar de un quinto m etodo (a partir del free cash ow ajustado al riesgo del negocio), aunque este no es un m etodo propiamente nuevo, sino que deriva de los anteriores: E. A partir del free cash ow ajustado al riesgo del negocio y de Ku (rentabilidad exigida a los activos) La f ormula (11) indica que el valor de la deuda (D) m as el de los recursos propios (E) es el valor actual de los free cash ows ajustados al riesgo del negocio (FCF\\Ku) esperados que generar a la empresa, descontados a la rentabilidad exigida a los activos (Ku): E0 + D0 = VA0 [Kut ; FCFt \\Ku] . (11) (10)
9 Las expresiones del valor actual del ahorro de impuestos debido al pago de intereses (DVTS) para una perpetuidad creciente a la tasa g est an recogidas en la subsecci on 4.3.
96
ndez Pablo Ferna Es f acil comprobar que la denici on de los free cash ows ajustados al riesgo del negocio10 (FCF\\Ku) es FCFt \\Ku = FCFt (Et1 + Dt1 ) [WACCt Kut ] . (12)
An alogamente, podemos hablar de un sexto m etodo (a partir del cash ow disponible para las acciones ajustado al riesgo del negocio), aunque este no es un m etodo propiamente nuevo, sino que deriva de los anteriores: F. A partir del cash ow disponible para las acciones ajustado al riesgo del negocio y de Ku (rentabilidad exigida a los activos) La f ormula (13) indica que el valor de las acciones (E) es el valor actual neto de los cash ows disponibles para las acciones ajustados al riesgo del negocio esperados (CFac\\Ku) descontados a la rentabilidad exigida a los activos (Ku): E0 = VA0 [Kut ; CFact \\Ku] . (13)
Es f acil comprobar que la denici on de los cash ows disponibles para las acciones ajustados al riesgo del negocio11 (CFac\\Ku) es CFact \\Ku = CFact Et1 [Ket Kut ] . (14)
Tambi en podr amos hablar de un s eptimo m etodo; a partir del capital cash ow ajustado al riesgo del negocio y de Ku (rentabilidad exigida a los activos), pero el capital cash ow ajustado al riesgo del negocio es id entico al free cash ow ajustado al riesgo del negocio (CCF\\Ku = FCF\\Ku). Por tanto, este m etodo ser a id entico a (E).
10 La 11 La
expresi on (12) resulta de igualar (11) y (1). expresi on (14) resulta de igualar (13) y (3). En el resto del art culo se omiten los sub ndices t y t 1.
97
Un ejemplo. La empresa Delta Inc. tiene las previsiones de balance y cuenta de resultados para los pr oximos a nos que se adjuntan en la tabla 1. A partir del a no 3 se prev e que el balance y la cuenta de resultados crecer an al 4% anual.
0 400 1.600 1.600 2.000 1.000 1.000 2.000 1 430 1.800 200 1.600 2.030 1.000 1.030 2.030 2 515 2.300 450 1.850 2.365 1.100 1.265 2.365 3 550 2.600 720 1.880 2.430 1.100 1.330 2.430 4 572,00 2.956,00 1.000,80 1.955,20 2.527,20 1.144,00 1.383,20 2.527,20
NOF (circulante neto) Activo jo bruto - amort. acumulada Activo jo neto TOTAL ACTIVO Deuda (N) Capital (valor contable) TOTAL PASIVO Cuenta de resultados Margen Intereses BAT Impuestos BDT (benecio neto)
300 120 180 63 117
500 120 380 133 247
572 132 440 154 286
603,20 132,00 471,20 164,92 306,28
Tabla 1: Previsiones de balance y cuenta de resultados de Delta Inc. A partir de las previsiones de balance y cuenta de resultados de la tabla 1 es inmediato obtener los ujos que se adjuntan en la tabla 2. L ogicamente, los ujos crecen al 4% a partir del a no 4.
0 CF acciones = Dividendos FCF CFd CCF 1 87,00 165,00 120,00 207,00 2 12,00 -10,00 20,00 32,00 3 221,00 306,80 132,00 353,00 4 253,08 294,88 88,00 341,08 5 263,20 306,68 91,52 354,72
Tabla 2: Previsiones de ujos de Delta Inc La beta de los activos (de las acciones de la empresa sin deuda) es 1. La tasa sin riesgo es 10%. El coste de la deuda es 12%. La tasa de impuestos es 35% y la prima de riesgo de mercado (risk premium ) es 8%. Por consiguiente, utilizando el CAPM, la rentabilidad exigida a los activos es de un 18%.12 Con estos par ametros, la valoraci on de las acciones de esta empresa, utilizando las f ormulas precedentes, aparece en la tabla 3.
12 Utilizamos
en este ejemplo el CAPM: Ku = RF + u PM = 10% + 8% = 18%.
98
f ormula 0 18,00% 21,74% 2.043,41 14,917% 1.043 1.043 1.000 2.043,41 16,972% 1.043 442,09 1.601 2.043,41 1.043 2.043,41 1.043 1.043 1 18,00% 21,30% 2.183,23 15,114% 1.183 1.183 1.000 2.183,23 17,038% 1.183 458,66 1.725 2.183,23 1.183 2.183,23 228,00 1.183 48,00 1.183 2 18,00% 21,01% 2.523,21 15,253% 1.423 1.423 1.100 2.523,21 17,084% 1.423 478,22 2.045 2.523,21 1.423 2.523,21 53,00 1.423 -27,00 1.423
ndez Pablo Ferna

3 18,00% 20,86% 2.601,29 15,336% 1.501 1.501 1.100 2.601,29 17,112% 1.501 495,00 2.106 2.601,29 1.501 2.601,29 376,10 1.501 178,10 1.501 4 18,00% 20,86% 2.705,34 15,336% 1.561 1.561 1.144 2.705,34 17,112% 1.561 514,80 2.191 2.705,34 1.561 2.705,34 364,18 1.561 210,18 1.561
Ku Ke E + D = VA(WACC; FCF) WACC (1) D = E E = VA(Ke;CFac) D = VA(CFd; Kd) D + E = VA(WACCBT ; CCF) WACCBT (6)D = E DVTS = VA(Ku; D T Ku) Vu = VA (Ku;FCF) Vu + DVTS (9) - D = E D+E = VA(Ku;FCF\\Ku) FCF\\Ku (11) - D = E CFac\\Ku E = VA(Ku;CFac\\Ku)
(1) (2) (3) (4) (6) (7)
(10) (9) (11) (12) (14) (13)
Tabla 3: Valoraci on de Delta Inc. La rentabilidad exigida a las acciones (Ke) aparece en la segunda l nea de la tabla13 . La f ormula (3) permite obtener el valor de las acciones descontando los ujos disponibles para las acciones a la rentabilidad exigida a las acciones (Ke)14 . An alogamente, la f ormula (4) permite obtener el valor de la deuda descontando los ujos para la deuda a la rentabilidad exigida a la deuda (Kd)15 . Otro modo de calcular el valor de las acciones es a partir de la f ormula (1). El valor actual de los free cash ows descontados al WACC (f ormula (2)) nos proporciona el valor de la empresa, que es el valor de la deuda m as el de las acciones16 . Restando a esta cantidad el valor de la deuda se obtiene el valor de las acciones. Otro modo de calcular el valor de las acciones es a partir de la f ormula (6). El valor actual de los capital cash ows descontados al WACCBT (f ormula (7)) nos proporciona el valor de la empresa, que es el valor de la deuda m as el de las acciones. Restando a esta cantidad el valor de la deuda se obtiene el valor de las acciones. El cuarto m etodo de calcular el valor de las acciones es a partir del Adjusted Present Value, la f ormula (9). El valor de la empresa es la suma del valor de la
rentabilidad exigida a las acciones (Ke) se ha calculado seg un la teor a de Modigliani-Miller, que veremos m as tarde. 14 La relaci on entre el valor de las acciones de dos a nos consecutivos es: Et = Et1 (1+Ket ) CFact . 15 El valor de la deuda coincide con el nominal (valor contable) de la tabla 1 porque hemos considerado que la rentabilidad exigida a la deuda coincide con su coste (12%). 16 La relaci on entre el valor de la empresa de dos a nos consecutivos es: (D + E)t = (D + E)t1 (1 + WACCt ) FCFt .
13 La
99
empresa sin apalancar (f ormula (10)) m as el valor actual del ahorro de impuestos debido a la deuda (DVTS)17 . Por u ltimo, al nal de la tabla 3 se calculan el cash ow disponible para las acciones y el free cash ow ajustados al riesgo del negocio (CFac\\Ku y FCF\\Ku) seg un las f ormulas (14) y (12). La f ormula (13) permite obtener el valor de las acciones descontando los ujos disponibles para las acciones ajustados al riesgo del negocio a la rentabilidad exigida a los activos (Ku). Otro modo de calcular el valor de las acciones es a partir de la f ormula (11). El valor actual de los free cash ows ajustados al riesgo del negocio descontados a la rentabilidad exigida a los activos (Ku) nos proporciona el valor de la empresa, que es el valor de la deuda m as el de las acciones. Restando a esta cantidad el valor de la deuda se obtiene el valor de las acciones. El ejemplo de la tabla 3 muestra que el resultado obtenido con las seis valoraciones es el mismo. el valor de las acciones hoy es 1.043. Como ya hemos comentado, estas valoraciones se han realizado seg un Modigliani y Miller. Las valoraciones utilizando otras teor as que se presentan en las siguientes secciones de este art culo se presentan en la subsecci on 5.3.
Breve rese na de los art culos m as relevantes sobre valoraci on de empresas por descuento de ujos
Existe una abundante literatura sobre la valoraci on de empresas por descuento de ujos. Comentamos aqu los art culos m as relevantes y hacemos especial hincapi e en los que proponen diferentes expresiones para el valor actual de los ahorros de impuestos por pago de intereses (DVTS). Modigliani y Miller (1958), (1961) y (1963) estudiaron el efecto del apalancamiento en el valor de la empresa. Publicaron sus famosas proposiciones, que siguen siendo punto de referencia en cualquier trabajo sobre nanzas de la empresa en general y sobre valoraci on en particular. Su proposici on 1 (Modigliani y Miller (1958), f ormula (3)) es que, en ausencia de impuestos, el valor de la empresa es independiente del endeudamiento, esto es, E0 + D0 = Vu , si T = 0.
Su segunda proposici on (Modigliani y Miller (1958), f ormula (8)) es que, en ausencia de impuestos, la rentabilidad exigida por los accionistas (Ke) aumenta en proporci on directa con el endeudamiento (la proporci on D/E) a valor de mercado: Ke = Ku + (D/E) (Ku Kd) .
17 Como se ha calculado rentabilidad exigida a las acciones (Ke) seg un la teor a de ModiglianiMiller, tambi en hemos de calcular el DVTS seg un Modigliani-Miller: DVTS = VA(Ku ; D T Ku).
100
ndez Pablo Ferna
En presencia de impuestos, su segunda proposici on (Modigliani y Miller (1963), f ormula (12.c)) es: Ke = Ku + D (1 T) (Ku Kd)/E . En presencia de impuestos, su primera proposici on, en el caso de una perpetuidad, se transforma en (Modigliani y Miller (1963), f ormula (3)): E0 + D0 = Vu + D T . DT es el aumento de valor debido al apalancamiento (DVTS). Sobre los dividendos armaron que a igualdad de impuestos sobre el cobro de dividendos y plusval as eran irrelevantes: el accionista es indiferente entre cobrar dividendos o vender acciones para obtener una remuneraci on prejada. Modigliani y Miller (1963) presentan varias f ormulas de valoraci on que utilizaremos en este trabajo: - Su f ormula (31.c) es WACC = Ku - Su f ormula (11.c) es WACCBT = Ku D T Ku Kd . E+D 1 TD . E+D
Tambi en arman en su f ormula (33.c) que en una inversi on que se puede nanciar totalmente con deuda, la rentabilidad exigida a la deuda debe ser igual a la rentabilidad exigida a los activos: si D/(D + E) = 100%, entonces Kd = Ku. Sin embargo, en la u ltima ecuaci on de Modigliani y Miller (1963), proponen calcular la estructura de nanciaci on objetivo de la empresa [D/(D + E)] utilizando valores contables de D y E, en lugar de valores de mercado. Myers (1974) fue el introductor del APV (adjusted present value ). Seg un Myers, el valor de la empresa apalancada es igual al valor de la empresa sin deuda (Vu) m as el valor actual del ahorro de impuestos debido al pago de intereses (DVTS). Myers propone calcular el DVTS del siguiente modo: DVTSt = T Kd Dt+1 Kd Dt Kd Dt+2 + + + 1 + Kd (1 + Kd)2 (1 + Kd)3 .
El argumento es que el riesgo de los ahorros de deuda es el mismo que el de la deuda. Y el valor de la empresa es:
N
APV = E0 + D0 = Vu + DVTSt =
t=1
FCFt Kd T Dt1 + . t (1 + Ku) (1 + Kd)t t=1
101
Arditti y Levy (1977) sugieren calcular el valor de la empresa descontando los Capital Cash Flows (ujo disponible para las acciones m as ujos para la deuda) en lugar del Free Cash Flow. Los Capital Cash Flows (CCF) se deben descontar al WACCBT (WACC antes de impuestos). Es f acil demostrar que:
D 0 + E0 =
t=1
FCFt
t
=
t=1
CCFt
t
(1 + WACCi )
i=1
(1 + WACCBTi )
i=1
siendo WACCBT su f ormula (2): WACCBTt = Ke Et1 Dt1 + Kd . Et1 + Dt1 Et1 + Dt1
El art culo de Arditti y Levy (1977) presenta un problema fundamental: calculan las ponderaciones de deuda (D/[E + D]) y de recursos propios (E/[E + D]) a valor contable, en lugar de a valor de mercado. Es por esta raz on por la que arman (pp. 28) que el valor de la empresa que se obtiene descontando los FCF es distinto que el que se obtiene descontando los CCF. Miller (1977) argumenta que s que existe una estructura o ptima de endeudamiento para el agregado de las empresas, pero no existe para cada empresa. Miller argumenta que debido al efecto clientela, el endeudamiento no a nade ning un valor a la empresa. Por consiguiente, seg un Miller, E + D = Vu. Tambi en introduce impuestos personales adem as de los impuestos de la empresa. La tasa de impuestos para la empresa es T, la tasa de impuestos personales sobre las acciones son TPA y la tasa de impuestos personales sobre la deuda es TPD . Seg un Miller, para una perpetuidad, el valor de la empresa sin deuda despu es de los impuestos personales es: Vu = FCF (1 TPA ) . Ku
Si la empresa tiene deuda de valor nominal N, su valor es: D = N Kd (1 TPD ) . Kd
Miller dice que la creaci on de valor debida a la deuda, en el caso de una perpetuidad, es: (1 T) (1 TPA ) D 1 . 1 TPD Pero contin ua diciendo (ver p agina 268) que un intento de una empresa por aumentar de valor endeud andose m as ser a incompatible con el equilibrio del mercado. El aumento de deuda originar a cambios en las rentabilidades exigidas a la deuda y a las
102
ndez Pablo Ferna
acciones y en los propietarios de las acciones, de manera que el valor de la empresa ser a independiente del endeudamiento. Miller contin ua su art culo diciendo que si TPA = 0, la oferta agregada de deuda debe ser tal que ofrezca un inter es R0 /(1T), siendo R0 la tasa que pagan instituciones libres de impuestos. Miller y Scholes (1978) muestran que, incluso si la tasa del impuesto sobre la renta es mayor que la tasa sobre plusval as, muchos inversores no pagar an m as que tasa sobre plusval as aplicada a los dividendos. Concluyen que los inversores ser an indiferentes entre cobrar dividendos o realizar plusval as si la empresa recompra acciones. Seg un ellos, el valor de la empresa no depender a de la pol tica de dividendos ni siquiera con impuestos personales y con impuestos sobre benecios. DeAngelo y Masulis (1980) extienden el trabajo de Miller. Considerando que la tasa marginal de impuestos es diferente para distintas empresas, predicen que las empresas utilizar an menos deuda cuantas m as posibilidades tengan de reducir el pago de impuestos por otros medios: amortizaci on, desgravaci on por inversiones. . . Miles y Ezzell (1980) sostienen que el APV y el WACC proporcionan distinto valor: salvo que el endeudamiento y, en consecuencia, Ke sean ex ogenos (no dependan del valor de la empresa en cada momento), el WACC tradicional no es apropiado para valorar empresas. Seg un ellos, la valoraci on de una empresa que quiere mantener un ratio D/E constante se debe realizar de distinto modo que si la empresa tiene un volumen prejado de deuda. En concreto, la f ormula [20] de su art culo dice que para una empresa con un objetivo de endeudamiento [D/(D + E)] jo, el free cash ow (FCF) se debe descontar a la tasa: D Kd T (1 + Ku) WACC = Ku . E+D 1 + Kd Llegan a esta f ormula partiendo de su f ormula (11), que para una perpetuidad creciente es: Et1 + Dt1 = Kd T L (Et1 + Dt1 ) FCFt Kd T Dt1 FCFt + = + . Ku g Kd g Ku g Kd g
L es el endeudamiento. Arman que la tasa correcta para descontar el ahorro de impuestos debido a la deuda (Kd T Dt1 ) es Kd para el ahorro de impuestos del primer a no, y Ku para el ahorro de impuestos de los a nos siguientes. La expresi on de Ke es su f ormula (22): Ke = Ku + D (Ku Kd) 1 + Kd (1 T) . (1 + Kd) E
103
Miles y Ezzell (1985) muestran en su f ormula (27) que la relaci on entre la beta apalancada y la beta de los activos (suponiendo que la deuda no tiene riesgo y que la beta de la deuda es cero) es L = u + T RF D u 1 E 1 + RF .
Chambers, Harris y Pringle (1982) comparan cuatro m etodos de valorar empresas por descuento de ujos: actualizar el ujo disponible para las acciones(CFac) a la tasa Ke (rentabilidad exigida a las acciones); actualizar el Free Cash Flow (FCF) al WACC (coste ponderado de deuda y acciones); actualizar el Capital Cash Flow (CCF) al WACCBT (coste ponderado de deuda y acciones antes de impuestos); y el APV de Myers (APV). Dicen que los tres primeros m etodos proporcionan el mismo valor si el endeudamiento es constante, pero que proporcionan distintos valores si no es constante. Tambi en arman que el APV s olo proporciona el mismo resultado que los otros tres m etodos en dos casos: en empresas con s olo un periodo, y en el caso de perpetuidades sin crecimiento. La raz on de esta discrepancia es que ellos calculan el ratio de endeudamiento D/[D + E] con valores contables, en lugar de valores de mercado. Su exhibit 3 es una prueba de ello: no puede ser que el WACC y Ke sean constantes. Si Ke = 11, 2%, como ellos proponen, el WACC correcto es 6,819% el primer a no (en lugar de su 5,81%) y aumenta los a nos siguientes; y el WACCBT correcto es 7,738 % el primer a no (en lugar de su 6,94%) y aumenta los a nos siguientes. calculando el ratio de endeudamiento (D/[D + E]) con valores de mercado, los tres procedimientos proporcionan el mismo valor. Harris y Pringle (1985) proponen en su f ormula (3) que WACCBT = Ku, y por consiguiente, su expresi on para el WACC es: WACC = Ku Kd T D . E+D
Tambi en proponen que el valor actual del ahorro de impuestos debido al pago de intereses (DVTS) se debe calcular descontando el ahorro de impuestos debido a la deuda (Kd T Dt1 ) a la tasa Ku: DVTSt = T Kd Dt+1 Kd Dt Kd Dt+2 + + + 2 (1 + Ku) (1 + Ku) (1 + Ku)3
Harris y Pringle proporcionan un ejemplo interesante. Para una perpetuidad con Ke = 17% , Kd = 11% , T = 46% y D = 30% , (D + E)
el resultado de Ku al que se llega por 4 procedimientos distintos es el siguiente:
104 a) seg un Modigliani y Miller (1963) y Myers (1974): b) seg un Miller (1977): c) seg un Miles y Ezzell (1980): d) seg un Harris y Pringle (1985):
ndez Pablo Ferna 15,87% 13,68% 15,26% 15,20%
Estos resultados de Ku provienen de utilizar las siguientes f ormulas: a) seg un Modigliani y Miller (1963) y Myers (1974): Ku = b) seg un Miller (1977): Ku = We Ke + Wd Kd (1 T) ; c) seg un Miles y Ezzell (1980): We Ke + Wd Kd (1 T) + Ku = T Wd Kd 1 (1 + Kd) T Wd Kd (1 + Kd) We Ke + Wd Kd (1 T) ; 1 T Wd
d) seg un Harris y Pringle (1985): Ku = We Ke + Wd Kd = WACCBT , siendo We = E/(D + E) y Wd = 1 We. Ruback (1995) supone en su f ormula (2.6) que L = u (D + E)/E d D/E. Es inmediato comprobar que con esta suposici on: WACCBT = Ku. Llega a unas f ormulas equivalentes a las de Harris y Pringle (1985). Lewellen y Emery (1986) muestran que en el caso de una perpetuidad sin crecimiento, el valor de la empresa apalancada de acuerdo a las f ormulas de Miles y Ezzell (1980) es el siguiente (ver su f ormula (7)): E0 + D0 = Vu + Kd T D + Kd T D/Ku . 1 + Kd
Tambi en muestran que en el caso de una perpetuidad sin crecimiento, el valor de la empresa apalancada de acuerdo a las f ormulas de Modigliani y Miller (1963) y Myers (1974) coinciden y es el siguiente (ver su f ormula (5)): E + D = Vu + T D . M as adelante, muestran que para una perpetuidad creciente a una tasa g , el valor de la empresa apalancada es:
n de empresas por descuento de flujos Valoracio a) seg un Modigliani y Miller (1963): E0 + D0 = Vu + b) seg un el APV de Myers (1974): E0 + D0 = Vu + c) seg un Miles y Ezzell (1980): E0 + D0 = Vu + D T Kd (1 + Ku) . (Ku g ) (1 + Kd) D T Kd ; Kd g D T Ku ; Ku g
105
Taggart (1991) proporciona un buen resumen de f ormulas de valoraci on sin impuestos personales y con impuestos personales. Propone que las f ormulas de Miles y Ezzell (1980) deben de utilizarse cuando la empresa se ajusta a su objetivo de endeudamiento una vez al a no y las de Harris y Pringle (1985) cuando la empresa se ajusta continuamente a su objetivo de endeudamiento. Damodaran (1994) argumenta18 que si todo el riesgo del negocio es soportado por las acciones, entonces la f ormula que relaciona la beta apalancada (L ) con la beta de los activos( u) es: L = u + (D/E) u (1 T). N otese que esta expresi on procede de la relaci on entre la beta apalancada, la beta de los activos y la beta de la deuda de Modigliani-Miller19 , suponiendo que la beta de la deuda es cero. Otro modo de calcular la beta apalancada en funci on de la beta de los activos es el siguiente: L = u + (D/E). Denominaremos a esta f ormula la f ormula de los pr acticos, porque se utiliza con mucha frecuencia por consultores y bancos de inversiones20 . Es obvio que seg un esta f ormula, a igualdad de u, se obtiene una superior que seg un Modigliani-Miller y que seg un Damodaran (1994).
18 pp. 31. Esta expresi on de la beta apalancada aparece en muchos libros y es frecuentemente utilizada por consultores y bancos de inversiones. 19 La relaci on entre la beta apalancada, la beta de los activos y la beta de la deuda de ModiglianiMiller es: L = u + (D/E)( u d)(1 T). 20 Dos de los muchos sitios donde aparece son: Ruback (1995), pp. 5; y Ruback (1989), pp. 2.
106
ndez Pablo Ferna
F ormulas principales de los art culos m as relevantes
En este apartado incluimos las f ormulas relevantes para la valoraci on que se derivan de los art culos mencionados en el apartado anterior. Algunas f ormulas aparecen expl citamente en los art culos. Otras se derivan a partir de las que aparecen en los art culos.
4.1
Relaci on de la rentabilidad exigida a las acciones (Ke) con la rentabilidad exigida a los activos (Ku)
Las distintas expresiones de Ke (a partir de Ku21 ) seg un los principales art culos ya comentados son22 : a) seg un Modigliani y Miller (1963): Ke = Ku + D (1 T) b) Seg un Myers (1974)23 : Ke = Ku + (D DVTS) Ku Kd . E Ku Kd . E
Para el caso de una perpetuidad creciente a una tasa g , DVTS = D T Kd/(Kd g ) y Ke es: Ku Kd Ke = Ku + D [Kd (1 T) g ] . E (Kd g ) c) Seg un Miller (1977): Ke = Ku + d) Seg un Miles y Ezzell (1980): Ke = Ku + D (Ku Kd) 1 + Kd (1 T) . (1 + Kd)E D [Ku Kd (1 T)] . E
21 La rentabilidad exigida a los activos (Ku) es igual a la rentabilidad exigida a las acciones en la empresa sin deuda. 22 Estas son las expresiones que proporcionan una valoraci on consistente: igual valor utilizando el WACC, el APV y el descuento del ujo disponible para las acciones. 23 Recuerde el lector que seg un Myers:
DVTSt = T
Kd Dt+1 Kd Dt Kd Dt+2 + + + (1 + Kd) (1 + Kd)2 (1 + Kd)3
n de empresas por descuento de flujos Valoracio e) seg un Harris y Pringle (1985) y Ruback (1995): Ke = Ku + f) seg un Damodaran (1994): Ke = Ku + g) seg un el m etodo de los pr acticos: Ke = Ku + D (Ku RF ) . E D (1 T) (Ku RF ) . E D (Ku Kd) . E
107
4.2
Distintas expresiones del WACC y del WACCBT
Las expresiones del WACC (coste ponderado de los recursos) correspondientes con los valores de Ke del apartado anterior son: a) seg un Modigliani y Miller (1963): WACC = Ku 1 b) Seg un Myers (1974): WACC = Ku DVTS (Ku Kd) + D Kd T . E+D TD . E+D
Para el caso de una perpetuidad creciente a una tasa g , DVTS = y WACC = Ku 1 E+D D Kd T D T Kd Kd g Ku g Kd g .
c) Seg un Miller (1977): WACC = Ku. d) Seg un Miles y Ezzell (1980): WACC = Ku 1 E+D Kd T 1 + Ku 1 + Kd .
e) Seg un Harris y Pringle (1985) y Ruback (1995): WACC = Ku D Kd T . E+D
108
ndez Pablo Ferna
f) Seg un Damodaran (1994): WACC = Ku D g) Seg un el m etodo de los pr acticos: WACC = Ku D RF Kd (1 T) . E+D T Ku (1 T) (Kd RF ) . E+D
Las expresiones del WACCBT (coste ponderado de los recursos, antes de impuestos) correspondientes con los valores de Ke del apartado anterior son: a) seg un Modigliani y Miller (1963): WACCBT = Ku b) Seg un Myers (1974): DVTS (Ku Kd) . E+D Para el caso de una perpetuidad creciente a una tasa g , WACCBT = Ku DVTS = y WACCBT = Ku c) Seg un Miller (1977): WACCBT = Ku + d) Seg un Miles y Ezzell (1980): WACCBT = Ku D T Kd (Ku Kd) . (E + D) (1 + Kd) D T Kd . E+D D T Kd Kd g D T (Ku Kd) . E+D
D T Kd (Ku Kd) . (E + D) (Kd g )
e) Seg un Harris y Pringle (1985) y Ruback (1995): WACCBT = Ku . f) Seg un Damodaran (1994): WACCBT = Ku D g) Seg un el m etodo de los pr acticos: WACCBT = Ku + D Kd RF . E+D (Kd RF ) T (Ku RF ) . E+D
109
4.3
Distintas expresiones del valor actual del ahorro de impuestos debido al pago de intereses (DVTS)
Las expresiones del valor creado por el endeudamiento, esto es, del valor actual del ahorro de impuestos debido al pago de intereses (DVTS) para una perpetuidad creciente a la tasa g son: a) seg un Modigliani y Miller (1963): DVTS = b) Seg un Myers (1974): DVTS = c) Seg un Miller (1977): DVTS = 0 . d) Seg un Miles y Ezzell (1980): DVTS = D Kd T 1 + Ku . (1 + Kd) (Ku g ) D Kd T . Kd g D Ku T . Ku g
e) Seg un Harris y Pringle (1985) y Ruback (1995): DVTS = f) Seg un Damodaran (1994): DVTS = D Ku T D (Kd RF ) (1 T) . Ku g D Kd T . Ku g
g) Seg un el m etodo de los pr acticos: DVTS = D Ku T D (Kd RF ) . Ku g
110
ndez Pablo Ferna
4.4
Distintas expresiones de la beta apalancada
Las distintas expresiones de la beta apalancada (L ) en funci on de la beta de los activos o beta de las acciones en la empresa sin apalancar ( u), seg un los distintos art culos, son: a) seg un Modigliani y Miller (1963)24 : L = u + b) Seg un Myers (1974): L = u + D DVTS ( u d) . E D (1 T) ( u d) . E
Para el caso de una perpetuidad creciente a una tasa g : L = u + D [Kd (1 T) g ] c) Seg un Miller (1977): L = u D+E D RF d (1 T) T E E PM . u d . E (Kd g )
d) Seg un Miles y Ezzell (1980): L = u + Kd D ( u d) 1 T . E 1 + Kd
e) Seg un Harris y Pringle (1985) y Ruback (1995): L = u + f) Seg un Damodaran (1994): L = u + g) Seg un el m etodo de los pr acticos: L = u +
24 Esta
D ( u d) . E
D (1 T) u . E
D u . E
f ormula coincide con la (2A.6) de Taggart (1991), porque el asume que d = 0
111
4.5
Distintas expresiones del ujo disponible para las acciones ajustado al riesgo de los activos
Las distintas expresiones de CFac\\Ku (ujo disponible para las acciones ajustado al riesgo de los activos)25 seg un los distintos art culos son: a) seg un Modigliani y Miller (1963): CFac D (Ku Kd) (1 T) . b) Seg un Myers (1974): CFac (Vu E) (Ku Kd) . c) Seg un Miller (1977): CFac D [Ku Kd (1 T)] . d) Seg un Miles y Ezzell (1980): CFac D (Ku Kd) 1 + Kd (1 T) . 1 + Kd
e) Seg un Harris y Pringle (1985) y Ruback (1995): CFac D (Ku Kd) . f) Seg un Damodaran (1994): CFac D (Ku RF ) (1 T) . g) Seg un el m etodo de los pr acticos: CFac D (Ku RF ) .
25 Sobre el signicado del ujo disponible para las acciones ajustado al riesgo de los activos, y del de free cash ow ajustado al riesgo de los activos ver Fern andez (1999) pp. 357.
112
ndez Pablo Ferna
5
5.1
Diferencias en la valoraci on seg un los art culos m as relevantes

Perpetuidad creciente con endeudamiento prejado del 30%
Aplicando las f ormulas anteriores a una empresa con FCF1 = 100, Ku = 10%, Kd = 7%, D/(D + E) = 30%, T = 35%, RF = 5%, y g = 5%; se obtienen los valores de la tabla 4. El valor de la empresa sin apalancar (Vu) es en todos los casos 2.000. N otese c omo, seg un Myers, Ke < Ku = 10%, lo que no tiene sentido. Seg un Myers, DVTS > D cuando g > Kd (1 T), en el ejemplo, cuando g > 4, 55%. Como veremos m as adelante, si g > 4, 55%, Ke < Ku, lo que no tiene sentido.
ModiglianiMiller 8,950% 10,836% 9,685% 2.531,65 2.000,00 1.772,15 759,49 531,65 103,42 Myers 8,163% 9,711% 8,898% 3.162,1 2.000,00 2.213,4 948,62 1.162,06 104,27 Miller 10,000% 12,336% 10,735% 2.000,0 2.000,00 1.400,0 600,00 0,00 102,70 MilesEzzell 9,244% 11,256% 9,979% 2.356,0 2.000,00 1.649,2 706,81 356,05 103,18 HarrisPringle 9,265% 11,286% 10,000% 2.344,7 2.000,00 1.641,3 703,40 344,67 103,17 Damodaran 9,340% 11,393% 10,075% 2.304,1 2.000,00 1.612,9 691,24 304,15 103,11 Pr acticos 9,865% 12,143% 10,600% 2.055,5 2.000,00 1.438,8 616,65 55,50 102,77
WACC Ke WACCBT E+D Vu E D DVTS CFac
Tabla 4: Ejemplo de valoraci on de una empresa. FCF1 = 100, Ku = 10%, Kd = 7%, D/(D + E) = 30%, T = 35%, RF = 5%, y g = 5%. Si introducimos cambios en el crecimiento, las tablas 5 a 11 muestran los par ametros fundamentales de la valoraci on en funci on del crecimiento g . La tabla 5 muestra que el valor de la empresa seg un Modigliani-Miller y seg un Myers son iguales para una perpetuidad (cuando no hay crecimiento). Con crecimiento, el valor de la empresa seg un Myers es superior al valor de la empresa seg un Modigliani-Miller. Todas las dem as teor as proporcionan valores inferiores a Modigliani-Miller. Seg un Myers el valor de la empresa es innito para crecimientos iguales o superiores a g = Kd[D(1 T) + E]/(E + D); en el ejemplo cuando g 6, 265%.
g 0% 1% 3% 4% 6% 7% ModiglianiMiller 1.117,3 1.257,9 1.680,7 2.020,2 3.389,8 5.128,2 Myers 1.117,3 1.266,2 1.750,2 2.207,5 9.434,0 Miller 1.000,0 1.111,1 1.428,6 1.666,7 2.500,0 3.333,3 MilesEzzell 1.081,7 1.212,9 1.601,4 1.906,8 3.082,2 4.455,5 HarrisPringle 1.079,3 1.209,9 1.596,2 1.899,3 3.062,8 4.415,0 Damodaran 1.070,7 1.199,0 1.577,3 1.872,7 2.994,0 4.273,5 Pr acticos 1.013,7 1.128,0 1.456,7 1.705,0 2.587,3 3.490,4
Tabla 5: Valor de la empresa (E + D) en funci on del crecimiento g (D/(D + E) = 30%).
113
La tabla 6 muestra que el WACC de la empresa seg un todas las teor as es independiente del crecimiento, salvo seg un Myers. Seg un Myers, el WACC desciende cuando aumenta el crecimiento y es igual al crecimiento cuando g = Kd [D(1 T)+E]/(E+D); en el ejemplo, cuando g = 6, 265%.
g 0% 2% 4% 6% ModiglianiMiller 8,95% 8,95% 8,95% 8,95% Myers 8,95% 8,82% 8,53% 7,06% Miller 10,00% 10,00% 10,00% 10,00% MilesEzzell 9,24% 9,24% 9,24% 9,24% HarrisPringle 9,27% 9,27% 9,27% 9,27% Damodaran 9,34% 9,34% 9,34% 9,34% Pr acticos 9,87% 9,87% 9,87% 9,87%
Tabla 6: WACC en funci on del crecimiento g (D/(D + E) = 30%). La tabla 7 muestra que el WACCBT de la empresa seg un todas las teor as es independiente del crecimiento, salvo seg un Myers. Seg un Myers, el WACCBT desciende cuando aumenta el crecimiento.
g 0% 2% 4% 6% ModiglianiMiller 9,69% 9,69% 9,69% 9,69% Myers 9,69% 9,56% 9,27% 7,80% Miller 10,74% 10,74% 10,74% 10,74% MilesEzzell 9,98% 9,98% 9,98% 9,98% HarrisPringle 10,00% 10,00% 10,00% 10,00% Damodaran 10,08% 10,08% 10,08% 10,08% Pr acticos 10,60% 10,60% 10,60% 10,60%
Tabla 7: WACCBT en funci on del crecimiento g (D/(D + E) = 30%). La tabla 8 muestra que el DVTS seg un Modigliani-Miller y seg un Myers son iguales para una perpetuidad (cuando no hay crecimiento). Con crecimiento, el valor del DVTS seg un Myers es superior al DVTS seg un Modigliani-Miller. Todas las dem as teor as proporcionan valores inferiores a Modigliani-Miller.
g 0% 2% 4% 5% 6% 7% ModiglianiMiller 117,3 188,8 353,5 531,6 889,8 1.794,9 Myers 117,3 215,4 540,8 1.162,1 6.934,0 Miller 0,0 0,0 0,0 0,0 0,0 0,0 MilesEzzell 81,7 130,4 240,1 356,0 582,2 1.122,2 HarrisPringle 79,3 126,5 232,7 344,7 562,8 1.081,7 Damodaran 70,7 112,4 206,0 304,1 494,0 940,2 Pr acticos 13,7 21,5 38,4 55,5 87,3 157,1
Tabla 8: DVTS en funci on del crecimiento g (D/(D + E) = 30%). La tabla 9 muestra el cash ow disponible para las acciones en el a no uno seg un las distintas teor as. La f ormula (5) para una perpetuidad se convierte en: CFac1 = FCF1 + D0 [g Kd (1 T)] . Como FCF = 100, el CFac es mayor cuanto mayor es el valor de la empresa, porque la deuda se ha jado en el 30% del valor de la empresa.
114
g 0% 1% 2% 3% 4% 5% 6% ModiglianiMiller 84,7 86,6 89,0 92,2 96,7 103,4 114,7 Myers 84,7 86,5 88,8 91,9 96,4 104,3 141,0 Miller 86,4 88,2 90,4 93,4 97,3 102,7 110,9 MilesEzzell 85,2 87,1 89,4 92,6 96,9 103,2 113,4 HarrisPringle 85,3 87,1 89,5 92,6 96,9 103,2 113,3
ndez Pablo Ferna

Damodaran 85,4 87,2 89,6 92,7 96,9 103,1 113,0 Pr acticos 86,2 88,0 90,3 93,2 97,2 102,8 111,3
Tabla 9: Cash ow disponible para las acciones en el periodo 1 (CFac1 ) en funci on del crecimiento g (D/(D + E) = 30%). La tabla 10 muestra que la rentabilidad exigida a las acciones seg un todas las teor as es independiente del crecimiento, salvo seg un Myers. Seg un Myers, Ke desciende cuando aumenta el crecimiento y es igual a Ku cuando g = Kd (1 T), en el ejemplo para g = 4, 55%. Esto, l ogicamente, no tiene ning un sentido.
g 0% 2% 4% 5% 6% 7% ModiglianiMiller 10,8% 10,8% 10,8% 10,8% 10,8% 10,8% Myers 10,8% 10,7% 10,2% 9,7% 8,1% Miller 12,3% 12,3% 12,3% 12,3% 12,3% 12,3% MilesEzzell 11,3% 11,3% 11,3% 11,3% 11,3% 11,3% HarrisPringle 11,3% 11,3% 11,3% 11,3% 11,3% 11,3% Damodaran 11,4% 11,4% 11,4% 11,4% 11,4% 11,4% Pr acticos 12,1% 12,1% 12,1% 12,1% 12,1% 12,1%
Tabla 10: Ke en funci on del crecimiento g (D/(D + E) = 30%).

g 0% 2% 4% 6% 7% 8% ModiglianiMiller 782,1 1.007,2 1.414,1 2.372,9 3.589,7 7.368,4 Myers 782,1 1.025,8 1.545,3 6.603,8 Miller 700,0 875,0 1.166,7 1.750,0 2.333,3 3.500,0 MilesEzzell 757,2 966,3 1.334,8 2.157,6 3.118,9 5.625,2 HarrisPringle 755,5 963,5 1.329,5 2.144,0 3.090,5 5.533,6 Damodaran 749,5 953,7 1.310,9 2.095,8 2.991,5 5.223,9 Pr acticos 709,6 890,0 1.193,5 1.811,1 2.443,3 3.753,4
Tabla 11: Valor de las acciones (Eo) en funci on del crecimiento g (D/(D + E) = 30%). Si cambia el endeudamiento, las tablas 12 a 16 muestran los par ametros fundamentales de la valoraci on en funci on del endeudamiento. La tabla 12 muestra el DVTS en funci on del endeudamiento seg un las diferentes teor as. El valor del DVTS seg un Myers es superior al DVTS seg un Modigliani-Miller. Todas las dem as teor as proporcionan valores inferiores a Modigliani-Miller. Se puede comprobar que el DVTS seg un Myers se hace innito para un endeudamiento D/(D + E) = (Kd g )/(T Kd), en nuestro ejemplo 81,63%.

g 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% ModiglianiMiller 0,0 150,5 325,6 531,6 777,8 1.076,9 1.448,3 1.921,6 2.545,5 3.405,4 4.666,7 Myers 0,0 279,2 649,0 1.162,1 1.921,6 3.161,3 5.547,2 12.035,1 98.000,0 Miller 0,0 0,0 0,0 0,0 0,0 0,0 0,0 0,0 0,0 0,0 0,0 MilesEzzell 0,0 106,1 224,1 356,0 504,7 673,3 866,3 1.089,4 1.350,0 1.658,7 2.030,1 HarrisPringle 0,0 103,0 217,3 344,7 487,6 649,0 832,9 1.044,1 1.289,5 1.577,8 1.921,6 Damodaran 0,0 92,1 193,0 304,1 427,2 564,1 717,4 890,2 1.086,4 1.311,3 1.571,4
115
Pr acticos 0,0 18,2 36,7 55,5 74,7 94,2 114,2 134,5 155,2 176,3 197,8
Tabla 12: Valor actual del ahorro de impuestos por pago de intereses (DVTS) en funci on del crecimiento g (D/(D + E) = 30%). La tabla 13 muestra que el WACC de la empresa seg un todas las teor as desciende con el endeudamiento, salvo seg un Miller. S olo seg un Myers el WACC llega a ser inferior al crecimiento (5%) (para endeudamientos superiores a D/(D + E) > (Kd g )/(T Kd), en nuestro ejemplo 81,63%).
g 0% 20% 40% 60% 80% 90% 100% ModiglianiMiller 10,00% 9,30% 8,60% 7,90% 7,20% 6,85% 6,50% Myers 10,00% 8,78% 7,55% 6,33% 5,10% 4,49% 3,88% Miller 10,00% 10,00% 10,00% 10,00% 10,00% 10,00% 10,00% MilesEzzell 10,00% 9,50% 8,99% 8,49% 7,99% 7,73% 7,48% HarrisPringle 10,00% 9,51% 9,02% 8,53% 8,04% 7,80% 7,55% Damodaran 10,00% 9,56% 9,12% 8,68% 8,24% 8,02% 7,80% Pr acticos 10,00% 9,91% 9,82% 9,73% 9,64% 9,60% 9,55%
Tabla 13: Coste ponderado de los recursos (WACC) en funci on del endeudamiento (g=5%). La tabla 14 muestra que el WACCBT de la empresa en funci on del endeudamiento: desciende con el seg un Modigliani-Miller, Myers y Miles-Ezzell; es constante (igual a Ku) seg un Harris-Pringle y asciende seg un Miller, Damodaran y los pr acticos.
D D +E 0% 20% 40% 60% 80% 90% ModiglianiMiller 10,00% 9,79% 9,58% 9,37% 9,16% 9,06% Myers 10,00% 9,27% 8,53% 7,80% 7,06% 6,69% Miller 10,00% 10,49% 10,98% 11,47% 11,96% 12,21% MilesEzzell 10,00% 9,99% 9,97% 9,96% 9,95% 9,94% HarrisPringle 10,00% 10,00% 10,00% 10,00% 10,00% 10,00% Damodaran 10,00% 10,05% 10,10% 10,15% 10,20% 10,23% Pr acticos 10,00% 10,40% 10,80% 11,20% 11,60% 11,80%
Tabla 14: Coste ponderado de los recursos antes de impuestos (WACCBT ) en funci on del endeudamiento (g=5%).
116
D D +E 0% 20% 40% 60% 80% 90% 95%
ndez Pablo Ferna
ModiglianiMiller 10,00% 10,49% 11,30% 12,93% 17,80% 27,55% 47,05%
Myers 10,00% 9,83% 9,55% 8,99% 7,30% 3,93%
Miller 10,00% 11,36% 13,63% 18,18% 31,80% 59,05% 113,55%
MilesEzzell 10,00% 10,73% 11,95% 14,40% 21,73% 36,38% 65,69%
HarrisPringle 10,00% 10,75% 12,00% 14,50% 22,00% 37,00% 67,00%
Damodaran 10,00% 10,81% 12,17% 14,88% 23,00% 39,25% 71,75%
Pr acticos 10,00% 11,25% 13,33% 17,50% 30,00% 55,00% 105,00%
Tabla 15: Rentabilidad exigida a las acciones (Ke) seg un el endeudamiento (g=5%).
D D+E 0% 20% 40% 60% 70% 80% 90% 100% ModiglianiMiller 2.000,00 2.325,58 2.777,78 3.448,28 3.921,57 4.545,45 5.405,41 6.666,67 Myers 2.000,00 2.649,01 3.921,57 7.547,17 14.035,09 100.000,0 19.512,2 8.888,89 Miller 2.000,00 2.000,00 2.000,00 2.000,00 2.000,00 2.000,00 2.000,00 2.000,00 MilesEzzell 2.000,00 2.224,07 2.504,68 2.866,33 3.089,36 3.350,03 3.658,75 4.030,13 HarrisPringle 2.000,00 2.217,29 2.487,56 2.832,86 3.044,14 3.289,47 3.577,82 3.921,57 Damodaran 2.000,00 2.192,98 2.427,18 2.717,39 2.890,17 3.086,42 3.311,26 3.571,43 Pr acticos 2.000,00 2.036,66 2.074,69 2.114,16 2.134,47 2.155,17 2.176,28 2.197,80
Tabla 16: Valor de la deuda y las acciones (Eo+Do) seg un el endeudamiento (g=5%).
5.2
Perpetuidad creciente con deuda prejada
La tabla 17 es id entica a la tabla 4, con la u nica diferencia de que se ja el nivel de deuda inicial en 759,49 (en lugar del ratio de endeudamiento en 30%). Aplicando las f ormulas anteriores se obtienen los valores de la tabla 17. El valor de la empresa sin apalancar (Vu) es en todos los casos 2.000. N otese c omo, seg un Myers, Ke < Ku = 10%, lo que no tiene mucho sentido.
ModiglianiMiller 8,950% 10,836% 9,685% 2.531,65 2.000,00 1.772,15 759,49 531,65 103,42 30,00% Myers 8,413% 9,764% 9,048% 2.930,38 2.000,00 2.170,89 759,49 930,38 103,42 25,92% Miller 10,000% 13,337% 10,930% 2.000,00 2.000,00 1.240,51 759,49 0,00 103,42 37,97% MilesEzzell 9,197% 11,372% 9,978% 2.382,59 2.000,00 1.623,09 759,49 382,59 103,42 31,88% HarrisPringle 9,216% 11,413% 10,000% 2.372,15 2.000,00 1.612,66 759,49 372,15 103,42 32,02% Damodaran 9,284% 11,568% 10,081% 2.334,18 2.000,00 1.574,68 759,49 334,18 103,42 32,54% Pr acticos 9,835% 12,901% 10,734% 2.068,35 2.000,00 1.308,86 759,49 68,35 103,42 36,72%
WACC Ke WACCBT E+D Vu E D DVTS CFac D/(D+E)
Tabla 17: Ejemplo de valoraci on de una empresa. FCF1 = 100, Ku = 10%, Kd = 7%, D = 759, 49, T = 35%, RF = 5%, y g = 5%.
117
Las tablas 18 a 24 muestran los par ametros fundamentales de la valoraci on en funci on del crecimiento g . La tabla 18 muestra que el valor de la empresa seg un Modigliani-Miller y Myers son iguales para una perpetuidad (cuando no hay crecimiento). El valor de la empresa seg un Myers es superior al valor de la empresa seg un Modigliani-Miller. Todas las dem as teor as proporcionan valores inferiores a Modigliani-Miller. Seg un Myers el valor de la empresa es innito para crecimientos iguales o superiores a Kd [D (1 T) + E]/(E + D) (en el ejemplo, para g 6, 265%).
g 0% 2% 4% 6% 7% 9% ModiglianiMiller 1.265,8 1.582,3 2.109,7 3.164,6 4.219,4 12.658,2 Myers 1.265,8 1.622,2 2.286,9 4.360,8 Miller 1.000,0 1.250,0 1.666,7 2.500,0 3.333,3 10.000,0 MilesEzzell 1.191,3 1.489,1 1.985,5 2.978,2 3.971,0 11.912,9 HarrisPringle 1.186,1 1.482,6 1.976,8 2.965,2 3.953,6 11.860,8 Damodaran 1.167,1 1.458,9 1.945,1 2.917,7 3.890,3 11.670,9 Pr acticos 1.034,2 1.292,7 1.723,6 2.585,4 3.447,3 10.341,8
Tabla 18: Valor de la empresa (E + D) en funci on del crecimiento g (D = 759, 49). La tabla 19 muestra que el WACC de la empresa seg un todas las teor as aumenta con el crecimiento, salvo seg un Miller. Seg un Myers, el WACC es igual al crecimiento cuando g = Kd [D (1 T) + E]/(E + D); en el ejemplo cuando g = 7%.
g 0% 2% 4% 6% 7% 8% 9% ModiglianiMiller 7,90% 8,32% 8,74% 9,16% 9,37% 9,58% 9,79% Myers 7,90% 8,16% 8,37% 8,29% Miller 10,00% 10,00% 10,00% 10,00% 10,00% 10,00% 10,00% MilesEzzell 8,39% 8,72% 9,04% 9,36% 9,52% 9,68% 9,84% HarrisPringle 8,43% 8,74% 9,06% 9,37% 9,53% 9,69% 9,84% Damodaran 8,57% 8,85% 9,14% 9,43% 9,57% 9,71% 9,86% Pr acticos 9,67% 9,74% 9,80% 9,87% 9,90% 9,93% 9,97%
Tabla 19: WACC en funci on del crecimiento g (D = 759, 49). La tabla 20 muestra que el WACCBT aumenta con g seg un Modigliani-Miller y Miles-Ezzell y desciende para el resto (salvo en Harris-Pringle, en que es constante).
g 0% 2% 4% 6% 7% 8% 9% ModiglianiMiller 9,37% 9,50% 9,62% 9,75% 9,81% 9,87% 9,94% Myers 9,37% 9,31% 9,19% 8,72% Miller 11,86% 11,49% 11,12% 10,74% 10,56% 10,37% 10,19% MilesEzzell 9,96% 9,96% 9,97% 9,98% 9,99% 9,99% 10,00% HarrisPringle 10,00% 10,00% 10,00% 10,00% 10,00% 10,00% 10,00% Damodaran 10,16% 10,13% 10,10% 10,07% 10,05% 10,03% 10,02% Pr acticos 11,47% 11,18% 10,88% 10,59% 10,44% 10,29% 10,15%
Tabla 20: WACCBT en funci on del crecimiento g (D = 759, 49).
118
ndez Pablo Ferna
La tabla 21 muestra que el DVTS seg un Modigliani-Miller y seg un Myers son iguales para una perpetuidad (cuando no hay crecimiento). Con crecimiento, el valor del DVTS seg un Myers es superior al DVTS seg un Modigliani-Miller. Todas las dem as teor as proporcionan valores inferiores a Modigliani-Miller.
g 0% 2% 4% 6% 7% 9% ModiglianiMiller 265,8 332,3 443,0 664,6 886,1 2.658,2 Myers 265,8 372,2 620,3 1.860,8 Miller 0,0 0,0 0,0 0,0 0,0 0,0 MilesEzzell 191,3 239,1 318,8 478,2 637,6 1.912,9 HarrisPringle 186,1 232,6 310,1 465,2 620,3 1.860,8 Damodaran 167,1 208,9 278,5 417,7 557,0 1.670,9 Pr acticos 34,2 42,7 57,0 85,4 113,9 341,8
Tabla 21: DVTS en funci on del crecimiento g (D = 759, 49). La tabla 22 muestra que la rentabilidad exigida a las acciones seg un todas las teor as disminuye con el crecimiento. Seg un Myers, Ke es igual a Ku (10%) cuando g = Kd (1 T) (en el ejemplo, 4, 55%). Esto, l ogicamente, no tiene ning un sentido.
g 0% 2% 4% 5% 6% 7% 9% 10% ModiglianiMiller 12,9% 11,8% 11,1% 10,8% 10,6% 10,4% 10,1% 10,0% Myers 12,9% 11,3% 10,3% 9,8% 9,1% Miller 27,2% 18,4% 14,6% 13,3% 12,4% 11,6% 10,4% 10,0% MilesEzzell 15,2% 13,1% 11,8% 11,4% 11,0% 10,7% 10,2% 10,0% HarrisPringle 15,3% 13,2% 11,9% 11,4% 11,0% 10,7% 10,2% 10,0% Damodaran 16,1% 13,5% 12,1% 11,6% 11,1% 10,8% 10,2% 10,0% Pr acticos 23,8% 17,1% 13,9% 12,9% 12,1% 11,4% 10,4% 10,0%
Tabla 22: Ke en funci on del crecimiento g (D = 759, 49).
5.3
Diferencias en la valoraci on para la empresa de la tabla 1
La tabla 3 conten a la valoraci on de la empresa de la tabla 1 seg un Modigliani-Miller. Aqu recogeremos la valoraci on de Delta Inc. seg un Myers (1974), Harris-Pringle (1985), Ruback (1995), Damodaran (1994) y el m etodo de los pr acticos.
DVTS = VA(Kd; D Kd T) Ke E WACC E+D WACCBT CFac\\Ku FCF\\Ku 0 514,92 20,61% 1.116,25 14,555% 2.116,25 16,540% 1 534,71 20,22% 1.259,28 14,721% 2.259,28 16,580% 57,90 237,90 2 556,88 20,17% 1.501,86 14,940% 2.601,86 16,716% -15,92 64,08 3 577,50 19,98% 1.583,79 14,987% 2.683,79 16,709% 188,41 386,41 4 600,60 19,98% 1.647,14 14,987% 2.791,14 16,709% 221,73 375,73
Tabla 23: Valoraci on de Delta Inc. seg un Myers (1974).

0 294,72 24,70% 896,05 15,785% 1.896,05 18,000% 1 305,77 23,82% 1.030,34 15,931% 2.030,34 18,000% 27,00 207,00 2 318,81 23,22% 1.263,80 16,046% 2.363,80 18,000% -48,00 32,00 3 330,00 22,94% 1.336,29 16,104% 2.436,29 18,000% 155,00 353,00
119
4 343,20 22,94% 1.389,74 16,104% 2.533,74 18,000% 187,08 341,08
DVTS Ke E WACC E+D WACCBT CFac\\Ku FCF\\Ku
Tabla 24: Valoraci on de Delta Inc. seg un Harris y Pringle (1985), y Ruback (1995).
DVTS Ke E WACC D+E WACCBT CFac\\Ku FCF\\Ku
0 350,86 23,46% 952,19 15,439% 1.952,19 17,590%
1 364,02 22,78% 1.088,58 15,606% 2.088,58 17,617% 35,00 215,00
2 379,54 22,32% 1.324,53 15,732% 2.424,53 17,637% -40,00 40,00
3 392,86 22,09% 1.399,14 15,799% 2.499,14 17,648% 163,80 361,80
4 408,57 22,09% 1.455,11 15,799% 2.599,11 17,648% 195,88 349,88
Tabla 25: Valoraci on de Delta Inc. seg un Damodaran (1994).
DVTS Ke E WACC D+E WACCBT CFac\\Ku FCF\\Ku
0 154,38 28,59% 755,71 16,747% 1.755,71 19,139%
1 160,17 27,04% 884,73 16,833% 1.884,73 19,061% 7,00 187,00
2 167,00 25,91% 1.111,99 16,906% 2.211,99 18,995% -68,00 12,00
3 172,86 25,46% 1.179,14 16,938% 2.279,14 18,965% 133,00 331,00
4 179,77 25,46% 1.226,31 16,938% 2.370,31 18,965% 165,08 319,08
Tabla 26: Valoraci on de Delta Inc. seg un el m etodo de los pr acticos.
120
ndez Pablo Ferna
El problema de fondo: el valor del ahorro de impuestos debido a los intereses (DVTS)
Denimos Gu como el valor actual de los impuestos en la empresa sin apalancar y GL como el valor actual de los impuestos en la empresa apalancada. KIU es la rentabilidad exigida a los impuestos en la empresa sin apalancar y KIL es la rentabilidad exigida a los impuestos en la empresa apalancada. IMPu son los impuestos de la empresa sin apalancar; IMPL son los impuestos de la empresa apalancada. FCFo es el free cash ow de la empresa sin impuestos y Kuo es la rentabilidad exigida a los activos de la empresa sin impuestos. Vuo es el valor de la empresa sin apalancar y sin impuestos. Por consiguiente, Vuo = VA[Kuo ; FCFo] . Si no hay costes derivados del endeudamiento, el valor total de la empresa sin deuda (valor de las acciones, Vu, m as valor actual de los impuestos, Gu) es igual al valor total de la empresa aplancada (valor de las acciones, E, m as valor de la deuda, D, m as valor actual de los impuestos, GL ): Vuot = Vut + Gut = Et + Dt + GL t . (15)
La f ormula (16) presenta la igualdad de la suma de los ujos de la empresa sin deuda y de la suma de los ujos de la empresa apalancada. FCFo = IMPu + FCF = IMPL + CFac + CFd = IMPL + CCF . De (15) y (16) se deriva: Vuo Kuo = = Vu Ku + Gu KIU = E Ke + D Kd + GL KIL (D + E) WACCBT + GL KIL . (17) (16)
El llamado valor actual neto del ahorro de impuestos debido al pago de intereses (DVTS) es: DVTSt = Gut GL t . (18) El DVTS es la diferencia de dos valores actuales netos, de dos ujos (el de los impuestos de la empresa sin apalancar y el de los impuestos de la empresa apalancada) que tienen, obviamente, distinto riesgo. En una perpetuidad creciente: DVTSt = Gut GL t = La relaci on entre IMPu y IMPL es: IMPut+1 IMPL t+1 = Dt Kd T . (20) IMPL t+1 IMPut+1 . KIU g KIL g (19)
121
L ogicamente, los impuestos de la empresa sin apalancar tienen menos riesgo que los impuestos de la empresa apalancada, y por consiguiente: KIU < KIL . Una l ogica limitaci on del valor del DVTS es que DVTS < GU . Despejando el valor de KIU y KIL de (15), (16) y (17) se obtiene: KIU KIL = = Vuo Kuo Vu Ku . Vuo Vu Vuo Kuo E Ke D Kd . Vuo E D (22) (21)
La relaci on entre KIU y KIL es: KIL = Gu KIU E Ke D Kd + Vu Ku . Vuo E D
En la empresa sin apalancar, la relaci on entre los impuestos y el benecio antes de impuestos es: IMPu = T BATu . (23) La relaci on entre el free cash ow y los impuestos de la empresa sin apalancar es IMPu = T FCF + H . 1T (24)
H es un par ametro que incorpora el aumento de necesidades operativas de fondos (necesidades de circulante) m as la compra de activos jos menos la amortizaci on del a no: H = (NOF) + (AF) amortizaci on = (NOF) + (AFN) , (25) siendo AFN los activos jos netos. En la empresa apalancada, la relaci on entre el cash ow para las acciones y los impuestos de la empresa sin apalancar es IMPL = T CFac + H D . 1T (26)
En una perpetuidad creciente a la tasa g : H = g NOF + g AFN 0 , IMPu FCF FCFo = + , Kuo g KIU g Ku g FCFo IMPL FCF = + . Kuo g KIL g WACC g
Vuo = Vuo =
122
ndez Pablo Ferna
En una perpetuidad creciente, KIU < Ku. Al ser Kuo el coste promedio de KIU y Ku, tambi en debe cumplirse que KIU < Kuo. Por consiguiente: KIU < Kuo < Ku. Seg un Myers, DVTSt = Gut GL t = IMPL t+1 D Kd T IMPut+1 = , Kd g Kd g Kd g
esto es: KIU = KIL = Kd, lo que no tiene sentido. Seg un Harris y Pringle, DVTSt = Gut GL t = IMPL t+1 D Kd T IMPut+1 = , Ku g Ku g Ku g
esto es: KIU = KIL = Ku, lo que no tiene sentido. En una perpetuidad sin crecimiento26 , resulta (seg un Modigliani Miller) que KIU = Ku; y KIL = Ke. Es obvio que en una perpetuidad sin crecimiento: GL = E T ; 1T Gu = Vu T . 1T
En las otras teor as que hemos visto, no es cierto que DVTSt = Gut GL t . No es veros mil una teor a que proporcione un valor del DVTS superior al de Modigliani-Miller, porque ello signicar a que el valor conjunto de la empresa apalancada (Et + Dt + GL t ) es superior al valor conjunto de la empresa sin apalancar (Vut + Gut ). Todas las teor as proporcionan un valor del DVTS inferior al de Modigliani-Miller, excepto la de Myers (1974) en algunos casos: con crecimiento proporciona valores del DVTS superiores al de Modigliani-Miller. Podemos considerar el DVTS de las diversas teor as como la diferencia entre el DVTS seg un Modigliani-Miller (DVTSMM ) y un coste del apalancamiento (CA) que introduce impl citamente cada teor a. E + D = Vu + DVTS = Vu + DVTSMM CA , CA = DVTSMM DVTS , siendo DVTSMM = VA[Ku ; DTKu]. Seg un esto, el coste del apalancamiento seg un las distintas teor as resulta: a) seg un Modigliani y Miller (1963): CA = 0. b) Seg un Myers (1974): CA = VA[Ku ; DTKu] VA[Kd ; DTKd] .
26 En una perpetuidad sin crecimiento, H = 0. La compra de activos jos es id entica a la amortizaci on y las necesidades operativas de fondos son constantes.
n de empresas por descuento de flujos Valoracio c) Seg un Miller (1977): CA = VA[Ku ; DTKu] . d) Seg un Miles y Ezzell (1980): CA = VA[Ku ; T D Kd] 1 + Ku . 1 + Kd0
123
e) Seg un Harris y Pringle (1985) y Ruback (1995): CA = VA[Ku ; DTKu DTKd] . f) Seg un Damodaran (1994): CA = VA[Ku ; D (Kd RF ) (1 T)] . g) Seg un el m etodo de los pr acticos: CA = VA[Ku ; D (Kd RF (1 T) + D(Ku RF ), T] . A partir de los resultados anteriores, o bien a partir de las expresiones del valor de la empresa en funci on del free cash ow ajustado al riesgo del negocio o del cash ow disponible para las acciones ajustado al riesgo del negocio, es f acil comprobar las diferencias en la valoraci on para una empresa con un nivel de deuda prejado: EModigliani y Miller EDamodaran = VA[Ku ; D(Kd RF (1 T)], EModigliani y Miller EPr acticos = VA[Ku ; D(Kd RF ) (1 T)+D T (Ku RF )], EModigliani y Miller EHarris y Pringle = VA[Ku ; D T (Ku Kd)], EModigliani y Miller EMyers = VA[Ku ; (D T DVTSM yers) (Ku Kd)], EHarris y Pringle EPr acticos = VA[Ku ; D (Kd RF ).
124
ndez Pablo Ferna
Anexo 1: abreviaturas utilizadas en el documento

d u L CCF CFac Dt DVTS Et Evct EVA FCF Beta de la deuda (Beta, or systematic risk, of the debt ). Beta de las acciones de la empresa sin apalancar (Beta, or systematic risk, of the unlevered rm ). Beta de las acciones de la empresa apalancada (Beta, or systematic risk, of the levered rm ). (Capital Cash Flow), cash ow disponible para los tenedores de acciones y de deuda. Cash ow available for all stakeholders: equity and debt. Cash ow disponible para las acciones (available for equityholders ). Valor de la deuda en t. Market value of debt. (Discounted Value of Tax Shields), valor actual del ahorro de impuestos debidos a la deuda. Valor de las acciones en t (Market value of equity ). Valor contable de las acciones en t (Book value of equity ). (Economic Value Added), benecio antes de intereses menos el valor contable de la empresa multiplicado por el coste promedio de los recursos. (Free Cash Flow), cash ow disponible para las acciones si la empresa no tuviera deuda (Cash ow available for equityholders in the hypothetical unlevered rm ). Free cash ow de la empresa sin impuestos (Cash ow available for equityholders in the hypothetical unlevered rm without taxes ). (growth), crecimiento. Valor actual de los impuestos de la empresa apalancada (Present value of taxes of the levered company ). Valor actual de los impuestos de la empresa sin deuda (Present value of taxes of the unlevered company ). Impuestos si la empresa est a apalancada (Taxes of the levered company ). Impuestos si la empresa no est a apalancada (Taxes of the unlevered company ). Rentabilidad exigida a la deuda (Required return on the debt ). Rentabilidad exigida a las acciones de la empresa apalancada (Cost of equity. Required return on the equity ows ).
FCFo g GL Gu IMPL IMPu Kd Ke
n de empresas por descuento de flujos Valoracio KIU KIL Ku
125
Tasa de descuento de los impuestos pagados por la empresa sin apalancar (Required return on the taxes of the unlevered company ). Tasa de descuento de los impuestos pagados por la empresa apalancada (Required return on the taxes of the levered company ). Rentabilidad exigida a las acciones de la empresa sin apalancar. Rentabilidad exigida a los activos (Required return on the rms unlevered ows. Unlevered, or all-equity cost of capital ). Rentabilidad exigida a los activos de la empresa sin impuestos (Required return on the rms unlevered ows without taxes ). (Net Operating Prot After Taxes). Benecio de la empresa sin apalancar (sin deuda). Tambi en se llama BAIDT (benecio antes de intereses despu es de impuestos). Prima de mercado = E (RM ) RF = rentabilidad exigida al mercado por encima de la tasa sin riesgo (Market risk premium ). Tasa de inter es sin riesgo (Risk free rate ). Tasa del impuesto sobre el benecio (Corporate tax rate ). Valor Actual (Net present value ). Valor de las acciones de la empresa sin deuda (Market value of the unlevered rm ). (Weighted Average Cost of Capital). Coste promedio ponderado de los recursos (deuda y acciones), utilizando en la ponderaci on el valor de mercado de la deuda y las acciones.
Kuo NOPAT
PM RF T VA Vu WACC
WACCBT (Weighted Average Cost of Capital Before Taxes). Coste promedio ponderado de los recursos sin tener en cuenta los impuestos.
Anexo 2: m etodos de valoraci on de empresas por descuento de ujos cuando el valor de mercado de la deuda (D) no coincide con su valor nominal (N )
Este anexo contiene las expresiones fundamentales de los cuatro m etodos fundamentales para valorar empresas por descuento de ujos, cuando el valor de mercado de la deuda (D) no coincide con su valor nominal (N )27 .
27 Estas
f ormulas proceden de Fern andez (1999), pp. 389-391.
126
ndez Pablo Ferna
Si el valor de la deuda (D) no coincide con su valor nominal (N ) es porque la rentabilidad exigida a la deuda (Kd) es distinta que el coste de la misma (r). Los intereses pagados en un periodo t son: It = Nt1 rt . El aumento de deuda en un periodo t es Nt = Nt Nt1 . Por consiguiente, el ujo para la deuda en un periodo t es: CFd = It Nt = Nt1 rt (Nt Nt1 ) . Por consiguiente, el valor de la deuda en t = 0 es:
D0 =
t=1
Nt1 rt (Nt Nt1 )

t
(1 + Kdt )
1
Es f acil demostrar que la relaci on entre el valor de la deuda (D) y su valor nominal (N ) es: Dt Dt1 = Nt Nt1 + Dt1 Kdt Nt1 rt . Por consiguiente, Dt = Nt + Dt1 Kdt Nt1 rt . El hecho de que el valor de la deuda (D) no coincida con su valor nominal (N ) afecta a varias f ormulas del apartado 1 de este trabajo. Las f ormulas (1), (3), (4), (6), (7), (9) y (10) siguen siendo v alidas, pero el resto de las f ormulas cambia. La expresi on del WACC en este caso es: WACC = E Ke + D Kd N r T . E+D (2 )
La expresi on que relaciona el CFac con el FCF es: CFact = FCFt + (Nt Nt1 ) Nt1 rt (1 T) . La expresi on que relaciona el CCF con el CFac y con el FCF es: CCFt = CFact + CFdt = CFact (Nt Nt1 ) + Nt1 rt = FCFt + Nt1 rt T . (8 ) Distintas expresiones del valor actual del ahorro de impuestos debido al pago de intereses (DVTS) Las expresiones del valor creado por el endeudamiento, esto es, del valor actual del ahorro de impuestos debido al pago de intereses (DVTS) cuando el valor de mercado de la deuda (D) no coincide con su valor nominal (N ) son: a) Modigliani y Miller (1963): DVTS = VA [Kut ; Dt1 Kut T (Nt1 rt Dt1 Kdt ) T] (5 )
n de empresas por descuento de flujos Valoracio b) Myers (1974): DVTS = VA [Kdt ; Nt1 rt T] . c) Miller (1977): DVTS = 0. d) Miles y Ezzell (1980): DVTS = VA [Kut ; Nt1 rt T] e) Harris y Pringle (1985) y Ruback (1995): DVTS = VA [Kut ; Nt1 rt T] . f) Damodaran (1994): 1 + Ku . 1 + Kd
127
DVTS = VA [Kut ; Nt1 rt T + Dt1 T (Kut RF t ) Dt1 (Kdt RF t )] . g) M etodo de los pr acticos: DVTS = VA [Kut ; Nt1 rt T Dt1 (Kdt RF t )] .
Referencias
[1] Arditti, F. D. y H. Levy (1977), The Weighted Average Cost of Capital as a Cuto Rate: A Critical Examination of the Classical Textbook Weighted Average, Financial Management (Fall), pp. 2434. [2] Chambers, D.R., R.S. Harris, y J.J. Pringle (1982), Treatment of Financing Mix Analyzing Investment Opportunities, Financial Management (Summer), pp. 2441. [3] Damodaran, A (1994), Damodaran on Valuation, John Wiley and Sons, New York. [4] DeAngelo, L. y R. Masulis (1980), Optimal Capital Structure under Corporate and Personal Taxation, Journal of Financial Economics 8, Marzo, pp. 329. [5] Fern andez, Pablo (1999), Valoraci on de Empresas, Ediciones Gesti on 2000. [6] Hamada, R.S. (1972), The Eect of the Firms Capital Structure on the Systematic Risk of Common Stock, Journal of Finance 27 (May), pp. 435452 [7] Harris, R.S. y J.J. Pringle (1985), Risk-Adjusted Discount Rates Extensions form the Average-Risk Case , J. of Financial Research (Fall), pp. 237244.
128
ndez Pablo Ferna
[8] Inselbag, I. y H. Kaufold (1997), Two DCF Approaches for Valuing Companies under Alternative Financing Strategies (and How to Choose Between Them), Journal of Applied Corporate Finance (Spring), pp. 114122. [9] Lewellen, W.G. y D.R. Emery (1986), Corporate Debt Management and the Value of the Firm, J. of Financial Quantitative Analysis (December), pp. 415 426. [10] Miles, J.A. y J.R. Ezzell, (1980) The Weighted Average Cost of Capital, Perfect Capital Markets and Project Life: A Clarication, Journal of Financial and Quantitative Analysis (September), pp. 719730. [11] Miles, J.A. y J.R. Ezzell, (1985) Reformulating Tax Shield Valuation: A Note, Journal of Finance 40, 5 (December), pp. 14851492. [12] Miller, M.H. (1977), Debt and Taxes, Journal of Finance (May), pp. 261276 [13] Miller, M., and F. Modigliani, (1961), Dividend Policy, Growth and the Valuation of Shares, Journal of Business 34, pp. 411433. [14] Miller, M., y M. Scholes, (1978), Dividend and Taxes, Journal of Financial Economics (Dec.), pp. 333364. [15] Modigliani, F., and M. Miller, (1958), The Cost of Capital, Corporation Finance and the Theory of Investment, American Economic Review 48, 261297. [16] Modigliani, F y M.H. Miller (1963), Corporate Income Taxes and the Cost of Capital: A Correction, American Economic Review (June), pp. 433443. [17] Myers, S.C. (1974), Interactions of Corporate Financing and Investment Decisions - Implications for Capital Budgeting, J. of Finance (March), pp. 125 [18] Ruback, Richard S. (1986), Calculating the Market Value of Risk-Free Cash Flows, Journal of Financial Economics (March), pp. 323339. [19] Ruback, Richard S. (1989), RJR Nabisco. Teaching note, Harvard Business School, 5-289-057. [20] Ruback, Richard S. (1995), A Note on Capital Cash Flow Valuation, Harvard Business School, 9-295-069. [21] Taggart, R.A. Jr. (1991) Consistent Valuation and Cost of Capital. Expressions With Corporate and Personal Taxes Financial Management (Autumn), 820. Pablo Fern andez IESE Camino del Cerro del Aguila, 3 28023-Madrid, Espa na e-mail: fernandezpa@iese.es
Redes neuronales para aplicaciones en negocios

An bal R. Figueiras Vidal1
Introducci on
En el a mbito de los negocios, como en todos los de aplicaci on de m etodos cient cos o t ecnicos, los problemas a resolver tienen una de las dos formas fundamentales (ocasionalmente se combinan estas): elegir entre un conjunto numerable de alternativas problemas de decisi on o de clasicaci on, o bien evaluar una cantidad no directamente accesible problemas de estimaci on: si el valor que se estima pertenece al futuro, se utiliza el t ermino predicci on. Los ejemplos de tales tipos de problemas son numeros simos: la concesi on o denegaci on de un cr edito es un ejemplo de los primeros (si bien en muchos casos conviene tambi en calicar la situaci on: t picamente estimando la probabilidad de mora), la predicci on de un valor futuro de una serie temporal de consumo de un producto es otro. No puede pretenderse aqu no ya exhaustividad, sino ni siquiera una visi on sistem atica: detecci on de fraude, conguraci on de cestas de compra, valoraci on de riesgo de una operaci on, optimizaci on de una pol tica de compras, etc., son ejemplos adicionales que permiten entrever las amplias dimensiones en que nos movemos; s e que el lector lo har a con m as facilidad que yo. Solo por mor de completitud har e notar que los casos de agrupamiento o segmentaci on, tan habituales en marketing, pueden considerarse como clasicadores que proceden de acuerdo con un criterio preestablecido. A diferencia de lo que ocurre en otras aplicaciones, en los casos de negocio no existe un modelo f sico que relacione lo que se puede observar con las alternativas a elegir o la variable a estimar: si, por citar un caso, en radar (tras un adecuado tratamiento de la se nal recibida) se sabe que la presencia o ausencia de un blanco en un lugar implica disponer de una observaci on de un valor E (seg un el caso) m as un ruido gaussiano, en concesi on de cr editos no hay tal modelo para relacionar la morosidad con las variables sociodemogr acas o econ omicas sobre las que se pregunta al cliente. De modo que, en negocios, nos encontramos con situaciones de dise no muestrales: se tienen ejemplos etiquetados (de resultado conocido por experiencia anterior), y a partir de ellos hay que construir un decisor o un estimador.
1 An bal R. Figueiras Vidal es catedr atico de la Universidad Carlos III de Madrid y Acad emico de N umero de la Academia de Ingenier a. Esta charla se imparti o en la sesi on del Seminario MEFFUAM de junio de 1998.
130
An bal R. Figueiras Vidal
Tradicionalmente se han aplicado, y se aplican, m etodos estad sticos para la resoluci on de estos problemas: desde las decisiones logit hasta la regresi on polin omica son bien conocidos, y su detalle y potencial tambi en, as como una innidad de recomendaciones pr acticas sobre su uso. Los m etodos que aqu se examinan, las Redes Neuronales, no son, sustancialmente, algo nuevo respecto a ellos: son simplemente el resultado de considerar o de visualizar de forma distinta posibles arquitecturas con par ametros libres (como los coecientes de una recta de regresi on) que, mediante un adecuado m etodo de entrenamiento (como la minimizaci on del error cuadr atico entre valores conocidos y los correspondientes a las ordenadas de la recta), establecen una m aquina que sirve para hab erselas con el problema: a la presentaci on de un nuevo conjunto de observaciones (ahora no etiquetadas, naturalmente) responden con la elecci on de una hip otesis (la atribuci on a una clase) o con un valor estimado (o predicho). Pero esa nueva forma de considerar y visualizar las arquitecturas tiene algunas implicaciones destacables.
Las Redes Neuronales
Formalmente, una Red Neuronal no es cosa distinta que un sistema paralelo no lineal con componentes repetidos en ramas, habitualmente dispuesto en capas. La gura 1 presenta el m as tradicional, el Perceptr on Multicapa, designado frecuentemente por sus siglas en ingl es (MLP, Multi-Layer Perceptron).
o
(capa oculta)
Figura 1: MLP Las observaciones x colocadas en su entrada progresan a trav es de ramas, en las que las componentes del vector se multiplican por valores de los par ametros de la red, y la suma de los resultados se somete a una operaci on no lineal en cada nodo: normalmente consiste en la toma de la tangente hiperb olica. Tal proceso se repite en una siguiente capa, hasta alcanzar una salida (o varias) que proporciona un indicador para la decisi on o un(os) valor(es) estimado(s).
131
Esta arquitectura, sobre cuya muy larga historia no voy a entretenerme, puede parecer arbitraria. No lo es: de un lado, constituye una extensi on por combinaci on progresiva de una arquitectura tipo logit; de otro, su forma remeda (no equivale a!) la de una porci on del sistema nervioso de un animal dotado de este; y, por u ltimo, su expresi on matem atica, ya cuando s olo hay una capa de nodos intermedios (ocultos), constituye un aproximador universal para cualquier correspondencia entre entrada y salida. Y lo tercero es lo que voy a enfatizar aqu : aproximador universal es, por ejemplo, un polinomio de orden suciente en las componentes de x: pero todo el mundo sabe de las dicultades originadas por la explosi on dimensional del n umero de componentes (n umero de par ametros) en los modelos polin omicos cuando se pretenden establecer correspondencias complicadas. Tal fen omeno no se interpone aqu en nuestro deseo de dise nar un buen decisor o un buen estimador; o, mejor dicho, no se interpone de forma tan violenta: s olo hace falta acertar en la elecci on del n umero de nodos ocultos para conseguir la potencia expresiva adecuada para el problema, y eso, sin resultar trivial, no es tan dif cil de determinar. Por otra parte, la forma de entrenar esta m aquina se simplica precisamente por el aspecto paralelo, repetitivo y en capas que se le da: es f acil comprender que se trata de una relaci on entradasalida de la forma o=f
j
wj fj
(2) i (1) wi xi ,
y que la derivaci on de tal expresi on mediante la Regla de la Cadena conduce a un algoritmo de b usqueda por pendiente de complejidad moderada: no creo que este sea el mejor lugar para exponer detalles, con lo que citar e su nombre, Algoritmo de Retropropagaci on, y citar e un texto [1] que, si bien lo presenta de la misma forma que otros muchos, lo discute (junto con otros temas relevantes para esta clase de m aquinas) de manera magistral. Arquitecturas como la del MLP brindan, adem as, una importante oportunidad: como quiera que lo verdaderamente deseable de cualquier decisor o estimador es una buena generalizaci on (que responda convenientemente ante ejemplos no utilizados en el entrenamiento), su comparativa compacidad representa una ventaja: es bien sabido que la caracter stica de generalizaci on depende del objetivo (anal tico) del entrenamiento, de la forma de aplicar el algoritmo para optimizar el objetivo, de la representatividad de las observaciones dedicadas a entrenar, de la relevancia y grado de independencia de las variables observadas, y de la parsimonia de la arquitectura seleccionada: respecto a este u ltimo factor, es claro que encontrarse problemas de dimensionado no extraordinariamente complejos facilita ganar generalizaci on a trav es de el. En cuanto a los dem as, se consideran y solventan de formas no muy distintas a las cl asicas. No pretendo ocultar que estas m aquinas tienen sus dicultades y limitaciones: aparte del dimensionado, su entrenamiento es lento y dif cil, ya que la dependencia
132
de la funci on objetivo de los par ametros de la m aquina da lugar a (h per)supercies de b usqueda muy irregulares. Pero, sobre todo, la actuaci on de las Redes Neuronales es muy dif cil de ser interpretada en los t erminos en que se suele explicar la actuaci on de otros clasicadores o estimadores: esta variable, de gran inuencia por venir afectada de un factor de magnitud elevada, puede arg uirse en el caso de una m aquina lineal (en t erminos simplicados, aclaro); pues bien, nada parecido puede decirse en el caso de tomar las tangentes hiperb olicas de diversas combinaciones lineales y combinar linealmente de nuevo los resultados. Esta limitaci on hace poco apto el empleo de Redes Neuronales en situaciones que puedan aparejar actuaciones forenses. No obstante lo anterior, debo advertir que se trabaja muy decididamente en el desarrollo de t ecnicas de interpretaci on (reducci on a reglas y referencia a prototipos, entre otros; aunque debo confesar que mis esperanzas personales se cifran m as en el desarrollo de aproximaciones local-global: en ellas se regionaliza el problema, y se dise nan soluciones de estructura global, como las polin omicas, para cada regi on: con la fundada esperanza de que el impl cito compromiso tama no de las regiones-sencillez de los problemas regionales permita explicaciones directamente inteligibles. Mis esperanzas se cifran, como digo, en esta opci on porque. . . yo personalmente dedico esfuerzo investigador en este tipo de dise nos: bien s e que quienes lo hacemos a un no hemos resuelto la dicultad de optimizar la regionalizaci on en funci on del objetivo general del problema observado, pero. . . falta poco!). Y, ya que hablamos de aproximaciones locales, no podemos olvidar la otra familia m as popular de Redes Neuronales: las Redes de Funciones Radiales de Base (RBFN, Radial Basis Functions Networks), cuyo esquema se presenta en la Figura 2.
o
(capa oculta)
Figura 2: RBF En este caso, los vectores de entrada se comparan directamente con una serie de vectores de referencia, ci , asociados a los nodos de la red: el resultado de la similitud (normalmente el cuadrado de la distancia eucl dea) se utiliza como argumento de una funci on decreciente (muchas veces una exponencial negativa), y a las salidas de los
133
nodos se les aplica un decisor o un estimador lineal convencional. Estas redes, tambi en, como el MLP, de extraordinaria potencia expresiva, tienen la gran ventaja de que, adem as de permitir entrenamiento por gradiente (u otros m etodos de b usqueda local), pueden entrenarse mediante agrupamiento (clustering) m as parametrizaci on convencional de los pesos de salida: lo que da lugar a que se puedan aprovechar, en el primero de los pasos, conocimiento previo sobre el problema a resolver (adem as, se evitan numerosas ca das en o ptimos locales). Por otro lado, a partir de la s olida teor a de la Minimizaci on del Riesgo Emp rico del matem atico ruso Vladimir Vapnik (hoy en los Laboratorios Bell norteamericanos), puede formularse el dise no de RBFN de modo tal que el dimensionado (n umero de nodos) y los pesos de salida se obtengan autom aticamente por Programaci on Cuadr atica: el resultado, conocido como M aquinas de Vectores de Soporte (SVM: Support Vector Machines) [2], constituye probablemente la opci on m as potente disponible por el momento para solucionar problemas de clasicaci on; las dicultades de carga computacional elevada y falta de adaptabilidad a posibles variaciones del entorno (de las condiciones del problema) se van solucionando con peque nos trucos algor tmicos: nuestro grupo de investigaci on de la Universidad Carlos III de Madrid ha construido algoritmos reconocidamente ecaces.
Una discusi on comparativa
Presentadas ya las Redes Neuronales y sus caracter sticas principales, procede una discusi on de sus ventajas e inconvenientes respecto a las alternativas existentes para el tratamiento de los problemas de decisi on y estimaci on. Empezar e diciendo que renuncio a proceder estrictamente de modo dicho en el caso de los m etodos estad sticos cl asicos: a mi juicio, y como ya he anticipado, no hay diferencias metodol ogicas signicativas, sino u nicamente la proposici on de nuevas arquitecturas, potentes y de manejo (relativamente) c omodo, como Redes Neuronales. S hay diferencias signicativas en los Sistemas Expertos, que se basan en la aplicaci on de reglas (del tipo si A y B, entonces C) en la soluci on de estos problemas; extrayendo las reglas del conocimiento de expertos o de un apropiado manejo de los propios datos. No voy a exponer mi propia opini on acerca de la capacidad de las reglas para representar el conocimiento experto, que es materia muy conceptual para este momento: s debo decir que los Sistemas Expertos tienen obvias ventajas forenses, a cambio de p esimas caracter sticas de generalizaci on (derivadas de la resoluci on de problemas de clasicaci on mediante el empleo de regiones t picamente hiperc ubicas: cuyo aspecto de puercoespines cuando se trabaja en muchas dimensiones propicia el que se separen muestras muy semejantes y se agrupen otras muy dispares) y de una casi absoluta imposibilidad de actuar adaptativamente. Este u ltimo defecto tambi en lastra los conocidos como Arboles de Decisi on (o de Estimaci on), m etodos en los que se van separando grupos de muestras mediante
134
la aplicaci on consecutiva de diversos tests, hasta llegar a grupos compuestos por muestras de una sola clase: v ease la Figura 3.
o1
nodo 1
o2 nodo 2
(test 1)
terminal 1 (clase A) terminal 2 (clase B)
(test 2)
terminal 3 (clase A)
Figura 3: Arboles La elecci on de los tests es dicultosa; las propiedades de generalizaci on, medianas (no s olo dependientes de los tests aplicados, como dicen algunos autores: el mero hecho de buscar la detenci on por separaci on no contribuye a su bondad). Sin embargo, como los Sistemas Expertos, son f acilmente interpretables (si lo son los test empleados). Por todo lo anterior, son recomendables para problemas estad sticos de complejidad intermedia: sobre todo, ante necesidades forenses. Los M etodos Basados en Memoria consisten en comparar los valores de las variables del caso bajo an alisis con los de ejemplos de soluci on conocida, y decidir de acuerdo con el caso m as semejante, o en virtud de lo sugerido por las soluciones del grupo de casos m as similares. Como se aplica a problemas muy variados, resultan cr ticas para obtener buenas prestaciones la selecci on de las variables y la medida de la similitud que se aplique: ambas cosas dif ciles de determinar en funci on de la b usqueda de las mejores prestaciones. Otra limitaci on intr nseca radica en que se trabaja de acuerdo con la losof a de los bien conocidos procedimientos de k vecinos m as pr oximos (k -NN, k -Nearest Neighbours), robustos pero no excesivamente potentes; pero cabe advertir que tal forma de actuar puede ampliarse sin dicultad seria aparente: por ello, opino que estos m etodos tienen un muy prometedor futuro; a un admitiendo que las versiones adaptativas requieren mucho esfuerzo de investigaci on adicional. Otras t ecnicas emergentes, como la L ogica Difusa y los Algoritmos Gen eticoEvolutivos, no son alternativas directas a las vistas: la primera es complementaria
135
para cualquiera de ellos, ya que proporciona una forma diferente de la habitual de representar y tratar variables, y los segundos constituyen un m etodo de b usqueda global, incorporable al dise no de las m aquinas que hemos revisado (eso s , como en su aplicaci on directa, con problemas en la generalizaci on alcanzable). Resumiendo: las Redes Neuronales son una muy recomendable alternativa para problemas de elevada complejidad y moderados requisitos de adaptatividad e interpretatividad: aunque cada d a mejoran sus caracter sticas frente a estas demandas.
Aplicaciones en negocios
Dicho lo anterior, no es extra no que las Redes Neuronales hayan sido ampliamente aplicadas en problemas de decisi on y de estimaci on surgidos en el mundo de los negocios. Ser a inadecuado seleccionar y discutir aqu casos particulares: cualquier lector puede acceder a fuentes m as extensas como [3] y [4], y apreciar all aplicaciones y resultados. Peros s conviene resaltar que, en general, la aplicaci on de las Redes Neuronales conduce a realizaciones con mejores prestaciones que sus alternativas: siempre que, desde luego, su dise no y aplicaci on se lleve a cabo por buenos conocedores de este a mbito (ya que, como se ha advertido, est a llenas de trampas para el usuario accidental), y, como en el caso de cualquier m etodo, se incorpore debidamente el conocimiento previo general y espec co que se tenga del problema bajo estudio. As lo hemos podido vericar los componentes del equipo de investigaci on que, en el departamento de Tecnolog as de las Comunicaciones de la Universidad Carlos III, hemos dise nado Redes Neuronales para su aplicaci on en problemas de concesi on de cr editos, segmentaci on de mercados, detecci on de fraude, y otros an alogos (adem as de aplicaciones en a mbitos de tipo puramente tecnol ogico, como predistorsionadores e igualadores en telecomunicaciones). Al mismo tiempo, este equipo de investigaci on ha abordado con estas herramientas problemas de car acter diferente del de los anteriores, que habitualmente se re unen bajo la denominaci on gen erica de Miner a de Datos: trabajando sobre muestras que no son las t picas variables sociales, econ omicas, f sicas, etc., sino representativas de otras informaciones, como las contenidas en un texto (aparici on de ciertos t erminos, con sus frecuencias y posiciones relativas, etc), y cuya aplicaci on se extiende desde la creciente disciplina de la Gesti on del Conocimiento hasta su empleo para la personalizaci on de los servicios y aplicaciones ofrecidos a trav es de las redes de telecomunicaciones: por no citar el important simo caso de la Inteligencia de Negocio, en que datos de mercado, noticias, anuncios, etc., proveen el material crudo utilizable para el estudio de la competencia y la toma de decisiones estrat egicas para la empresa (o tambi en de instituciones u organizaciones de orientaci on no empresarial: es desolador constatar que las administraciones espa nolas se encuentran en estado de casi absoluto desconocimiento de estas posibilidades y t ecnicas; claro est a que no hay muchas diferencias con otros a mbitos. . . )
136
Tendencias
No ser a justo negar al lector informaci on sobre las tendencias de mayor relevancia en el campo de las Redes Neuronales: a un forzado a la s ntesis, lo har e gustoso. A mi juicio, y adem as de lo relevante de las variaciones y extensiones de las ya citadas SVM, proporcionar an resultados de importancia: Los trabajos de desarrollo de metodolog as y algoritmos para construir asociaciones de redes neuronales, en que se busca aprovechar ventajas de diferentes tipos de arquitecturas, fraccionar la resoluci on de un problema en tareas parciales m as asequibles, combinar lo obtenido de acuerdo con criterios diferentes, etc. V ease [5]. En particular, es muy alta mi conanza en los procedimientos en los que se realiza un agrupamiento (clustering) y una posterior aplicaci on de una m aquina a los grupos (sea acretiva, sea interpolativamente). La dicultad de dise nar el agrupamiento inicial para optimizar las prestaciones del sistema global, o de agrupar en operaci on tras un dise no en el que el agrupamiento se realiza impl citamente optimizando la asignaci on de muestras a las m aquinas del nivel de salida, se puede atacar con ecaces heur sticos (sub optimos), de los que nuestro equipo de investigaci on analiza varios. Personalmente considero que estas aproximaciones de tipo local m as global brindan una magn ca v a para resolver el dilema entre aproximaci on mediante funciones globales o mediante funciones locales de siempre abierto en las matem aticas. La aplicaci on de conceptos evolutivos, o, para ser m as exactos, probabil sticos, en el dise no de (complejas) m aquinas de decisi on o de estimaci on. Como quiera que muchas arquitecturas admiten de forma natural el empleo de algoritmos de crecimiento por poner un ejemplo sencillo: la construcci on de un modelo polin omico de alto orden mediante la repetida aplicaci on de modelos polin omicos de dos variables y orden dos, recurrir para crecer a una selecci on determinista de opciones en funci on de las prestaciones inmediatamente obtenidas representa un serio riesgo de obtener prestaciones nales limitadas por malas elecciones intermedias. Si se introduce la aleatoriedad en el crecimiento (con proporcionalidad de las oportunidades de aparici on de cada opci on y sus prestaciones inmediatas como en el caso de los Algoritmos Gen eticos), el manejo de poblaciones de soluciones as construidas brinda opci on de salir de tales trampas en la construcci on. Posibilidades y m etodos en esta l nea necesitan a un, no hay que decirlo, de un proceso de maduraci on. Debo se nalar tambi en que caben combinaciones de las l neas que he citado, cuyo detalle reviste una complejidad que no propicia su exposici on en estas p aginas.
137
Consideraciones nales
Lo que antecede constituye, en apretad sima s ntesis, el resultado de mis experiencias de ya bastantes a nos en el empleo de Redes Neuronales en la resoluci on de problemas de negocio (y de otros: no debo ocultar que mantengo la opini on de que el trabajo en diversos campos de aplicaci on otorga la inapreciable ventaja de aprovechar el transplante de conceptos, m etodos y procedimientos), junto con unas referencias a fuentes m as extensas de casos concretos cuya presencia aqu ocluir a el espacio preciso para cuestiones generales que creo mucho m as relevantes. Si bien s e que su lectura puede despertar la curiosidad, tambi en conozco que puede provocar escepticismo, e incluso rechazo. Tales sentimientos se combaten con la vericaci on de las reales posibilidades de las Redes Neuronales: salvo que nazcan del temor de que pongan en compromiso la importancia de nuestro propio saber o nuestro propio trabajo. Si bien est a en la naturaleza humana la aparici on de este temor, me atrevo a alegar que deriva de una inseguridad poco justicada: es maniesto que el conocer y disponer de un mejor arsenal de herramientas no da na la posici on de quien emplea dichas herramientas, ni siquiera de quien est a al cargo de gestionar el resultado del trabajo de quienes las emplean. Lo que s produce peligroso y vano riesgo es renunciar a u tiles que puedan resultar denitivos para la mejora del negocio. No hay que temer a las Redes Neuronales: como todas sus alternativas, cl asicas o emergentes, son una ayuda para las tareas de los humanos, y no un sustituto. Y ya considerada la primera dicultad que realmente encuentra la aplicaci on de las Redes Neuronales, se me permitir a recordar una obviedad: de donde no hay, no se puede sacar. Lo que, traducido a lo que ahora nos ocupa, dice que sin datos, no hay ayuda posible. Muchas organizaciones no perciben que hoy incluso se puede hablar de sobreabundancia de datos e informaciones: de ah viene realmente la expresi on Sociedad de la Informaci on. Muchas organizaciones no reparan en que desperdiciar la informaci on supone desventaja: se quedan en el estadio anterior, sin aspirar a integrarse en la Sociedad del Conocimiento. Conocer signica interpretar y asimilar informaci on. Esa informaci on ha de ser buscada, seleccionada, mantenida, almacenada,. . . Despu es, analizada y explotada: pero si no se toman las providencias posibles para lo primero, no hay opci on para lo segundo. La informaci on es materia prima: el conocimiento, producto. Agradecimientos En la concepci on y orientaci on de estas p aginas han tenido decisiva inuencia muchos colegas. Me permito citar, entre ellos, a quienes han discutido conmigo a diario sobre este tema: los profesores Antonio Art es, Angel Navia y Jes us Cid, de mi Universidad.
138
Referencias
[1] C. M. Bishop: Neural Networks for Pattern Recognition. Oxford: Oxford University Press, 1995. [2] B. Sch olkopf, K-K. Sung, C.J.C. Burges, F. Girosi, P. Niyogi, T. Poggio, V. Vapnik: Comparing Support Vector Machines with Gaussian Kernels to Radial Basis Function Classiers, IEEE Trans. on Signal Proc. 45 (1997), pp. 2758-2765. [3] A. S. Weigend, Y. Abu-Mostafa, A.-Paul N. Refenes (eds): Decision Technologies for Financial Engineering. Singapore: World Scientic, 1997. [4] A.-P. N. Refenes, A. N. Burges, J. E. Moody (eds): Decision Technologies for Computational Finance. Dordrecht, The Netherlands: Kluwer, 1998. [5] A. J. C. Sharkey (ed): Combining Articial Neural Nets. Ensemble and Modular Multi-Net Systems. London: Springer-Verlag, 1999.
An bal R. Figuieras Vidal Departamento de Tecnolog as de las Comunicaciones Universidad Carlos III de Madrid Avenida de la Universidad, 30 20912 Legan es (Madrid) e-mail: arfv@tsc.uc3m.es
Volatilidad estoc astica, volatilidad incierta

nchez Calle1 Antonio Sa
Resumen En el modelo de Black-Scholes para la evoluci on del precio de activos nancieros se hacen varias hip otesis que son una primera aproximaci on a la realidad y cuyo objetivo principal es hacer el modelo simple y calculable. Que esas suposici on no son compatibles con los datos de mercado se hace aparente en la sonrisa de volatilidades, volatility clusters, fat tails, etc. En un intento de solventar estas dicultades se han propuesto ciertos cambios al modelo de B-S; del que nos vamos a ocupar hoy es de modicar la hip otesis = constante.
1
1.1
Black-Scholes
Hip otesis del modelo
no hay costes de transacci on, no hay restricci on en ventas al descubierto; los activos son innitamente divisibles; los activos no pagan dividendos; el tipo para prestar y tomar prestado es el mismo; las transacciones tienen lugar en tiempo continuo; ...
Comenzamos con un repaso r apido a Black-Scholes.
El mercado est a exento de fricci on y es continuo:
El mercado es eciente, es decir, no hay oportunidades de arbitraje. Hay un activo sin riesgo, la cuenta corriente Bt , cuya evoluci on es dBt = rBt dt , y, por tanto, B0 = 1
Bt = ert .
1 Antonio S anchez Calle es Catedr atico de An alisis Matem atico de la Universidad Aut onoma de Madrid. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de septiembre de 1998.
140
nchez Calle Antonio Sa El activo con riesgo, St , sigue una evoluci on del tipo dSt = dt + dWt , St donde y son constantes ( = 0) y Wt es un movimiento browniano. Como 2 St sigue una distribuci on lognormal de par ametros ( consecuencia, )t y S0 2 t. Es decir, Xt = log
def
St 2 )t, t Normal ( S0 2
1.2
Consecuencias
La hip otesis de ausencia de oportunidades de arbitraje permite valorar derivados usando carteras autonanciadas: una cartera est a dada por dos procesos, t y t , que representan el n umero de activos y bonos, respectivamente. El valor de esa cartera en el instante t es Vt = t St + t Bt . Una cartera autonanciada se crea haciendo un desembolso inicial y reajustando la composici on de la cartera a partir de ese momento sin hacer ning un desembolso adicional (ni positivo ni negativo). Eso se traduce en que su cambio de valor procede u nicamente del cambio en la cotizaci on del activo o en el valor del bono; es decir, dVt = t dSt + t dBt . Una cartera autonanciada est a determinada por su valor inicial y por una cualquiera de sus partes, t o t . Por ejemplo, dado un t , la cartera autonanciada correspondiente es la soluci on a dVt = r(Vt t St ) dt + t dSt . La ausencia de oportunidades de arbitraje tiene varias consecuencias inmediatas: dos carteras autonanciadas que tengan el mismo valor en alg un instante han de tener el mismo valor en todo momento. Si un producto derivado tiene payo X a vencimiento T y hay una cartera autonanciada que cumple VT = X , entonces el valor de ese producto derivado en cualquier instante anterior t < T deber a ser Vt . Esa u ltima observaci on proporciona un m etodo para valorar derivados: encontrar una cartera autonanciada que reproduzca el payo en el instante T y tomar como valor del derivado en el instante t el que tenga la cartera en ese mismo instante. n: lo anterior nos dice en principio c Observacio omo valorar y adem as c omo cubrir el riesgo en cada momento formando la susodicha cartera. Pero, c omo se encuentra esa cartera?; o, de otra manera, cu al es el t ? En un momento lo recordaremos.
stica, volatilidad incierta Volatilidad estoca
141
1.3
Propiedades del modelo de Black-Scholes
Entre las propiedades de este modelo se pueden destacar: este modelo es completo en el sentido siguiente: es posible reproducir un payo arbitrario mediante una cartera autonanciada. El valor en el instante t de un derivado con payo del tipo f (ST ) depende u nicamente de la cotizaci on St del activo en ese instante. Adem as, si V (St , t) es ese valor, entonces la funci on V (S, t) verica la ecuaci on de Black-Scholes: 1 2 V (S, t) + 2 S 2 V (S, t) r V (S, t) = 0 , V (S, t) + r S 2 t 2 S S y se tiene t = V (St , t) (= la delta) S
Una forma alternativa de calcular la cartera autonanciada es utilizar lo que se conoce como probabilidad riesgo-neutro o probabilidad-martingala equivalente, que se obtiene reasignando probabilidades a las posibles evoluciones del subyacente de forma que se reeje la ausencia de oportunidades de arbitraje. Desde el punto de vista de la evoluci on eso signica que dSt = r dt + dWt , St donde Wt es un browniano para esa nueva probabilidad Q. En t erminos de esta Q, el valor de la cartera autonanciada con payo X en T es X Vt = Bt EQ Ft = er(T t) EQ [X Ft ] ; BT esto es, el valor esperado del payo conocida la informaci on del mercado hasta el instante t, descontado a fecha t.
1.4
Volatilidad impl cita
La famosa f ormula de B-S da el valor de una call de vencimiento T (jo en lo que sigue) y precio de ejercicio K : Cbs (K, ) = S (d+ ) K erT (d ) , con 1 1 S d = log T, rT Ke 2 T
142
nchez Calle Antonio Sa
y donde S es el valor del activo subyacente en t = 0 y (x) es la funci on de distribuci on de la Normal(0,1), x 2 dt et /2 . (x) = 2 Es f acil ver que la funci on Cbs (K, ) es inversible, y su inversa da lugar a lo que se conoce como volatilidad impl cita: si c(K ) es el precio de mercado de una call con precio de ejercicio K, entonces la volatilidad impl cita en ese precio de mercado es la k tal que Cbs (K, k ) = c(K ) . Si el modelo de B-S fuese correcto, haciendo esto para distintos strikes se deber a obtener, esencialmente, un valor independiente de K , pero eso no es as . Nota: realmente tenemos una supercie de volatilidades (K, T ).
Variando la
Comenzamos con un repaso somero de los modelos de volatilidad estoc astica: suponemos que St sigue una evoluci on dSt = t dt + t dWt , St con t (y t ) un proceso aleatorio adaptado a la informaci on disponible en el mercado hasta ese momento, Ft . Nota: Normalmente se toma como Ft la informaci on contenida en el browniano hasta ese instante, o lo que es lo mismo, la dada por la evoluci on del activo hasta ese momento. En este caso el modelo es completo aunque demasiado general si no se imponen condiciones adicionales sobre el proceso t .
143
Un aparte sobre incompletitud En el caso en que la informaci on en Ft sea m as amplia que la contenida en el browniano, el modelo es, en general, incompleto y eso signica que habr a payos que no se pueden reproducir por una cartera autonanciada. Eso conlleva una complicaci on a la hora de valorar ciertos productos. Una posibilidad sensata a la hora de valorar un producto derivado, desde el punto de vista del vendedor, consiste en tomar su payo X (vencimiento T ) y considerar todas las carteras autonanciadas que cumplen VT X . Se elige como valor el nmo de todos los V0 s. Este procedimiento hace pertinentes ciertas preguntas: existe alguna cartera autonanciada que domine el payo ? (es decir VT X ). Si la respuesta es S I, corresponde el nmo de los V0 s a una cartera autonanciada? Si existe esa cartera optima en el instante t = 0, sigue siendo o ptima en instantes posteriores? Y, si es as , la estrategia de cobertura de esa cartera es calculable?
Como hemos mencionado, el modelo planteado arriba es demasiado general para obtener resultados concretos, y se suelen hacer ciertas hip otesis adicionales sobre t . Por ejemplo, se supone que t de la forma (St , t), con (S, t) conocida. Volatilidad estoc astica: se supone que t verica una ecuaci on diferencial estoc astica de la forma d t = t dt + t dZt , t on con el que rige la con Zt un movimiento browniano que tiene cierta correlaci evoluci on de St (dWt dZt = dt). Volatilidad incierta: se supone que t es desconocida pero satisface una condici on del tipo 0 < m(St , t) t M (St , t) .
2.1
= (St , t) conocida
dSt = (St , t) dt + (St , t) , dWt ; St
Suponemos que St sigue una evoluci on del tipo
es decir, la volatilidad depende u nicamente de la cotizaci on del subyacente en ese instante. En este caso gran parte de lo que es cierto en el modelo de B-S sigue cumpli endose; por ejemplo,
144
nchez Calle Antonio Sa El modelo sigue siendo completo, aunque ahora St ya no es lognormal. El valor de un derivado con payo f (ST ) sigue siendo de la forma V (St , t), donde V (S, t) satisface una EDP similar: 1 2 V (S, t) + 2 (S, t) S 2 V (S, t) r V (S, t) = 0 , V (S, t) + r S t 2 S 2 S que es tratable por los m etodos cl asicos de aproximaci on num erica.
Adem as tiene una ventaja importante: se puede adaptar la (S, t) a los datos de mercado en el instante actual (t = 0), al menos te oricamente. Si C (K, T ) representa el valor de una call con strike K y vencimiento T (que suponemos conocido para todo K, T ) entonces 2 T C (K, T ) (K, T ) = 2 2 . K K C (K, T ) Problema: la volatilidad impl cita es realmente una funci on (t, t+T, K ), donde t = instante actual, y estamos ajustando el modelo para un t = 0. Desafortunadamente, la supercie de volatilidades no es muy estable al variar t y el modelo pronto se desajusta. Adem as, en qu e situaciones es razonable suponer que la volatilidad depende u nicamente del nivel del activo?
2.2
Din amica conjunta para St , t
Tratamos ahora de modelos en los que la din amica que rige la evoluci on de St , t es de la forma dSt = (St , t , t) dt + t dWt , St dt = (St , t , t) dt + (St , t , t) dZt , t on , con || < 1 y la forma donde Wt y Zt son movimientos brownianos de correlaci de los coecientes se obtiene haciendo hip otesis adicionales sobre el modelo. Las dicultades en estos modelos no son s olo computacionales; estos modelos no son completos debido a que la volatilidad no se negocia (aunque el mercado de opciones se puede utilizar como un suced aneo). Hull-White (ver Hull y White (1987)) Es uno de los primeros modelos, y en el se supone b asicamente que la volatilidad sigue un proceso independiente al del activo: dSt = (St , t , t) dt + t dWt , St dt = (t , t) dt + (t , t) dZt , t = 0.
145
Esas hip otesis permiten concluir, entre otras cosas, que conocida la varianza media V =
def
1 T
T 0
2 t dt ,
la distribuci on de ST /S0 es lognormal como en el caso de B-S, tomando como volati lidad V (tambi en analizan el caso = 0 num ericamente). Desafortunadamente, para explicar las sonrisas se necesita hacer otras hip otesis adicionales que ellos mismos consideran poco sensatas. Zhu-Avellaneda (ver Zhu y Avellaneda (1998)) La hip otesis adicional es que las calls a muy corto plazo siguen Black-Scholes (para la volatilidad instant anea). Eso hace que si F (S, , t) denota el precio de un derivado, entonces formando una cartera con 1 unidad del derivado, unidades de activo y unidades de la call a corto plazo, eliminando la parte aleatoria se puede concluir que F verica una EDP especial que se traduce en una evoluci on riesgo-neutro dSt St dt t = = r dt + t dWt , 1 V t dt + V dZt 2
y este es el modelo que se estudia. Una de sus principales virtudes es que predice sonrisas de volatilidad. Un ligero defecto es que, para el caso de correlaci on negativa, la volatilidad explota en tiempo nito, aunque en general el tiempo de explosi on es grande comparado con los vencimientos habituales.
2.3
Volatilidad incierta
En los modelos de volatilidad incierta (ver Avellaneda, Levy y Par as (1995) y Lyons (1995)) la hip otesis b asica es que la volatilidad es desconocida y s olo se conocen ciertas restricciones sobre su comportamiento, en general sobre su tama no. Suponemos que St sigue una evoluci on dSt = t dt + t dWt , St pero sobre t s olo se sabe que, por ejemplo, 0 < min t max , con min y max constantes.
146
n de Derivados Supongamos que tenemos un payo f (ST ) a venciValoracio miento T que queremos valorar en el instante t. Punto de vista del vendedor: quiere poder hacer frente a todas las posibles evoluciones. Cada elecci on de admisible da lugar a una cartera autonanciada, Vt que reproduce el payo en T. Es natural, pues, pedir como prima la cantidad W + (St , t) = Punto de vista del comprador: comprar f (ST ) vender f (ST ). Cada elecci on de admisible da lugar a una cartera autonanciada, Vt que reproduce el payo f (ST ). Entonces Vt reproduce f (ST ). Por tanto, puesto que sup(a) = inf(a), un comprador ofrecer a W (St , t) = por ese producto. Una pregunta inmediata es c omo se calculan W ? La respuesta, cuya justicaci on es un uso elemental de la f ormula de It o, es que ambas satisfacen la EDP
2 1 2 W (S, t) + a2 W (S, t) r W (S, t) W (S, t) + r S S t 2 S 2 S W (S, T ) def min max def min max
sup
Vt .
inf
Vt
= 0 = f (S )
donde para W + hay que usar max min max min si si 2 W + (S, t) 0 , S 2 2 W + (S, t) < 0 , S 2 2 W (S, t) 0 , S 2 2 W (S, t) > 0, . S 2
a+ =
def
y para W a+ =
def
si si
147
Desde un punto de vista intuitivo es una ecuaci on razonable: por ejemplo el vendedor (caso +), al considerar el valor de un derivado tipo call, con payo convexo 2 (es decir S > 0) observa que ese valor aumenta con la volatilidad, luego su peor situaci on es cuando esa volatilidad sea max . Por otro lado, para un derivado con 2 S < 0, por ejemplo (activo call ), el mayor valor corresponde a la volatilidad m as peque na min . Esto es precisamente lo que recoge la a+ ! n: W corresponden a carteras autonanciadas para posibles evolucioObservacio def 2 dadas por t = a (S W ). nes de la volatilidad t Comentarios: La cobertura de una cartera no es lineal en sus componentes. Si f1 (ST ) y f2 (ST ) son payos en T y tomamos f12 (ST ) = f1 (ST ) + f2 (ST ), entonces
+ (S, t) W12 W12 (S, t) def
+ + W1 (S, t) + W2 (S, t) , W1 (S, t) + W2 (S, t) .
Para el caso de derivados con convexidad ja, por ejemplo calls y puts, W + y W corresponden a los valores de Black-Scholes ordinarios tomando como volatilidad max y min , respectivamente. En el caso de convexidad mixta, por ejemplo en call spreads, eso no es as . De hecho la valoraci on usando el supremo (sobre ) de las valoraciones por BlackScholes para cualquier volatilidad constante con min max se queda corta. Ver los gr acos en Avellaneda, Levy y Par as (1995).
+ Si la t que el mercado sigue es distinta de la t , entonces la cartera autonanciada (para esa t ) que en el instante inicial vale W + (S0 , 0) y cuya componente en el activo viene dada por t = S W + (St , t) tendr a un valor nal superior a f (ST ). En ese caso se podr a pensar en alg un sistema de cash-back.
Existe un modelo discreto usando a rboles trinomiales (procedente de m etodos num ericos para la EDP). Una modicaci on posible al modelo anterior consiste en permitir cubrir el riesgo de volatilidad usando el mercado de opciones sobre el subyacente. Pero eso. . . es otra historia !
148
Referencias
[1] Avellaneda, M., Levy, A. y Par as A., Pricing and hedging derivative securities in markets with uncertain volatilities, Applied Math. Finance 2, 73-88 (1995). [2] Avellaneda, M. y Par as, A., Managing the volatility risk of portfolios of derivative securities: the Lagrangian Uncertain Volatility model, Applied Math. Finance 3, 21-52 (1996). [3] Hull, J. y White, A., The pricing of options on assets with stochastic volatilities, J. of Finance 42, 281-300 (1987). [4] Lyons, T., Uncertain volatility and the risk-free synthesis of derivatives, Applied Math. Finance 2, 117-133 (1995). [5] Zhu, Y., Avellaneda, M., A risk-neutral stochastic volatility model, International Journal of Theoretical and Applied Finance 1, 289-314, (1998).
Antonio S anchez Calle Departamento de Matem aticas Universidad Aut onoma de Madrid Campus de Cantoblanco, s/n 28049-Madrid e-mail: antonio.sanchez@uam.es
Short-Term Options with Stochastic Volatility: Estimation and Empirical Performance

n and Gonzalo Rubio1 Gabriele Fiorentini, Angel Leo
Abstract This paper examines the stochastic volatility model suggested by Heston (1993). We employ a time-series approach to estimate the model and we discuss the potential eects of time-varying skewness and kurtosis on the performance of the model. In particular, it is found that the model tends to overprice out-of-themoney calls and underprice in-the-money calls. It is also found that the daily volatility risk premium presents a quite volatile behavior over time; however, our evidence suggests that the volatility risk premium has a negligible impact on the pricing performance of Hestons model2 .
Introduction
Given the Black-Scholes (1973) (BS henceforth) assumptions, all option prices on the same underlying security with the same expiration date but with dierent exercise prices should have the same implied volatility. However, the well known smiles and smirks suggest that the BS formula exhibits systematic biases in pricing options. There have been various attempts to deal with this apparent failure of the BS valuation model. In principle, as explained by Das and Sundaram (1999), the existence of the smile eect may be attributed to the presence of excess kurtosis in the conditional return distributions of the underlying assets. It is clear that excess kurtosis makes
1 Gabriele Fiorentini es Profesor Titular del Departamento de Fundamentos de An alisis Econ omico en la Universidad de Alicante. Angel Le on es Profesor Titular del Departamento de Econom a Financiera de la Universidad de Alicante. Gonzalo Rubio es Catedr atico de Fundamentos del An alisis Econ omico de la Universidad del Pa s Vasco. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de octubre de 1998. 2 We thank Stewart Mayhew, Eduardo Schwartz, Enrique Sentana, Ignacio Pe na, Gregorio Serna, Rafael Salinas, Jos e Luis Fern andez, Eliseo Navarro, Alejandro Balb as, Rafael Santamar a and two anonymous referees for their helpful comments and useful discussions. Earlier versions of this paper were presented at the European Finance Association in Helsinki (1999), EFM Association Conference in Paris (1999), Universidad Carlos III, II Foro Finanzas-Segovia, and MEFF Institute for Derivative Assets. Gabriele Fiorentini, Angel Le on and Gonzalo Rubio acknowledge the nancial support provided by Direcci on Interministerial Cient ca y T ecnica (DGICYT) grants PB96-0339, PB980979 and PB97-0621 respectively. The three authors have beneted from the research grant received from the Instituto Valenciano de Investigaciones Econ omicas (IVIE). The contents of this paper are the sole responsibility of the authors.
150
n and Rubio Fiorentini, Leo
extreme observations more likely than in the BS case. This increases the value of outof-the-money and in-the-money options relative to at-the-money options, creating the smile. Meanwhile, if the pattern shown by data contains a clear asymmetry in the shape of the smile, i.e. a smirk pattern, then this may be due to the presence of skewness in the distribution which has the eect of accentuating just one side of the smile. Given this evidence, extensions to the BS model that exhibit excess kurtosis and skewness have been proposed in recent years along two lines of research: Jumpdiusion models under a Poisson-driven jump process, and the stochastic volatility framework are the key developments in the theoretical option pricing literature3 . A recent and important attempt to summarize alternative option pricing models is carried out by Bakshi, Cao and Chen (1997). They are able to derive a closed-form jump-diusion model that includes previously studied models. It allows not only stochastic volatility, but also stochastic interest rates and stochastic jumps. Moreover, following Bates (1996), they use a cross-sectional framework to implement their model, and analyze the performance and hedging behavior of the nested option pricing models. Das and Sundaram (1999) also examine the extent to which these models are able to capture the observed anomalies discussed in literature. Bakshi, Cao and Chen (1997) nd that both stochastic volatility and jumps are important for pricing. However, they suggest that recognizing stochastic volatility alone produces the best hedging performance. On the other hand, Das and Sundaram (1999) argue that, generally speaking, stochastic volatility models yield better pricing results than jumps, although none of them is able to explain all patterns of kurtosis, skewness and volatility smiles found in empirical pricing literature. Stochastic volatility option valuation start with the bivariate diusion processes of Hull and White (1987), Scott (1987) and Wiggins (1987). In these models the volatility risk premium is not rewarded and thus removed from the valuation equation. Further, the correlation between the volatility and the stock return is zero in Hull and White (1987) and Scott (1987). Heston (1993) provides a closed-form solution for a European call option without imposing the restrictions of zero correlation and zero price of volatility risk by using Fourier inversion methods. The key objective of the paper is to analyze the empirical performance of the stochastic volatility model proposed by Heston (1993) relative to the BS framework for options on the Spanish IBEX-35 stock exchange index. At the same time, the empirical and theoretical behavior of the parameters characterizing the diusion process assumed for the instantaneous variance is studied. Their behavior is discussed
3 Alternatively, Corrado and Su (1996), and Backus, Foresi, Li and Wu (1997) adapt a GramCharlier series expansion of the normal density function to obtain skewness and kurtosis adjustment terms for the BS formula. Eberlein, Keller and Prause (1998) introduce the hyperbolic density to account for excess kurtosis and skewness, and they are even able to obtain a closed option pricing formula under this assumption. Rosenberg (1998) suggests the so called exible density function methodology to estimate risk-neutral densities implied by option pricing data.
Short-Term Options with Stochastic Volatility
151
relative to the appropriate skewness and kurtosis underlying in Hestons model. This provides us with insights clarifying the reasons behind the poor performance found for the stochastic volatility option pricing model. Since the Spanish option market shows a very limited number of exercise prices traded simultaneously and where liquidity is also less generalized than in the US market, then there are serious consequences for the empirical implementation of the cross-sectional estimation of implied parameters proposed by Bates (1996) and Bakshi, Cao and Chen (1997), that is; there are just not enough prices available in order to estimate jointly all parameters embedded in Hestons model. This forces us to turn to estimate the parameters in two separate steps. First, we estimate the parameters of the stochastic volatility, which are required inputs of Hestons model, by employing the indirect inference estimation technique of Gourieroux, Monfort and Renault (1993) on a time-series of the underlying hourly return. Second, the price of the volatility risk and instantaneous variance are backed out by minimizing the sum of squared pricing errors between the option model and market prices as in Bakshi, Cao and Chen (1997). Besides this practical argument, it should also be pointed out that the cross-sectional approach may easily ignore relevant information in the original series that may not be embedded in the option prices The remainder of the paper is organized as follows: Section 2 contains a brief summary of the Spanish options market features and the options data. The theoretical model, i.e. Hestons model, employed in this paper appears in Section 3. The empirical results regarding the time series estimation of the stochastic volatility parameters, the theoretical discussion on the relationship between these parameters and the appropriate skewness and kurtosis in Hestons framework appears in Section 4. The implied volatility graphs of the daily implied variance, the estimation and testing of the volatility risk premium through option prices is shown in Section 5. In Section 6, the out-of-sample pricing performance is carried out. Finally, in Section 7, we conclude with a summary.
2
2.1
The Spanish IBEX-35 Index Options

Market description
The Spanish IBEX-35 index is a value-weighted index comprising the 35 most liquid Spanish stocks traded in the continuous auction market system. The ocial derivative market for risky assets, which is known as MEFF, trades a futures contract on the IBEX-35, the equivalent option contract for calls and puts, and individual option contracts for blue-chip stocks. Trading in the derivative market started in 1992. The market has experienced tremendous growth from the very beginning. Relative to the volume traded in the Spanish continuous market, trading in MEFF represented 40% of the regular continuous market in 1992, 156% in 1994, and 170% in 1995. The number of all traded contracts in MEFF relative to the contracts traded in the CBOE reached 20% in 1995.
152
The IBEX-35 option contract is a cash settled European option with trading during the three nearest consecutive months and the other three months of the March-JuneSeptember-December cycle. The expiration day is the third Friday of the contract month. Trading occurs from 10:30 to 17:15. During the sample period covered by this research, the contract size is 100 Spanish pesetas times the IBEX-35 index, and prices are quoted in full points, with a minimum price change of one index point or 100 pesetas4 . The exercise prices are given in 50 index point intervals. It is important to point out that liquidity is concentrated in the nearest expiration contract. Thus, during 1995 and 1996 almost 90% of crossing transactions occurred in this type of contracts. Finally, it should be noticed that options and futures contracts are directly associated. The futures contract has exactly the same contract specications as the IBEX-35 options. This will allow us to employ the futures price rather than the spot price in our empirical exercise. In fact, this is what is usually done by practitioners.
2.2
The data
For this paper, our database is comprised of all call options on the IBEX-35 index traded daily on MEFF during the period January 3, 1996 through April 30, 1996. Given the concentration in liquidity, our daily set of observations includes only calls with the nearest expiration day. Moreover, we eliminate all transactions taking place during the last week before expiration (to avoid the expiration-related price eects). As usual in this type of research, our primary concern is the use of simultaneous prices for the options and the underlying security. The data, which are based on all reported transactions during each day throughout the sample period, do not allow us to observe simultaneously enough options with the same time-to-expiration on exactly the same underlying security price but with dierent exercise prices. In order to avoid large variations in the underlying security price, we restrict our attention to the 45-minute window from 16:00 to 16:45. It turns out that, on average and during our sample period, almost 25% of crossing transactions occur during this interval. Moreover, care was also taken to eliminate the potential problems with articial trading that are most likely to occur at the end of the day. Thus, all trades after 16:45 were eliminated so that we avoid data which may reect trades to inuence market maker margin requirements. At the same time, using data from the same period each day avoids the possibility of intraday eects in the IBEX-35 index options market. These exclusionary criteria yield a nal daily sample of 768 observations. Table 1 describes the sample properties of the call option prices employed in this work. Average prices, average relative bid-ask spread and the number of available calls are reported for each moneyness category. Moneyness is dened as the ratio of the exercise price to the futures price. A call option is said to be deep out-of-the-money if the ratio
4 This
has recently been changed to 1,000 pesetas.
153
K/F belongs to the interval (1.03, 1.08); out-of-the-money if 1.03 > K/F 1.01; atthe-money when 1.01 > K/F 0.99; in-the-money when 0.99 > K/F 0.97; and deep-in-the-money if 0.97 > K/F 0.90. As we already discussed, there are 768 call option observations, with OTM, ATM and ITM options respectively representing 51% , 32% and 17% . The average call price ranges from 12.88 pesetas for deep OTM options to 185.42 pesetas for deep ITM options. The average relative bid-ask spread goes exactly in the opposite direction to the average price. In particular, it ranges from 0.39 for deep OTM options to 0.10 for deep ITM calls. The implied volatility for each of our 768 options is estimated next. Note that we take as the underlying asset the average of the bid and ask price quotation given for each futures contract associated with each option during the 45-minute interval5 . Recall that we are allowed to use futures prices given that the expiration day of the futures and options contracts systematically coincides during the expiration date cycle. Moreover, note that dividends are already taken into account by the futures price. To proxy for riskless interest rates, we use the daily series of annualized repo T-bill rates with either one week, two weeks or three weeks to maturity. One of these three interest rates will be employed depending upon how close the option is to the expiration day. Finally, as discussed by French (1984), volatility appears to be a phenomenon that is basically related to trading days. However, interest rates are paid by the calendar day. Thus, in order to estimate the implied volatility of each option in our sample, we employ Blacks (1976) option pricing formula adjusted by two time measures to reect both trading days and calendar days until expiration. These implied volatilities will be used later as the basis for comparison with Hestons implied instantaneous volatilities.
Hestons stochastic volatility option pricing model
Heston (1993) obtains a closed-form solution for the price of a European call option on an asset with stochastic volatility. Heston works with Fourier transforms of conditional probabilities that the option expires in-the-money. The characterization of these probabilities is achieved through their characteristic function. The stochastic volatility model proposed by Heston generalizes Geometric Brownian Motion by allowing the volatility of the return process itself to evolve stochastically over time in a square root mean-reverting fashion: dSt = St dt + Vt St dW1t , (1) dVt = ( Vt ) dt + Vt dW2t , dW1t dW2t
5 It
dt ,
might be that lack of liquidity in the futures market is responsible for the lack of variation in the price of the underlying asset during the 45-minute window. However, this is not the case. In fact, the futures market is at least as liquid as the spot market in terms of comparable measures of trading volume.
154
where is the instantaneous expected rate of return of the underlying asset, Vt is the instantaneous stochastic variance, is the long-term mean of the variance, governs the rate at which the variance converges to this mean, and represents the volatility of the variance process. The parameters of the variance process, , and are all strictly positive constants. W1t and W2t are each a standard Brownian motion allowed to be instantaneously correlated. Thus increases (decreases) in volatility could be related to the level of the underlying asset. The risk-neutral probability measure incorporates the market price of volatility, denoted as , to distinguish the objective probability measure from the risk-neutral one. The volatility risk premium is assumed to be proportional to the instantaneous variance, Vt , and its sign arises from the (sign of) correlation between the Brownian processes assumed for the instantaneous variance and the (aggregate) consumption. The model is given by: dSt = r St dt + Vt St dW1 t, (2) dVt = ( Vt ) dt + Vt dW2 t,
dW1 t dW2t
= dt ,
where = + and = / ( + ). Let c (S, , t) be the value of an European call option where S St and Vt to abbreviate. Hestons formula is given by: c (S, , t) = St P1 K er (T t) P2 , (3)
where P1 and P2 are two risk-neutralized probabilities having the same interpretation as in the standard BS expression. In the application below, we use future options so that the actual version of the formula we employ is given by: c (F, , t) = er (T t) (Ft P1 K P2 ) , where F is the future price on the underlying spot price, and Pj (x, , T t ; ln [K ]) = Prob xT ln [K ] xt = x, t = , (5) (4)
where, x ln [Ft ], j = 1 or 2 (the probability of the event {x ln [K ]} depends on whether we chose the futures for j = 1, or the riskless rate for j = 2), and Pj is, therefore, the conditional probability that the option expires in-the-money. These probabilities depend on the vector of parameters (, , , , ) given by the process assumed by Heston under the original probability. It is important to note that the expressions for these probabilities are slightly dierent from the original values given by Heston (1993) since we are using futures. The actual formulae employed in this paper are provided in Appendix A.
155
In the empirical implementation of the model, we could always go to option prices and implicitly infer the parameters under the risk neutral probabilities. Given the limited number of daily option prices available to the 45-minute window from 16:00 to 16:45, the cross-sectional approach is just not possible if we wish to estimate all the parameters implied in Hestons model for calculating daily call prices. Our estimation approach consists of a two step procedure. First, we employ the original asset data (a time series of the spot IBEX-35 return) to estimate the parameters from the true process, i.e. equation (1), by adjusting the discretization biases throughout the indirect inference estimation methodology which is discussed in the following section. Second, we estimate the volatility risk premium and daily instantaneous variance from option prices. Denitively, this methodology needs two dierent sources of data for estimating the parameters, both time series data and cross section data.
4
4.1
Indirect estimation
Parameter estimation of the stochastic variance process
To estimate the process under the original probability measure given by (1), we have to assume that available data are discrete-time observations of a continuous time process. If we apply regular econometric methods to discrete-time approximations, we would have a serious estimation -discretization- bias in our results. In order to avoid this bias, we employ the indirect estimation proposed by Gouri eroux, Monfort and Renault (1993). The procedure consists of two steps. First of all, by maximum likelihood techniques, we estimate an appropiate auxiliary model. Secondly, the estimates of the auxiliary model are compared with estimators based on simulations of the path of the continuous time process given by (1). More specically, we have to introduce a discrete time analogue of (1) corresponding to a small time unit , such that 1/ is an integer. This is done by Euler approximation. Then for a given value of the parameters, we simulate the process, and obtain simulated values for the observation dates by merely selecting the values corresponding to integer indexes. This yields an accurate simulation of V as long as is suciently small. In Appendix B, we precisely describe the steps necessary for the indirect estimation procedure. With respect to the empirical results, we rst estimate the in-sample period from January 2, 1994 to January 2, 1996 with continuously compounded hourly returns from the Spanish IBEX-35 index. The calculation of returns is based on the last recorded logarithmic index levels over consecutive hourly intervals. It is well known that the rst hourly return incorporates adjustments to the information which has arrived overnight, and therefore presents a higher average return variability than any other hourly return. This basically implies that this rst return is not an hourly return, and we consequently delete it from the estimation. Our nal sample length consists of 2,450 hourly observations. The results for the in-sample period are reported in Table 2. We consider four alternative combinations of simulations and frequencies and the Euler discretization
156
specication6 . The results contained in the rst column are obtained simulating the process once (N = 1), and 2, 450 10 data points since the frequency used is equal to = 1/10. In the second column of Table 1, we simulate (1 , 2 ) 10 times (N = 10), and we generate 10 series of size 2, 450 10 given that the frequency is again = 1/10. Hence, our nal estimates are based on the following minimization: min
1 10
10
i
i=1
In the third column, the process is simulated once (N = 1), but the calibration of the Nagarch (1,1) model is performed with more data. In particular, the length of the vector used in the estimation is 2, 450 10. Given that the frequency is = 1/10, we have a total of 2, 450 10 10 data points. Note, of course, that once we go back to the Nagarch (1,1) model using simulated data of equal frequency as the real data, we have 2, 450 10 observations. This is the methodology employed in the rolling procedure for the out-of-sample estimation which will be used in testing Hestons option pricing model under Euler discretization. Finally, the results reported in the fourth column simulate the process once (N = 1), but the frequency is established at = 1/50. The estimations contained in Table 2 suggest that, independently of the alternative procedure employed, the results tend to be quite similar. There seems to be some minor dierence in the estimator of the volatility of the variance process, and the estimator of the correlation coecient between the shocks when we use the frequency at = 1/50. In any case, independently of the procedure employed, the estimate of the correlation coecient is, surprisingly, positive and close to zero. Given the asymmetry coecient found in the Nagarch (1,1) structure, we would have expected a negative correlation coecient to reect the asymmetry generally observed in literature to reect that agents seem to react more to bad news than to good news7 . For the out-of-sample estimation, the same process is estimated 80 times using systematically 2,450 past observations. This rolling procedure of the indirect inference is necessary to yield estimates of the parameters involved in Hestons expression for each day between January 3, 1996 and April 30, 1996. Thus, the daily changing estimates of these parameters are used as inputs in Hestons option pricing formula. Figure 1 presents the evolution of the parameters associated with the stochastic variance process throughout the out-of-sample period. As we can observe from the gure,
(1997) discretization was also carried out but the results were similar to Euler discretization. 7 Duan (1997) shows that both the Glosten, Jagannathan and Runkle (1993) -GJR (1,1)- and Nagarch (1,1) models converge to the same limiting diusion process for the stochastic volatility, i.e. the mean-reverting Geometric Brownian motion. Le on and Sentana (1998) shows that the relationship between the parameters of this Brownian process and the ones corresponding to the Nagarch (1,1) model is exactly given by: = 2 1 + 2 2 . Then, it must be the case that sign() = sign( ). Of course, this limiting result does not hold for the square-root mean reverting process.
6 Nowman
157
the long-term volatility, , the rate of convergence of the instantaneous variance to the long-term average, , and the volatility of the variance process, , remain quite stable over the out-of-sample period. However, the coecient of correlation between the shocks for the stock and the variance, , increases continuously. As before, it is interesting to observe that the correlation remains positive throughout the out-ofsample period.
4.2
Skewness, kurtosis and the parameters of the stochastic volatility process
The time-varying behavior of the correlation coecient found above may have a serious impact on the capacity of Hestons model to explain option pricing data. It suggests that, in this case, we may have a similar problem than the one we have when assuming constant volatility in the BS context. If correlation between prices and volatility changes continuously over time, skewness of the underlying asset may also exhibit time-varying behavior. This is clearly a potential and relevant problem for models with stochastic volatility. On the other hand, according to the estimates shown in Figure 1, it seems that the behavior of the volatility of volatility is rather stable over time. This suggests that accounting for changing kurtosis of the underlying asset may not be as crucial as taking into consideration changing skewness. Das and Sundaram (1999) obtain closed-form expressions for conditional and unconditional skewness and kurtosis under Hestons stochastic volatility model. Their expressions, for a given frequency of t = 1 can be seen in Appendix C. Given our rolling estimates of and from January 1996 to April 1996, we can daily estimate (C1) and (C2) from Appendix C, so that we may observe how the characteristics of the distribution of the underlying asset skewness and kurtosis change with both the correlation coecient and the volatility of volatility. Figure 2 depicts the conditional and unconditional skewness over the sample period. As we can easily observe, their behavior closely follows the pattern of the correlation coecient of Figure 1. As expected, skewness mainly arises from the correlation between changing prices and stochastic volatility of the underlying asset. The problem is that, of course, Hestons model assumes a constant unconditional skewness over time8 . In Le on and Rubio (2000), it is shown that, all else being constant, the relationship between skewness and the correlation coecient is positive; i.e. (SKEW ) > 0 for both the conditional and unconditional cases. This explains the behavior of skewness in Figure 2. Therefore, if (as is in fact the case) the correlation coecient
8 Given the similarity between the behavior of unconditional and conditional skewness, the impact of the instantaneous stochastic variance on the conditional skewness must be negligible relative to the eect of the correlation coecient.
158
tends to be rather unstable, an option pricing model with a stochastic dierential equation for correlation would be welcomed9 . Given this evidence, we should not expect to nd a good performance of the stochastic volatility model when we compare observed market prices with theoretical prices. We will come back to this issue later in the paper. Figure 3 contains the same evidence for the kurtosis case. In principle, the impact of time-varying kurtosis on the misspecication of Hestons model seems to be much less severe than the inuence of the correlation coecient. Its pattern over time is much more stable than . This is related to the behavior of the volatility of volatility, , in Figure 1. This is the parameter that allows for kurtosis in the stochastic volatility option pricing model, and it does not seem to change much over time. Once again, in Le on and Rubio (2000), it is shown that the relationship between kurtosis and is positive, i.e, (KU RT ) > 0, for both the conditional and unconditional cases. As a summary, time-varying skewness may be the key issue to analyze if we want to understand the failure of the tests we report below. Its consequences may be much more serious than the potential eects of changing kurtosis. However, this point will be claried in Section 6.
Estimating the implied variance and the volatility risk premium
Volatilities from the BS model and both the instantaneous variance and the volatility risk premium from Heston model are estimated every day from option prices, specifically all available call options transacted over the 45-minute interval from 16:00 to 16:45 during the period January 3, 1996 to April 30, 1996.
5.1
Estimation procedure
Note that, once we have estimated for each day the parameters of the variance process, , , and, using the rolling indirect inference procedure, we still need to estimate the volatility risk premium, , and the instantaneous variance, Vt , before we can actually price a given call option under Hestons model. Therefore, it seems reasonable to expect that we may use cross-sectional data to implicitly infer the parameters that minimize the sum of squared errors (SSE) in a given day of the sample. Given the set of parameters of the indirect estimation procedure obtained for a particular day t in the sample, t = t , t , t , t , and for each option i (i = 1, . . . , n) and each day t, we dene the pricing error as: eit Vt , ; t = cit (Ki ) cit (Ki ) ,
9 See
(6)
Nandi (1998) for an exhaustive discussion of the importance of the correlation coecient
159
where cit (Ki ) is the theoretical price of call i in day t, and cit (Ki ) is the corresponding observed market price. We then want to nd the instantaneous variance, Vt , and the risk premium parameter, , to solve:
n
SSEt min
{Vt ,}
eit Vt , ; t
i=1
(7)
Direct inspection of the quadratic form in (7) shows that it is highly valley shaped and, therefore, it is extremely at along a direction corresponding to a nontrivial combination of and Vt . As a consequence, derivative based minimization methods are expected to perform poorly since the numerically computed gradient is very unstable in a neighbourhood of a minimum. This is conrmed by some experiments that we performed with the Newton-Raphson method. In order to avoid the above problem, a derivative-free minimization algorithm is called for. Since the function in (7) only depends on two parameters, the downhill simplex method of Nelder and Mead (1965) seems to be a natural candidate. We have robustied the method by choosing randomly several starting triplets and in those cases of convergence to dierent local minima their minimum was selected as the global solution. So, we obtain daily estimates of both and Vt from January 3, 1996 to April 30, 1996; i.e., a sample of 80 days. Finally, a series for daily implied volatilities estimated in the corresponding BSs versions of (6) and (7) is also obtained.
5.2
The volatility risk premium
In Figure 4, we can see that the estimated daily values of further suggest that the volatility risk premium varies over time in the IBEX-35 futures option market. Notice that this time-varying behavior of volatility risk premium is not consistent with Hestons model since must be a constant value as we can see in equation (2). The estimated values of ranges from 0.252 to 0.333. The mean (median) for the time-series of is 0.021 (0.027) and the standard deviation is 0.113. The skewness and the excess kurtosis are 0.527 and 0.747 respectively. The p-value of the Jarque-Bera test (normality test) for the time-series of is 0.085, so the normal distribution is rejected at the 10% signicance level. The null hypothesis of = 0 was tested under three dierent methods. First, assuming that values were drawn from a normal distribution, a Student t test was performed, so the Student t ratio was 1.58 (p-value= 0.1183). Second, a nonparametric test for the population median of values, specically the sign test10 , was performed whose p-value was 0.1544. Third, we also test H0 : = 0 by obtaining accurate condence intervals for based on the bootstrap-t condence interval (see Efron and Tibshirani, 1993). By selecting 10,000 independent bootstrap samples, each consisting of 80 data values
10 Since the empirical distribution is skewed, the sign test is more appropriate than the popular nonparametric Wilcoxon signed-rank test because the last one assumes that the underlying distribution is symmetric (see Rohatgi, 1984).
160
drawn with replacement from the series of , then the 90% , 95% and 99% bootstrap-t condence intervals for the mean are respectively: [0.0426; 0.0021], [0.0467; 0.007] and [0.0551; 0.0175]. Summing up, we conclude that the evidence does not support rejection of the null hypothesis, i.e. the volatility risk premium is zero. The nding of a zero volatility risk premium for Ibex-35 index options is opposite to those of Guo (1998), Sarwar and Krehbiel (2000) using currency options, Lin, Strong and Xu (1999) using FTSE 100 index options. Nevertheless, our result is consistent with Bakshi, Cao and Chen (1997) who nd that the maximum likelihood estimates of and are statistically indistinguisable from their respective S& P 500 option-implied counterparts, i.e. and . Since the parameter in Hestons model is a constant value and given that we can impose that = 0 for the volatility risk premium embedded in Heston option prices on the Spanish Ibex-35 index, according to the conclusions from the above tests, then we can solve equations (6) and (7) and obtain a daily estimate of instantaneous variance from January 3, 1996 to April 30, 1996. We also repeat the above procedure for two alternative values, specically = 0.50 and = 0.50, in order to analyze the sensibility of the estimated instantaneous variance under dierent magnitudes of which are in accordance with our sample of values. Denitively, in Section 6 we will test the out-of-sample pricing performance for Hestons model under the three candidate values of and compare each with BSs performance.
5.3
Implied volatility graphs
Figure 5 contains the evolution for several series of implied volatilities from January 3, 1996 to April 30, 1996 for each version of Hestons model, i.e. for = 0.5, 0, 0.5 and for a dierent daily value, and also, of course, for the BS formula. It can be easily appreciated that Hestons volatility, independently of the volatility risk premium assumed, tends to be higher than BSs volatility. A priori, this may have serious implications for pricing. Also, note that the four implied volatility series under Hestons model are very similar and for = 0.5, 0, 0.5, in general, the larger the required volatility risk premium imposed, the (very slightly) larger the instantaneous volatility. To obtain a general picture of the potential misspecications of the option pricing models employed in this research, we report the average pattern of implied volatilities across degree of moneyness. The results are shown in Figure 6. In the BS case, we back out the implied volatility of each call option and for each day of the above period using the procedure discussed in sub-section 2.2. Then, the equally-weighted implied volatility for each moneyness category and each day in the sample period is calculated. There is a U-shaped pattern with a hump in the middle in Figure 6. This suggests that the BS model tends to underprice deep OTM and deep ITM calls. This is the typical smile pattern of the Spanish option market analized by Pe na, Rubio and Serna (1999). Any reasonable alternative model to BS
161
must be able to properly price deep OTM and deep ITM call options. Of course, Hestons stochastic volatility model, given by equations (4) and (5), is a potential and particularly interesting candidate. For = 0.5, 0, 0.5, we may analyze the pattern of implied volatilities across alternative degrees of moneyness. The evidence reported in this regard in Figure 6 suggests a rather asymmetric smile in Hestons model independently of the volatility risk premium imposed. It seems that Hestons approach tends to underprice (deep) ITM calls and overprice (deep) OTM calls. We now turn to formal tests of the alternative option pricing specications considered in this paper.
Out-of-sample pricing performance
In order to test the out-of-sample pricing performance for each model analyzed in this work, we employ two years of rolling data to estimate, by indirect inference, the parameters of the stochastic variance process assumed in (1). Given these estimates, and a chosen volatility risk premium, , we use all call options available in our 45minute window to compute for each day from January 2, 1996 to April 29, 1996 the instantaneous variance that minimized the squared error between the theoretical value and the market price of the call options according to (6) and (7). We then compute the theoretical price of each option using the previous days instantaneous variance and the corresponding parameters of the stochastic volatility process. For the BS case, the previous days implied volatility that minimized the squared error between the theoretical value and the market price of the options is used to obtain the theoretical BS price of each option in the sample. In this way, we have 768 pricing errors for each of the calls available from January 3, 1996 to April 30, 1996, and for each of the models analyzed. These pricing errors are the basis for our analysis. Table 3 reports two measures of performance for the alternative model specications. Panel A contains the absolute pricing error which is the sample average of the absolute dierence between the model price and the market price for each call in the testing sample period. This statistic is reported for each moneyness category and for all calls in the sample. In Panel B, the reported percentage pricing error is the sample average of the theoretical price minus the market price, divided by the market price. Again, this statistic is calculated for each moneyness category and for all calls in the sample. Overall, in absolute terms, Hestons option pricing model tends to value slightly better than BS. The absolute pricing error over all calls is approximately 2.8 pesetas for Hestons model independently of the volatility risk premium assumed, and 3.2 pesetas for the BS model. It is quite important to notice that the level of the volatility risk premium does not seem to have any inuence on the performance of Hestons stochastic volatility model. The pricing errors obtained under alternative values
162
are practically identical. There might be some evidence in favor of a positive risk premium, but we can safely conclude that option prices do not seem to be sensitive to the volatility risk premium. This is an important empirical result. Kapadia (1998) shows that the expected value of the delta-hedged gain, under a stochastic volatility model where volatility risk is not priced, is equal to zero. This is exactly the same result that holds under BS. Consequently, if we may assume that volatility risk is not priced, the analysis of the dynamic hedging performance of both models is clearly facilitated11 . It should be pointed out that the overall slightly better performance of Hestons model is not maintained throughout all moneyness categories. In particular, Hestons model tends to value ATM and ITM calls better than BS. However, the opposite result holds for OTM calls. This is also the case when we analyze the percentage pricing error in Panel B. Hestons model, regardless of the volatility risk premium imposed, tends to overvalue OTM calls and, at the same time, the model undervalues ITM calls. However, the percentage pricing for ATM calls is practically zero. In fact, ATM calls are clearly more consistent with a stochastic volatility model than with the well known lognormal assumption. BS, on the other hand, tends to undervalue deep OTM and deep ITM calls. This is consistent with a U-shaped volatility smile. In Hestons case, the evidence points towards a sneer rather than a regular smile.
6.1
The statistical signicance of the out-of-sample performance
Overall, at least over the sample period studied and regardless of , Hestons model seems to overvalue, while BS tends to undervalue call options. Unfortunately, however, the simple statistics reported above do not help in making inferences in terms of the statistical signicance of improvement when we contrast one model versus another. In this paper, the statistical signicance of performance for out-of-sample pricing errors is assessed by analyzing the proportion of theoretical prices lying outside their corresponding bid-ask spread boundaries12 . The following Z-statistic for the dierence between two proportions is employed in the tests. The statistic is given by: Z= p1 p2
p1 (1p1 ) n1 p2 (1p2 ) n2
where p1 is always the proportion of BS prices outside the bid-ask boundaries, and p2 is the equivalent proportion for alternative Hestons specications. n1 and n2 are sample sizes corresponding to these proportions. The statistic is asymptotically distributed as a standardized normal variable.
11 On the other hand, if volatility risk is priced, the average delta-hedged gain is proportional to the magnitude and sign of the volatility risk premium. Comparisons of the dynamic hedging performance between Hestons model and BS may be much more complicated than the analysis carried out by Bakshi, Cao, and Chen (1997). 12 See Corrado and Su (1996).
163
The empirical results are reported in Table 4. Given that we are also interested in knowing whether a given theoretical valuation model undervalues or overvalues market prices, the Z-statistic is also calculated to obtain the proportion for which the theoretical model yields a price below the bid quote, and the proportion for which the model gives a price above the ask quote. If a theoretical model tends to undervalue market prices, it would yield a higher proportion of prices below the bid quote. If, on the other hand, the model tends to overvalue market prices, it would have a higher proportion of prices above the ask quote. When we consider all call options together, the p-value for statistiscal improvement of Hestons model over the BS formula is equal to 0.068. The proportion of BS prices lying outside the bid-ask boundaries is 48% , while the proportion of Hestons model, regardless of the volatility risk premium, is about 42% . As we see, there is a slight improvement, but we might interpret the results as rather disappointing. It is not clear at all that, given the costs of implementation, Hestons approach is worthwhile in terms of practical applications. It is also the case that 35% of BS prices are below the bid quote. Given that Hestons model yields approximately 18% of prices below the bid, we can conclude that BS, on average, signicantly undervalues market prices relative to Hestons approach. At the same time, Hestons prices are 24% of total observations above the ask quote. We can also conclude that Hestons model tends to signicantly overvalue market prices relative to the BS formula. Thus, mispricing associated with BS is basically related to the tendency of the model to yield prices below market values. However, Hestons mispricing is a consequence of the models tendency to oer prices above their corresponding market values. When we classify all call options by moneyness, similar conclusions are obtained. A call option is said to be OTM if K/F > 1, and a call is classied as ITM when K/F < 1. In general, within a given category of moneyness, neither model is statistically superior to the other. However, once again, BS mispricing comes from the tendency of the model to undervalue either OTM or ITM calls relative to Hestons model (and relative to market prices). The opposite is true for Hestons formula relative to BS (and market prices). It should be pointed out that in the latter case, the main problem arises when we value OTM with Hestons formula. There seems to be a strong tendency in Hestons model to overvalue this type of options. This was also reected in Table 3. It was mentioned above that the percentage pricing error of Hestons model for ATM calls turns out to be quite small. It may be the case that Hestons formula works particularly well for ATM options. Calls are said to be ATM when 1.01 > K/F 0.99. The exercise described above is repeated for this type of options. Overall, as expected, Hestons model tends to yield lower proportions of prices lying outside the bid-ask spread than in previously analyzed cases. However, the p-value for the dierence relative to BS proportions is just 0.078. As before, BS tends to yield a statistically signicant higher proportion of prices lower than the bid quote relative to Hestons
164
prices, and Hestons model presents a statistically signicant higher proportion of prices above the ask quote relative to BS prices. Now, however, Hestons formula has similar proportions above the ask or below the bid. Regarding ATM calls, Hestons misspecication cannot be explained by either overvaluation or undervaluation of call prices.
6.2
The structure of pricing errors
Given the poor empirical performance of both BS formula and Hestons stochastic volatility model, a further analysis trying to understand the structure of pricing errors of these models would seem to be called for. Following the evidence reported by Pe na, Rubio, and Serna (1999) and Bakshi, Cao, and Chen (1997), we use a simple regression framework to study the relation between the percentage pricing errors and factors that are either option-specic or market dependent. We rst take as given an option pricing model, and let ei t be the i-th call options percentage pricing error on day t. Finally, we run the following regression for the whole sample period: eit = + 1 Xit + 2 it + 3 spt + 4 volt + 5 termt + 6 skewt + 7 kurtt + it , (8) where X is the moneyness of the i-th call at time t as dened by the ratio between the exercise price (K ) and the futures price (F ); i t is the annualized time to expiration of the i-th call on day t; sp is the average relative bid-ask spread of all calls and puts transacted between 16:00 and 16:45 on date t; volt is the annualized standard deviation of the IBEX-35 index returns computed from 1-minute intradaily returns; termt is the yield dierential between the annualized ten-year government bond and the annualized one-month repo Treasury bill; skewt is the conditional skewness, and kurtt the conditional kurtosis. They are given by expression (C1)13 from Appendix C. Table 5 contains the regression results based on the entire sample period and 768 call options, and where the standard error for each coecient estimate is adjusted by the White (1980) heteroskedasticity-consistent estimator. The explanatory variables employed in the regressions tend to be statistically signicant. However, there are important dierences between the percentage errors associated with either BS or Heston. In particular, a key point of the results is the statistical signicance of the coefcient estimates related to skewness and kurtosis, when we consider Hestons model under any of the three volatility risk premia used in the analysis. This is a very important result. As we argued in sub-section 4.2, the assumption of constant correlation between stochastic variance and price changes, and even the assumption of
13 The same regression was run using one day lagged values for the market dependent variables. Very similar results were found. The actual regressions employ excess kurtosis as an explanatory variable.
165
constant volatility of variance under Hestons model do not seem to be the appropriate assumptions to adequately explain the behavior of option prices even allowing for stochastic volatility. Note that, on the other hand, the coecient associated with kurtosis is not statistically signicant in the BS case. However, as in Hestons case, the skewness bias is also relevant to explain the BS pricing errors. Table 5 also shows that the annualized standard deviation of the IBEX-35 index slightly explains the percentage pricing errors independently of the model employed in the estimation. On average, pricing errors tend to be lower the higher the volatility of the index. However, the statistical signicance of the coecients is very weak. Traditional biases are not corrected for any of the models. The bias associated with moneyness has the opposite sign in both types of models. As expected, given previous results, Heston tends to price OTM options worse than ITM calls. However, on average, the opposite result is found for BS. Moreover, the bias related to time to expiration seems to be relevant for both types of models. On average, the percentage pricing errors are larger the longer the time to expiration. The inuence on percentage pricing errors of both the yield dierential between interest rates and the average spread for all calls and puts transacted over the 45 minute window are, interestingly enough, dierent for Heston and BS expressions. In the BS case, the higher the long-term yield relative to the short-term rate, the lower the percentage pricing error. However, this result disappears when we allow for stochastic volatility under any of Hestons specications. Let us analyze the spread variable. Regressions of a similar type were run including the relative bid-ask spread at date t of the call i. This is, contrary to the results reported in Table 5, a contract-specic variable. Surprisingly, the estimated coecients are never signicant regardless of the model considered. By doing this, we are really incorporating a transaction cost variable directly associated with the liquidity of each individual option. Again, this variable does not seem to be signicant in explaining percentage pricing errors. On the other hand, however, when we include the average spread over all call and put options for a particular day t, the estimated coecient becomes positive and signicant in Hestons case. This aggregate variable may indicate the average consensus about the uncertainty of trading throughout the option market. It may be understood as the average adverse selection confronting market makers in trading options on the Spanish index. As we see from Table 5, the larger the average spread -larger average adverse selection among traders- the higher the percentage pricing error in Hestons stochastic volatility models. The impact of this type of uncertainty does not seem to be relevant in explaining the percentage pricing errors of BS. In short, the pricing errors from Hestons framework have some moneyness, maturity, average (aggregate) bid-ask spread, skewness and kurtosis related biases. On the other hand, the BS case presents some moneyness, maturity, yield dierential and
166
skewness associated biases. Neither model seems to capture appropriately the underlying distribution characteristics of the underlying asset. Further research is clearly justied.
Conclusions
This paper introduces a two-step procedure, based on both time series and cross section data, for estimating all the parameters that we need to compute Heston call price. This estimation approach should be particularly useful in thin markets where a single cross-sectional estimation approach may be dicult to implement, for instance Spanish options data on the Ibex-35 stock index. Moreover, to employ just a crosssectional procedure may ignore relevant information that may be included in the original series but not in the option prices. The empirical results, however, are quite disappointing. On average, over all call options available in our sample, Hestons model improves the (poor) performance of BS just marginally. It is clear that this extremely limited improvement cannot justify the implementation costs involved in the estimation of Hestons approach14 . The overall rejection of Hestons model coincides with recent ndings by Bakshi, Cao and Chen (1997) and Chernov and Ghysels (1998) for options written on the S& P 500 index. We are quite convinced that the ultimate reasons behind the performance failure of Hestons model are closely related to the time-varying skewness and kurtosis found in the data. In particular, the assumption of a constant correlation coecient between returns and stochastic volatility should be relaxed if we really want to have a richer model. Unfortunately, of course, the complexities needed to price options seem to increase without bounds. It may be the case that simple nonparametric methodologies are able to incorporate the missing (realistic) factors in our option pricing models. It is also found that the daily volatility risk premium presents a quite volatile behavior over time. However, our evidence suggests that the volatility risk premium has a negligible impact on the pricing performance of Hestons model. A potentially interesting area of research might be related to endogenously incorporating liquidity costs in option pricing models with either stochastic volatility, stochastic jumps or both. Once again, this approach may be extremely demanding from a theoretical point of view but it would probably be welcomed.
14 Hedging performance of alternative models is not analyzed in this paper. It is possible that the hedging improvement under Hestons stochastic volatility model might be clearly superior to BS.
167
Appendix A: Hestons stochastic volatility model using futures options

Given that we use futures options, the actual version of the formula we employ is given by: c (F, , t) = er (T t) (Ft P1 K P2 ) , where F is the future price on the underlying spot price, K is the exercise price and the probabilities are given by: 1 1 + 2
Pj =
Re
0
ei ln[K ] fj i
d ;
j = 1, 2 , 1,
where Re(y ) is the real part of the function y ; i is the imaginary number i = and fj (x, , T t ; ) = exp [C (T t ; ) + D(T t ; ) + i x] , where15 C (T t ; ) = D(T t; ) = g = a 2 (bj i + d) (T t) 2 ln , 1 g ed (T t) 1g ,
bj i + d 1 ed (T t) 2 1 g ed (T t) bj i + d , bj i d
d = a b1 b2 u1
( i bj )2 2 (2 uj i 2 ) ,
= , = + , = + , 1 1 ; u2 = = 2 2
verifying that C (0) = D(0) = 0, and where C (T t ; ) and D(T t ; ) (and therefore the probabilities Pj , j = 1, 2) depend on the vector of parameters, (, , , , ) given by the processes assumed by Heston under the original probability.
15 As we have already pointed out, the expression for C (T t ; ) below is slightly dierent than the original value given by Heston (1993) since we are using futures.
168
Appendix B: Indirect estimation procedure

1. Let xt ln St ; then equation (1) becomes dxt = Vt 2 dt + Vt
1/ 2
dW1t , dW2t ,
(B1) (B2)
dVt = ( Vt ) dt + Vt dW1t dW2t = dt .
1/ 2
The above expressions are used to obtain the set of estimators (, , , , ). 2. Let us consider next the Euler discretization16 of both (B1) and (B2) with frequency : xt = xt + Vt 1/ 2 + Vt 1/2 1t , 2 (B3) (B4)
Vt = + (1 ) Vt + t , where t = 1/2 Vt 1t + (1 2 )1/2 2t , (1t , 2t ) N (0, 1) . 3. Our auxiliary model is given by the Nagarch (1,1) model17 : Rt ht = + t ; t = ht
1/ 2 iid 1/ 2
t ;
t N (0, 1) ,
1/ 2 2
iid
(B5)
= + ht1 + t1 + ht1
where Rt xt xt1 and represents the relation between the shocks and the conditional variance. The set of parameters to be estimated by maximum likelihood is given by (, , , , ). 4. Now we discuss the steps for the indirect estimation itself: [4.1] Estimation of the auxiliary Nagarch (1,1) model with the observed data ( = 1) = , , , , .
16 Bakshi, Cao and Chen (2000) also use the Euler discretization for the method of simulated moments which is employed to estimate the structural parameters of the continuous stochastic volatility (SV) process of the underlying asset. They also use a square root process for the SV model. 17 See Le on and Mora (1999) for the behavior of alternative specications within the GARCH family in the Spanish stock market.
169
[4.2] Let 0 be a initial value for .

1 , so (1t , 2t ) is a (T / ) 1 dimensional [4.3] We simulate (1t , 2t ) with = 10 vector, i.e. (1t , 2t ) : t = 1, 2, . . . , T / .
[4.4] Given [4.2] and [4.3], simulate with Euler discretization equation (B4) to generate Vt , and then equation (B3), where the set of simulated values from (B3) is denoted by: X (0 , ) {xt ; t = 0, , 2, . . . , T / } . [4.5] In order to work with the same frequency as the real data, take from X (0 , ) the following values: (0 , ) {x0 , x10 , x20 , . . . , xT } . [4.6] Now, we go back to the auxiliary model and with the data from [4.5], we have a new set of Rt in (B5). We estimate the Nagarch model with the simulated data at real frequency and get (0 , ) which is the vector of ML estimators 1 of the Nagarch model with = 10 for (0 , ). [4.7] We have the same number of parameters in and , so that we, in fact, minimize a distance with I as the weighting matrix18 . Then, if = (0 , ) = indirect estimators are 0 = END. [4.8] If = (0 , ) = GO TO [4.2]. So, let 1 be another new value for , continue the process and stop if = (1 , ); otherwise come back to [4.2] and repeat the process until convergence.
18 Since the model is exactly identied, the results are unaected by the choice of the weighting matrix.
170
Appendix C: (Un)conditional skewness and kurtosis

a) The conditional case: skew = 3e/2 3 1 + 2 (2 2e + + e ) (1 + e ) [ (1 e + e ) + (e 1)] A1 A2 B ,
3/ 2
, (C1)
kurt = where
A1 = 1 + 4e 5e2 + 4 e + 2 e2 + 42 6e 6e2 + 4 e + 2 e2 + 2 e A2 = 2 1 e2 + 2 e + 8 2 2 e 2 e2 + 2 e + 2 e , B = 2 [ (1 e + e ) + (e 1)] and where Vt . b) The unconditional case: skew = 3 1 e + e 3/2 e (C2)
2
kurt =
2 3 1+ 1 e + e + 4 2 [2 2 e + + e ] 2 e
Appendix D: Tables and Figures

Table 1: Sample Characteristics of Ibex-35 Futures Options
Average prices, average relative bid-ask spread and the number of available calls are reported for each moneyness category. All call options transacted over the 45 minute interval from 16:00 to 16:45 are employed from January 2, 1996 to April 30, 1996. K is the exercise price and F denotes the futures price of the IBEX-35 index. Moneyness is dened as the ratio of the exercise price to the futures price. OTM, ATM, and ITM are out-of-the-money, at-the-money, and in-the-money options respectively. Moneyness (K/F ) 1.03-1.08 1.01-1.03 0.99-1.01 0.97-0.99 0.90-0.97 Average Price 12.88 28.68 57.20 98.54 185.42 50.52 Average BidAsk-Spread 0.3903 0.2205 0.1491 0.1273 0.0987 0.2015 Number of Observations 116 273 245 108 26 768
DEEP OTM CALLS: OTM CALLS: ATM CALLS: ITM CALLS: DEEP ITM CALLS: ALL CALLS:
171
Table 2: Indirect Inference In-Sample Estimation (January 1994-December 1995)

The parameters of the following processes: dSt dVt dW1t dW2t = = = Vt St dW1t , ( Vt ) dt + Vt dW2t , St dt + dt ,
are estimated using the indirect inference technique; is the instantaneous expected rate of return of the underlying asset, Vt is the instantaneous stochastic variance, is the long-term mean of the variance, governs the rate at which the variance converges to this mean, represents the volatility of the variance process, and is the instantaneous correlation. The auxiliary model employed in the estimation is the following Nagarch(1,1) model19 : Rt ht = = + t ; t = ht
1/2
t ;
t N (0, 1) ,
1/2 2
iid
+ ht1 + t1 + ht1
where Rt is the hourly rate of return of the IBEX-35 index, and is the asymmetric parameter of the Nagarch. Euler discretization technique is employed in the estimation. Moreover, alternative frequencies and simulations are also used.
N = 1; = 1/10 Length = 2,450 0.619 12.09 0.029 1.631 0.044 N = 10; = 1/10 Length = 2,450 0.696 12.77 0.034 1.864 0.038 N = 1; = 1/10 Length = 2, 450 10 0.702 11.92 0.034 1.857 0.020 N = 1; = 1/50 Length = 2,450 0.868 11.73 0.025 1.542 0.087
Notes: 1. First column: The process is simulated once (N = 1), so that we employ 2, 450 10 data points since = 1/10, where 2,450 is the number of hourly returns available from January 1994 to December 1995. 2. Second column: The process is simulated 10 times (N = 10), so that we generate 10 series of size 2, 450 10 since, as in the rst column, = 1/10. 3. Third column: The process is simulated once (N = 1), but now we employ 2, 450 10 10 data points since = 1/10. Thus, the calibration of the Nagarch is done with more data: 2, 450 10. 4. Fourth column: The process is simulated once (N = 1), so that we employ 2, 450 50 data points since = 1/50. 5. is annualized and is given in percentage terms. 6. represents the standard deviation of the long-term variance. It is annualized and is given in percentage terms.
7. is annualized and is given in percentage terms.

19 The estimates of the Nagarch(1,1) parameters are = 0.00412, = 0.00301, = 0.07680, = 0.89025 and = 0.1225.
172
Table 3: Out-Of-Sample Pricing Error for Alternative Option Pricing Models. Absolute Pricing Error and Percentage Pricing Error
Two years of rolling daily data are employed to estimate by indirect inference the parameters of the stochastic volatility process under Hestons model. Given these estimates, and a chosen volatility risk premium, , we use all call options transacted over the 45 minute interval from 16:00 to 16:45 to compute for each day from January 2, 1996 to April 29, 1996, the instantaneous variance that minimized the squared error between the theoretical value and the market price of the options. We then compute the the theoretical price of each option using the previous days instantanteneous variance and the corresponding parameters of the stochastic volatility process. For the Black-Scholes case, the previous days implied volatility that minimized the squared error between the theoretical value and the market price of the options is used to obtain the theoretical price of each option in the sample. The reported absolute pricing error is the sample average of the absolute dierence between the model price and the market price for each call in a given moneyness category. The reported percentage pricing error is the sample average of the theoretical price minus the market price, divided by the market price. K is the exercise price and F denotes the futures price of the IBEX-35 index. Moneyness is dened as the ratio of the exercise price to the futures price. OTM, ATM, and ITM are out-of-the-money, at-the-money, and in-the-money options respectively.
Panel A: Absolute Pricing Error

Moneyness (K/F ) DEEP OTM CALLS: OTM CALLS: ATM CALLS: ITM CALLS: DEEP ITM CALLS: ALL CALLS: 1.03-1.08 1.01-1.03 0.99-1.01 0.97-0.99 0.90-0.97 Black-Scholes (Pta) 1.829 2.641 3.847 4.690 4.357 3.249 =0 2.033 2.815 2.874 3.593 3.809 2.859 Heston (Pta) = 0.5 = 0.5 2.030 2.038 2.819 2.774 2.854 2.896 3.586 3.600 3.795 3.823 2.852 2.853
Panel B: Percentage Pricing Error

Moneyness (K/F ) DEEP OTM CALLS: OTM CALLS: ATM CALLS: ITM CALLS: DEEP ITM CALLS: ALL CALLS: 1.03-1.08 1.01-1.03 0.99-1.01 0.97-0.99 0.90-0.97 Black-Scholes (Pta) -7.091 1.189 -3.894 -3.285 -2.560 -2.439 =0 5.916 8.429 -0.053 -1.351 -2.251 3.607 Heston (Pta) = 0.5 = 0.5 6.009 5.673 8.583 8.310 -0.021 -0.089 -1.331 -1.371 -2.243 -2.260 3.689 3.513
173
Table 4: Nonparametric Testing for Alternative Option Pricing Models

The statistical signicance of performance for out-of-sample pricing errors is assessed by analyzing the proportion of theoretical prices lying outside their corresponding bid-ask spread boundaries. The following Z-statistic for the dierence between two proportions given by: Z= p1 p2 p1 (1 p1 )/n1 + p2 (1 p2 )/n2
is employed in the tests, where p1 is always the proportion of Black-Scholes prices outside the bid-ask boundaries, and p2 is the equivalent proportion for alternative Hestons model specications. n1 and n2 are sample sizes corresponding to these proportions. The statistic is asymptotically distributed as a standardized normal variable. All call options transacted over the 45 minute interval from 16:00 to 16:45 from January 3, 1996 to April 30, 1996 are used in the tests below. K is the exercise price and F denotes the futures price of the IBEX-35 index. Moneyness is dened as the ratio of the exercise price to the futures price.
Categories All Options
p(Bid > c> Ask) p(c < Bid) p(c > Ask)
BS
Heston ( = 0) 0.4249 0.1837 0.2412
Z-stat. (p-value) 1.825 (0.068) 6.730 (0.000) -5.253 (0.000) 1.421 (0.155) 6.694 (0.000) -4.864 (0.000) 1.166 (0.244) 2.418 (0.015) -2.294 (0.022) 1.762 (0.078) 5.134 (0.000) -4.272 (0.000)
Heston ( = 0.5) 0.4249 0.1821 0.2428
Z-stat. (p-value) 1.825 (0.068) 6.804 (0.000) -5.319 (0.000) 1.421 (0.155) 6.791 (0.000) -4.940 (0.000) 1.166 (0.244) 2.418 (0.015) -2.294 (0.022) 1.763 (0.078) 5.395 (0.000) -4.341 (0.000)
Heston ( = 0.5) 0.4233 0.1901 0.2332
Z-stat. (p-value) 1.882 (0.060) 6.434 (0.000) -4.919 (0.000) 1.421 (0.155) 6.399 (0.000) -4.566 (0.000) 1.1656 (0.244) 2.303 (0.021) -2.123 (0.034) 1.765 (0.077) 5.189 (0.000) -4.231 (0.000)
0.4761 0.3487 0.1274
OTM Options
0.4821 0.3184 0.1637
0.4347 0.1351 0.2995
0.4347 0.1329 0.3018
0.4324 0.1419 0.2905
ITM Options
0.4615 0.423 0.0385
0.4011 0.3022 0.0989
0.4011 0.3022 0.0980
0.4011 0.3077 0.0934
ATM Options
0.4583 0.3981 0.0602
0.3750 0.1806 0.1944
0.3749 0.1800 0.1987
0.3735 0.1893 0.1915
Notes: OTM Options: K/F > 1. ITM Options: K/F < 1. ATM Options: 1.01 > K/F 0.99. c stands for cMODEL .
174
Table 5: Percentage Pricing Errors and Explanatory Variables

For a given option pricing model, the following regression is employed to explain the percentage pricing errors of all call options transacted over the 45 minute interval from 16:00 to 16:45 from January 3, 1996 to April 30, 1996: eit = + 1 Xit + 2 it + 3 Spt + 4 Volt + 5 Termt + 6 Skewt + 7 Kurtt + it , where eit is the percentage pricing error of the ith call on date t; X is the moneyness of the ith call at time t as dened by the ratio between the strike price (K ) and the futures price (F ); it is the annualized time to expiration of the ith call on day t; Sp is the average relative bid-ask spread of all calls and puts transacted between 16:00 and 16:45 on date t; Volt is the annualized standard deviation of the IBEX-35 index returns computed from 1-minute intradaily returns; Termt is the yield dierential between the annualized tenyear government bond and the annualized one-month repo Treasury bill; Skewt in the conditional skewness and Kurtt is the conditional (excess) kurtosis. The t-statistic reported in parenthesis in based on Whites heteroskedasticity consistent estimator of standard errors. A total of 768 call options are employed in the regressions. Coecient Constant Moneyness (X) Time to expiration ( ) Spread (Sp) Volatility (Vol) Term structure (Term) Skewness (Skew) Kurtosis (Kurt) Black-Scholes 0.683 (2.17) -0.746 (-2.20) 0.025 (2.67) 0.225 (1.48) -0.998 (-1.69) -0.129 (-4.65) 1.545 (2.65) 0.217 (0.77) Heston ( = 0) -0.563 (-1.79) 0.715 (2.02) 0.027 (2.84) 0.404 (2.69) -0.976 (-1.70) -0.008 (-0.26) 1.238 (2.15) -0.827 (-2.38) Heston ( = 0.5) -0.590 (-1.87) 0.737 (2.08) 0.026 (2.77) 0.409 (2.72) -0.986 (-1.72) -0.007 (-0.24) 1.227 (2.13) -0.805 (-2.31) Heston ( = 0.5) -0.550 (-1.77) 0.694 (1.98) 0.027 (2.84) 0.406 (2.71) -0.941 (-1.64) -0.009 (-0.30) 1.223 (2.13) -0.808 (-2.33)
175
Figure 1: Indirect Inference (January 96-April 96)
Figure 2: Time varying Skewness (January 96-April 96)
176
Figure 3: Time varying Kurtosis (January 96-April 96)
Figure 4: Daily Volatility Risk Premium (January 96-April 96)
177
Figure 5: Daily Implied Volatilities (January 96-April 96)
Figure 6: Smiles (January 96-April 96)
178
References
[1] Backus, D., Foresi, S., Li, K, and L. Wu (1997). Accounting for biases in BlackScholes, Working Paper, Stern School of Business, New York University. [2] Bakshi, G., Cao, C., and Z. Chen (1997). Empirical performance of alternative option pricing models, Journal of Finance 52, pp. 2003-2049. [3] Bakshi, G., Cao, C., and Z. Chen (2000).Pricing and hedging long-term options, Journal of Econometrics 94, pp. 277-318. [4] Bates, D. (1996). Jumps and stochastic volatility: exchange rate processes implicit in Deutsche mark options, Review of Financial Studies 9, pp. 69-107. [5] Black, F. (1976). The pricing of commodity contracts, Journal of Financial Economics 3, pp. 167-179. [6] Black, F. and M. Scholes (1973). The pricing of options and corporate liabilities, Journal of Political Economy 81, pp. 637-659. [7] Chernov, M. and E. Ghysels (1998). What data should be used to price options?, unpublished manuscript, Pennsylvania State University. [8] Corrado, C., and T. Su (1996). Skewness and kurtosis in S& P 500 index returns implied by option prices, Journal of Financial Research 19, pp. 175-192. [9] Das, S., and R. Sundaram (1999). Of smiles and smirks: A term-structure perspective, Journal of Financial and Quantitative Analysis 34, pp. 211-239. [10] Duan, J. (1997). Augmented GARCH (p,q) process and its diusion limit, Journal of Econometrics 79, pp. 97-127. [11] Eberlein, E., Keller, U. and K. Prause (1998). New insights into smile, mispricing, and value at risk: The hyperbolic model, Journal of Business 71, pp. 371-405. [12] Efron, B. and R.J. Tibshirani (1993). An introduction to the bootstrap. Chapman & Hall. [13] French, D. (1984). The weekend eect on the distribution of stock prices: implications for option pricing, Journal of Financial Economics 13, pp. 547-559. [14] Glosten, L., Jagannathan, R. and D. Runkle (1993). On the relation between the expected value and the volatility of the nominal excess return on stocks, Journal of Finance 48, pp. 1779-1801. [15] Gouri eroux, C, Monfort, A., and E. Renault (1993). Indirect inference, Journal of Applied Econometrics 8, pp. 85-118.
179
[16] Guo, D. (1998). The risk premium of volatility implicit in currency options, Journal of Business & Economic Statistics, 16, pp. 498-507. [17] Heston, S. (1993). A closed-form solution for options with stochastic volatility with applications to bond and currency options, Review of Financial Studies 6, pp. 327-344. [18] Hull, J. and A. White (1987). The pricing of options on assets with stochastic volatilities, Journal of Finance 42, pp. 281-300. [19] Kapadia, N. (1998). Do equity options price volatility risk? Empirical evidence, Working Paper, University of Massachusetts. [20] Le on, A. and J. Mora (1999). Modelling conditional heteroskedasticity: application to the IBEX-35 stock return index, Spanish Economic Review 1, pp. 215-238. [21] Le on, A. and E. Sentana (1998). From discrete to continuous time and back again, unpublished manuscript, CEMFI, Madrid. [22] Le on, A. and G. Rubio (2000) Smiling under stochastic volatility, unpublished manuscript, University of Alicante and University of Pa s Vasco. [23] Lin, Y., N. Strong and X. Xu (1999). Pricing FTSE 100 index options under stochastic volatility, working paper 99/018, The Management School, Lancaster University. [24] Nandi, S. (1998). How important is the correlation between returns and volatility in a stochastic volatility model? Empirical evidence from pricing and hedging in the S& P 500 index options market, Journal of Banking and Finance 22, pp. 589-610. [25] Nelder, J.A. and R. Mead (1965) A simplex method for function minimization, Computer Journal 7, pp. 308-313. [26] Nowman, K. (1997). Gaussian estimation of single-factor continuous time models of the term structure of interest rates, Journal of Finance 52, pp. 1695-1706. [27] Pe na, I., Serna, G, and G. Rubio (1999). Why do we smile? On the determinants of the implied volatility function, Journal of Banking and Finance 23, pp. 11511179. [28] Rohatgi, V.K. (1984). Statistical inference. John Wiley & Sons Inc. [29] Rosenberg, J. (1998). Pricing multivariate contingent claims using estimated risk-neutral density functions, Journal of International Money and Finance 17, pp. 229-247.
180
[30] Sarwar, G. and T. Krehbiel (2000). Empirical performance of alternative pricing models of currrency options, The Journal of Futures Markets 20, pp. 265-291. [31] Scott, L.O. (1987). Option pricing when the variance changes randomly: Theory, estimation and an application, Journal of Financial and Quantitative Analysis 22, pp. 419-438. [32] White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity, Econometrica 48, pp. 817-838. [33] Wiggins, J.B. (1987). Option values under stochastic volatilities: Theory and empirical estimates, Journal of Financial Economics 19, pp. 351-372.
Gabriele Fiorentini Departamento de Fundamentos del An alisis Econ omico Facultad de Ciencias Econ omicas y Empresariales Universidad de Alicante Campus San Vicente Raspeig Apartado de Correos 99, 03080-Alicante, Spain e-mail: gafi@merlin.fae.ua.es Angel Le on Valle Departamento de Econom a Financiera Facultad de Ciencias Econ omicas y Empresariales Universidad de Alicante Campus San Vicente Raspeig Apartado de Correos 99, 03080-Alicante, Spain e-mail: aleon@ua.es Gonzalo Rubio Departamento de Fundamentos Facultad de Ciencias Econ omicas Universidad del Pa s Vasco Avenida L. Aguirre, 83 48015 Bilbao, Spain e-mail: jepruirg@bs.ehu.es
May static pricing models be useful when pricing catastrophe-linked derivatives?

s, In aki R. Longarela and Julio Lucia1 Alejandro Balba
Abstract This paper deals with the PCS Catastrophe Insurance Option Contracts, providing empirical support on the level of correspondence between real quotes and standard nancial theory. The highest possible precision is incorporated since the real quotes are perfectly synchronized and the bid-ask spread is always considered. A static setting is assumed and the main topics of arbitrage, hedging and portfolio choice are involved in the analysis. Three signicant conclusions are reached. First, the catastrophe derivatives may be very often priced by arbitrage methods, and the paper provides some examples of practical strategies that were available in the market. Second, hedging arguments also yield adequate criteria to price the derivatives and some real examples are provided as well. Third, in a variance aversion context many agents could be interested in selling derivatives to invest the money in stocks and bonds. These strategies show a suitable level in the variance for any desired expected return. Furthermore, the methodology here applied seems to be quite general and may be useful to price other derivative securities. Simple assumptions on the underlying asset behavior are the only required conditions.
Introduction
New investment and nancing opportunities, and innovative risk management techniques involving derivatives have been developed to allow individuals and corporations to cost-eectively reallocate funds and transfer risks to other parties. A growing concern on catastrophe losses has particularly brought attention to catastrophe derivatives and their potential nancing and risk sharing benets for the insurance industry. The PCS (Property Claim Services) Catastrophe Insurance Options Contracts launched by the Chicago Board of Trade (CBOT) on September 29, 1995 are among the most signicant catastrophe derivatives. These are standardized option contracts
1 Alejandro Balb as es Catedr atico de Econom a Financiera de la Universidad Carlos III de Madrid. I naki R. Longarela es Doctor en Econom a por la Universidad Carlos III de Madrid y Associate Professor of Finance en la Universidad de Estocolmo. Julio Lucia es Profesor Titular en el Departamento de Econom a Financiera y Matem atica de la Universidad de Valencia. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de noviembre de 1998.
182
s, Longarela And Lucia Balba
based on indices that track the insured losses, as estimated by PCS, resulting from catastrophic events that occur in a given area and period. Previously, the CBOT had traded catastrophe futures and options contracts on an index provided by the Insurance Services Oce (ISO). Moreover, the CBOT planned to list PCS SingleEvent Catastrophe options in 1998 to broad its product oering. In 1997 the Bermuda Commodities Exchange (BCOE) also began trading derivative securities based on the Guy Carpenter Catastrophe Index, an index of losses from climate events in US. In this paper we focus on the CBOTs PCS options. Previous literature on these particular contracts and other related catastrophe derivatives can be roughly divided into two major categories, according to their main objective. The rst group of papers concentrates on pricing issues. They view catastrophe derivatives as nancial instruments and, accordingly, they take a nancial approach to valuing (see Cummins and Geman (1995), Geman and Yor (1997), among others; Tomas (1998) suggests an actuarial approach). They theorize on the dynamic stochastic behavior of the relevant underlying variables in order to obtain the desired pricing result. From a theoretical point of view this line of research is really important and very promising. From a practical point of view there are some diculties due to market imperfections (bidask spread, other transaction costs, short-selling restrictions, illiquidity that makes a continuous trading rather dicult, etc.) and some specic properties shown by the underlying indices (their stochastic behavior, the absence of any underlying security available for trade, etc.).This motivates the existence of a second group of papers devoted to describe the contracts and illustrate their most signicant applications to both, insurance and capital markets (e.g. see DArcy and France (1992), Canter et al. (1996), Litzenberger et al. (1996), OBrien (1997) and Jaee and Russell (1997)). They explore the potential benets of using catastrophe derivatives for the insurance industry, and they compare these securities to other competitive alternatives such as reinsurance and catastrophe-linked bonds. Some papers analyze the potentially attractive new investment opportunities provided by the catastrophe-linked assets from any investors perspective. Most of all these papers stress the traders need for an understandable and reliable complete pricing methodology for these innovative securities. The present paper may be included in the second group, but the standard static asset pricing models are applied. We consider real bid and ask prices of catastrophelinked derivatives and we test their adequacy with static nancial theory. Examining static valuation minimizes the impact of real market imperfections, and problems deriving from the nature of the underlying variables may be solved if one prices an arbitrary derivative by only bearing in mind the interest rates and the prices of other derivative securities. Thus, we can apply the main topics of asset pricing, arbitrage, hedging and portfolio choice, in a model where bonds and derivatives are the only marketed assets.2
2 Stocks, whose returns show an insignicant correlation with the PCS indices (see Canter et al. (1996)), can be included too.
Pricing catastrophe-linked derivatives
183
In order to use static theory, we will consider a two periods model characterized by the current date, the derivatives expiration date, their current bid and ask prices and their nal payos. The analysis is independently implemented once a day. Once the context has been xed, we will start by analyzing the existence of arbitrage portfolios. There are two dierent perspectives. First, we test the situation of an investor who incurs in the cost of the bid-ask spread, i.e. he/she sells at the bid and buys at the ask price. As expected, we will nd it impossible to form any arbitrage portfolio. Second, we explore the position of any agent who post one of the available prices (i.e. either he/she buys at the bid or sells at the ask for a given asset and incurs on the cost of the bid-ask spread for the rest of assets). If an arbitrage portfolio were available in this context, any other agent could oer a better price and still retain some of the arbitrage gains. Competition among traders willing to earn money without any risk should lead to a more reduced spread. As will be shown, arbitrage portfolios may be available from this second perspective and, consequently, some relative misspricings may be found, narrower spreads may be possible, and traders can sometimes improve real bid or ask quotes without any type of risk or, equivalently, they can price by arbitrage methods. When a concrete derivative can not be priced by arbitrage methods, we explore the existence of hedged portfolios containing this derivative. In particular, there may exist some hedged portfolios with a slightly lower guaranteed positive return than the risk-free return, but with a possible return far larger than it (under some conditions). Our results show that interesting portfolios of this type can actually be formed in some cases. Again, competition among traders trying to exploit the attractive benets of these portfolios should lead to reductions of the spread. Previous studies, Canter et al. (1996), Litzenberger et al. (1996), utilized empirical evidence concerning the insignicant correlation of the PCS national index with the S&P 500 index (see also Litzenberger et al. (1996) and the references contained therein for more evidence on this regard). Canter et al. (1996) stress the diversication benets open to investors participating in the new securitized insurance risk. Litzenberger et al. (1996) illustrate the attractiveness of including some hypothetical catastrophe bonds in diversied stocks or bonds portfolios in terms of the new risk/returns opportunities oered. This paper follows the theoretical framework proposed by Fisher Black and Robert Litterman based on the Capital Asset Pricing Model (C.A.P.M.) including the calculation of some necessary parameters from historical data of insurance losses and premiums. While our study on portfolio choice is closely related to these previous studies it takes a dierent point of view. We focus on the investment in catastrophe insurance options market. Suppose that an investor, possibly attracted by the accompanied diversication benets, adds insurance risk to his/her traditional portfolio of stocks, bonds and real estate. How should he/she eciently combine PCS options in this insurance portfolio with the risk-less asset in order to obtain the desired expected
184
return with a minimum variance? We will try to answer this question with the assistance of two important theorems concerning the static pricing theory. We also need the real probability distribution for the underlying insured loss index, and a linear pricing rule compatible with the real quotes. The probability distribution is obtained via simulation and historical catastrophe data. The linear pricing rule originates from a risk-neutral probability measure attained by applying a methodology proposed by Rubinstein (1994) and a number of others. Once the probability measure and the pricing rule are established we shall look for minimum variance portfolios. An interesting result seems to hold. For investors whose risk is not correlated with the PCS indices (i.e., investors that are not insurers) it may be very useful to sell catastrophe derivatives and to invest the money in other kind of assets, like bonds or stocks. Summarizing, our empirical results conrm the potential interest of catastrophelinked derivatives. They are useful to insurers because, in some sense, they can be regarded as a special type of reinsurance. Besides, they may also be interesting for other type of Financial Institutions (banks, for instance) because, if arbitrage and hedging arguments lead to low bid-ask spreads, these institutions can adequately diversify their portfolios by selling derivatives. Consequently, the high level of risk due to catastrophic events may be appropriately diversied among large numbers of investors who trade reinsurances in a nancial market. Finally, the applied methodology highlights two interesting properties. First, its usefulness to traders in providing practical criteria and investment strategies. Second, it may be quite general and can be implemented to analyze other kind of securities. Very weak assumptions are then required. Arbitrage and hedging arguments will hold if one is able to identify the underlying uncertainty, i.e., the underlying variables if we are working with derivatives. Variance aversion and C.A.P.M.-type arguments will work well if the probability measure, aecting the underlying uncertainty, can be determined.3 The remainder of the paper is organized as follows. Section 2 briey reviews the main theoretical results we rely on to carry out our empirical analysis of PCS options quotes. Section 3 summarizes the foremost characteristics of PCS options and all our data. In Section 4 we present the concrete methodology we adopt in our empirical research of PCS options quotes and provide our results. The paper ends with some concluding remarks in Section 5.
3 When the usual C.A.P.M. is tested, it is not possible to describe the underlying probability space, and researchers have to obtain information about it by studying correlations within the set of available securities. Nevertheless, in this case, the underlying PCS indices behavior has been directly analyzed and the derivatives quotes and returns were not used for this purpose.
185
Theoretical Background
Throughout the paper we will consider a static setting to analyze how the theory of portfolio selection and dierent asset pricing models may be applied to PCS option contracts. Thus, rst of all, we must summarize the general framework and the basic assumptions that lead to the most important theoretical results on asset pricing. A brief review of these topics is the main purpose of the present section. Later, we will provide the way this theory applies in this article to study the market of PCS option contracts. We focus on the two periods approach characterized by the present date t0 , a future date t1 , n securities denoted by S1 , S2 , . . . , Sn , their bid prices at t0 denoted by v1 , v2 , . . . , vn , their ask prices c1 , c2 , . . . , cn , and the future prices (or nal payos) at t1 which depend on a nite number of states of the world W1 , W2 , . . . , Wk and are given by the matrix A = (aij ), i = 1, 2, . . . , k, j = 1, 2, . . . , n, being aij 0 the price of Sj if the state Wi takes place.4 i > 0 will denote the probability of Wi , i = 1, 2, . . . , k, and will denote the whole probability measure. The inequalities cj vj for j = 1, 2, . . . , n are clear, and we will accept the convention vj = 0 (cj = ) if there is no bid (ask) price available for Sj . The inequalities cj > 0 and vj 0, j = 1, 2, . . . , n, will also be assumed. The rst security S1 will be a riskless asset (its nal payo is 1 and does not depend on the state of the world) and c1 = v1 > 0.5 As usual, the riskless return is given by R = 1/v1 . The row matrix x = (x1 , x2 , . . . , xn ) will represent the portfolio composed of xj units of Sj , j = 1, 2, . . . , n, and xj 0 (xj 0) must hold if vj = 0 (cj = ). Its n current (at t0 ) price will be P (x) = j =1 pj xj being pj = cj (pj = vj ) if xj 0 6 (xj 0), where we assume the convention 0 = 0 if cj = and xj = 0. Its price at t1 depends on the state of the world and is given by the column matrix Axt where xt is the transpose of x.
+ For an arbitrary portfolio x, we will consider the portfolios x+ = (x+ 1 , . . . , xn ) and x = (x1 , . . . , xn ) composed of the purchased and sold securities respectively. To be precise, x+ j = max{xj , 0} and xj = max{xj , 0} for j = 1, 2, . . . , n.
The prices of the purchased and sold assets will be denoted by C (x) and V (x) respectively, and are given by
n
C (x) =
j =1
cj x+ j
and
V (x) =
j =1
vj x j .
4 Almost all the results still hold for a matrix A whose elements are also negative, but the constraint aij 0 makes things a little easier and is always fullled in our empirical analysis. 5 Once again, this assumption can be avoided in a general framework, but it is useful and fullled in this article. 6 I.e., agents can buy or sell any security but prices are larger if they buy.
186
The relationships P (x) = C (x) V (x) and x = x+ x are clear. We will follow the approach proposed for instance by Prisman (1986) or Ingersoll (1987) to introduce the concept of arbitrage.7 Denition 1 The portfolio x is said to be an arbitrage portfolio of the second type, or a strong arbitrage portfolio, if P (x) < 0 and Axt 0 , or P (x) = 0 and Axt >> 0. The portfolio x is said to be an arbitrage portfolio of the rst type, or a weak arbitrage portfolio, if P (x) = 0 and Axt > 0. Let us consider a simple numerical example in order to illustrate what is meant by weak and strong arbitrage. Suppose that A= 1 1 1 0 0 1 1 2 ,
the bid prices are v1 = 1, v2 = 0.7, v3 = 0.4 and v4 = 2.6, and the ask prices are c1 = 1, c2 = 0.8, c3 = 0.5 and c4 = 3. Then, x = (1, 1, 1, 0) is an arbitrage portfolio of the second type because its current price is 0.1 and its nal payos 0 are . The portfolio x = (1.1, 1, 1, 0) is also a strong arbitrage portfolio 0 0.1 because its current price and nal payos are 0 and respectively. Finally, 0.1 x = (0, 2, 2, 1) is a weak arbitrage portfolio with current price equal to 0 and future 1 payos equal to . 0 Previous literature has characterized the absence of arbitrage by the existence of state prices or discount factors (see for instance Chamberlain and Rosthchild (1983), Ingersoll (1987) or Hansen and Richard (1987)). The following result is a minor extension that incorporates the bid-ask spread and may be easily proved by readapting classical proofs (see also Jouini and Kallal (1995)). Theorem 1 There are no arbitrage opportunities if and only if there exists a vector d = (d1 , d2 , . . . , dk ) of discount factors such that di > 0, i = 1, 2, . . . , k and
k
vj
i=1
aij di i cj
for j = 1, 2, . . . , n.
(1)
There are no arbitrage opportunities of the second type if and only if there exists a vector d = (d1 , d2 , . . . , dk ) of discount factors such that di 0, i = 1, 2, . . . , k, and (1) holds.8
7 In what follows, Axt 0 denotes that all the elements in this matrix are larger than or equal to 0. Analogously, Axt >> 0 denotes that all the elements are larger than 0, and Axt > 0 denotes that elements are larger than or equal to 0, but at least one element is strictly positive. Similar notations will appear in similar cases. 8 An analogous result holds if the probability measure is not specied. In such a case, the discount factors di > 0 ( 0) must verify vj k i=1 aij di cj . The proof is trivial since one can dene di = di i .

k
187
Let us remark that (1) leads to 1/R =
i=1
di i . If we set i = 1, 2, . . . , k, (2)
i = R di i , then i 0 and
k i=1
i = 1, and therefore, = (1 , 2 , . . . k ) can be considered as a 1 E (Sj ) cj R
probability measure. Furthermore, (1) leads to vj (3)
for j = 1, 2, . . . , n, being E (Sj ) the expected value of Sj at t1 computed with the probability measure instead of . This is the reason why is called a Risk Neutral Probability Measure, and Theorem 1 shows that its existence (and positiveness) is the necessary and sucient condition to guarantee the absence of arbitrage of the second type (of any kind). The latter theorem provides a very well known and important condition to ensure the absence of arbitrage. Nevertheless, if the arbitrage occurs, it will be interesting to measure, in monetary terms, the degree of arbitrage. This measurement will be useful to improve bid or ask real prices for PCS option contracts, and will also permit to analyze other type of market imperfections. For instance, owing to transaction costs, the presence of arbitrage could be only apparent but not real. The following result summarizes some properties of the measures developed by Balb as and Mu noz (1998). Theorem 2 Suppose that the set X of arbitrage strategies of the second type is non void. Then, problems max
xX
P (x) V (x)
and
max
xX
P (x) C (x) + P (x)
achieve an optimal value at the same portfolio x . The ratios in the theorem above represent relative arbitrage prots (P (x)) with respect to the price of the sold assets (V (x)) or the total volume of trade (C (x) + V (x), price of the purchased and sold assets without considering a negative sign for the sold assets). For instance, if we consider the strong arbitrage portfolio x = (1, 1, 1, 0) introduced in the previous numerical example, the price of the sold assets is V (x ) = 1.1, the price of the purchased assets is C (x ) = 1 , the total volume of trade is V (x ) + C (x ) = 2.1 and the price of x is P (x ) = C (x ) V (x ) = 0.1 . Consequently, the relative arbitrage prots with respect to V (x ) and C (x ) + V (x ) are 0.1/1.1 and 0.1/2.1. Consider the portfolio x whose existence is guaranteed by Theorem 2. The disagreement measures m and l are dened by m= P (x ) V (x ) and l = P (x ) , C (x ) + V (x )
188
or zero if no arbitrage opportunities of the second type exist. Measures m and l vanish if and only if there are no arbitrage opportunities of the second type. When the arbitrage occurs m (l) yields available relative arbitrage gains with respect to the value of the sold (interchanged) assets. The inequalities 0 l m 1 may be proved, and the level of violation of the arbitrage absence grows up as the measures move from m z 0 to 1. The relationship l = 2 z 2 m holds, and thus since [0, 1] z [0, 1] is an increasing one to one function, both measures provide equivalent information. Further details may be found in Balb as and Mu noz (1998) or Balb as et al. (1998).9 Let us remark that Theorems 1 and 2 permit to analyze and detect arbitrage portfolios without previously specifying the exact nature of the strategy to be used. For instance, many nancial papers empirically test the existence of violations for the usual put-call parity or the relationship between spot and future prices. This type of test can not be implemented when analyzing PCS options, in which case all the available securities must be simultaneously considered and the market globally tested. Let us turn now to hedging strategies and arbitrage portfolios of the rst type. Assume that the model does not allow arbitrage portfolios of the second type. Then, R is the highest return than can be guaranteed. However, an investor may be interested in a hedging portfolio whose guaranteed return is very close to R but provides larger returns in some states of the world. Let us x a concrete security Sj0 , and consider the usual way of hedging the purchase of this security, i.e., solve the problem10
xj0 =1 , Axt 1
min
P (x)
(4)
If the solution is attained at x, the absence of strong arbitrage ensures that P (x) > 0, and 1/P (x) is the optimal (maximum) guaranteed return if a unit of Sj0 is bought.11 Notice that the solution x of (4) dominates the riskless security and, consequently, x is really a hedged portfolio. Moreover, P (x) 1/R because, otherwise, investors could implement strong arbitrage (against our assumptions) by selling the riskless asset and buying x. Furthermore, if P (x) = 1/R then either x replicates the riskless asset or the previous strategy is weak arbitrage. Thus, there are arbitrage portfolios of the rst type such that xj0 = 1 if R = 1/P (x) and Axt > 1. An analogous analysis may be done to hedge the sale of Sj0 . Just write xj0 = 1 in (4) instead of xj0 = 1. Obviously, not only hedging portfolios, but also arbitrage of the rst type can be detected by computing all the hedging portfolios when j0 moves from 1 to n.
9 Once again, these results are also veried in a model where is not specied. Another procedure, useful to detect arbitrage portfolios, may be found in Garman (1976). 10 Recall that Axt 1 means that all the elements in the column matrix Axt are larger than or equal to 1. 11 Clearly, the return 1/P (x) is maximum if and only if the price P (x) is minimum.
189
The last part of this synopsis focuses on individual portfolio selection and variance aversion. Assume that there are no arbitrage opportunities (of any sort) in the model. Then, Theorem 1 shows that the arbitrage absence still holds for some concrete linear pricing rule such that vj j cj , j = 1, 2, . . . , n, being j the price of Sj provided by .12 Besides, may be considered as a positive real valued linear operator over the space span(A), span of the columns of A.13 Then, the Riesz Representation Theorem of linear operators in Hilbert spaces allows us to establish the following result (see Chamberlain and Rosthchild (1983)). Theorem 3 There exists a unique discount factor d such that
k
aij di i = j
i=1
for j = 1, 2, . . . , n, and dt belongs to span(A). We will assume that dt is not the payo of a riskless asset. This hypothesis is not restrictive (it only arms that the market is not risk-neutral and, consequently, = ) and will always hold in our empirical test. In order to achieve an easier notation, denote by Sj the j th -column of A, j = 1, 2, . . . , n, and let us identify each feasible portfolio x with its nal payo Axt = y k span(A). Denote by (Axt ) = (y ) = i=1 yi di i its current price provided by . k In particular, (dt ) = i=1 d2 i i > 0. Dene by R(y ) = y/ (y ) the return (provided by ) of any y span(A) such that (y ) > 0, and consider its expected value E (R(y )) and standard deviation (R(y )).14 Then, the statement below, whose proof is a consequence of the Projection Lemma of Hilbert spaces (see for instance Due (1988)) provides the optimal portfolios in a variance-averse model. Theorem 4 For any y span(A) such that (y ) > 0, there exists a linear combination of dt and the riskless asset, S1 + dt , such that (i) i) 0 (ii) (y ) = (S1 + dt ) (iii) E (R(y )) = E (R(S1 + dt )) (iv) (R(y )) (R(S1 + dt ))
for instance, the linear pricing rule provided by the second term of (3). the space of k 1 column matrices that can be obtained by linear combinations of the columns of A. 14 Recall that y and R(y ) may be considered random variables and, therefore, E (R(y )) =
13 I.e., 1 (y ) k i=1 12 Take,
yi i and (R(y )) =
1 (y )
k i=1 [yi
(y )E (R(y ))]2 i
1/2
190
Hence, for a desired expected return, the minimum variance is attained by selling the portfolio x such that Axt = dt and investing the price of x, along with the own capital, in the riskless asset.15 Theorem 4 will play a crucial role in our analysis because it allows to compute the minimum variance portfolios without deriving correlations within the set of available PCS options. In fact, these portfolios are given by the vector d of discount factors and, as pointed out by (2), d is the density between a risk neutral measure , and the initial probability measure .
Markets and Data
CBOTs PCS Catastrophe options are standardized contracts based on PCS indices that track the insured losses resulting from catastrophic events that occur in a given area and a risk period, as estimated by PCS. When PCS estimates that a natural or man-made event within the US is likely to cause more than $25 million in total insured property losses and determines that such eect is likely to aect a signicant number of policyholders and property/casualty insurance companies, PCS identies the event as a catastrophe and assigns it a catastrophe serial number. PCS provides nine loss indices daily to the CBOT: a national index, ve regional indices, and three state indices (National, Eastern, Northeastern, Southeastern, Midwestern, Western, Florida, Texas and California loss indices). Each PCS loss index represents the sum of current PCS estimates for insured catastrophic losses in the area and loss period covered, divided by $100 million, and rounded to the nearest rst decimal point. The loss period is the time during which a catastrophic event must occur in order for the resulting losses to be included in a particular index. Most PCS indices have quarterly loss periods, some of them (California and Western) have annual loss periods, and one of them (National) has both quarterly and annual risk periods. Following the loss period, there exists a development period (twelve months) during which PCS continues estimating and reestimating losses for catastrophes occurring during the loss period. The development period estimates aect PCS indices and determine the nal settlement value of the indices. Catastrophe options are available for trading till the end of the development period. They are European and cash-settled (each point equals $200 cash value). They can be traded as either small-cap or large-cap contracts. These caps limit the amount of losses that are included under each contract: insured losses from $0 to $20 billion for the small contracts and losses from $20 to $50 billion for large contracts.
15 Under the usual CAPM assumptions, the considered securities are stocks and, under the suitable hypotheses on their stochastic behavior, the portfolio x is composed of a long position in the riskless asset and short positions in the stocks. Since the coecient must be negative, the Market Portfolio consists of stocks in a long position.
191
In practice, traders prefer negotiating call spreads and so further limiting their associated payos. Sophisticated combinations traded as a package, that include several expirations dates and indices, are also available.16 Catastrophe options bid and ask quotes and the current value of the indices are daily provided by the CBOT. Premiums are quoted in (index) points and tenths of a point (each point equals $200). Strikes values are listed in integral multiples of ve points. Our empirical work studies two periods: from February 25 to April 20, 1998 and from June 23 to July 30, 1998. The quotes used in the empirical analysis were provided by the CBOT and correspond to synchronized bid and ask quotes posted at the end of each day. For each of the days considered, we have also included a risk-free asset. Its prices were obtained from the coupon-only strips quotes reported by The Wall Street Journal.17 Treasury-bills could have been used instead but strips maturities were much closer to the options expiration dates.18 In order to learn about the distributional properties of the catastrophe waiting times and their associated amount of insured losses, our data also includes a 25-year (1973-1997) catastrophe record provided by PCS. This record included all catastrophes occurred in each state with indication of its serial number, begin and end dates, causes and PCSs estimates of insured losses. The monetary value of losses was converted into 1997 dollars by using the Producer Price Index reported by the Bureau of Labor Statistics (US Department of Labor). We restricted our observations of the value of the insured losses to the sample period 1990 1997 for several reasons that will be analyzed in the fourth section (Subsection 4.3).
Empirical Research: Methodology and Results
From now onwards, we will rst specify the methodology employed in the empirical analysis and then we will present the obtained results. Our analysis only targets those derivatives with a single underlying index and a unique expiration date. We group those derivatives with the same expiration date and the same underlying index that are available for trading. For a given we require day a minimum of four assets in each set. For the rst period this ltering left us with derivatives associated to the following indices: National Annual-98 (36 valid days), California Annual-98 (36 days), Eastern September-98 (12 days) and Southeastern September-98 (10 days). For the second period we have the National Annual-98 (27 valid days), Eastern September-98 (10) and Southeastern September-98 (10).19 Table
16 Henceforth, for short, we will merely say derivative or option to refer to a single option or a package of ones. 17 Exact values were obtained through linear interpolation of midpoints of the bid-ask prices by using the closest maturities to the option expiration dates. 18 Some of our computations were also implemented with T-bills returns. Our main results remained unchanged. 19 The expiration dates for the 98 annual contracts and the September 98 contracts are 12/31/99 and 09/30/99, respectively.
192
1 summarizes the number of relevant derivatives satisfying the above criteria and the number of their quotes available in the two periods.20 This constitutes our whole sample for Subsections 4.1 and 4.2.
Table 1: Overview of the PCS Catastrophe Options Sample. This table describes our sample of derivatives. For each specic index, we require a minimum of four tradable securities in order to include a given day in the analysis. The rst column gives cumulated gures corresponding to the entire periods, and the subsequent ones summarize the daily number of derivatives and quotes.
In Subsections 4.3 and 4.4, a large amount of data on waiting times and their associated amount of losses is required in order to perform reliable simulations and, therefore, the characteristics of our historical data set compel us to exclusively concentrate on the National Annual-98 Index.
4.1
Pricing by strong arbitrage methods
The price of the PCS derivatives will be analyzed once a day during each tested period. Hence, under the notations of the second section, the date t0 will always be the corresponding day, while securities S2 , S3 , . . . , Sn will be PCS option contracts (call or put spreads, butteries etc.), available this day, and with the same underlying index W and expiration date t1 .21 Their bid and ask prices are perfectly synchronized and provided by CBOT. S1 will be a pure discount bond available at t0 and such that its maturity is as close to t1 as possible. Of course, all the data and parameters (dates t0 or t1 , securities, prices etc.) depend on the concrete day under revision. Let W1 be the current value of the index and denote by W2 , . . . , Wr the striking prices, corresponding to Sj , j = 2, 3, . . . , n. The future state of the world will be determined by the nal (at t1 ) value of W (any real number greater than W1 and rounded to the nearest rst decimal point), and the matrix A of nal payos may be
20 We have also detected that it was possible to synthetically produce some other Eastern September-98 options based on their corresponding Northeastern and Southeastern options for some days in our sample. To be consistent, as the latter two negotiated independently, we decided not to include the synthesized Eastern options as other derivatives available for trading in our sample. 21 i.e., the underlying index and the loss and development periods coincide for all the considered derivatives.
193
easily computed. In fact, all the elements in its rst column (payos of the riskless asset) are equal to 1, and the rest of the columns are given by the usual dierences between W and Wi , i = 1, 2, . . . , r. It is obvious that, for an arbitrary strategy x, its nal payos verify the constraints Axt 0, Axt > 0 or Axt >> 0 if and only if these constraints are fullled when the settlement value of W belongs to the set {W1 , W2 , . . . , Wr }.22 So, the absence or existence of arbitrage may be tested under the assumption that these elements are the only possible states of the world. Furthermore, this simplication neither modies the value of the disagreement measures m and l, nor aects the results when hedging or weak arbitrage portfolios are being computed. Consequently, we will permit W to attain all the feasible values only when testing portfolio choice models. Once the available derivatives, their real bid-ask prices provided by CBOT, the r states of the world and the matrix A are xed, we can compute the measure m and the portfolio x introduced in Theorem 2. If m = 0, there are arbitrage opportunities. This case has never appeared along the tested periods. Next, we x an arbitrary option Sj0 , j0 = 2, 3, . . . , n, and consider an agent who can buy this derivative by paying the price vj0 .23 If the new values for m and x show the presence of arbitrage and the prots represented by m are high enough to overcome the market frictions, it may be concluded that the market allows us to price Sj0 by arbitrage methods. An agent can oer a new bid price vj0 (such that vj0 vj0 cj0 and, therefore, better than the current bid price vj0 ) without any kind of risk. The position will be hedged by implementing the arbitrage portfolio x if a new investor accepts the new bid price. Analogously, one can analyze if the ask price cj0 may be improved. Just consider that cj0 equals both, the bid and the ask price, and compute the new solutions for m and x .24 The above procedure can be applied for all the available securities (for j0 = 2, 3, . . . , n) in order to test how often the market allows us to price by strong arbitrage methods. The empirical results are conned to Table 2.
22 Strategies 1 and 2 below illustrate this fact. Notice that the same property holds if one writes 1 instead of 0 in the right side of above inequalities. 23 i.e., the bid-ask spread vanishes for the j th security. The rest of the prices is not modied. Of 0 course, this analysis has not been implemented in cases where vj0 = 0. 24 This analysis has not been implemented in cases where c j0 = .
194
Table 2: Second Type Arbitrage Opportunities. The bid-ask spread has been removed for each option at a time and its price was set equal to alternatively the bid quote and the ask quote when possible. The table summarizes the resulting arbitrage opportunities of the second type and their associated optimal gains as quantied by the m measure. First two columns show the no. of days for which there are some arbitrage opportunities. Subsequent columns give statistics computed over those days with arbitrage opportunities (m = 0).
This particular type of arbitrage is quite often detected. It should be noted that these results seem to reveal that the price setting process might be improved. Hedging (with arbitrage portfolios) would be feasible. The arbitrage prots are quite large and this should be used by investors to oer new prices. For the National Annual-98 index, arbitrage opportunities appear in seven out of 36 days for the rst period (see Table 2) and in up to three dierent cases. The maximum value of m is equal to .1398 (this corresponds to an l value of .0752). For the second period and the same index, arbitrage is feasible every day for up to three dierent available premium quotes. This time the maximum value of m is .375 (l = .2308). This reects a riskless benet that amounts to a 37.5% of the total monetary value of the sold assets (or a 23.08% of the total monetary value of all traded assets). With respect to other indices, California and Eastern include a unique position that allows for arbitrage hedging in the rst period (the maximum value of m is .2105 and .0698, respectively) and the same may be said about the Eastern index in the second period (maximum m = .1023). No misspricings were found for the Southeastern index. In any event, the number of available positions were notably low for these last three indices (see Table 1). Thus, note that for a signicant percentage of days, agents could analyze the bid-ask spread and oer more ecient prices in some cases without assuming any kind of risk. This fact should lead to smaller spreads. For illustration purposes, we show in Table 3 the optimal (maximum m value) second type arbitrage portfolio detected on date 07/24/98 for the Eastern September98 Index (Strategy 1). Figure 1 plots the portfolio payos pattern for dierent levels of the nal index value.
195
Table 3: Optimal Second Type Arbitrage on July 24, 1998 for the Eastern September-98 Index. This table shows the optimal second type arbitrage opportunity corresponding to date 7/24/1998 and the Eastern September-98 Index. CA 20 40 and PU 50 stand for a call spread and a put with relevant exercise prices as indicated, respectively. All derivatives available for trading together with their bid and ask prices are reported. The price for a zero-coupon bond (risk-less asset) with a maturity value of one point is also given. All prices are expressed in points with a value of $200. Bold face is used to indicate those assets involved in the detected arbitrage portfolio, and the bought or sold units are given in parenthesis beside the aected price (a negative sign indicates a sale). The last two rows give the portfolio price and the m value. This arbitrage was detected when the bid quote was set equal to the ask for the PU 50 derivative. The same arbitrage strategy was detected for 9 days. Asset Bond CA 20 40 CA 40 60 CA 150 200 PU 50 Portfolio Price Value of m Bid 0.93867188 3 (4) 2.5 (1) 2 30 10.6328 0.1017 Ask 0.92867188 (100) n.a. 3.5 n.a. 45 (2)
Figure 1. Final Payos of Strategy 1
Strategy 1 may be interpreted as follows: The bonds can be purchased at 93.87 and calls can be sold at the bid prices so the net portfolio costs the owner 79.37 now (t0 ).
196
Later (t1 ) the owner cashes the bonds and pays the call owners. It is always adequate to cover the liability for selling the put portfolio. Therefore, the owner can ask a price of 79.37 for a portfolio of two puts (39.68 per put) and still have a non negative payo without paying anything now. This is a weak arbitrage and getting more (45) per put contract provides the owner with an up front arbitrage prot. We can also check the value of m and l. Since the price of the purchased assets (bonds) is 93.87 and the price of the sold assets is (assuming that both puts are sold at 45 per put) 104.5, then m= 104.5 93.87 104.5 93.87 = 0.101722 and l = = 0.05358 104.5 104.5 + 93.87
The strategy shows that an ask price could have been reduced from 45 to 39.68 index points at least. If the ask price is lower than 39.68 the strategy does not work, but a new strategy might appear. Thus, if we are interested in the lowest possible ask price (actually, it is 39.68 is this case) we have to apply the general procedure provided by Theorem 2 instead of analyzing the previous strategy. Figure 1 also shows that the nal global payo of the strategy may be easily determined by means of the nal payos obtained when the index nal value W belongs to the set of striking prices {0, 20, 40, 50, 60, 150, 200} (connect with segments the corresponding points). Moreover, the nal payo is never lower than zero if and only if the property holds when W belongs to this set. This is the reason why we can simplify the set of states of the world. Let us leave Strategy 1 and go back to the general case. Since the oering of better quotes is sometimes feasible, it is interesting to measure the highest adjustments that could have been implemented in the bid (an increase) and ask (a decrease) premiums. To this aim we followed the next algorithm. Focusing on one of the quotes which gave rise to the above arbitrage opportunities, we appropriately moved up or down the implied quote only for a tick and then searched again for arbitrage opportunities. This process was iterated until reaching a total removal of the riskless arbitrage hedging. We carried over this algorithm for each price independently. The corresponding price and spread nal adjustments are given in Table 4. For the National Annual derivatives, our results show price changes that range from 2.5% to 100% along with spread reductions ranging from 5% to 56.52%. Signicant adjustments were also possible for the other indices. It should be mentioned that some renements of this procedure point out that a more adjusted set of prices may still be reached. If the above algorithm is not carried out ceteris paribus, that is, if we keep the nal adjusted premium before moving to the next one, we nd that new arbitrage hedging strategies could appear, thereby leading to possible further reductions in the spread.25
25 As
the ordering might be relevant in this case, we do not report our results.
197
In short, although the market quotes studied here do not permit to gain arbitrage prots to anyone obliged to incur in the cost of the bid-ask spread, better relative pricing by (strong) arbitrage methods is possible in the PCS options market for the studied periods. This is important for two reasons. First, this information is useful to traders since the whole arbitrage portfolio may be shown. Second, frictionless pricing theory suggests that competition among traders should lead to a situation with correct relative prices, i.e., a set of quotes that should exhaust any exploitable possibility of making money without any sort of risk.
Table 4. Bid-Ask Spread Reduction This table shows the bid-ask spread reduction that can be implemented for each derivative in order to remove the second type arbitrage strategies previously detected. Those assets involved are reported on the left side; additionally, in parenthesis we indicate whether changes correspond to the ask quote (a), bid quote (b) or both bid and ask quotes (b/a). CA 40 60 stands for a call spread with exercise prices 40 and 60. PS 40 60 stands for a put spread with relevant exercise prices as indicated. The price change and the spread reduction are both given in ticks (i.e. $20 or one tenth of a point) and in percentage terms. For each asset some descriptive statistics have been computed over those days and quotes for which changes were possible.
Some factors related to the implementation of the detected arbitrage strategies and not considered so far might explain those relative misspricings. One of them is the existence of transaction costs. In relation to this, it should be kept in mind that measures m and l represent relative arbitrage prots, and the levels achieved by these measures are high enough to reect gains after discounting transaction costs.
198
Another factor is due to the number of units of each asset needed to implement some arbitrage strategies. This number might not be available for trading. We do not have any piece of information about the volume associated to the quotes gathered by the CBOT constituting our sample. Nevertheless, a comparison between the volume corresponding to the transactions made during our sample periods and the number of derivatives needed to implement the detected arbitrage strategies lead us to think that in most cases this does not seem to be a real problem. Some nal reasons might be related to margin rules or illiquidity but, anyway, the empirical results seem to be signicant for researchers and traders.
4.2
Weak arbitrage and hedging portfolios
Suppose that for a xed j0 {2, 3, . . . , n} it is still obtained m = 0 after assuming that vj0 is the ask price (respectively, cj0 is the bid price). Then, problem (4) (respectively, after the modication xj0 = 1) has been solved in order to analyze how the real bid price vj0 (ask price cj0 ) can be improved. This case will hold when the achieved solution guarantees a return R or very close to R. If so, investors can oer a new bid (ask) vj0 vj0 cj0 and the solution of (4) provides a portfolio that will almost guarantee the riskless return R if a new agent accepts the new price. Furthermore, this strategy could lead to great returns in some states of the world, and thus, it could be interesting for many investors. Following this procedure hedging strategies were obtained and they were grouped into rst type arbitrage opportunities (with a guaranteed return equal to R and payos greater than one in at least one state of the world) and other optimal hedging strategies. Both, the guaranteed net return and the maximum possible net return, were computed for each detected position available for hedging and mean values are given in Table 5. We also report the corresponding mean values after substractioning the return guaranteed by the risk-free asset. For some states of the world, extraordinarily large returns might be obtained (e.g., there were rst type arbitrage opportunities that involved selling one call spread 80/100 and gave rise to a possible net return of 2, 215.63). The minimum net return equals that of the risk-free asset for almost all cases. Thus, an important part of the available positions might have been hedged by means of weak (and strong) arbitrage or other optimal strategies leading in many cases to possible returns exceeding largely that of the risk-free asset. Note that this has been feasible even in a situation in which the underlying index is not tradable, and put derivatives are seldom available. Considerations akin to the ones pointed out at the end of the previous subsection, regarding the proper interpretation of these results, also apply here. Again, for illustration purposes, Table 6 shows the optimal hedging portfolio (weak arbitrage) on date 07/01/98 for the National Annual-98 Index (Strategy 2). Figure 2 plots the portfolio payos pattern for dierent levels of the nal index value.
199
Table 5. Optimal Hedging Portfolios: CA 40 60 stands for a call spread with exercise prices 40 and 60, and similarly for the other possible exercise prices. CB denotes a buttery call spread with relevant exercise prices as indicated. For each derivative, the number of days for which a hedging strategy was available is given and in parentheses it is indicated whether the hedged derivative is bought (b) or sold (s) at the optimal hedging portfolios. Guaranteed and maximum returns in average terms along with the corresponding excesses over the risk free rate, R, are also given.
Table 6. Optimal Hedging Portfolios: (Weak Arbitrage) on July 1, 1998 for the National Annual-98 Index. This table shows the optimal hedging portfolio detected on date 7/1/1998 for the National Annual-98 Index. CA 30 50 stands for a call spread with exercise prices 30 and 50, and similarly for the other possible exercise prices. CB denotes a buttery call spread with relevant exercise prices as indicated. The reported portfolio is a weak arbitrage portfolio which permits to hedge the purchase of the CA 60 80 derivative. All derivatives available for trading together with their bid and ask prices are reported. The price for a zerocoupon bond (risk-less asset) with a maturity value of one point is also given. All prices are expressed in points with a value of $200. Bold face is used to indicate those assets involved in the detected hedging portfolio, and the bought and sold units are given in parenthesis (a negative sign indicates a sale). The last row gives the portfolio price. The same portfolio was available for 5 consecutive days.
200
Figure 2. Final Payos of Strategy 2
Thus, an important part of the available positions might have been hedged by means of weak (and strong) arbitrage or other optimal strategies leading in many cases to possible returns exceeding largely that of the risk-free asset. Note that this has been feasible even in a situation in which the underlying index is not tradable, and put derivatives are seldom available. Considerations akin to the ones pointed out at the end of the previous subsection, regarding the proper interpretation of these results, also apply here. The interpretation of Strategy 2 may be as follows: The purchase of ve butteries and the sale of one CA 40 60 provides an income equal to 2.5 index points, the bid price for the CA 60 80. Suppose that this call is bought at 2.5. Then the payos associated to the sold assets are dominated (strictly in some states of the world) by the payos associated to the purchased ones. So, the whole portfolio price is zero but provides (strictly in some states) positive payos (weak arbitrage). Besides, if we add the bond to this portfolio, the nal payos dominate the bond payos (strictly in some states), but we have to pay now just the bond price.
4.3
Evaluating the index real distribution and the risk-neutral probability measure
To assess catastrophe options from an actuarial point of view, an analysis of the distribution of possible future values of the underlying indices is required. There are essentially two approaches to form such probability assessments. One is to use computer simulation of scenarios based on a vast amount of meteorological, seismological and economic information. The other relies on statistical modelling based on historical data. This subsection is partially devoted to the analysis of the distributional properties of the National Annual Index to be used in the rest of the paper and for this matter we concentrate on the statistical analysis of historical catastrophe data. We develop a nonparametric simulation procedure in order to obtain the expected nal payos.
201
This method does not require any distributional assumption, instead, it lets the data talk. Furthermore, it relies on the empirical distribution of waiting times and their associated losses thereby avoiding the traditional shortage of data that is faced when using exclusively the empirical distribution of the nal historical values of the index. In addition, from a nancial perspective, once we have exhausted the arbitrage and hedging pricing approaches, risk considerations come into place and therefore the use of probability assessments is also necessary. As stated by Theorems 3 and 4, the underlying index real distribution is required in order to solve minimum variance problems. In this case, variables and parameters (dates t0 or t1 , the riskless return, securities, prices etc.) are introduced by the procedures already mentioned, but the set of states of the world must be enlarged. Now this set must incorporate all the index feasible nal (at the expiration date t1 ) values and not only the derivatives striking prices. Fix a day t0 and, consequently, let us assume that all the parameters are xed. To determine the nal distribution of the underlying index W , we proceed as follows: First of all, we consider the empirical distribution of random variables T , time between two consecutive catastrophes, and L, losses caused by a specic catastrophe. It is assumed that T and L can achieve several values with probabilities according to the empirical frequencies obtained from the real data described in the third secs1 tion. Later, we simulate several values T1 , T2 , . . . , Ts of T till i=1 Ti t1 t0 and s i=1 Ti > t1 t0 , and s 1 values L1 , L2 , . . . , Ls1 of L. Each specic result Li is incorporated if and only if Li 25 million of dollars, and we take Li = 0 otherwise. If W0 is the index value at t0 , the simulation process provides the total value s1 W = W0 + i=1 Li where each Li has been previously translated into index points. The whole simulation process is repeated a high number of times in order to attain a numerical distribution of W . The risk neutral probabilities, dened in (2), have been determined too. We have followed the general method proposed by Hansen and Jagannathan (1997).26 Hence, x a day t0 and all the parameters of the problem. Suppose that the simulation process has already been implemented and, therefore, the (real) probability measure = (1 , 2 , . . . , k ) is known. Then, the (risk-neutral) measure is obtained by k k minimizing i=1 (i i )2 among the row-matrixes such that i=1 i = 1, i 0, i = 1, 2, . . . , k, and (3) holds. Once the measures and have been determined, we can give two theoretical prices per security. The rst one, E (Sj )/R, j = 1, 2, . . . , n, is the Pure Premium, usual in Actuarial Sciences. The second (see (3)), E (Sj )/R, is obtained from a nancial point of view by considering real quotes and applying the most important topics on static asset pricing models.
26 The Hansen and Jagannathan (1997) method extends the procedure provided by Rubinstein (1994) and by Jackwerth and Rubinstein (1996) to study the eect of the volatility smile on the risk neutral probability measure.
202
The procedure above has been implemented for all the possible values of the current date t0 (every day of our sample periods). We restrict our observations of the value of insured losses to the sample period 1990 1997 for several reasons. Figure 3 shows the annual number of catastrophes and quantity of their associated losses since 1973. As the gure suggests, while the annual number of catastrophes remains reasonably stable along the period, there seems to be a general positive trend, and a structural change in the behavior of the value of insured losses associated with each catastrophe since hurricane Hugo hit US in 1989. However, these patterns may be only illusory. Insured losses are aected by multiple variables ignored so far such as population growth, development, changes in insurance coverage, number of premiums, and ination. When a sucient long period of time is considered, the adjustments of the loss series for these variables tend to homogenize it, approximating past losses to more recent ones.27 Thus, a reasonable approximation to the adjusted series can be obtained by concentrating on recent (unadjusted) losses.28 Moreover, any adjustment of the series, which become necessary when a long period of losses is considered, might be methodologically questionable.
Figure 3. Number of Catastrophes and Total Amount of Loss (1973-1997)
With regard to the number of catastrophes, the whole sample since 1973 has been used. This is expected to surmount the diculty of getting good estimates based on small samples for the probability of low frequency events such as catastrophes. The simulation process described above has been implemented in order to estimate the probability distributions of the nal (end of 1999) value of the National Annual-98
27 An illustration of this eect may be found in Litzenberger et al. (1996). These authors adjust historical loss ratios for both, the increase in population and the market penetration of catastrophe coverage. 28 Anyway, the results of the simulation process only show slight modications if adjusted data and long periods are considered.
203
index for every day in both periods. A number of 100, 000 replications has been used for each day. Results are given in Figure 4. As the gures show, the probability mass is mainly accumulated around the index levels lying approximately in the intervals 20 -100 and 160-240. The distributions are bimodal or even trimodal because (probably) the number of catastrophes must be entire. Note also that as days go by and the nal date approaches, the probability mass tends to concentrate due to the accompanying uncertainty reduction.
Figure 4. Probability distributions of the index nal value. First period (top) and second period (bottom)
204
Once the required distributions have been estimated, the next step is to compute the discounted expected options payos (pure premiums). These are given in Table 7. Notice that the call spreads 100/120 and 120/140 have almost identical discounted expected payos. This is due to the general lack of probability mass in the interval 100-140. In general, pure premiums lie around options real quotes.
Table 7. Theoretical Prices for the National Annual-98 Index Derivatives. This table shows the average theoretical prices corresponding to the National Annual-98 derivatives available for trading on any day during the sample periods. CA 40 60 stands for a call spread with exercise prices 40 and 60, and similarly for the other possible exercise prices. CB denotes a buttery call spread with relevant exercise prices as indicated. The rst two columns give the mean pure (actuarial) premiums and the mean risk-neutral (nancial) prices. The third column displays the absolute value of the dierence between the previous two columns and the last two columns show the mean real bid and ask quotes for comparison purposes. All gures are given in points, each point with a value of two hundred dollars. We also report the mean Euclidean distance between the measures and (standard deviation in parenthesis) in the last row of each panel.
205
We now use this result to infer some conclusions on the individual prices of these options, based on our estimated future probabilities. First, mean midpoints of the bid and ask quotes being around mean pure premiums seem to indicate that, in average terms, transactions in this market could have been made at reasonable prices, close to the actuarial fair ones during the sample periods. These are good news for those participating in the market that hedge their catastrophe insurance risks: this market seems to be an attractive alternative to traditional reinsurance, for instance, as it oers reinsurance at fair prices. However, those willing to participate in the (re)insurance market by selling options seeking for attractive returns over the risk-free rate might nd it dicult to get them, at least when trading with individual options. Second, for our sample periods, conclusions inferred from spread midpoints may substantially change when the real bid-ask spread is taken into account . On the one hand, in general, those hedgers able to buy at bid prices will nd this market more attractive than the usual reinsurance (for most options mean bid prices are lower than the mean actuarial ones) but things might turn the other way round for hedgers that buy at the ask. On the other hand, keeping aside, at the moment, risk considerations, investors seeking for high returns should try to buy close to the bid or sell close to the ask. Third, investors searching for new investment/nancing opportunities also attend to risk considerations in their decisions. Analyzing risk and return on an option by option basis hardly makes sense as it is in a portfolio choice context where risk/return opportunities are fully understood (it is in this context where risk-pooling benets, for instance, come into place). These are the topics considered in the next subsection. Let us now analyze the risk-neutral probabilities. We solve the minimization program that gives their possible values and we use them to obtain the corresponding theoretical risk-neutral prices. These are also given in Table 7.29 We also report the average value of the objective function for both periods, and the resulting discrepancies between pure (actuarial) premiums and prices calculated with the risk-neutral probabilities. As these gures show, it might well be concluded that real prices, as summarized by the linear pricing rule extracted from them, are reasonably close to their fair value. Lane and Movchan (1999) also consider real market mid-year 1998 prices and compute a compatible risk neutral probability measure.30 They follow a dierent and interesting approach because they do not use any real distribution . Instead, they impose the constraints (3) and minimize the dierences between the prices of real transactions and the theoretical prices provided by the second term of (3). This methodology could show some advantages with respect to the one we followed because
29 Risk-neutral prices certainly verify the bid and ask quote restrictions day by day, even though the reported mean risk-neutral prices do not have to lie inside the mean bid-ask spread; the bid/ask quote may not be present whenever these prices are computed (we use the term risk-neutral prices as has become usual in nance). 30 We thank an anonymous referee for pointing out this reference.
206
we could commit errors when evaluating (although the authors assume a specic type of distribution for the relevant random variables). However, bearing in mind our purposes, we preferred to obtain from for several reasons. First, one of the present paper main objectives is to give well diversied portfolios and (as pointed out by Theorem 4), consequently, the measure is required. Second, the theoretical prices obtained by Lane and Movchan are very close to the risk-neutral prices provided in Table 7. Third, as will be pointed out in next subsection, the main conclusions concerning well diversied portfolios seem to be more robust if we take as close to as possible.
4.4
Looking for well diversied portfolios
Consider an arbitrary date t0 , the probability measures and obtained for t0 in the previous subsection, and the linear pricing rule such that j = E (Sj )/R, j = 1, 2, . . . , n. Then, the (unique) discount factor d of Theorem 3 may be easily found by means of the following conditions t d span(A)
k i=1
aij di i = j ,
j = 1, 2 , . . . , n
The condition dt span(A) leads to a simple linear system of equations. In fact, consider the matrix B = (A, dt ) (i.e., add the column dt to A) and impose that the ranges of A and B must be identical. According to Theorem 4, the minimum-variance strategies are obtained by selling the portfolio x = ( x1 , x 2 , . . . , x n ) such that Axt = dt . The portfolio x = 31 (0, x 2 , . . . , x n ) will also provide a very useful information. Depending on the sign of n its theoretical price j =2 j x j , we know when variance-averse individuals must sell or buy derivatives in their reinsurance-linked portfolios. As it will be shown later, it turned out that quite often investors in the PCS options market must sell derivatives (the sign is positive).32 For the empirical analysis we rst excluded those redundant derivatives for each day. See Table 8 for a summary of the main results. With regard to the x portfolio, this was mainly composed of bonds (87.98% and 68.36% in average terms for the rst and the second periods, respectively) and its mean (gross) return equals .9460 and .9096. The portfolio x has a positive value for 77.78% of the days in the rst period and 81.48% in the second. Thus, risk-averse investors should view the PCS options market as a very protable source of capital that allows them to nance their investments in other markets.
the portfolio x once the bond has been excluded that we are considering investors willing to participate in the catastrophe insurance market through PCS options in order to benet from their new risk/return opportunities when included in diversied stock and bond portfolios. Of course, insurers, who must hedge their proper liabilities when the nal index value becomes large, follow dierent portfolio criteria.
32 Notice 31 I.e.,
207
Table 8. Well Diversied Portfolios. The rst two rows show the risk/return characteristics of the entire portfolio by providing the expected return and its standard deviation, respectively (returns dened as payos divided by the risk neutral price). The third row gives the weight of x (the derivatives portfolio) over the entire diversied portfolio, while the fourth row shows the price of x in index points. Finally, the number of days in which x has positive price is indicated in the last row.
Table 9. An Example of Two Well Diversied Portfolios. This table shows the asset weights of the derivatives portfolio x for two selected days. Weights were calculated as value invested in the asset divided by the portfolio value (a negative sign indicates a sale). The corresponding risk-neutral prices are also reported. The weights for these derivative portfolios over the entire (bond included) portfolios are 0.29 and 0.63 for panels A and B, respectively.
208
Let us remark that the last conclusion concerning the sign of the price of x seems to be robust with respect to the measure used in the analysis. In fact, when is obtained by minimizing the distance to , we are minimizing the Risk Premium, dierence between the risk-neutral price and the actuarial pure premium. Thus, if we take a new measure for which (3) holds but such that = , the risk premium will probably increase and short positions in PCS options will probably be more interesting for traders. In order to further illustrate the relationship between probabilities and , and the portfolio x payos (appropriately standardized), these variables are plotted at Figure 5 for a representative day of both periods. It is clear that the portfolio nal payo becomes signicantly negative only for states of the world (index nal values) with slight probability.
Figure 5. Standardized Payos of Portfolio x on February 25, 1998 (top) and July 30, 1998 (bottom), along with the corresponding Risk-Neutral Probabilities
209
Concluding Remarks
All along the paper it has been shown that arbitrage arguments, in a static setting, very often allows us to price catastrophe-linked derivatives and reduce their bidask spread. Though the empirical literature concerning the existence of arbitrage in real nancial markets usually focuses on concrete well-known arbitrage portfolios (put-call parity, relationship between spot and future prices, etc.), this procedure does not apply to catastrophe-linked derivatives. We have followed more complex methodologies that are based on the main principles of asset valuation and provide arbitrage portfolios without previously specifying the exact nature of the arbitrage strategy to be used. This seems to be a signicant dierence with respect to other nancial papers. Furthermore, the procedure to detect arbitrage portfolios has been given and some illustrative examples have been presented. Hedging arguments have also been applied and, again, they have shown many possibilities to price these derivatives. The hedging portfolios have been computed by using general procedures too, rather than usual particular methods that only apply in special situations. Concrete examples of hedging portfolios have been given. The Theory of Portfolio Selection also yields suitable strategies to invest. Moreover, if the bid-ask spread is reduced by arbitrage, the real quotes available in the market show very signicant particularities. Linear pricing rules compatible with the quotes usually imply theoretical prices quite close to the actuarial pure premiums. However, even though a well diversied portfolio (in a variance aversion context) is composed of dierent catastrophe-linked derivatives in short and long positions, its price is usually negative (i.e., the total price of the sold derivatives is greater than the total price of the purchased ones) and this capital must be invested in shares and bonds. Catastrophe-linked derivatives can usually be added to a portfolio of other assets to the mean-variance advantage of the portfolio holder. This result emanates from the fact that catastrophe derivatives outcomes are uncorrelated with bonds or stock outcomes. But the conclusion presented in the previous paragraph provides an important additional information. In a well diversied portfolio, the price of the catastrophe derivatives is usually negative. The last comment lead to signicant implications. Insurers can consider the market in order to buy reinsurances and hedge their liabilities. The price paid may be adequate paid with respect to the pure premiums. On the contrary, variance averse investors whose risk does not depend on the indices can use catastrophe-linked derivatives to compose portfolios with negative price that must be invested in other type of assets. This makes the market very attractive because it allows us to appropriately diversify among many investors the risk due to catastrophic events. However, we must notice that these properties hold for linear pricing rules and, therefore, it is important to reduce the real bid-ask spreads observed in the market. As mentioned above, this is possible by arbitrage and hedging arguments.
210
Well diversied portfolios have also been determined by applying a very general method. In fact, the optimal mean-variance strategies are given by densities between risk neutral measures and initial probability measures, rather than by correlations associated with the returns of the available catastrophe-linked derivatives. The applied methodology seems to reveal two interesting advantages. It is useful for traders because practical criteria and strategies to invest are provided. Moreover, it seems to be general enough to apply in many other contexts. Only two properties are needed. The underlying uncertainty must be easily identied (arbitrage and hedging), and the probability space that explains its behavior, along with the risk neutral probability, needs to be determined (variance aversion). Acknowledgments Research supported by DGICYT (Reference Number: PB95-0729-C02-02), Comunidad Aut onoma de Madrid (project on Complex Decision Problems) and Fundaci on Caja de Madrid. This paper is a simplied version of Balb as Longarela and Lucia (1999) and correspond to those preliminary results presented by the authors at Seminario de Matem atica Financiera (organized by MEFF-RV and Universidad Aut onoma de Madrid, November 1998). We would like to thank Vicente Meneu (Universidad de Valencia ) and Cara Eugenia Garc ia (Universidad Carlos III de Madrid ) for very useful comments on the article. We are also very grateful to L.P. Hemond (Chicago Board of Trade) and G. Kerney (Property Claim Service) for providing us several databases. Mr. Kerney also helped us to clarify interesting aspects of the PCS indices. Important information to understand some properties of the PCS options contracts and the quotes data was kindly provided by Michael J. Tomas, from CBOT. The usual caveat applies.
References
[1] Balb as A., P.J. Guerra and M.J. Mu noz (1998). Measuring the Arbitrage Opportunities in an Intertemporal Dynamic Asset Pricing Model. Applied Decision Analysis, Kluwer, Boston, 159172. [2] Balb as A., I.R. Longarela and J. Lucia (1999). How Financial Theory Applies to Catastrophe-Linked Derivatives. An Empirical Test of Several Pricing Models. Journal of Risk and Insurance, 66, no. 4, 551582. [3] Balb as A. and M.J. Mu noz (1998). Measuring the Degree of Fulllment of the Law of One Price. Applications to Financial Market Integration. Investigaciones Econ omicas, 22, no. 2, 153177.
211
[4] Canter M.S., J.B. Cole and R.L. Sandor (1996). Insurance Derivatives: A New Asset Class for the Capital Markets and a New Hedging Tool for the Insurance Industry. Journal of Derivatives, 4, no. 2, 89105. [5] Cummins J.D. and H. Geman (1995). Pricing Catastrophe Insurance Futures and Call Spreads: An Arbitrage Approach. The Journal of Fixed Income, 4, no. 4, 4657. [6] Chamberlain G. and M. Rothschild (1983). Arbitrage, Factor Structure and Mean-Variance Analysis of Large Assets. Econometrica, 51, no. 5, 12811304. [7] DArcy S.P. and V.G. France (1992). Catastrophe Futures: A Better Hedge for Insurers. The Journal of Risk and Insurance, 59, no. 4, 575601. [8] Due D. (1988). Security Markets: Stochastic Models, Academic Press. [9] Garman M.B. (1976). An Algebra for Evaluating Hedge Portfolios. Journal of Financial Economics, 3, no. 4, 403427. [10] Geman H. and M. Yor (1997). Stochastic Time Changes in Catastrophe Option Pricing. Insurance Mathematics and Economics, 21, no. 3, 185194. [11] Hansen L.P. and R. Jagannathan (1997). Assessing Specication Errors in Stochastic Discount Factor Models. The Journal of Finance, 52, no. 2, 567 590. [12] Hansen L.P. and S.F. Richard (1987). The Role of Conditioning Information in Deducing Testable Restrictions Implied by Dynamic Asset Pricing Models. Econometrica, 55, no. 3, 587613. [13] Ingersoll J.E. (1987). Theory of Financial Decision Making, Rowman & Littleeld Publishers Inc. [14] Jackwerth J.C. and M. Rubinstein (1996). Recovering Probability Distributions from Option Prices. The Journal of Finance, 50, no. 6, 16111631. [15] Jaee D.M. and T. Russell (1997). Catastrophe Insurance, Capital Markets and Insurable Risk. The Journal of Risk and Insurance, 64, no. 2, 205230. [16] Jouini E. and H. Kallal (1995). Martingales and Arbitrage in Securities Markets with Transaction Costs. Journal of Economic Theory, 66, no. 1, 178197. [17] Lane M. and O. Movchan (1999). The perfume of the premium II. Derivatives Quarterly, 5, no. 3, 2740. [18] Litzenberger R.H., R.D. Beaglehole and C.E. Reynolds (1996). Assessing Catastrophe Reinsurance-Linked Securities as a New Asset Class. Journal of Portfolio Management, Special Issue, 7686.
212
[19] Prisman E.J. (1986). Valuation of Risky Assets in Arbitrage Free Economies with Frictions. The Journal of Finance, 41, no. 3, 54556. [20] OBrien T. (1997). Hedging Strategies Using Catastrophe Insurance Options. Insurance Mathematics and Economics, 21, no. 2, 153162. [21] Rubinstein M. (1994). Implied Binomial Trees. The Journal of Finance, 69, no. 3, 771818. [22] Tomas M.T. (1998). A Note on Pricing PCS Single-Event Options. Derivatives Quarterly, 4, no. 3, 2328.
Alejandro Balb as Departamento de Econom a de la Empresa Universidad Carlos III C/ Madrid, 126 28903Getafe (Madrid) balbas@emp.uc3m.es Julio Lucia Departament dEconomia Financera i Matem` atica Edici Dptal Oriental (5 e pis) Campus del Tarongers Universitat de Valencia 46071Valencia julio.j.lucia@uv.es
Derivatives as Tradeable Assets

Terry J. Lyons1
Introduction
Consider a market with a heavily traded asset, where a secondary market for vanilla call options has developed, and where the voltility in the price of the fundamental asset has a stochastic uctuations. This paper aims to develop a model and methodology for the joint behviour of the prices of the call option and underlying asset. In consequence, we are able to provide a more systematic approach to hedging and pricing other less commonly traded derivatives. The existence and market price of the traded derivative should strongly inuence the hedging and pricing behaviour of a bank or intermediary selling the OTC derivative 2 ; it provides a new opportunity to hedge risk, it introduces a danger of arbitrage, and it changes the market price of the OTC derivative via the market practise of using implied volatilities. Our objective can be summarised as the identication of low dimensional models, complete in market priced assets, where the liquid derivatives and the underlying assets are independent tradeable assets. To be useful, a model must be consistent3 , risk adjusted, and provide a reasonably explicit description for the dynamics of the traded assets; it should allow one to reduce the hedging/pricing question to standard and computationally feasible calculations in the spirit of the classical theory of arbitrage theory. OUr models have all these properties.
1.1
Stochastic Volatility an academic approach
The existing academic literature approaches our question indirectly, looking at models for stochastic volatility and the market premium for risk. We are not happy with the details of such an approach, as the relationship between the hard to measure market premium of risk and the volatility of the liquid derivative, and the underlying assets
1 Terry J. Lyons es profesor del Departamento de Matem aticas del Imperial College de Londres. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de diciembre de 1998. 2 For simplicity we use the term OTC derivative when we refer to the derivative an intermediary is interested in pricing and selling, and a traded derivative when we refer to the derivative with a market price. 3 Force the pricing relationship between derivatives and the underlying that arise because of arbitrage. See later for the denition.
214
Terry J. Lyons
seems too distant to be reliable. As the full matrix of joint volatilities of the traded assets is the essential quantity that inuences price it seems wise to model these directly. To understand this point consider a basket of underlying assets and the idealistic situation where a Markov model has been constructed which perfectly describes the random evolution of their prices and their volatilities: dSt dt = (St , t ) St dt + (St , t ) St dWt , = b (St , t ) dt + (St , t ) dVt , (1)
where is an extra state variable that captures all that is important to the evolution of the system. In general the state variable will not be directly tradeable. Let us add to our assumption on the model (1) that the dimension of the additional state variable t in this model can be matched to the number of traded options Oi |i = 1, ..., d . However, on its own, and even if the model perfectly captured the behaviour of our system, (1) will not price any of the assets Oi ; information concerning attitudes to risk must be added. In his pioneering paper, Vasicek [5], observed that although prices of the Oi are not determined by such a model, arbitrage assumptions force the prices of dierent derivatives products to be mathematically interrelated, and this insight points the way to the explicit identication of the extra ingredient that must be identied making the prices unique. 1.1.1 The market premium for risk
The Markovian assumptions embodied in (1) ensure that at any time t the two state variables (St , t ) contain all the information that is available at time t concerning the future evolution of St . It is therefore at least plausible to assume that the price of any option or derivative on S should be a function of (St , t ) alone. The arbitrage arguments of Vasicek show how the existence of such a family of pricing functions is essentially equivalent to the assumption that the market attaches a consistent market price to risk. We will not re-derive the equivalence here, but to set up notation, we will explain how the model (1) leads to a huge range of dierent pricing model and hedging strategy, each suggesting its own model for the joint volatility of (St , Ot ) . Let b (St , t ) be any bounded function, which we refer to (somewhat loosely) as the price of risk. Consider the change of measure from P to Pb so that under the new measure, the model for (St , t ) satises the stochastic dierential equation: dSt dt = (St , t ) dWt , = ( S t , t ) b ( St , t ) + b (St , t ) dt + (St , t ) dVt . (2)
Providing do not become zero Pb is absolutely continuous with respect to P, so that any argument (e.g. about hedging) holds almost surely with respect to the one probability measure will also hold almost surely with respect to the other.
215
i OT
Suppose for simplicity that the terminal time for each option Oi is T and that b = F i (ST ). Dene Ot by setting
b Ot = Eb [F (ST ) |Ft ] .
b It follows from the Markov property, that Ot is a function Ob (St , t , t) of the state b and t it will variables. Exploiting the equality of dimension between the variables Ot b generically the case that O (St , , t) is at least locally invertible. To make our point b , St , t more simply, we further assume that it is globally invertible so that t = b Ot is its inverse. b In this case, the new variables Ot when taken with St are also a complete set of b state variables. Indeed, (Ot , St ) will be Markov and we can recover its joint volatility b , St ) is a martingale by dierentiating the inverse function. Under Pb the process (Ot and satises the SDE
dSt
b dOt
b = 0dt + St , b Ot , St , t b = St , b Ot , St , t b + St , b Ot , St , t
St dWt
Ob b , St , t , t dVt St , b Ot Ob b , St , t , t dWt St , b Ot S
This system of tradeable assets is obviously complete, and any contingent claim which b ) is functionally dependent on the underlying assets St (or even on the derivatives Ot b can be expressed in a unique way as a stochastic integral against the assets (Ot , St ). Therefore our choice of b has lead to a consistent model with known volatility for the joint evolution of the prices for our xed choices of options and the underlying. Assuming we have correctly identied market price of risk so that the volatility of the prices corresponds to reality, we can provide a consistent risk free hedging strategy to price any given contingent claim. 1.1.2 The essential diculty
So what is the problem with the Vasicek approach set out above? 1. There is an obvious diculty in choosing the function b sensibly, this becomes particularly tricky when one realises that stable hedging and pricing that is robust to small changes in the modelling relies on understanding the joint volatility b of the traded assets (Ot , St ). The riskless hedge is completely determined by the volatility of the tradeable assets, in our case (St , Ot ). Get their volatility wrong, and the second order eects of hedging will cause the portfolio value to drift away from its planned value. The approach outlined above implicitly forces
216
Terry J. Lyons
one to identify this market price of risk function b, solve the PDE to nd the price function Ob for this model, compute the inverse, and the derivative, to nally get the volatility of the option (St , Ot ). Unfortunately, this relationship is not explicit or local. Changes in the values of b away from the current values of the state variables will change the volatility of (St , Ot ) ; moreover, the eect is indirect and even obscure. The lack of a transparent relationship between parameters and the critical volatilities and lack of measures of the errors in the joint volatility arising from dierent price of risk models means that they cannot currently be regarded as robust. 2. A second related problem concerns the state variable . If it has an intrinsic meaning as representing some economic factors, then it is going to evolve and have a value; in theory once the market premium of risk has determined the i values of (St , t ) each of the options Oj has its price completely determined. In general, this will inevitably produce a conict with market prices. Recalibration replaces this diculty with an ever changing model for the volatility and hedging losses.4
1.2
Practitioner approaches
Probably the most common approach adopted by practitioners to accommodate the existence of prices for traded derivatives is calibration. One can consider models which are complete in the underlying assets but which have a number of undetermined parameters. By setting the parameters so that their model produces prices for traded derivatives that agree with market prices they hope that the Black Scholes framework will adequately reect the interrelated movement of the OTC option and the underlying asset and permit it to be hedged. In other words, implied volatility is used to accommodate the reality of derivatives priced independently by the market, in the hope that the market will incorporate all relevant historical pricing data relating to the volatility of the underlying. This approach gives no protection against the implied volatility moving. An intermediary may well use some vega hedging in an attempt to minimize the impact on portfolio values when the calibration implied by the derivatives undergoes changes. This practitioner approach is computationally feasible and of course this is very important, but it also has considerable limitations. Without a vega hedge, there is complete exposure to model risk and a shift in the implied volatility can result in a signicant change in the value of the portfolio. But as we will see later, vega hedging is an imperfect attempt to avoid this diculty.
4 Our approach in this paper is only free of such criticism if one uses historical data to calibrate the joint volatility model as the prices of assets and traded derivatives are independently incorporated. However, if one had many traded derivatives and some were used to calibrate our model (e.g. to match smile) then our approach would be subject to the same criticism.
217
Moreover, the implicit assumption that the market in the underlying assets is complete raises a serious intellectual obstruction to tackling our main question: how should we hedge when we have traded options. This diculty arises because the implicit consequence of this assumption is that the liquid option can be synthesized as a portfolio in the underlying asset alone. One is forced to conclude that the liquid derivative adds nothing to the market and is redundant when one tries to hedge. The stochastic volatility approach has more intellectual validity and avoids this second conict. However, it generally lacks computational bite and cannot always be translated into a realistic hedging strategy. The risk from the need to recalibrate does not in general go away.
1.3
This paper
This paper aims to provide a more limited and more computationally valid mathematical framework for analyzing and determining the appropriate price for an OTC derivative, and a hedging strategy to synthesize the same, in the presence of nearby liquid derivatives. In outline one could say that our approach is guided by the principle that once a product is heavily traded in the market, its price has its own independent volatility and (in so far as it inuences the pricing of our OTC options) should be included in our vector of prices. The problem is to model that volatility reasonably accurately using models that are complete in the visible traded assets including the options. Of course we will never be able to completely characterize the joint volatilities of the options and underlying correctly. But we are not so concerned by this for the following two reasons. We see the situation as parallel to the primeval days when Black and Scholes had just arrived and where historical models of volatility suciently accurate to price and hedge derivatives. The pricing of OTC option is now essentially a second order process as liquid derivatives can be used to oset much of the risk in a contract and really one only has to price a smaller residual contract; substantial errors in this residual price will not radically aect the price of the original contract of interest. It is for this reason that banks are willing to enter OTC contracts at all under the surface they hope they are really entering contracts with much smaller total downside and pretty uncertain price, but with the uncertainty compensated by large prot margins to compensate for the high risks. Although we leave the investigation of this aspect for another paper, it is quite possible to introduce non-linear pdes to take into account the uncertainty of the volatilities we model, following [2]. We expect this more robust approach to have similar features to the intuition we mention above. Residual contracts will see a big spread between buy and sell reecting the uncertainties that cannot be hedged away. However, the nearness of many OTCs to existing derivatives will result in the price of a typical OTC being small. The spreads in the prices of the residuals are a reection of uncertainties, but will always remain within arbitrage limits.
218
Terry J. Lyons
The structure of the paper can be summarized as follows: 1. We rst review in a less discursive way, two standard industrial approachs to the existence of liquid derivatives: calibration with implied volatilities and vega hedging. 2. Then, we analyze the above approach for the minimalist example where there is a single stock, cash can be borrowed and lent at zero rate of interest, and there is exactly one liquid derivative with a price in the market. The liquid derivative will be a simple European call option of xed strike price and maturity. 3. Staying with the same simple framework we propose approximate volatility models for the joint volatility of the call and the underlying. We will call such a volatility model consistent if, with probability one for the associated risk neutral measure, the paths of the call and the underlying coalesce appropriately at the maturity time. We illustrate this by showing that our main model is consistent, but that it is a delicate matter as small perturbations to it are not. The prices of contingent claims can therefore be calculated and hedged using the standard known volatility/no arbitrage paradigm. However, the price of the liquid derivative and the underlying asset are no longer deterministically related, and the hedge involves them both in a dynamic way. 4. The basic calculations in (3) do not depend on the underlying model being lognormal motion, nor are they dependent on there only being one derivative traded in the market. Combining Dupires approach[1] with ours, we believe that the case of a market with traded call options at many dierent strikes and could also easily be accommodated.
2
2.1
Standard practise explored

A simplistic example
A very simple example should focus attention. Suppose a single security is freely traded, and has price processes St and that in addition there is a single option freely traded on the market with price process OK,t , where for simplicity, the option is a European call, with payout at a xed maturity T given by (ST K )+ . Now suppose that a client approaches an intermediary wishing to purchase or sell a similar option on the same underlying, with the same maturity time T , but with a strike price K = K . There are a number of alternatives that can might taken in developing the price OK ,t for this new option. 1. In the rst case5 one could collect large amounts of data relating to the volatility of the underlying security, using this historical data, to estimate volatility for the new option; then use the Black and Scholes model.
5 We
do not suggest this is sensible or done in practise.
219
2. In the second case one could back out an implied volatility from the traded option under the assumption that it satises the Black and Scholes model for some choice of the volatility, and use this implied volatility as a substitute for the experienced volatility in the Black and Scholes formula to price the new option. Hybrids are also possible. Of course the rst approach ignores the marketed option and leaves the bank open to arbitrage and is really a non-starter although it might be an excellent approach in an immature market where options are not freely traded. The second approach, eectively the industry standard, is an interpolation which ignores historical data and cannot directly accommodate more market prices than there are parameters in the model. Hedging the resulting contract poses additional questions. Some are not easily settled in the classical theory. Common sense suggests that there is a huge dierence between a hedging strategy that holds the security and numeraire alone, and one that also uses the traded option. Unfortunately, the classical theory following on from applying Black and Scholes predicts that the traded option can itself be synthesized using the underlying security and so gives no guidance about how much of the traded option the bank should hold when hedging the OTC option. In practise this diculty in hedging is often nessed, in this case one could purchase a unit of the traded option, and then use the underlying to dynamically hedge the residual liability using the standard theory and implied volatility. For the simple contracts studied here, this approach, with its static hedge in the traded derivative is obviously a much more stable and lower risk strategy than hedging the original claim in terms of the underlying alone. The residual contract has a maximum value of |K K | and if this is small, even quite substantial errors in hedging the residual will not seriously aect the price of the original contract. However, it is not obvious, even in this simple example that the static hedge is the best one. In theory and for K > K, one could develop a hedging strategy by taking note of the fact that the call option with strike K > K is an option on an option with the strike K. So that providing we could identify a model for the volatility of OK, t we could dynamically hedge the derivative OK , t in terms of OK, t and cash.
2.2
The generic approach
More generally and for more complex products, we can caricature the industry standard approach as follows: Identify and isolate components of the contract that can be priced separately, and also identify the products available in the market which are nancially close to the OTC contract we are trying to price. Use the market prices of these to calibrate a model by backing out the volatility of the underlying assets. The calibrated model is then used to price the new contingent claim and to
220
Terry J. Lyons
value the existing portfolio. The deltas predicted by this calibrated model are used to decide the mix of cash and the underlying securities required to hedge the portfolio. Also recognize that if the market changes its view of volatility the prices of the traded options will move independently of the underlying securities to reect this, and that such changes are not hedged at all using the approaches outlined above. So introduce a second level of hedge (known as vega) where traded options are introduced into the portfolio so that to rst order, the portfolio is neutral to movements in the specied implied volatilities as well as the underlying. Both approaches have problems attached to them. The former obviously leaves the intermediary wide open to movements in the market view of volatility. Vega hedging brings more subtle problems. If one were to analyze it mathematically, one would appreciate that it produces a risk free hedging strategy only if the volatility of the implied volatility is zero, for in any other case the second order eects of hedging will produce a portfolio whose value drifts away from the desired stationary value. It can also have hidden instabilities. None the less one might consider it a plausible approach in the case where the vol of vol was small. 2.2.1 Vega hedging our simplistic example
Before proceeding to the main part of the article, we briey look at the consequences of using implied volatility and vega hedging in the simplistic example of a liquid European option with strike K and an OTC European option with strike K introduced in Section (2.1). Without the stabilising eects of transaction costs etc. vega hedging exhibits a phase transition making the computation more interesting and acting as a general warning of instability. We take advantage of this example to set up notation that will hold throughout the article.
2.3
Notation
Suppose that the price process St for the securities is modelled by a random Markov process i i dSt = ai (St ) dt + i, j (St ) Sj dWtj with risk adjusted probability measure P so that under this law
i i = i,j (St ) St dWji . dSt
It is well known that for this model, and under regularity conditions to ensure integrability, and conservativeness, the price pF (s, t) of European option with payout F (ST ) at a predictable maturity time T and with the current time being t and the current stock price being St is given by pF (s, t) = E [ F (sT )| st = s]. The standard Black Scholes model corresponds to the case where i, j = if i = j and zero otherwise, and
221
there is only one stock. In this case we call the scalar the volatility. A call option with strike price K corresponds to a payout F (s) = (s K )+ . Denition 1 Let p (s, ) denote the standard Black-Scholes arbitrage free price of an option based on a security with geometric Brownian trajectories, strike price 1, current price, volatility 1, and having units of time to run till maturity. Scaling arguments yield from p the price of an option under dierent volatility and strike assumptions. Because the price process is characterized by dst = st dWt standard scaling arguments show E (sT K ) |st = s = Kp s/K, 2 (T t) and using the representation for st = s0 e Vt 2 t where V is another standard Brownian motion one can integrate to obtain p in the well known closed form p (s, ) := 1 2 1 + s Erf r + 2 log (s) 2 2 + s Erf r + 2 log (s) 2 2 . (3)
1 2
The closed form is useful for numerical calculations, but for much of what we do in this paper it is better to use the symbolic form Kp s/K, 2 (T t) . The function p satises the well known Black Scholes pde with the following normalisation 1 2 2 p p s = 0. 2 s2 (4)
Varying between 0 and we see that any market price of the option satisfying the arbitrage bounds (st K )+ OK, t st is attained for a unique (OK, t ) [0, ]. Denition 2 We say (OK, t ) is the estimated operational time to maturity associated to the current price of an option with security price s, option price OK, t . It is the unique solution in to the equation. Kp (s/K, ) = OK, t . =
(OK, t) (T t)
is the implied volatility determined by option price.
Denition 3 The actual operational time to maturity is given by the integral of the empirical volatility between now and maturity
T t
2 (St ) dt.
The justication for the term operational time should be apparent to anyone with experience of the classical Black and Scholes model. It is a (random) variable reecting the turbulence to be faced. In the classical model, the standard discrete proof demonstrates how one can hedge perfectly by reapportioning ones assets every time
222
Terry J. Lyons
the underlying share goes up or down by a set percentage. The standard continuous result is obtained by taking the limit. However, the whole approach only works if one knows, in advance how many such step changes in price there will be before maturity. One does not need to know how regularly, or irregularly these changes occur, only how many there are going to be. In our language, operational time measures the number of such intrinsic ticks and in the standard model these come regularly and operational time is proportional to time; but in the real world where the volatility of the security may change, the amount of operational time before maturity will not be known in advance. In this case, the option price can be regarded as a representation of the market view of exactly how much operational time there will be from time t maturity.
2.4
Pricing and hedging
The market approach of using implied volatility we described in remark 2 from section (2.1) can be summarized as saying that one would value the second option at K p (s/K , (OK ,t )). There are then two standard approaches to hedging it. The rst uses only the underlying, and assumes (or hopes?) that the implied volatility will not change, and also assumes that the experienced volatility will coincide with the implied volatility. In other words, it values the portfolio of stock and the option at V (s, ) = s K p (s/K , (OK, t )) and proposes a hedge of = p(1,0) (s/K , (OK, t )) units of stock. This immunizes the portfolio to rst order against movement of the underlying given that the implied volatility remains constant. Providing empirical volatility coincides with the implied volatility6 , the fact that p in the pricing formulae solves the Black Scholes pde ensures that second order price movements are also cancelled out on average which is enough to hedge. Movements in the implied volatility are unhedged as are dierences between the empirical and the implied volatility. It is a stable hedge, in the sense that the amount of the stock held is always between nothing and one unit, swinging between the two according to the extent that security is out or in the money, and becoming close to one or other of these extremities as the estimated operational time goes to zero. A vega hedge7 would aim to hold units of the underlying, and units of the option with strike K so that the portfolio is neutralised against movements in price of the underlying or of volatility. In other words so that the total derivative of the portfolio in the price of the underlying and implied volatility (keeping time xed, is zero). Because of the way that implied volatility and time always occur together we
6 I am told that the market seems to place the implied volatility on the high side of reality, reecting the fact that the intermediaries tend to be short call options. 7 In all that follows, we will assume that we have fully discounted and that the return on a bond is zero. This achieves the usual simplication of presentation without loss of content.

V s V
223
see this is the same as asking that in (s, ) co-ordinates, where
and
are both zero
V (s, ) = s + Kp (s/K, ) K p (s/K , ) whereas the conventional hedge above would demand that the derivative set = 0. The vega hedge an easily be computed from the above.
s K p(0,1) K , p(1,0) s K p(0,1) K , s K , s K, s K, V s
= 0 and
+ p(1,0)
s , K
K p(0,1) K p(0,1)
and from the closed form of the solution one can deduce the portfolio suggested by the hedge. In the following graphs we show how the mix between the underlying stock and the option as operational time goes to zero in the case where K < K and vice versa. Meanwhile in the case where one is in the money, but K < K one gets a much better stability in the money. However, there remains a less serious out of the money singularity which could cause some diculty. This instability seems to be generic and can be observed in stochastic volatility models as well.
Figure 1: The exotic derivative has strike 1.1 while the liquid derivative has strike 1.0. Operational time ranges from .01 to 1.5 in steps of .2. The suggested hedge goes arbitrarily short in the stock as maturity approaches.
224
Terry J. Lyons
Figure 2: The exotic derivative has strike 0.9 while the liquid derivative has strike 1.0. Operational time ranges from .01 to 1.5 in steps of .2. The suggested hedge holds positive bounded amounts of the stock and option as maturity approaches.
3
3.1
The joint volatility of the call and the underlying

Overview
Vega hedging is a hybrid. It aims to hedge to second order in the underlying security, but only delta hedges in the direction of uctuations in the price of the liquid option. But those uctuations exist or one would not be interested in vega hedging! It is the nature of markets and market makers that they impose a price on any traded security. In the absence of complete knowledge of the volatility of the underlying, the price of a liquid derivative must uctuate in relation to any given Black Scholes predictions. The traded derivative has a price of its own. It is not a deterministic function of the underlying assets, and in eect it becomes a new asset in its own right. If one could model the joint volatility of the underlying price vector (St , OK,t ) for the traded option and the underlying together we could use the classical Black and Scholes paradigm to price and hedge a new contingent claim using them both, however we would no longer use the classical Black Scholes formula to price the contingent claim, but a new function that comes from solving the new pde. Although this approach seems far more natural than the vega hedging, some obvious issues need to be addressed before it can really be considered as a starter. 1. The behaviour of an option and the underlying are clearly not completely independent, and at the maturity of the option there is no independence at all. Therefore any model for the joint volatility of the pair should force this terminal relationship for the risk neutral measure without further assumptions.
225
2. There needs to be some basic and reasonably natural models for the volatility of the pair that can play the role of the geometric Brownian motion in the classical case. 3. Time to maturity must play an essential role in the model for the volatility, at least with respect to the volatility of the option price. We initiate the study of these issues below, and demonstrate the existence of a workable class of models. The approach provides a framework for pricing in a mature market where many derivatives have market prices, which takes account of those prices, and also provides rational hedging strategies, indicating the correct mix of derivatives and underlying assets. In addition, including market priced options in the hedging asset mix means that their calibration is taken into account automatically via the state of the system, rather than via the parameters. It is therefore possible, without any conict with market prices, to use historical data to improve the model for the joint volatility.
3.2
Connection with classical stochastic volatility papers
Before proceeding to the details of our approach, we make one nal remark. There are numerous papers on stochastic volatility, reecting the importance of the topic. Our approach, which constructs some quite explicit models for stochastic behaviour of the implied volatility, is not primarily directed to modelling in a stochastic way, the empirical volatility of the underlying, although this is a bi-product. Our primary interest is to identify the volatility of derivatives with prices in the market, along with that of the underlying. The volatility of our new price process (St , OK,t ) will still be uncertain and the methods of stochastic empirical volatility could be applied in this setting if required. However, the residual nature of most contracts, and large errors one expects, indicate that an approach based on non-linear pde and uncertain volatility might be simpler and quite adequate.
3.3
Modelling stochastic implied volatility
Suppose that we have two tradeable assets, a security with price St and a European call option Ot then we may always re-scale the units of the stock so that the strike price is 1. Our objective is to identify sensible models for volatility of the pair (St , Ot ) and to calculate prices based on these models. Now the process (St , Ot ) is forced by arbitrage constraints to live in domain E = {(s, O) |1 + O s O, O 0} and any value in E is potentially possible at any time although some values are obviously more likely than others.
226
Terry J. Lyons
Figure 3: The domain E . Let p(s, ) be the function dened in (3), giving the Black-Scholes price of an option with current stock price s and with volatility 1, units of time to run. Recall that if Ot is the price of option maturing at time T , then the equation Ot = p (s, ) always has a unique solution for any point in E . We regard s and as a new parameterisation for E , as the map (s, ) (s, p (s, )) is one to one, taking R+ R+ onto E . To get started, we make the following modelling assumptions: the volatility of the price of stock s is controlled by the value of the estimated operational time but is independent of the volatility of . the implied volatility determined by the option and stock prices agrees with the volatility of the stock price. Both are only assumed to keep the mathematics relatively simple but seem reasonable to rst approximation at least. We have not had the chance to rigorously test them against market data. More complicated assumptions would not radically change the picture only the numerical complexity of extracting a solution. The assumptions restrict us to a model of the form dSt dt = = t st dVt + t dt T t g (T, t) dWt + t dt (5) (6)
As both assets are assumed tradeable, one knows from the classical complete market paradigm that to obtain arbitrage prices, the next step is to choose t and t so as to make the measure risk neutral, in other words, chosen so that so that St and Ot = p (St , t ) are both martingales, and take expectations. The risk free nature of these prices make it standard that prices do not dependent on the real values of these parameters. However, there is a new feature our model must satisfy, and which we have not seen previously. It is clear that as time approaches maturity, the prices of the option and the underlying must become more closely aligned and compatible with the relationship determined by the payo of the contingent claim.
227
Denition 4 We say that such a model as (6, 5) is consistent if at the terminal time + (ST 1) = OT . In other words if the rst exit time of the process (Ot , St ) from E occurs at time T and at that time the process leaves E through the lower boundary. As an initial remark in the direction of understanding the condition of consistency we remark that our models are always almost consistent, in the sense of the next result. Lemma 1 In any model in the above class, (St , Ot ) will converge to point on one of the three lines bounding E at maturity time. Proof. Since (St , Ot ) is a positive martingale in each co-ordinate, it must converge as t tends to T . Suppose it converges to some point in the interior of E . It follows from standard martingale arguments that the Martingale must have nite quadratic variation along almost every sample path. Moreover, the map from R+ R+ onto E is smooth with locally bounded and invertible derivative. So the quadratic variation of (St , t ) provides a lower bound for that of (St , Ot ) and must also be nite. However the assumption that (St , t ) converges to an interior point ensures that
T u=t
u Su T u
du =
giving a contradiction. From this we conclude that the process (St , Ot ) must converges to a boundary point as t tends to T.
3.4
Fixing the price
In this section we will do the computation giving the values for t , t for which (st , P (st , t )) is martingale. Observe rst that the requirement that st is a martingale implies that t 0 since Vt is already a Brownian motion, and hence a martingale. So all the interest is in identifying the correct form for t . For this it seems substantially simpler to work in a fairly general way as the explicit form of the Black and Scholes formula does not seem to play a big role. We introduce a notation for the innitesimal generator for the diusion (s, t ): L= 1 2 T t s2 1 2 2 2 g 2 + t + + s2 2 t (7)
Then our requirement is to choose t so that one has Lp (s, ) = 0. By hypothesis p (s, ) is the Black-Scholes solution and satises (4) so that Lp = 1 2 2 g 2 + t + 2 T t p (s, )
and thus the condition that Lp = 0 translates into t + T t p log = 2 2 g
228
p
Terry J. Lyons
It remains to compute
log
and we will have an explicit form for .
We could just go directly into computation using the formula (3) but it is better and applies more generally to observe that if p is any solution to the equation 1 2 2 s 2 s2 p=0
2
p 2 then so is , and this function can be rewriten as 1 2 s s2 aswell. However, from p this it is obvious that the boundary data corresponding to the solution for our particular p is given by a delta function and is hence relatively easy to compute without computation!
We have
T t
p g2 log = t 2
So far we have not used the explicit form of the call option Black and Scholes solution, but on doing this one gets the following result. Proposition 1 The value of making the process (St , Ot ) risk neutral is given by 2 K ln s0 1 1 1 1 2 = + g 2 8 2 2 2 T t which is deduced from
K ln s p 1 1 0 log = + 8 2 2 2 2
using the identity

+ T p log = 2 2 g
from above. Corollary 1 The risk adjusted process has the innitesimal generator L = 1 2 T t
2
s2
2 s2
s log K 2 2 2
1 2 2 + g (, T t) + 2 2 + t
1 1 2 8
229
Proof. We must be careful about normalisations to get a consistent answer. Dene dst = st dVt or st = s0 e Vt 2 t where V is a Brownian motion with variance t at time t, then p (s, T t) = E (sT K ) |st = s and p (s0 , T t) = t = = so that p (s0 , ) =
0 +
1
1 d2 p s2 2 ds2 1 d + E s2 2 (s K ) |st = s0 2 ds 1 d2 + (s0 , s, T t) s2 2 (s K ) ds 2 ds 0
1 d2 + (s0 , s, ) s2 2 (s K ) ds 2 ds
where (s0 , s, T t) is the density of the Log normal distribution. That is, it is the 1 density of s0 e Vt 2 (T t) at s. A simple calculation shows this to be 1 1 (s0 , s, ) = exp s 2
1 2
+ log s log s0 2
Suppose that f is a smooth function of compact support at innity and bouned at zero then integrating by parts:
0
1 d2 + f (s) s2 2 (s K ) ds 2 ds
= = = = =
2 1 + d (s K ) f (s) s2 ds 2 ds2 0 d2 1 (s K ) 2 f (s) s2 ds ds K 2 d 1 d (s K ) f (s) s2 ds 2 ds ds K 1 d f (s) s2 ds 2 K ds 1 2 K f (K ) 2
and combining these two remarks we have that p (s0 , ) = = 1 2 K (s0 , K, ) 2 2 1 ( + 2 ln (K/s0 )) 1 2 K exp 4 8
230
Terry J. Lyons
and that p log = = 1 2 ( + 2 ln (K/s0 )) ( + 2 ln (K/s0 )) 2 8 2 2 ( + 2 ln (K/s0 )) ( + 2 ln (K/s0 )) 1 2 8 2

2
K K 2 ln s + 2 ln s 1 0 0 = 2 8 2 K 2 4 ln s 1 0 = 2 8 2 and the result follows by substitution. 2 K ln s 1 1 0 + = 8 2 2 2
3.5
Models for g
If represents number of ticks of operational time till maturity then do we have any intuition as to how it should behave and how volatile this estimate should be. Suppose then that you regard the volatility of market as a measure of the number n of market actions till maturity, where n is large, then the number that will arrive over a time period of length t is approximately nt nt (mean variance.) So it seems quite reasonable that the market view of the operational time should swing by this sort of amount over a unit of time, suggesting a model of the form dSt = t st dVt T t (8) 2 dt T t (9)
dt
K 1 1 ln s0 1 1 dWt + + 2 8 2 2 2
for initial experiments. The crucial point is that the volatility of the operational time is heavily dependent on the time to maturity. Moreover we have not just assumed that the volatility of the underlying security is stochastic more subtly we have assumed that the option has its own vol depending on maturity. A heuristic example of what we have in mind comes from the process of downloading a le across the Internet. Most web browser programmes provide a continual estimate for the time till download of the le it is often based on some compromise between the current and integrated rate of data transfer, and is remarkably frustrating as it can go up as well as down. The relative volatility can get bigger as one gets near the completion of the transfer. The model above is obviously consistent, as the nal term ensures that the process , for small time intervals near maturity, is bounded above with probability one; and so combining with the lemma we proved above, about the convergence of the process to one of the three boundaries of the region E , it follows that it must converge to the lower boundary as required for consistency.
231
3.5.1
A second example A second model sets g = the volatility

T t
T t ,
and seems rather similar, as we do not expect
to uctuate by orders of magnitude as the contracts approach
maturity. This second model has attractive features from a mathematical perspective. dSt = t st dVt T t (10)
2
dt
K 1 1 ln St 1 1 dWt + + T t 2 8 2 2 2
(11) dt T t T t
In this second example, we may transform this process into a nicer one by introducing a random time change. Let dt du = T t then we can rewrite (10,11) as dSt dt u = st dV
2
K 1 1 ln St 1 1 = dWu + + 2 8 2 2 2
(12) (13)
1 du
The fascinating remark is that this model is not consistent. There is always a small chance that the process will escape to innity.
4
4.1
Wider issues
Robust approaches to hedging
It is therefore open to the bank to devise consistent models for the volatility of underlying assets and derivatives at once. It goes without saying that this paper treats a very special case and is not the complete picture. There are a number of ways in which to try to improve it. The rst is to note that it is not vital to get the correct volatility model for the joint behaviour. This is because of the way these approaches to pricing are used in the market. The point is that real contracts have quite small residual components after a crude hedge with derivatives, so that the price of the contract is not ultra-sensitive, and using a robust conservative hedging strategy permitting a wide possible range of volatilities need not make a large dierence to the price of a contract. It is also easy to impliment.
232
Terry J. Lyons
4.2
Broader classes of model and many strikes
In general there are a number of tradeable derivatives, and those that are traded freely are often change with only those broadly on the money having liquidity. Moreover dierent derivatives have dierent maturity. We certainly do not have a clear model to choose in every example although the principle of looking for sensible and consistent models is obviously a sensible one. However, we recall the approach of Dupire where he uses derivatives at all strikes and maturities to predict a price dependent model for the volatility of the security. This approach allows one to easily generalise the results of this paper to derivatives with dierent strikes.
4.3
Other relevant work
Like the present paper, Zhu and Avelleneda [6] also use a derivative as an additional stochastic state variable while explicitly retaining market completeness, although from a somewhat dierent standpoint. They pre-specifying a lognormal process for the instantaneous volatility of a single underlying asset, and then make the connection with traded derivatives by identifying this volatility with the implied volatility of short maturity call options, which necessarily therefore have the same implied volatility.
References
[1] Dupire, B., Pricing and hedging with smiles, in Mathematics of derivative securities (Cambridge, 1995), Publ. Newton Inst. 15, 103111. Cambridge Univ. Press, Cambridge, 1997. [2] Hobson D., Stochastic Volatility, Chapter 14 in Statistics in Finance, eds. David J Hand, Saul D. Jacka, Arnold Applications in Science. [3] Hull J., White A. (1982), The Pricing of Options on Assets with Stochastic Volatilities, Journal of Finance 42, 2, 281300. [4] Lyons, T.J. Uncertain Volatility and the Risk-Free Synthesis of Derivatives, Applied Mathematical Finance 2 (1995), 117133. [5] Vasicek, O.A., An Equilibrium Characterisation of the Term Structure, Journal of Financial Economics 5 (1977), 177188. [6] Zhu Y., Avellaneda M., A Risk Neutral Stochastic Volatility Model, International Journal of Theoretical and Applied Finance, 1, 2, 289310.
Terry J Lyons Department of Mathematics Imperial College, 180 Queens Gate London SW7 2BZ England e-mail: t.lyons@ic.ac.uk
Information Transmission around Block Trades on the Spanish Stock Exchange

M. A. Mart nez, M. Tapia y J. Yzaguirre1
Abstract This paper investigates the informational eects of large transactions, or Block Trades (BT), in the Spanish Stock Exchange (SSE). In the SSE, in the open market period, this topic was not facilitated as in other markets till 1998. In this way, the SSE provides a special environment for analyzing the information transmission of these specic transactions. We assume that information can be better reected by changes in true asset value, proxied by the midpoint of bid-ask best quotes. Therefore, we will look at changing true asset value orders instead of trades. We study three dierent eects around BTs: price, liquidity and information transmission. To capture them, we consider three dierent endogenous variables: true asset value returns, relative spreads and adverse selection spread component. With this approach, we nd no clear eects of BTs. The main result of the paper is that there seems to be an increase in information asymmetries when we look at the adverse selection spread component in some of the dierent subsample classications (buyer, seller and sweeping BT), but there is no signicant permanent eect on returns. This result could be related to insiders trading in the market. In sharp contrast with adverse selection evidence, we also observe a temporary increase in liquidity around BTs. These changes reect temporary liquidity eects related to other spread components (order processing costs and inventory costs). This opposite evidence could be explained by the fact that there are no special market participants such as specialists or dealers and as a result our market participants are not required to absorb temporary order imbalances2 .
1 Miguel Angel Mart nez es Profesor Asociado del Departamento de Fundamentos del An alisis Econ omico de la Universidad del Pa s Vasco. Mikel Tapia es Profesor Titular del Departamento de Econom a de la Empresa de la Universidad Carlos II de Madrid. J. Yzaguirre trabaja en Calidad de Mercado, Sociedad de Bolsas. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de enero de 1999. 2 We are grateful to G. Rubio, I. Pe na and J. Marhuenda for helpful comments and to participants at the XXIII Simposio de An alisis Econ omico, II Encuentro de Econom a Aplicada, II Workshop in Finance de Segovia, 1999 European Financial Management Association Annual Meeting, V Encontro Galego de Novos Investigadores de An alise Econ omica and Seminars at Universidad de las Islas Baleares, Universidad de Alicante and Universidad Carlos III. We would also like to acknowledge nancial support from Direcci on General Interministerial Cient ca y T ecnica (DGICYT), project PB94-1373, PB97-0621 and PB98-030. The contents of this paper are the sole responsibility of the authors.
234
Mart nez, Tapia e Yzaguirre
Introduction
Information transmission through order ow is an important issue in nancial research. The general markets eciency assumption is based on this point. According to theoretical nancial literature on information, the value of private information depreciates quickly (see, for example, Foster and Viswanathan (1990)). Thus, informed investors prefer large transactions (Block Trades) in order to get into a valuable position as soon as possible. On the other hand, it is also known that informed investors, in order to conceal their superior private information, are interested in camouaging their desired trades into small or medium size trades (Kyle (1985)). However, given the major increase in institutional trading and the internationalization of investors in recent years, Block Trades (BT) are observed all over the world. From the empirical point of view, is not clear whether these BTs may be understood as strategic trading motivated by information or whether they may be viewed only as a consequence of institutional investors balancing their portfolios.3 Most of the empirical research into BTs focuses on information transmission by looking at permanent and temporary eects of BTs on asset prices or returns. The permanent part is interpreted as being information motivated, whereas the temporary one is associated with price pressure or liquidity costs. Scholes (1972), Kraus and Stoll (1972), Holthausen, Leftwich and Mayers (1987, 1990) and Chan and Lakonishok (1993, 1995) are interesting examples of this issue on the NYSE. Both eects (permanent and temporary) seem to be present and the sign depends on the type of BT. However, the results depend on the sample and the methodology used in the study. Similar analyses for order driven markets can be found in Ball and Finn (1989), for the Sidney Stock Exchange, and Riva (1996), for the Paris Bourse. Gemmill (1996), for the London Stock Exchange, has recently analyzed the liquidity eects of BTs under dierent publication rules. In related literature, Seppi (1992) and Daley, Hughes and Rayburn (1995) among others, investigate the extent to which block price changes around quarterly earnings announcements. This paper investigates the impact of BTs in the Spanish Stock Exchange (SSE). BT eects should be greater in a market that does not facilitate this type of transactions; i.e. where BTs are dealt like small trades. This market microstructure characteristic can lead investors to avoid this type of trades because of the eort of crossing a BT and evident risk of interference. For this reason, we expect transacted BTs to be very informative. The SSE oers us a particularly appropriate testing ground for examining these issues. In the SSE, in the open market period, this topic was
3 Formal models of information disclosure through BTs can be found in Easley and OHara (1987) and Seppi (1990). Easley and OHara (1987) show how BTs signicantly increase the probability market participants attach to the existence of private information. Seppi (1990) develops a model where, under not very restricted circumstances, information-based BTs are traded in a partial-pooling equilibrium.
Information Transmission around Block Trades
235
not facilitated as in other markets till 1998.4 In this way, the SSE provides a special environment for analyzing the information transmission of these specic transactions. In order to analyze whether these transactions transmit information, we propose a new approach. In sharp contrast with previous BT research, we assume that information can be better reected by changes in true asset value, proxied by the midpoint of bid-ask best quotes. Looking at these intrinsic value changes instead of price changes, we avoid the eects of liquidity (noninformative) trades. These kinds of transactions modify asset prices without aecting their true value (the so-called bid-ask bounce).5 At the same time, this allows us to consider very informative bid-ask changes which do not result from a new transaction (so that, no new price is established), but which reect worthy changes in the investors preferences for assets. Therefore, we will look at changing true asset value orders instead of trades. Related with information eects, market microstructure literature analyzes how prices absorb information, measured by changes in the adverse selection spread component.6 Adverse selection can be understood as a measure of information asymmetries. Thus, if we observe a decrease in adverse selection component around BTs, we could conclude that BTs transmit information diminishing information asymmetries between agents. This adverse selection component must be dierentiated from liquidity in general. Therefore, we will also analyze the behavior of relative spreads around BTs to detect changes in liquidity not related to information transmission. These changes are related to order processing costs and inventory costs. As previous studies have done, we will also analyze the impact of BTs on true asset returns. This new approach we propose, together with the special market microstructure we analyze, makes this study innovative in current literature on information transmission around BTs. The main result of the paper is that there seems to be BT information transmission when we look at the adverse selection spread component in some of the dierent subsample classications, but there is no signicant permanent eect on returns. We also observe changes in liquidity around BTs, but this eect is related to temporary spread component. The remainder of the article is organized as follows. Section II reviews briey the microstructure of the SSE and, particularly, the block trading process. The dataset and sampling rules are presented in section III. Section IV discusses the methodology used and results obtained in the analysis. Finally, section V oers some concluding remarks.
4 Examples of these special BT devices are the upstairs market in the NYSE (Hasbrouck, Soanos and Sosebee (1993)) and the broader bid-ask spread in the Paris Bourse (Riva (1996)). 5 Seppi (1992) also points out that the conclusions obtained by looking at price changes may be aected by the potential presence of a variety of price pressure eects. 6 A good reference is OHara (1995).
236
Institutional Settings of the Spanish Stock Market
The electronic continuous market for equities in SSE is a purely order driven market. Through this system, 142 companies are traded. The main characteristic is a single order book for every stock. We nd three main periods in the daily market: a) Preopening period (from 9:00h to 10:00h): In this period, introduction, modication and cancellation of limit orders are allowed. Depending on supply and demand, the system calculates a preopening price in real time. At 10:00h the system assigns shares to orders at prices better than or equal to the opening price. b) Open market period (from 10:00h to 17:00h): During this period limit and market orders are introduced. If a counterpart is found they are automatically executed. If not, the order remains on the book until an incoming order ts it, or the order is canceled. In this period prices change in real time depending on the ow of buy and sell orders. c) Special operations period (from 17:00h to 20:00h): In this period it is possible to report pre-agreed trades with an eective volume bigger than 20% of the daily turnover. So, this period is specically designed for BT transactions. Traders can also use the preopening period to trade BTs. Introduction of pre-agreed large trades in this period consists in buying at the maximum possible price of the day (15% more than the closing price of the previous day) and selling the same quantity of shares with a 15% reduction on the closing price. With this behavior investors can be sure to trade BTs at the opening price. We will not use these closed market BTs in this study.7 This is because we want to identify clear information signals. We could hardly identify BT information eects because of the overnight problem in the opening prices. In order to observe the information eects of BTs, we will focus on the open market period. In this way, we avoid other news that could aect the opening asset price during the closed periods. Investors willing to trade BTs in the open market suer two handicaps. First, traders must introduce a limit order to execute a BT, so it is impossible to cross a transaction outside the limit order book. Second, it is not possible to trade outside the spread of best buy and sell prices. As a result, it is very dicult to trade large blocks of shares in this period. In this context, a BT trader can face two dierent market situations:
7 During the period analyzed, 11% of the number of BTs was traded in the preopening period, and 15% in the special operations period. Regarding the eective volume, the percentages are 20% and 16% , respectively.
237
(i) When there is a level of prices available between best buy and best sell (spread bigger than tick size), traders quickly introduce pre-agreed sell and buy limit orders for the same amount of shares at the price available inside the spread. (ii) When there is no such available price and traders do not want to wait, they sweep the necessary orders to open the spread and get a price available inside it. This sweeping activity is particularly necessary for stocks that are so liquid that it is very dicult to nd an available price. Obviously, it imposes an additional cost. Crossing both types of BTs, when one side order has been introduced, there is always the possibility that another limit order may arrive and the pre-agreed BT cannot be completely crossed.8 We call this issue interference risk.
Data
Data on all orders on the SSE in the open market during the one-year period from May 1996 to April 1997 were collected from SSE les. As indicated, we selected only orders which change true asset value. From now on, these orders will be termed orders. In spite of the possible existence of other orders between any two of them, they will be considered in our analysis as consecutive. So, the k th order changing the true asset value and the following one will be referred as the k and k + 1 orders, respectively. The available information for each of these orders includes: time of occurrence (stamped to the nearest second), date, bid, ask, transaction price and number of shares transacted since the previous order.9 The value of the SSE Index (IBEX-35) for each second was also obtained from SSE les. As a description of the SSE, Table 1 presents some summary statistics about the size distribution of all trades crossed in the SSE during the period considered.
8 Since November 6th 1998, a new device for reporting and trading BTs in the open market period has been operative. This feature allows market members, as other European markets already do, to trade BTs outside the best bid-ask spread of the book. In any case, this possibility is set according to a certain relationship with market prices. Specically, there are currently two ways to trade a block: (i) For stocks belonging to the IBEX-35 Index (the 35 most liquid stocks on the SSE), members can report arranged blocks to the Exchange. As a result, interference risk has been eliminated. Minimum required amount of shares for trade is 5% of the daily turnover in the last quarter of the year and 100 million pesetas (0.59 million euros). In this context, the spread is the on line weighted average price of the six best levels of bid and ask. (ii) For all the stocks on the SSE, market members can introduce orders bigger than 10% of the daily turnover in the last quarter with a deviation of 15% from last closing price and 250 million pesetas (1.5 million euros). Here there is no time and price priority rule and members can select any order. 9 For orders which do not produce transactions, we consider the price of the corresponding previous transaction. For the rst order of the day we use the accumulated volume of shares transacted in the preopening period.
238
< 1 mil. > 1 mil. and < 10 mil. > 10 mil. and < 50 mil. > 50 mil. < 5% > 5% and < 10% > 10% and < 20% > 20% and < 40% > 40% All trades
number of trades (%) 2,413,137 (60.28) 1,335,059 (33.35) 226,666 (5.66) 28,420 (0.71) 22,048 (0.55) 2,002 (0.05) 1,989 (0.05) 1,099 (0.03) 1,282 (0.03) 4,003,282
trading volume (%) 810,396 (5.9) 4,348,807 (31.6) 4,435,662 (32.2) 4,174,179 (30.3) 2,048,370 (14.9) 421,378 (3.1) 442,291 (3.2) 350,160 (2.5) 911,980 (6.6) 13,769,044
Table 1: Number and eective trading volume (in millions of pesetas) of all trades crossed in the open SSE during the period May 1996-April 1997, sorted by trading volume. Those with trading volume greater than 50 million are additionally sorted by their percentage of the average trading daily volume. The percentage of the total is in parentheses. As can be observed, the mean trading volume is 3.4 million pesetas. Trades of 1 million or less represent more than 60% of all trades, but these trades only represent 5.9% of eective trading volume. Throughout this paper, we dene BTs as any trade whose value is over 50 million pesetas and, at the same time, is greater than 20% of the average eective trading daily volume for the respective asset.10 According to this denition, there were 2,381 BTs during this period. They represent 9.1% of trading volume, but only 0.06% of the total number of trades.
3.1
Sampling rules
In order to select our sample of BTs, some lters were applied to the available rms and BTs. Firstly, we only consider BTs involving to the 50 most liquid rms. This restriction allows us to use a highly continuous trading sample. In this way, disturbing nontrading eects are eliminated. We exclude a BT if there is a payment or stock split (or any payment in the rm) in the 13 calendar-days window for each BT (6
10 This cut-o was chosen because it is the institutional requirement for specially communicated trades on the SSE.
239
calendar-days before and 6 calendar-days after).11 These BTs are likely to be noninformationally motivated, as Choe and Masoulis (1992) point out. BTs for which additional blocks occurred in the stock during the same 13 calendar-days window are also excluded. In this way, selected BTs are not aected by the close presence of another BT. For reasons of data availability (motivated by the estimation period chosen) we also exclude BTs occurring less than 14 calendar-days after the beginning of the period analyzed and 14 calendar-days before the end. Finally, we analyze only blocks occurring between 11:00h and 16:00h. The rst and the last hour of the trading day are excluded because of the disturbing eects of opening and closing trades. Many large transactions at opening cannot be considered as BTs: They are merely a large number of individual transactions crossed together and printed as one transaction. Likewise, transactions during the last hour may incorporate end-of-the-day eects (see Amihud and Mendelson (1986) and Harris (1986)). It must be said that some of the BTs selected according to these criteria did not appear in the original sample of orders changing the asset true value proxy. However, we decided to include them because their information eects could operate with some delay or advance.12 After applying all these sampling rules, the number of BTs we nally consider is reduced to 195, in 41 rms. They represent 1.3% of the trading volume during the whole of the period analyzed. BT trading volume ranges from 51 to 27,668 million pesetas and the mean value is about 947 million pesetas. The analyses will be performed individually for each BT. The estimation period we consider is a 29 calendar-days window for each BT (14 calendar-days before and 14 calendar-days after BT).13 It is clear, considering the dierences between assets, that the number of orders in this xed period is very dierent from one asset to another. The range goes from 235 orders for the least liquid asset to 4,460 for the most liquid, with 1,487 being the average number for all BTs in the sample.
3.2
Descriptive statistics
Unfortunately, our dataset does not identify the party initiating the large transaction. However, as is clear from empirical literature on BTs, the signs of the expected eects dier for buyer and seller-initiated transactions. A buyer-initiated BT is expected to produce a permanent increase in the asset price, whereas the inverse eect is expected for a seller-initiated BT. In order to sort BTs as buyer or seller-initiated, we calculate
11 There is nothing special in this gure. The only interest is to separate BT eects as far as possible from others. 12 We will observe this possibility when traders choose to introduce the BT order not in the rst level of book prices. If there is enough time another order may arrive and when the BT is crossed we will not observe a change in true asset value. 13 The estimation period must be long enough to provide precise estimates of parameters and short enough to keep the number of trades manageable. We consider this period as one which appropriately meets both requirements.
240
the dierence between the BT price and the true value proxy at the previous trade. If this dierence is positive, we classify the BT as buyer-initiated, whereas if it is negative we classify it as seller-initiated. BTs whose price equals the previous asset true value are classied as indeterminate-initiated.14 The dataset identies most BTs according to an inside the spread or sweeping classication. BTs not included in either of these types are considered as not classied.15 Intuitively, we expect stronger eects in sweeping BTs because of the additional cost they impose. BTs were also sorted by whether or not they change the asset true value. As above, we expect greater eects in BTs that change the asset true value. Additionally, BTs are classied in four groups according to their trading volume. Each group has about the same number of BTs, with BB being the group with the biggest BTs, SS the group with the smallest, and BS and SB the medium size group. We expect a direct relationship between information transmission and BT size. Tables 2 and 3 illustrate some of the distinguishing features of the BTs in the sample. Table 2 shows the sample composition regarding the side initiating the BT, type and changes or not in asset true value. As can be observed in panel A, the sample distribution is very similar regarding the side initiating the BT, especially in the volume transacted. The number of indeterminate-initiated BTs seems to be greater than the other types for small and medium BTs. Panel B shows that the largest BTs by volume transacted are traded inside the spread, whereas the BTs not classied seem to be the small ones. However, the number of BTs in each group is very similar. Panel C shows that the biggest BTs change the asset true value. But this relationship is reversed for the other size BTs. Table 3 describes the day-of-the-week and hour-of-the-day distribution of the BT sample. The rst value in each cell is the percentage of the number of BTs and the second is the corresponding trading volume. We nd a clear seasonal pattern in our sample. First, from the microstructure of the SSE, it is clear that investors tend to use the less competitive hours of the day to cross large transactions. We see in table 3 that the 13:00-14:00h period is the time of the trading day when the biggest BTs are crossed. We also observe dierences in days of the week. Surprisingly, on Friday (the day of the week when futures contracts expire) we observe no special derivatives eect, whereas we see a large volume activity during the rst part of the week.
14 This criterion has been used previously by Blume, Mackinlay and Terker (1989) and Hausman, Lo and Mackinlay (1992), among others. The tick test algorithm (which classies a transaction by looking at the previous transactions price) proposed in Lee and Ready (1991), is a less informationconsuming method. Hausman, Lo and Mackinlay (1992) consider the true value rule considerably more accurate. 15 These are BTs whose limit orders were introduced in the book but not at the rst level of prices, and they await execution.
241
SB SS Vol. (%) BTs (%) Vol. (%)
BB BTs (%) Vol. (%)
BS BTs (%) Vol. (%)
BTs (%)
PANEL A Buyer-init. Seller-init. Indeterminate-init. 0.35 0.39 0.26 0.32 0.45 0.23 0.26 0.25 0.49 0.32 0.33 0.35 0.21 0.29 0.50 0.34 0.39 0.30 0.22 0.16 0.61 0.33 0.34 0.33
PANEL B Inside the spread Sweeping Not classied 0.47 0.31 0.22 0.76 0.20 0.04 0.34 0.33 0.33 0.49 0.43 0.08 0.35 0.32 0.33 0.48 0.40 0.13 0.30 0.31 0.39 0.39 0.53 0.08
PANEL C Change in true asset value No change in true asset value 0.67 0.33 0.65 0.35 0.47 0.53 0.35 0.65 0.51 0.49 0.35 0.65 0.52 0.48 0.37 0.63
Table 2: Size distribution of our sample in number of BTs and trading volume (in percentage terms). Regarding trading volume, BTs are classied in four groups, including the biggest in BB and the smallest in SS. In panel A, BTs are classied according to the side of the market initiating the BTs (buyer, seller or indeterminate initiated), in panel B they are classied according to type (inside the spread, sweeping or not classied) and in panel C according to whether they change the asset true value or not.
11:00-12:00 12:00-13:00 13:00-14:00 14:00-15:00 15:00-16:00 All periods
MON (%) 5.64 6.81 6.15 5.45 4.62 7.79 2.05 0.51 2.56 1.43 21.03 21.99
TUE (%) 5.64 3.17 8.72 7.63 4.10 7.98 6.67 12.10 3.59 4.25 28.72 35.13
WED (%) 3.59 1.96 4.10 2.11 5.64 17.84 2.05 1.43 2.05 2.56 17.44 25.91
THU (%) 2.05 0.86 3.59 3.94 1.54 1.20 1.54 0.43 2.56 2.40 11.28 8.82
FRI (%) 1.03 0.18 6.15 2.45 6.15 3.89 6.15 1.29 2.05 0.36 21.54 8.16
All days (%) 17.95 12.96 28.72 21.58 22.05 38.71 18.46 15.75 12.82 10.99
Table 3: Day-of-the-week and hour-of-the-day distributions (in percentages terms) of the

BT sample. The rst value is the percentage of the number of blocks and the second one is the corresponding trading volume.
242
Methodology and Results
There are certain features that characterize our dataset. First, orders are sampled at irregularly spaced random intervals (whenever changes in true value occur), so observations are unlikely to be identically distributed, since some of them are very closely spaced in time while others may be separated by hours. Second, asset prices are always quoted in discrete units or ticks (discreteness). Among the existing models of stock price discreteness, ordered probit is the only specication that can easily capture the impact of explanatory variables on price changes while also accounting for price discreteness and irregular transaction intervals.16 However, the use of ordered probit specication comes up against a major problem with illiquid stocks. We must limit tick movements because of the necessary degrees of freedom in the estimation procedure. So this method is not useful for our sample. Therefore, in order to diminish the discreteness problem, we will use returns instead of prices. Furthermore, to solve the irregular random intervals problem, we will use two alternative specications: the use of dierences in time between consecutive orders as an explanatory variable and the use of a time adjustment for our exogenous and endogenous variables. As mentioned in the introduction, one of the variables we focus on is changes in asset true value. The true value idea is taken from market microstructure literature. Glosten and Milgrom (1985) advocate the use of the midpoint of bid-ask quotes as a proxy for the true value. For asset j , the true value after the k th order is denoted by mjk , and is obtained as: mjk = (Ajk + Bjk )/2 , (1) where Ajk and Bjk are the ask and bid prices of asset j on the k th order, respectively. The point here is that if large trades convey valuable information, agents revise their estimation of the true price and their subsequent orders will modify the book quotes. These modications are considered informative (whether or not there is a new transaction), because they represent changes in the amount investors are willing to pay or to receive for assets. We use continuously compounded returns of relative change in the true value proxy as the information variable. This variable is denoted by Rjk . BTs imply an important change in Normal Market Size. In addition to information transmission, BTs can involve temporary changes in liquidity. The idea is that BTs can aect investors optimal portfolio or related variables and impose an inventory cost. These liquidity eects of BTs are analyzed with regard to changes in relative spread. Many market microstructure articles focus on relative spread to study liquidity eects around dividend or earning announcements.17 The relative spread for asset j after the k th order is denoted by Sjk , and is dened by: Sjk = 2
16 A 17 Lee
Ajk Bjk . Ajk + Bjk
(2)
description of this estimation procedure can be found in Hausman, Lo and Mackinlay (1992). et al. (1993) and Rubio and Tapia (1996) are representative examples of this literature.
243
Additionally, we consider that BTs can aect some variables such as accumulated volume and dierences in time between orders. One conclusion of market microstructure literature is that market activity can be measured by trading volume. Some papers have shown that it is important to control for some activity variables when we want to measure the information ow.18 As Seppi (1992) indicates, when we look at BTs we may consider a proxy of activity. In this way, volume appears as one appropriate variable reecting information arrival. We denote Voljk the square root of accumulated number of shares traded on asset j between orders k 1 and k .19 We denoted further as Diftjk the square root of the time elapsed in seconds between orders k 1 and k on asset j .20 Engle and Lange (1997) show that this variable can signal changes in order ow regime. So we also look at these variables, looking for changes in regime around BTs. Preliminary evidence of BT eects on previous variables is shown in table 4.
Sjk -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 2.77 5.55 1.56 -1.93 0.63 -7.28 -4.31 -13.44* -5.11 -25.45* -5.18 2.13 -6.98 -6.33 -5.88 -9.77* -10.71* -13.75* -11.41* -11.26* -6.82 Diftjk 16.69 -10.11 3.74 8.00 9.03 58.39* 28.65* 101.36* 74.35* 118.87* 44.52* 75.39* 54.37* 53.76* 58.32* 56.42* 56.35* 28.21** 51.11* 53.99* 22.16 Voljk -25.43* -2.45 -24.02* -26.50** -22.09* 8.67 -2.13 4.41 42.44** 35.92 5793.02* 316.71* 20.57 -14.70 -0.27 4.92 -12.53 -16.45 3.71 42.33** 16.56 Rjk Rj 0.41E-05 -1.33E-05 -6.93E-05 -2.53E-05 2.90E-05 -4.65E-05 0.76E-05 -3.78E-05 -0.67E-05 -4.13E-05 -3.66E-05 1.62E-05 -3.72E-05 1.17E-05 -4.51E-05 0.04E-05 -6.79E-05 -3.28E-05 -3.25E-05 -6.29E-05 2.12E-05
Table 4: For the characteristics of relative spread, accumulated volume, and dierences in
time, we show the percentage changes, averaged across all BTs, according with the following statistic: Cjk Kjk = 1 100 , Cj where C j is the average of each characteristic. For returns we use the statistic: Kjk = Rjk Rj . The asterisk indicates signicance at 5% and double asterisk at 10%.
18 Previous research (Lee, Mucklow and Ready (1993) for the NYSE, and Rubio and Tapia (1996) for the SSE) has found clear eects of trade volume on relative spread. Therefore, we will consider volume as a control variable. 19 We use the square root to avoid the outlier problem. 20 When a change of day occurs, we use the time from the market opening.
244
In this table, we show percentage changes in relative spread, dierences in time and accumulated volume dividing each observation by its average by calculating the following statistic: Cjk Kjk = 1 100 , (3) Cj where C is S , Dift and Vol. For returns we use the statistic: Kjk = Rjk Rj . (4)
The average of these statistics across all BTs is calculated for ten orders just before and after them. The cross-sectional distribution of each average is used to study the signicant level of the event. We can observe dierent evidence in table 4. First, relative spreads seem to decrease before and after BTs. This indicates an increase in liquidity. This eect is especially important just after BTs. According to market microstructure theory, this reduction may be caused by a reduction in information asymmetries or trading cost. We do not observe any signicant variation in returns around BTs, but there is a decrease in volume before BTs which, could indicate that agents are waiting for BTs to arrive. The only abnormal volume is the next BT order. This could be a sign of agents updating their demands and portfolios. The positive and signicant numbers we nd in time dierences show that time between orders increases just before and after a BT. Again, this could be an indication of investors waiting for trading and updating their expectations. However, this evidence is contrary to insider trading behavior, as is shown in Engle and Lange (1997) and theoretical papers that indicate that insiders would use noisy trading intervals to camouage their trades. So the preliminary evidence around BTs shows dierent behavior of relevant variables such as spreads, volume and dierences in time. However, the observed eects on our variables may be due to variables aecting them other than BT information transmission. In order to isolate the BT eect, we need to control the endogenous variables considered for alternative inuential variables around BTs. The control variables we use are well known in nancial literature. As we have pointed out, volume appears to be one appropriate control variable for information arrival. Therefore, we use Vol as an independent variable in the regression analysis. Three lags of this variable are considered in order to allow for some delay in its eects. In order to avoid the disturbing overnight eect, we also consider an end-of-the-day dummy variable.21 This variable, denoted by Dend, equals 1 if k th order on asset j is the rst order of the day and 0 otherwise. We also take into account market return as an exogenous variable. We take the IBEX-35 Index as our market index. We take the nearest in seconds value for each order in the sample period. Its return is denoted by RIBEX . We also use three lags of this variable in order
21 It has been well-documented that overnight returns dier substantially from intraday returns (Amihud and Mendelson (1987) and Stoll and Whaley (1990)).
245
to allow for some delay in its eects. The aforementioned Dift is also considered as a control variable. Finally, to pick up eects around BTs, we consider 21 dummy variables (a window of 10 orders before and after each BT) denoted DBT . Each dummy equals 1 for the order occurring orders after the BT, and 0 otherwise. The order corresponding to the BT itself is considered as the reference order, = 0. So, after controlling by the aforementioned variables, the coecients show us the eect of BTs on our endogenous variables before and after they occur. As pointed out in the introduction, we consider three dierent endogenous variables: true asset value returns, relative spreads and adverse selection spread component. They capture price, liquidity and information transmission eects respectively. Because no two rms have an identical timing of orders, we cannot estimate our regressions as a multivariate system across all BTs, so we run one time-series regression for each BT. The coecients are therefore averaged over all of them and over the dierent subsamples considered. If BTs are relevant for these variables, we will observe signicant coecients for the appropriate BT dummy variables. These are the relevant variables in our analysis. The remaining variables are included only to control for external eects.
4.1
Returns Evidence
Next, we show the regression for each BT used to analyze the BT eects on true asset value returns. The time-series regression for each BT j is:
3 3 3
Rk
= +
=1
R +
=0
RIBEX +
=0 10
Vol (5)
+ Dift + Dend +
=10
DBT + k ,
where we use three lags of the endogenous variable and DBT stands for the dummy variable employed to pick up eects around BTs.22 First column of table 5 shows the results of the above regressions. We only report the results for the total BTs sample. First, we observe mean reversion in returns. This expected result is consistent with other results in literature. Secondly, clock time measured by Dift is also signicant. Other control variables seem relevant and coecient signs are as expected (RIBEX , Vol, Dend). So the use of these variables to control seems to be justied. Next, we show the closest BT dummy coecients. In general, they are not statistically signicant. The contemporary coecient is negative and signicant. The most striking result is that in the dierent subsample classications this coecient does not
22 The
range of observations for each regression goes from 235 to 4,460. We run 195 regressions.
246
change its sign or is not statistically relevant23 . This is especially important in the buyer and seller classication. This is not consistent with previous BT studies or with our intuition. In the total sample results, this negative eect of the contemporaneous BT dummy is almost oset by the eect of two orders later. In the end, it seems that there is no signicant permanent eect on returns. The reason for this result could be the specic problems that traders face in the SSE in crossing a BT. These problems could cause BT price not to be the real price. The idea is that investors willing to buy (sell) a BT would pay (renounce) an additional fee that is not observed by market participants. In this environment, BT prices could be not informative. Alternatively in order to control for irregular interval problem, we also calculate time returns according with the expression: Tark = (1 + Rk )1/Diftk 1 . The analogous regressions we now run for each BT are:
3 3
(6)
Tark
+
=1 3
Tar +
=0
TarIBEX
10
+
=0
Tavol + Dend +
=10
DBT + k ,
(7)
where TarIBEX is calculated in the same way as Tar, whereas Tavol is Vol/Dift. With this specication, the results are slightly dierent. In general, the control variables are not relevant or their coecients are lower than before, and BT dummies are not signicant. Although we cannot construct a statistical test to evaluate the appropriateness of time adjustment, by looking at adjusted R squared we can conclude that, in general, adjustment with Dift as an exogenous variable is better than Tar adjustment. This is why we do not include these results.
23 The
subsample results can be obtained from the authors by request.
247
S 0.00121* -2.41E-05* 1.59E-05* 0.421E-05* 0.0720E-05* 0.40441* 0.21956* 0.05397* -0.169E-05* 0.00224* -4.41E-05 8.45E-05 -24.0E-05 26.2E-05 -55.5E-05* -0.00311* -29.8E-05 -12.2E-05 31.1E-05** 18.7E-05 -4.79E-05
CONS R(1) R(2) R(3) RIBEX RIBEX(1) RIBEX(2) RIBEX(3) VOL VOL(1) VOL(2) VOL(3) S(1) S(2) S(3) DIFT DEND -5 -4 -3 -2 -1 0 1 2 3 4 5
R 1.5588E-05 -0.32* -0.07* -0.04* 0.40* 0.16* 0.10* 0.05* 0.132E-05* 0.00298E-05 -0.0166E-05 -0.0194E-05 -0.470E-05* 52.493E-05* 1.5507E-05 8.8251E-05 -7.297E-05 -17.42E-05 3.8204E-05 -72.46E-05* -11.36E-05 58.358E-05* -8.642E-05 19.328E-05 -13.33E-05
Adverse Selection 0.622E-05 0.000290E-05 -0.0131E-05 0.384E-05 -0.0381E-05 -0.0255E-05 0.509E-05 -0.0783E-05 0.0112E-05 5.64E-05 0.753E-05 0.0724E-05** -0.0206E-05
VOL 13.49* 0.02* 0.06* 0.04* 1.36* 64.22* -2.93 2.86 -1.99 9.79* -8.01* 348.24* 4.38 -26.57* -19.45* -5.95* -2.43*
Table 5: For each BT in the sample, three time series regressions are run with three dierent
specications. In particular the regressions are:
3 3 3 10
Rk Sk
= +
= 1 3
R +
=0
RIBEX +
Vol + Dift + Dend +

=0 10 =10
DBT + k
+
= 1 3
S +
=0
Vol + Dift + Dend +

=10 10
DBT + k ,
Volk
+
= 1
Vol + Dift + Dend +

=10
DBT + k ,
where RIBEX is the return of IBEX-35, Vol is the square root of accumulated volume between orders changing asset true value, Dift is the square root of time elapsed between orders, Dend is a dummy variable for end-of-the-day eects and DBT stands for the dummy variable employed to pick up eects around BTs. Two time series regressions are run with two dierent specications:
3 3
Volk Rk
+
= 1
R +
= 1 10
Vol + Dift + Dend + k ,
+ k +
=10
k DBT + uk .
Coecients are cross sectional averaged across all of them. White (1980) standard errors are used.
248
4.2
Information Transmission Evidence
To test the information transmission hypothesis, we look at the adverse selection spread component. The way in which we estimate this component is taken from Foster and Viswanathan (1993). These authors measure adverse selection as the returns response to unexpected volume. Given their model, we estimate the following regression:
3 3
VOLk = +
=1
R +
=1
VOL + Dift + Dend + k ,

10
(8)
Rk = + k +
=10
k DBT + uk .
(9)
The rst equation estimates the unexpected volume for each change in true return through residuals. The second equation measures the reaction of returns including as explanatory variables these residuals and BT dummies. In this context, coecient measures mean adverse selection and coecients measure abnormal adverse selection around BTs. Results are included in the second column of table 5. We can observe that the adverse selection component, measured as the coecient of residuals, is not signicant. The only signicant coecient is the one associated with four orders after BT. These results are consistent with Admati and Peiderer (1988) model where liquidity traders pool their trades. So insiders only act in these periods and not in the middle of the day, when they would be detected. So, BTs are not as informative as expected. As can be observed in table 6, when we look at dierent subsample classications, the results are slightly dierent. The contemporary BT dummy is signicantly positive for buyer and seller BTs but not for indeterminate BTs. This is consistent with the sign of the initiator party. The same dummy is also signicant and positive in sweeping BTs. This result is also consistent because of the additional cost this type of BTs imposes. Both results are indicative of information transmission. There exists an increase in adverse selection spread component for these subsamples These results are consistent with the presence of insiders trading with these types of orders. It seems that the knowledge of the BT initiator party increases information asymmetries among traders. Additionally, the implied greater cost of sweeping BT seems to be a signal of the quality of the information transmitted to the market.
249
Buyerinit. CONS
Seller-init.
Indeterminate init.
Inside the spread
Sweeping
Not ed
classi-
-5 -4 -3 -2 -1 0 1 2 3 4 5
-4,69E-06 4,16E-09 -8,85E-07 5,27E-06 7,32E-07 -8,34E-08 -1,87E-05 4,11E-06* -1,66E-07 2,69E-06 -9,93E-07 5,92E-07 3,49E-09
9,02E-06 4,64E-09 1,73E-07 -2,14E-07 -1,95E-06* -1,35E-07 -5,33E-07 1,79E-06* -9,73E-07 5,76E-08 3,98E-07 1,16E-06 -6,53E-07
1,07E-05 1,19E-09 1,14E-07 5,39E-06 -8,77E-08 -4,20E-07 2,17E-05** -5,02E-06 9,00E-07 1,19E-04 1,65E-05 5,44E-07 -6,29E-08
1,15E-05 -1,31E-09 8,59E-07 3,73E-06 4,66E-07 -1,55E-07 1,01E-05 -6,44E-06 1,27E-06 1,42E-04 1,97E-05 4,45E-07 -8,57E-08
9,20E-07 -3,10E-12 2,92E-07 4,55E-06 -4,53E-07 -1,18E-07 -9,44E-06 2,65E-06* -5,80E-07 1,46E-06 -3,24E-08 7,41E-07 4,11E-09
1,52E-05 4,16E-08** -7,55E-06** -2,13E-07 -3,94E-06* -1,61E-06 7,47E-05 3,99E-06 -9,08E-07 1,76E-06 -1,60E-06 1,95E-06 -2,13E-06
BB
CONS
BS 6,66E-06 -3,22E-09 -1,01E-06 -2,78E-06 -5,66E-07 -2,71E-07 5,61E-06 9,50E-07 2,78E-06* 4,11E-06 4,00E-07 -4,12E-08 -1,28E-08
SB -2,70E-05* 1,12E-08 2,92E-07 5,99E-06 -1,02E-06 -7,65E-07 2,45E-05 -7,30E-06 -9,08E-07 3,76E-06 -1,41E-06 7,95E-07 -8,28E-07
SS 1,27E-05 1,91E-09 -2,37E-08 8,04E-06 -5,82E-08 3,73E-07 -8,22E-06 1,77E-06 -1,46E-06 2,17E-04 3,06E-05 1,99E-06 -1,29E-08
change in true asset value
no change in true asset value
-5 -4 -3 -2 -1 0 1 2 3 4 5
3,18E-05* 1,89E-09 2,24E-07 4,15E-06 1,09E-07 -3,66E-07 -1,14E-06** 1,32E-06 1,45E-08 -3,63E-07 3,90E-07* 1,59E-07 1,72E-08
1,17E-05** 5,17E-09 3,42E-08 4,19E-06 -2,55E-07 -2,02E-07 1,06E-05 -3,81E-06 7,80E-07 9,77E-05 1,30E-05 4,36E-07 -3,62E-07
-9,93E-07 -1,02E-10 -3,49E-07 3,37E-06 -5,47E-07 -3,24E-07 -2,25E-06 3,21E-06* -7,70E-07 1,82E-06 3,07E-07** 1,11E-06 6,75E-10
Table 6: For each BT in the sample, two time series regressions are run:
3 3
Volk Rk
+
= 1
R +
= 1 10
Vol + Dift + Dend + k ,
+ k +
=10
k DBT + uk ,
where Vol is the square root of accumulated volume between orders changing asset true value,Dift is the square root of time elapsed between orders,Dend is a dummy variable for end-of-the-day eects and DBT stands for the dummy variable employed to pick up eects around BTs. The coecients are cross sectional averaged across all of them. White (1980) standard errors are used.
250
As a last test of information transmission, we consider volume as an endogenous variable. Volume will measure abnormal activity around BTs. In this case, this would be a signal of insiders around BTs and information ow in the market.24 The regressions for each BT are:
3 10
Volk = +
=1
Vol + Dift + Dend +

=10
DBT + k .
(10)
Results are in the third column of table 5. Before the BT we see an unclear pattern, with a negative coecient just before the BT but a positive one two orders before. However, after BTs there is a signicant decrease in market activity that could be related to the presence of insiders. Those insiders could lead the rest of market participants to decrease transacted volume. This is associated with previous ndings on adverse selection.
4.3
Liquidity Evidence
For relative spread, Sk , the time-series regression run for each BT is shown by the following expression:
3 3 10
Sk = +
=1
S +
=0
Vol + Dift + Dend +

=10
DBT + k . (11)
We show the results in the last column of table 5. The lagged variables are positive and signicant. As expected, we observe an autorregresive process in this variable. Another important variable is volume. We observe a negative contemporaneous coefcient and positive lagged ones. Negative relationship has been documented in other research into the SSE (Rubio and Tapia (1996)). This evidence is also consistent with Admati and Peiderers (1988) model and, at the same time, is contrary to the results of Lee, Mucklow and Ready (1993) for the US market. The most important result related to liquidity is the negative and signicant BT dummy coecients just before and contemporary with BT arrival. This is related to an increase in liquidity. The increase in liquidity before BT can be explained by the necessary introduction of pre-agreed BT limit orders for the same amount of shares at the price available inside the spread. After BTs there is a decrease in liquidity, so part of the eect is temporary. This result is related to a decrease in temporary spread components such as inventory cost and operative cost. This is relevant because these coecients have been obtained taking into account volume as a control variable. Looking at the SSE, this is a stronger result because previous research did not nd any eect on relative spread after controlling for volume.25
Admati and Pfeiderer (1988). and Tapia (1996) show that relative spreads do not change in the SSE around dividend announcements when they control for activity variables such as volume and number of transactions.
25 Rubio 24 See
251
Concluding Remarks
This study analyzes the role of BTs in the SSE. The contribution is the use of orders that change true asset value. Moreover, we apply this methodology to a market where BTs are not facilitated. Thus, BTs are dealt on the SSE like small trades. This market microstructure characteristic gives us a special testing ground. We study three dierent eects around BTs: price, liquidity and information transmission. To capture them, we consider three dierent endogenous variables: true asset value returns, relative spreads and adverse selection spread component. With this approach, we nd no clear eects of BTs. There is no signicant permanent eect on returns in the dierent subsample classication, which is contrary to previous evidence and to our intuition. In related papers, other authors have obtained clear eects of BTs on prices depending on BT type. We suspect that the reasons for these dierences could be related to methodology and SSE market microstructure. To discover the source of these dierences, this methodology should be applied to other markets with block trading facilities. In addition to previous studies, we analyze adverse selection as a measure of information asymmetries and as a consequence of information transmission. It seems that there is an increase in information asymmetries when we look at adverse selection spread component in the dierent subsample classications (buyer, seller and sweeping BT). This result could be related to insiders trading in the market. In sharp contrast with adverse selection evidence, we also observe a temporary increase in liquidity around BTs. These changes reect temporary liquidity eects related to other spread components (order processing costs and inventory costs). This opposite evidence could be explained by the fact that there are no special market participants such as specialists or dealers and as a result our market participants are not required to absorb temporary order imbalances.
252
References
[1] Admati, A. and P. Peiderer: A Theory of Intraday Patterns: Volume and Price Variability, The Review of Financial Studies 1 (1988), pp. 3-40. [2] Amihud, Y. and H. Mendelson: Asset Pricing and the Bid-Ask Spread. Journal of Financial Economics 17 (1986), 2, pp. 223-249. [3] Amihud, Y. and H. Mendelson: Trading Mechanisms and Stock Returns: An Empirical Investigation. Journal of Finance 42 (1987), pp. 533-553. [4] Ball, R. and F. Finn: The Eect of Block Transactions on Share Prices. Australian evidence. Journal of Banking and Finance 13 (1989), pp. 397-419. [5] Blume, M., C. Mackinlay and B. Terker: Order Imbalances and Stock Price Movements on October 19 and 20, 1987. Journal of Finance 44 (1989), pp. 827-848. [6] Chan, L. and J. Lakonishok: Institutional Trades and Intraday Stock Price Behavior. Journal of Financial Economics 33 (1993), pp. 173-199. [7] Chan, L. and J. Lakonishok: The Behavior of Stock Prices around Institutional Trades. Journal of Finance 50 (1995), pp. 1147-1174. [8] Choe, H. and R. Masoulis: Measuring the Impact of Dividend Capture Trading: A Market Microstructure Analysis. Working paper, Vanderbilt University, 1992. [9] Daley, L., J. Hughes and J. Raiburn: The Impacts of Earnings Announcements on the Permanent Price Eects of Block Trades. Journal of Accounting Research 33 (1995), 2, pp. 317-334. [10] Easley, D. and M. OHara: Prices, Trade Size and Information in Security Markets. Journal of Financial Economics 19 (1987), pp. 69-90. [11] Engle and Lange (1997): Measuring, Forecasting and Explaining Time Varying Liquidity in the Stock Market. Discussion Paper 97-12, University of California, San Diego. [12] Foster, F. and S. Viswanathan: A Theory of Interday Variations in Volumes, Variances and Trading Costs in Securities Markets The Review of Financial Studies 3 (1990), pp. 593-624. [13] Foster, F. and S. Viswanathan: Variations in Trading Volume, Return Volatility, and Trading Cost: Evidence on Recent Price Formation Models Journal of Finance 48 (1993), pp. 187-211. [14] Gemmill, G.: Transparency and Liquidity: A Study of Block Trades on the London Stock Exchange under Dierent Publication Rules. Journal of Finance 5 (1996), pp. 1765-1790.
253
[15] Glosten, L. and P. Milgrom: Bid, Ask and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders. Journal of Financial Economics 14 (1985), pp. 71-100. [16] Harris, L.: A Transaction Data Study of Weekly and Intradaily Patterns in Stock Returns. Journal of Financial Economics 16 (1986), 1, pp. 99-117. [17] Hasbrouck, J., G. Soanos and D. Sosebee: NYSE Systems and Trading Procedure, NYSE Working Paper. 93-01, 1993. [18] Hausman, J., A. Lo and C. Mackinlay: An Ordered Probit Analysis of Transaction Stock Prices. Journal of Financial Economics 31 (1992), pp. 319-379. [19] Holthausen, R., R. Leftwich and D. Mayers: The Eect of Large Block Transactions on Security Prices. Journal of Financial Economics 19 (1987), pp. 237-267. [20] Holthausen, R., R. Leftwich and D. Mayers: Large Block Transactions, the Speed of response, and Temporary and Permanent Stock-Price Eects. Journal of Financial Economics 26 (1990), pp. 71-95. [21] Kraus, A. and H. Stoll: Price Impacts of Block Trading on the New York Stock Exchange. The Journal of Finance 27 (1972), pp. 569-588. [22] Kyle, A.: Continuous Auction and Insider trading. Econometrica 53 (1985), pp. 1315-1335. [23] Lee, C., B. Mucklow and M. Ready: Spreads, Depths and the Impact of Earnings Information: An Intraday Analysis. Review of Financial Studies 6 (1993), pp. 345-374. [24] Lee, C. and M. Ready: Inferring Trade Direction from Intraday Data. The Journal of Finance 46 (1991), 2, pp. 733-747. [25] OHara, M.: Market Microstructure Theory. Blackwell Publisher, Massachusetts, (1995) [26] Riva, F.: Block Trading on the Paris Bourse Central Market: An Empirical Study., Proceedings of Organization and Quality of Equity Markets, Paris (1996). [27] Rubio, G. and M. Tapia: Adverse Selection, Volume and Transactions around Dividend Announcements in a Continuous Auction Markets. European Financial Management 2 (1996), 1, pp. 39-67. [28] Scholes, M.: The Markets for Securities: Substitution versus Price Pressure and the Eects of Information on Share Prices. Journal of Business 45 (1972), pp. 179-211.
254
[29] Seppi, D.: Equilibrium Block Trading and Asymmetric Information. The Journal of Finance 45 (1990), 1, pp. 73-94. [30] Seppi, D.: Block Trading and Information Revelation around Quarterly Earnings Announcements. The Review of Financial Studies 5 (1992), 2, pp. 281-305. [31] Stoll, H. and R. Whaley.: Stock Market and Volatility. Review of Financial Studies 3 (1990), pp. 37-71. [32] White, H.: A Heteroscedasticity-Consistent Covariance Estimator and a Direct Test for Heteroscedasticity. Econometrica 48 (1980), pp. 817-839.
M. A. Mart nez Departamento de Fundamentos del An alisis Econ omico Facultad de Ciencias Econ omicas Universidad del Pa s Vasco Avenida Lehendakari Aguirre, 83 48015 Bilbao, Spain e-mail: jepmasem@bs.ehu.es Mikel Tapia Departamento de Econom a de la Empresa Universidad Carlos III de Madrid C/ Madrid, 126 28093 Getafe (Madrid), Spain e-mail: mtapia@emp.uc3m.es J. Yzaguirre Calidad de Mercado Sociedad de Bolsas C/ Alfonso XI, 6 (Madrid) e-mail: j-yzaguirre@develnet.es
Market integration measures

s and Vicente Meneu1 Angel Pardo, Alejandro Balba
Abstract Many market integration measures are operationalized to compute their numerical values during a period characterized by the lack of stability and market turmoil. The results of the tests give their degree of eectiveness, and reveal that the measures based on the principles of asset valuation, versus statistical measures, more clearly yield the level of integration of nancial markets. Besides, cross market arbitrage-linked measures and equilibrium models-linked measures provide complementary information and reect dierent properties, and consequently, both types of measures may be useful in practice2 .
Introduction
The integration of nancial markets is a usual issue of special interest that has been the objective of numerous papers comparing stock markets, bond markets, foreign exchange markets, commodity markets and derivative markets. A high degree of integration among markets indicates that prices are formed in a correct way and, therefore, agents interested in well-diversied portfolios and appropriate risk-return ratios will concentrate on the available assets without taking into account the concrete markets. Furthermore, if derivative markets are involved, low cost operations can be carried out possibly helping to attract more investors to hedged or deferred positions, increasing market liquidity. Conversely, a low degree of integration implies quite dierent pricing rules with the subsequent eect on the diversication process. It also can dissuade hedged positions and leads to arbitrage strategies that generate risk-less prots derived from discrepancies in prices. Finally, some agents can implement speculation strategies that take prots from greater predictive power of one market over another.
1 Angel Pardo Tornero es Profesor del Departament de Economia Financera i Matem` atica de la Universitat de Val` encia. Alejandro Balb as es Catedr atico de Econom a Financiera de la Universidad Carlos III de Madrid. Vicente Meneu es Catedr atico de Econom a Financiera de la Universitat de Val` encia. Los tres profesores forman parte de un grupo de investigaci on en Econom a Financiera, Optimizaci on y Modelizaci on Matem atica. Esta charla se imparti o en la sesi on del Seminario MEFFUAM de febrero de 1999. 2 We thank Alfonso Novales and Ignacio Olmeda for helpful comments. We are grateful to Jorge Yzaguirre from Spanish Stock Exchange and Remedios Romeo from MEFF-RV for providing the data necessary to perform this study. Research funded by DGYCIT (reference number: PB95-0729C02-02) and University of Valencia (project UV98-2716).
256
s, Meneu and Pardo Balba
In spite of this, a rigorous denition of what is understood by integrated markets does not exist in the nancial literature and it is only commonly accepted an intuitive but imprecise idea: two nancial markets are integrated when they evolve in a combined way. Many authors try to formalize the concept and provide numerical and analytic integration measures. Several questions arise. For instance, under which conditions are they equivalent? Is any of them superior to the rest? Do the responses to these questions depend on the concrete setting? Theoretical and empirical approaches may be applied to answer these questions. This paper presents an empirical test that analyzes the eectiveness of a large number of measures in a situation of clear disintegration. To do that, the measures have been classied into two major categories accordingly to their nature. The rst group contains those measures introduced by statistical and econometric methods, while the second focuses on the asset pricing theory. It should be pointed out that this type of classication has never been previously proposed. The purpose is to analyze the convenience of considering the nancial theory to introduce an integration measure. In short, twelve integration measures are revised; four based on statistical techniques and eight on the basic principles of asset valuation. In order to guarantee the lack of integration, the Spanish index IBEX-35 and its derivative market have been chosen during a period characterized by market turmoil. These markets have shown a high degree of eciency, as pointed out by Lee and Mathur (1999), but a large number of cross-market arbitrage opportunities were available during the Asian crash of October, 1997 (Balb as et al. 1997). The applied procedure guarantees the maximal precision since perfectly synchronized high frequency data have been used to compute the value of the market integration measures. Two main results are reached. First, measures based on the principles of asset valuation provide an absolutely similar degree of integration, minute-by-minute, during the tested period, while the rest of measures contradict each other. This seems to imply a serious objection for the statistical measures that are not able to give an unied conclusion and, consequently, asset-pricing models could yield a more successful way to measure the level of market integration. Second, the measures based on theoretical approaches may be subdivided in measures based on cross-market arbitrage and measures based on equilibrium models. Cross-market arbitrage-linked measures seem to be more adequate when derivatives are involved and hedging or deferred strategies are the focus of the analysis. Besides, equilibrium-linked measures are useful when studying very incomplete markets and seeking for well-diversied portfolios, although the data of long periods are required in order to compute some of these measures. Thus, arbitrage and equilibrium arguments apply in dierent settings and reect dierent properties, what justies that both sorts of measures may be considered to analyze integration levels.
Market Integration Measures
257
The outline of the paper is as follows: Section 2 is devoted to summarize some integration measures based on statistical and econometric techniques. Section 3 studies the measures based on cross-market arbitrage. Section 4 reviews equilibrium-linked measures. Section 5 describes the data and the trading conditions on Spanish Financial Markets during the rst days of the Asian crash (October 1997). In Section 6 the measures are applied to the IBEX-35 futures market. Section 7 concludes the paper.
Measures based on Statistical and Econometric Techniques
The rst set of measures is based on statistical and econometric methods. The most widely used and intuitive measure is the cross correlation of contemporary returns of the compared markets. So, he rst measure of economic integration can be stated as: Measure 1. The spot-futures market integration can be measured by the correlation coecient between simultaneous returns in the markets. The higher the correlation coecient, the stronger the market integration is (Kempf and Korn, 1998). A correlation coecient near to one would indicate perfect integration between both markets since they incorporate the information in the same manner. A zero or negative correlation coecient would imply segmentation. This measure is usually accompanied by the cross-correlation coecient analysis of the non-contemporary returns of both markets. The purpose is to check if they are not autocorrelated. Through this analysis, another measure of the degree of market integration is dened. Measure 2. Spot-futures market integration can be measured by the crosscorrelation coecient between simultaneous returns and the cross-correlation coefcients for the spot return with the futures return at dierent lags. The higher the contemporaneous coecient and the minor lagged correlation coecient, the stronger the market integration is. We would like to emphasize that cash market frictions grant some comparative advantages to the futures market, which have led to some authors to analyze if prices in the futures market lead or lag those in the spot market. Consequently, dierent works studied the dynamics between the price returns of market indexes and their future contracts applying either the causality of Granger (1969) (see Kawaller et al. (1987) and Ng (1987)) or the causality of Sims (1972) (see Stoll and Whaley (1990), Chan (1992) and Abhyankar (1995)). For this last case, the regression equation is:
p
ct = +
k=p
k ftk + ut
(1)
where ct and ft indicate the returns of the cash and of the derivative assets at the date t, respectively, and k is the number of lags. The coecients with negative
258
(positive) subscripts indicate lag (lead) coecients. The degree of market integration is deduced, in this case, from the k values. Measure 3. The markets are integrated if the contemporary variable coecient k is greater than zero and all other coecients are not dierent from zero. Signicant values for the coecients at lags k would indicate that the returns in the futures markets tend to lead those in the spot market, and signicant values for the coecients at leads k would indicate that the futures market tends to lag the spot market. It should be pointed out that measures 1, 2 and 3 are characterized for the use of returns (rst dierences in prices). This causes some inconvenience when spot and futures prices form a cointegrating vector.3 For this reason, the development of cointegration techniques at the end of the eighties resulted in a new integration measure based on prices and not on returns. Measure 4. Two markets are integrated if a cointegrating structure between them exists. The study of the integration between the derivative market and its underlying asset through cointegration analysis lies on the relationship between arbitrage and cointegration. Pricing based on arbitrage must duplicate one asset with another (or a combination of other) asset(s). Hence, if the derivative asset follows a certain trend, the arbitrage activity should cause the underlying asset to share the same trend. Consequently, as Arshanapalli and Doukas (1997, pages 258-259) pointed out, cointegration [...] would imply that the deviation from the common equilibrium path should cause price realignments, restoring the original equilibrium. On the other hand, lack of cointegration between the index futures and the underlying cash market would suggest that the underlying forces which are required to integrate the two markets into one market are rather weak. The Granger Representation Theorem establishes that if the price series of the two comparative markets are cointegrated, the short-run adjustments of the series with regard to the equilibrium level are included in an error correction model. If spot and derivative asset prices are cointegrated, then either spot prices lead derivative prices, or derivative prices lead spot prices or a combination of the two eects exits. For this reason, measure 4, together with measures 2 and 3, have been applied to measure the integration between markets and to check a possible lead-lag relationship.4 All the measures based on statistical and econometric techniques study the integration between two concrete markets in a time interval and require a wide sample
components (price series) of the vector xt are said to be cointegrated of order d, b if all components of xt are integrated of order d, I (d), and there exists a vector = 0 such that linear combination is integrated of order d b, I (d b), where b > 0. The vector that allows a lineal combination of variables I (d) with a integration order smaller than d is called a cointegration vector. If exists a bivariate model that uses only rst dierences will be mispecied (Engle and Granger, 1987). 4 See Sutclie (1997, pages 162-172) for a complete survey of the empirical studies of leads and lags between spot and futures prices and Stephen and Whaley (1990) and De Jong and Donders (1998) for the case of relationships between options and stock markets.
3 The
259
period. However, they do not provide information on the strategy to implement in order to take advantage of the lack of integration, nor do they consider the transaction costs that would be incurred when carrying it out. We should stress that all of them share the characteristic of exclusively reecting movements in returns or in market prices of the comparative assets, without taking into account any valuation model. As a result, measures based on the principles of asset valuation are of a particular interest since they can be applied at any moment in time and, in some cases, they provide an optimum arbitrage strategy.
Measures based on Cross-Market Arbitrage
The rst studies to outline the integration between derivative markets and their underlying assets following the basic principles of asset valuation placed great emphasis on checking the fulllment of the Law of One Price (LOP).5 Consequently, the derivative asset has to be duplicated (or the underlying asset from the market price of the derivative asset) and the theoretical price has to be compared with its market price. Since the asset and its replica must oer the same payos, the equality between these prices indicates the fulllment of the LOP. On the other hand, the mismatches of the market prices with the theoretical ones allow a risk-less benet to be obtained through the purchase of the cheap asset and the sale of the expensive one. Of the above-mentioned, an integration measure p that compares the deviation between the theoretical spot price (Ct ) and the spot price (Ct ) can be dened as: p= Ct . Ct (2)
In this case, the presence (absence) of market integration is studied by means of fulllment (or not) of the Law of One Price. Measure 5. If p is equal to one, the LOP is fullled and the markets are integrated. Conversely, if p is bigger (smaller) than one the cash index is undervalued (overvalued) with regard to its replica obtained from the derivative contract and the risk-less bond. This measure allows the incorporation of the transaction costs involved in arbitrage strategy. It also permits the study of market integration between the futures market and the underlying market, between the options market and the cash market and between the futures and the options markets if these last two have a futures contract as an underlying asset. Recently, Chen and Knez (1995) have introduced a new approach to analyze the degree of market integration. They dene the concept of market integration in a weak and a strong sense and establish the corresponding measures. Two markets are integrated in a weak sense if the LOP is fullled between them.
5 See Protopapadakis and Stoll (1983), Cornell and French (1983a and 1983b), Modest and Sundaresan (1983) and Modest (1984).
260
Measure 6. Consider two markets A and B in which the LOP holds separately. The weak integration measure g (A, B ) is dened as the smallest dierence between each markets family of state prices and it is calculated as g (A, B ) =
dA DA , dB DB
min
dA dB
(3)
where DA and DB are the sets of state prices for each market and |||| is the Euclidean norm. If g (A, B ) = 0, the markets are integrated in a weak sense, while if g (A, B ) > 0, the markets are not integrated and arbitrage opportunities exist. Since the fulllment of the LOP does not imply the absence of arbitrage opportunities (AAO),6 Chen and Knez restrict the concept of integration and establish that two markets are integrated in a strong sense if cross-market arbitrage opportunities do not exist between them. They dene a new integration measure: Measure 7. Consider two markets A and B in which there are not arbitrage opportunities in either market. The integration measure in a strong sense is dened as the smallest dierence between the positive state prices and it is calculated as a(A, B ) =
+ + dA DA , dB DB
min
dA dB
(4)
+ + and DB are the sets of positive state prices for each market. If a(A, B ) = 0, where DA the markets are integrated in a strong sense, while markets are not integrated and arbitrage opportunities exist as long as a(A, B ) > 0.
Measures g and a represent an important advance on the measures based on statistical and econometric techniques and on measure p, since they inform of market integration by considering all possible arbitrage portfolios and they are not based on concrete strategies. Nevertheless, the two integration measures proposed by Chen and Knez are based on dierences in state prices and, therefore, they do not allow the transaction costs to be discounted. Balb as and Mu noz (1998), following the approach by Chen and Knez, propose a new integration measure m based on monetary terms. They use the benets that can be obtained from the optimal arbitrage strategy, if it exists. To obtain this measure, they consider a two period model, t and T , and a unique market that incorporates all the markets that they compare. n assets are traded at a price pi , with i = 1, 2, . . . , n at the date t. A portfolio x is dened as x = (x1 , x2 , . . . , xn ), where the xi indicates the bought (positive sign) or sold (negative sign) units of the asset i. Any market portfolio has a price at tgiven by:
n
P (x) =
i=1
xi pi .
6 The absence of arbitrage opportunities implies the fulllment of the LOP. In general, the reciprocal is not true. See Ingersoll (1987, p.59).
261
All the assets take prices that are known at moment T assuming a discrete source of uncertainty K . If the states of nature are H , only one of them can occur at T . The price of portfolio xat moment T is given by:
n
x (k ) =
i=1
xi i (k ) ,
where i (k ) indicates the pay-o of the asset i in the state of the nature k . Theorem 3 of Balb as and Mu noz (op.cit. p. 163) proves that when LOP fails, then there exists a solution x for the following optimization problem: Maximize f (x) = subject to xi pi iSx xi pi for every k K ,
n i=1
n i=1 n i=1
xi i (k ) = 0 xi pi < 0 ,
where Sx represents the set of the sold assets of portfolio x (i.e., xi < 0). The numerator of the objective function is the value of the arbitrage portfolio and the denominator is the aggregate amount of the sales, both expressed in monetary units at moment t. The quotient can be interpreted as the ratio between the benet obtained from the arbitrage strategy x and the value of the sold assets. The rst constraint implies that at the moment T the portfolio has a pay-o equal to zero in all states of nature. The second constraint looks for portfolios that provide an income at the moment t. Notice that the set of opportunities of the problem is the set of possible arbitrage portfolios. If the solution is reached at x , the integration measure is dened by m = f (x ) and takes values between 0 and 1. As a result, the new integration measure is: Measure 8. If m is equal to zero, the LOP holds and the markets are integrated. If m takes values greater than zero, arbitrage opportunities exist and the markets are not integrated. It is possible that arbitrage opportunities exist even when m is equal to zero. To detect them, it is only necessary to modify the sign of the rst constraint of the previous problem, imposing the search for a portfolio whose payos at T are bigger than or equal to zero in all the states of nature. In this case, the optimal value will be denoted by M . Therefore, following the terminology of Chen and Knez (1995), m and M could be considered the weak and strong integration measures proposed by Balb as and Mu noz (1998). Thus, we have a new integration measure: Measure 9. If M is equal to zero, the markets are integrated in a strong sense. If M takes a value greater than zero, arbitrage opportunities exist and the markets are not integrated.
262
The measures m and M denote the integration of the market in a global sense, since they consider all the possible arbitrage strategies and they choose the optimal one. Moreover, these measures do not need to make assumptions about the fulllment of the LOP or about the AAO on each market, because all the analyzed assets are included in only one market. The use of integration measures based on prots instead of state prices facilitates the consideration of the transactions costs paid when carrying out an arbitrage strategy. If l is dened as the quotient between the prot of the arbitrage portfolio and the total value of the exchanged assets, a relationship between l and m (Balb as and Mu noz (op.cit, p. 165)) can be stated: l= m . 2m
Assume that the total transactions costs (T ) incurred in arbitrage related strategies are proportional to the sum of the purchase (P ) and sold (S ) quantities and dene the ratio T C as the ratio T /(P + S ). The dierence between l and T C indicates the unitary prot obtained from the arbitrage once the transaction costs have been discounted. The consideration of the transaction cost is a fundamental aspect when determining if the markets are integrated or not. Arbitrage opportunities can exist, indicating that the markets are not integrated (m > l > 0), but they cannot be exploited since the prot would not compensate the transaction costs (T C > l). It is interesting to highlight that if l is equal to T C we have: m = 2 (T C ) , 1 + TC
where m is an implicit measure of integration that indicates the minimum value that m must take so that arbitrage opportunity exists. Or, in an alternative sense, the maximum value that m can take so that the market is integrated. Therefore, signicant results are obtained with the Balb as and Mu noz integration measures: the composition of the optimal portfolio and the possibility to discount the transaction cost.
Measures based on Equilibrium Models
The third group of measures is based on the principles of asset valuation but they rest on equilibrium models. Garbade and Silber (1983) collected this feature in the measure they suggested for testing the integration level between the cash and futures markets. These authors specied a dynamic equilibrium model and they established that the degree of market integration is a function of the elasticity of supply of arbitrage services that can be measured from the following model: Ct Ct = + Ct1 Ct1 + t , (5)
263
where Ct is the natural logarithm of the theoretical spot price, Ct is the natural logarithm of the observed spot price and is an inverse measure of the elasticity of supply of arbitrage services. In the context of equation [. . . ], measures the rate of convergence of cash and futures prices (op.cit. p. 294) and is the measure of integration 10: Measure 10. If is small, both markets are integrated and prices will converge quickly. If is equal to one, both markets are not linked and the futures and spot prices will follow uncoupled random walks. It is important to note that although Garbade and Silber have provided a model to estimate the rate of convergence of cash and futures prices which reects the corresponding level of index arbitrage activities, they do not furnish a statistical test for the signicance of the estimated coecients, [...] which has profound implications in the testing for market linkage (Wang and Yau, 1994, p. 461). Hence, Wang and Yau (op.cit.), Yadav (1992) and Kempf and Korn (1998) have outlined the estimation of testing for the presence of a unit root in the mispricing series, dened as the dierence between theoretical and market prices. Thus, we have a new measure of integration derived from the previous one that would be obtained by testing for the presence of a unit root in the following model:
p
Mt = 0 + Mt1 +
i=1
p Mtp + t ,
(6)
where Mt = Ct Ct and shows the mean reversion in mispricing. Its value is the integration measure 11: Measure 11. If there is not a unit root in the mispricing series (i.e., < 0), markets are linked. If there is a unit root (i.e., = 0), spot and futures price series are not related and the markets are not linked. The higher the mean reversion parameter , the stronger the market integration is. In this case, the integration is again wholly related to the existence of arbitrage opportunities. If the previous mispricing was positive (the spot price was underpriced), arbitrage activity would force the change in the mispricing to be negative (the underpricing would decline) and vice versa.7 Yadav et al. (1994) and Dwyer et al. (1996) have generalized the mean reversion analysis by applying a cost of carry model with nonzero transaction costs to motivate estimation of threshold models between futures and cash indexes. Their results suggest that the speed of convergence of the basis to its equilibrium value depends on the level of mispricing. Bessembinder (1992) proposed the latest measure we review. This author establishes that assets and futures markets are integrated if expected returns on portfolios
7 In perfectly integrated markets this measure is not dened since the mispricing series takes a zero value for all t (see Kempf and Korn, 1998).
264
consisting of asset and futures positions are identical to expected returns on asset-only portfolios of identical systematic risk.
a The relationship between Et1 (Rit ), the expected next period return on the ith asset, and its systematic risk is stated as a a Et1 (Rit ) = 0t + it 1 1t ,
(7)
a where 0t is a cross sectional constant, it 1 1t is a 1 n vector of conditional sensitivities of i asset to each of n economic variable and 1t is a n 1 vector of risk premiums at time t.
The behavior of futures prices in a model of capital market equilibrium obeys the relation f f Et1 Rit (8) = it 1 1t ,
f where it 1 1t is a 1 n vector of conditional sensitivities of percentage change in futures prices j to the n economic variable.
Since (7) holds for spot prices and (8) holds for futures, expected returns of portfolios composed of assets and futures are also given by (7). Thus, market integration implies that the futures premium and the expected excess return on the spot asset dier only if the systematic risk of the spot and futures markets dier. To evaluate this, conditional betas are estimated and are used to make cross sectional regressions of the form
Rpt = 0t + 0 t dp + n i=1
ipt + it it ipt dp + pt ,
ipt is the estimated where Rpt is the return on equity portfolio or futures contract p, beta for portfolio p with respect to the ith economic variable, and dp is a dummy variable equal to zero for spot assets and equal to unity for futures contracts. The hypothesis that futures markets are fully integrated with assets markets is checked by testing that risk premiums are uniform across assets and futures markets and that the intercept for futures contract is zero. Measure 12. The markets are integrated if we cannot reject that the estimates of f it and the estimates of 0 t 0t + 0t are equal to zero. In short, measures 10, 11 and 12 only study the integration between the spot and futures markets. They have similar characteristics to the measures based on econometric techniques but, unlike them, they have to take into account some equilibrium model. Table I shows the dierent nancial integration measures described in Sections 2, 3 and 4 and a summary of some of their characteristics.
265
Data and Trading Conditions
The IBEX-35 futures market (MEFF-RV) began to trade in January 1992 and since then it has consolidated itself as one of the most important in Europe.8 The stock and futures markets open at 10.00 a.m. and close at 5.00 p.m. and 5.15 p.m. respectively. Both markets are electronic and the priority for crossing a transaction is determined by price. If prices are equal, priority is given to the arrival time of the order. The data used in this study make reference to the period between the 22nd and 30 of October 1997 (7 market sessions). The minute to minute prices of the IBEX-35 index and the midquote of the bid and ask price of the futures contract on IBEX35, with expiration on the 21st of November 1997, have been obtained from the Market Information System (MIS) of MEFF-RV. The Sociedad de Bolsas provided the dividends paid out by the IBEX-35 index shares and the interest rates have been obtained from the Servicio de Series Temporales del Banco de Espa n a .9
th
On Thursday, October 23rd , 1997, the Hang Seng index of the Hong Kong Stock Exchange suered a fall of 10.41% , which in turn caused a generalized drop in the European markets. The IBEX-35 index fell 2.49% and 0.79% on the 23rd and 24th , respectively (see the third line of Table II). On the 27th of October the Hong Kong, the Indonesian and the Taiwanese Stock Exchanges received a large number of sale orders, which caused an important fall in market prices. Its eect was reected, again, in the Spanish market, rst, through increments in the trading volumes and in the intraday volatility of the markets and, second, through a decline in the IBEX-35 of 4.40% (fth column of Table II). The convulsions in the stock and futures markets worldwide continued during the market session of the 28th of October. The Spanish stock market stopped trading from 4.46 p.m. to 5.30 p.m. due to the spectacular increase in prices in New York (at 4.40 p.m., Spanish time). For the rst time in the history of the Spanish electronic market, an adjusting period was instigated from 5.30 p.m. to 6.00 p.m. in order to allow adjustments in the shares market. At 6.00 p.m. the open session began again and concluded a half an hour later. However, the derivatives market was closed from the 5.04 p.m. to 6.43 p.m., at which time trading began once more and continued uninterruptedly until 7.40 p.m. The results of a session with so many incidents can be summed up under three points: rstly, the stock and the derivative markets on IBEX-35 registered record maximum daily volumes; second, the size of the relative bid-ask spread of the futures contract took values that duplicated the bid-ask spread in stable periods; and, lastly, the intraday volatility of the minute by minute IBEX35 index was 0.125% , while the volatility of the IBEX-35 future was 0.235% (sixth column of Table II).
8 MEFF-RV, in 1994 and 1995, was the stock index futures market with the largest number of futures contracts traded worldwide (Sutclie, 1997, p.59, Table 3.4.). 9 Because intraday data were not available for interest rates, the daily middle rate corresponding to the repo operations carried out with Spanish Treasury Bonds has been chosen for all the minutes of the same day.
266
Market instability continued on the 29th of October. The shares market began trading 39 minutes late, due to the excessive volume of orders that had been placed during the adjusting period. The vendors stopped providing information about the spot index from 0.22 p.m. until the 0.49 p.m. Finally, the Spanish stock market rose by 5.66% , the biggest daily rise in the last six years, with an intraday volatility of over than 0.07% (seventh column of Table II). To sum up, the beginning of the crisis of the Asian nancial markets at the end of October, 1997 caused large variations in the closing prices, high intraday volatilities and unprecedented trading volumes in the spot and futures markets. All this justies the choice of this period for the study of the nancial integration between the two markets, comparing a stable subperiod (the 22nd , 23rd and 24th of October) with an unstable subperiod (the 27th , 28th , 29th and 30th of October).10 The delays, stops and extensions of the trading session in several market sessions have led to the adjustment of the sample period for the seven days. Consequently, the degree of nancial integration between stocks and stock index futures has been determined daily.
6
6.1
Results
Measures based on statistical and econometric techniques
The cross correlation analysis of the minute-by-minute returns for IBEX-35 spot and IBEX-35 futures is presented in Table III. The contemporary correlation coecients (measure 1) are signicant at the 1% level every day, except the 29th of October. For this day we cannot reject the segmentation hypothesis between markets (H0 : spot,f ut = 0) at the 1% level. In the non-contemporary cross correlation analysis (measure 2) we observe that, rstly, every day presents a cross correlation between spot price changes and oneminute lagged futures price changes (spot,f ut(1) ) signicant at the 1% level and higher than the contemporary correlation. Secondly, the coecients with k > 0 are signicant only starting from the 28th . These results suggest that new information tends to be reected rst in futures market in stable periods, while during the Asian crisis a bi-directional eect is observed. According to the second measure, therefore, the 22nd , 23rd and 24th show the highest degree of integration. Before estimating a bivariate model to determine the degree of market integration that measure 3 proposes, it is important to remind that this measure uses returns as variables (rst dierences in prices). If the cash and futures series were cointegrated,
10 The integration between the derivative markets of the S&P 500 market index and their underlying asset in stable and volatile periods has been studied in various works in which the sample period is centered around the crash of October of 1987. Harris (1989) studies the behavior of the basis; Kleidon (1992) and Kleidon and Whaley (1992) carry out a cross correlation analysis of the series of returns of the cash and derivative markets; Wang and Yau (1994) analyze the mean reverting of the residuals of the cointegration equation while Arshanapalli and Doukas (1997) study the integration using cointegration and error correction models.
267
a bivariate model expressed in rst dierences would not be well-specied (Engle and Granger (1987)). Hence, we are going to analyze if a cointegration relationship exists between the cash and futures prices on the IBEX-35 index. The null hypothesis of a single unit root is tested for each of the IBEX-35 spot and futures prices using the non-parametric test of Phillips and Perron.11 Although the null hypothesis is not rejected for the price series at the 5% level, it is for the series in dierences at the 1% level. Once proven that both series are integrated of the same order, we have tested for the existence of a stationary linear combination of them (measure 4) by applying the multivariate methodology proposed by Johansen (1988 and 1991) and by Johansen and Juselius (1990). Table V reports the cointegration results. The null hypothesis of no cointegration is rejected at 10% level and, therefore, we cannot reject the existence of at least one cointegration vector. These results dissuade the use of integration measure number 3. We highlight in Table V the fact that the null hypothesis is rejected at the 10% level on the 29th of October and at 5% level on the 27th and 28th , while the remaining days it is rejected at the 1% level. Therefore markets appear to have been highly integrated under normal trading conditions. After detecting the existence of a cointegration vector, an error correction model has been estimated for each day.12 The model that has nally been constructed, according to the Johansen procedure (1988 and 1991) is as follows:
p p
lpct lpft
= c1 + 1 zt1 +
i=1 p
a1i lpcti +
i=1 p
a2i lpfti + u1t , b2i lpfti + u2t ,

i=1
= c2 + 2 zt1 +
i=1
b1i lpcti +
where lpct and lpft indicate the natural logarithm of the prices of the last transaction of the series of the IBEX-35 and the midquote of the futures contract; c1 and c2 are constant; p indicates the number of lags and zt1 is the error correction term that is obtained from the following expression zt1 = 1 lpct1 c 2 lpft1 ,
this section, the term integration is used in an econometric sense. A series is integrated of order one if it contains a unit root. A series of this type becomes stationary or integrated or order zero when taking rst dierences. 12 The Schwartz Bayesian Criterion has been used to determine the number of lags of the error correction models. Subsequently, we proved the presence of serial correlation. If correlation did not appear, the chosen number of lags was that proposed for the criterion. Conversely, if serial correlation problems were detected, the number of lags was increased until eliminating the correlation. We should also point out that the number of lags proposed was the same for the both equations and both variables.
11 In
268
where 1 and 2 indicate the parameters of the cointegrating vector.13 The estimates of error correction coecients in the spot (1 ) and the futures (2 ) equations are presented in Table VI. 1 is signicant, negative and higher than 2 in absolute value for every day, while 2 is only signicant on the 27th of October. These results suggest that the spot market responds to the deviation from longrun equilibrium in (t 1) for every day except for the 27th , where a simultaneous adjustment is observed in the spot and futures markets. Furthermore, the absolute value of 1 diminishes strongly on the 28th and 29th . This indicates a smaller response of the spot market to the disequilibrium between spot and futures prices during the Asian crisis.14 In short, the measures based on statistical and econometric techniques contradict each other when determining the absence or presence of market integration. For example, according to measure 1 (spotf ut ) the markets are more integrated on the 28th than the 29th (Table II) while, according to measure 4, the markets are not integrated on the 28th and they are on the 29th at the 5% level (Table VI).
6.2
Measures based on cross-market arbitrage
The study of nancial integration measures based on the basic principles of assets valuation traditionally starts by measuring the degree of fulllment of the LOP. Although the absence of arbitrage opportunities (AAO) is stronger than LOP (AAO implies LOP but the converse fails in a general framework), they are equivalent conditions in the particular case of a stock index and its replica.15 Consequently, the study of the fulllment of the LOP between the futures market on IBEX-35 and its underlying asset (and the risk-less asset) in fact embraces the study of all the possible arbitrage opportunities. Furthermore, Pardo (1998) proved the equalities g = a and m = M in this particular context. Measures p, g and m have been calculated for each minute of the days considered and are summarized in Table VII (second, third, fourth and fth lines). If we do not consider the transaction costs all the measures indicate market disintegration. The maximum disintegration is observed during the 28th , 29th and 30th . On the 28th measure p takes the maximum and the minimum values of the period and, also, measures g and m reach their highest values. The greatest integration level is detected on the 22nd , 23rd and 24th . We also highlight the fact that the minutes with overvaluations of spot prices with regard to the futures prices (Ct > Ct ) are greater than those of the
13 The models have been estimated for each day and they do not include intercept in the cointegration equation on the 29th and 30th , while the rest of the days include intercept in the cointegrating equation and in the error correction vector. 14 The eects of infrequent trading in stocks are modeled through the methodology proposed by Jokivuolle (1995) to proxy for the true index prices. The results do not dier signicantly from those obtained without carrying out this adjustment and they are available upon request. 15 See Appendix.
269
undervaluations (Ct > Ct ) on both the Asian crash days and the other days (sixth and seventh line). As previously explained, measures p and m allow transaction costs to be discounted. Therefore, we can analyze if disparities in prices of the asset and its replica are or not explained by them. Transaction costs have been considered for each day taking into account market fees, commissions and market impact costs in the spot and futures markets.16 Having carried out this correction, the measures p and m lead to similar results. Complementarily, the eight and ninth lines of Table VII show the number of detected opportunities of direct and reverse cash-and-carry arbitrage strategies. In the days prior to the crash, all (except one) of the deviations between cash and futures prices are explained by the transaction costs (arbitrage opportunities do not exist and markets are integrated). However, during the 28th and 29th , most of the deviations are not explained by transaction costs (arbitrage opportunities exist and markets are not integrated). In these circumstances, we still detect the prevalence of inverse arbitrage opportunities except on October 30th when the number of direct arbitrage opportunities is greater. To sum up, the level of integration shown by the integration measures based on cross-market arbitrage coincides as much in stable periods as in volatile periods.
6.3
Measures based on equilibrium models
Finally, we have calculated the dierence between the natural logarithms of the theoretical spot price and of the observed spot price and we have tested whether the mispricings follow a mean reversion process (measure 11).17 We used the Augmented Dickey-Fuller (ADF) test (Dickey and Fuller, 1981) to check the presence of mean reversion in the mispricings. The results are reported in Table VIII. The estimated ADF values support the absence of a unit root, with the exception of October 27th and 28th . Hence, the mispricing series for both days are non-stationary and, according to measure 11, the markets are not integrated. The remaining days, the mispricing series behaves as a stationary series and, therefore, the spot and futures markets are integrated. Note that the parameter is bigger in absolute value the days outside the Asian crisis period. This indicates a bigger convergence from prices to the equilibrium level during those days and the presence of certain disintegration during the Asian crash.
16 The estimated transaction costs oscillate between 19 and 22 basic points, on the 22nd and 28th of October, which implies a m value of 0.0038 and 0.0043, respectively. 17 Miller et al. (1994) indicate that the mean reversion of the changes in the base is a statistical illusion, caused by the infrequent trading of stocks within the index. The empirical evidence obtained by Neal (1996) contradicts the previous results. This author examines 837 S&P 500 index arbitrage trades and he observes a relationship between arbitrage and mean reversion of the mispricing series.
270
The integration measure of Bessembinder (measure 12) has not been calculated for two reasons. First, because this measure needs long time series with low frequency and, second, because our empirical application includes only one futures markets and inference with regard to asset pricing models can be sensitive to the exclusion of securities from the cross-sectional analysis (Bessembinder, 1992, pages 639-640). If we compare these results with those obtained with the measures based on statistical and econometric techniques, we can conclude that spot and futures markets were integrated in the stable subperiod (the 22nd , 23rd and 24th ). Nevertheless, the results on market integration are partially contradicted in the volatile subperiod (the 27th , 28th , 29th and 30th ).18 The comparison with the results achieved when applying the measures based on cross-market arbitrage shows that the lower the mean reversion parameter ( ), the greater the existence of arbitrage opportunities is.
Conclusions
The paper empirically tests the eectiveness of a large number of market integration measures, and the analysis justies the convenience of classifying them into two major categories: statistical measures and measures related to the theory of asset pricing. A large number of measures are operationalized and their values are computed during a period characterized by disintegration and the eect of the Asian Crisis of October 1997. The results clearly reveal that measures based on statistical and econometric techniques contradict each other in volatile periods, and that the level of integration shown by measures based on cross-market arbitrage coincides as much in stable as in volatile periods. Besides, measures based on cross-market arbitrage and measures based on equilibrium models complement each other. In short, when measures based on theoretical approaches show integration, measures based on statistical and econometric techniques do, but the converse is not true in a general framework. Hence, pricing models must be taken into account when a measure of market integration is being developed. These measures may be dened by arbitrage methods or by equilibrium arguments. The rst group is appropriate if derivative markets are involved or if hedging strategies are the main purpose of the analysis. Instead, the second is useful to study well-diversied portfolios in incomplete markets. Anyway, there are certain measures that can be applied in both types of settings.
18 See
the values of spotf ut and in Tables II and VIII.
271
Appendix
The fulllment of the Law of One Price (LOP) and the absence of arbitrage opportunities are equivalent properties in some restricted contexts. Let us consider two dates t < T and three securities denoted by S1 , S2 and S3 . S1 will be a risk-less asset, S2 a risky one and S3 a futures contract on S2 with T maturity. Suppose that S2 does not pay any dividend between t and T and denote its price by I (t) > 0 at t and by I (T ) 0 at T . It is clear that I (t) must be a concrete numerical value while I (T ) must be a random variable. As usual, r > 0 will represent the interest rate between t and T and, consequently, 1/(1 + r) and 1 are the prices of S1 at t and T respectively. Finally, denote by F (t, T ) the future (at T ) price of S2 that can be guaranteed by S3 . Lemma 1 Under latter assumptions, there are no arbitrage opportunities in the model if and only if the Law of One Price holds. Proof. Assume that LOP holds. Then, I (t) = F (t, T ) or I (t) (1 + r) = F (t, T ) (1 + r) (9)
Let x = (x1 , x2 , x3 ) be an arbitrary portfolio composed by xi units of Si (i = 1, 2, 3) and denote by P (t) and P (T ) its numerical and random prices at t and T respectively. If x were an arbitrage portfolio, then P (t) 0 and P (T ) 0 should hold. Hence, the proof will be nished should the fulllment of the LOP and latter inequalities lead to P (t) = P (T ) = 0. Obviously, P (t) = and P (T ) = x1 + x2 I (T ) + x3 (I (T ) F (t, T )) = x1 (x3 F (t, T )) + I (T ) (x2 + x3 ) . Since P (T ) 0 must hold for any nal value of the random variable I (T ) price (or payo) of S2 , the following inequalities have to be fullled: x1 (x3 F (t, T )) 0 , x3 x2 . (9) and (11) lead to x1 x3 I (t) (1 + r) . We obtain from (10) that x1 x2 I (t) (1 + r) , and, thus, bearing in mind (12), x1 x2 I (t) (1 + r) x3 I (t) (1 + r) . Therefore, (13), (15), (11) and (12) must be equalities and P (t) = P (T ) = 0. (15) (14) (13) 0, future (11) (12) x1 + x2 I (t) 0 , (1 + r) (10)
272
Table I Integration measures between a nancial market and its derivatives markets
Measures based on statistics and econometric techniques
Series in dierences
Series in levels
Measures based on the asset pricing theory Measures based on cross-market arbitrage Measures based on equilibrium models Fulllment of the Law of Fulllment of the Law of One Price One Price and Absence of Garbade and Silbers Arbitrage Opportunities model (1983) Relationship between future contract and its underlying asset Put-call parity Measures of weak and strong integration of Chen and Knez (1995) Measures of weak and strong integration of Balb as and Mu noz (1998) Mean reversion of the mispricing series Tests of futures markets integration of Bessembinder (1992) Characteristics
Cross-correlation technique Granger causality (1969) Sims causality (1972)
Engle-Granger test for cointegration (1987) Cointegration procedure proposed by Johansen (1988 and 1991) and Johansen and Juselius (1990) Characteristics Integration level for a xed date Lead-lag relationships cannot be established
Characteristics
Integration for one period Lead-lag relationships can be established
Medium, last transaction or closing prices
Bid and ask, medium, last transaction or closing prices Transaction costs can be taken into account
Transaction costs cannot be taken into account
Integration for one period Lead-lag relationships cannot be established Medium, last transaction or closing prices Transaction costs cannot be taken into account
Relationships between concrete markets
Concrete strategies
Transaction costs cannot (resp., can) be taken into account by measures of Chen and Knez (resp., Balb as and Mu noz) Non-predetermined strategies
Concrete strategies
Table II Statistics for the ibex-35 cash index and the ibex-35 futures contract
The rst column shows the variables whereas the results for the corresponding day appear on the remaining columns. The second row shows the number of observed returns. The third row gives the close to close variation of the IBEX-35 stock index. The fourth (fth) row gives the IBEX-35 stock index (IBEX-35 futures contract) volatility obtained as the standard deviation of the minute to minute returns. The sixth row provides the spot returns autocorrelation coecient (spot ) and its p-value appears in parenthesis. The seventh row gives the autocorrelation coecient of the futures returns (f ut ) and its p-value appears in parenthesis. The eighth row shows the relative bid-ask spread of the futures contract (Sf ut ). The ninth and tenth rows give the transaction volume on the spot and futures market in millions of pesetas and number of contracts, respectively.
VARIABLE Observations Variation spot f ut spot f ut Sf ut V olspot V olf ut
22 415 -0.40% 0.046% 0.049% 0.077 (0.115) 0.115 (0.019) 0.05 138433.35 21525
23 415 -2.49% 0.063% 0.069% 0.085 (0.083) 0.037 (0.455) 0.05 124031.96 30179
24 415 -0.79% 0.056% 0.047% 0.197 (0) 0.007 (0.88) 0.05 82628.55 22351
27 415 -4.40% 0.074% 0.084% 0.192 (0) 0.139 (0.004) 0.06 102194.46 38374
Asian crisis 28 401 -4.18% 0.125% 0.235% 0.147 (0.003) 0.175 (0) 0.11 218310.71 93130
29 379 5.66% 0.078% 0.117% 0.233 (0) 0.131 (0.01) 0.07 151003.4 54816
30 407 1.12% 0.090% 0.118% 0.327 (0) 0.245 (0) 0.06 148770.57 61231
273
274
Table III Cross-correlation of minute-to-minute intraday returns Cross-correlation of minute-to-minute intraday returns for stock index and stock index futures. The rst and last columns show the number of lags (k ). The rest of the columns gives the cross correlation spot,f ut(k) for the corresponding day. The tstatistic appears in parenthesis. The numbers in bold are signicant at the 1% level.
Asian crisis 28 0.189 (3.789) 0.168 (3.368) 0.177 (3.548) 0.211 (4.223) 0.224 (4.492) 0.136 (2.723) 0.193 (3.859) 0.116 (2.315) 0.120 (2.403) 0.018 0.364 -0.020 -0.398
k -5 -4 -3 -2 -1 0 1 2 3 4 5
22 -0.001 (-0.024) 0.138 (2.809) 0.154 (3.127) 0.210 (4.276) 0.360 (7.338) 0.246 (5.005) 0.091 (1.852) 0.060 (1.220) -0.023 (-0.477) 0.031 (0.623) -0.045 (-0.921)
23 0.036 (0.735) 0.027 (0.548) 0.047 (0.947) 0.110 (2.231) 0.448 (9.126) 0.333 (6.782) -0.008 (-0.163) -0.011 (-0.230) -0.013 (-0.261) 0.085 (1.734) 0.001 (0.014)
24 -0.001 (-0.014) 0.051 (1.037) 0.023 (0.477) 0.229 (4.667) 0.478 (9.729) 0.334 (6.800) 0.007 (0.147) 0.096 (1.954) 0.046 (0.937) -0.054 (-1.106) -0.017 (-0.336)
27 0.083 (1.691) 0.064 (1.306) 0.123 (2.512) 0.249 (5.081) 0.476 (9.701) 0.297 (6.052) 0.125 (2.544) 0.070 (1.428) 0.024 (0.493) 0.013 (0.261) -0.075 (-1.536)
29 0.012 (0.232) 0.123 (2.389) 0.164 (3.199) 0.277 (5.395) 0.235 (4.579) 0.116 (2.266) 0.101 (1.972) 0.074 (1.439) 0.154 (3.004) 0.110 (2.134) -0.060 (-1.166)
30 0.057 (1.156) 0.077 (1.547) 0.208 (4.190) 0.408 (8.233) 0.512 (10.323) 0.430 (8.669) 0.193 (3.900) 0.002 (0.042) 0.004 (0.073) -0.015 (-0.305) -0.179 (-3.607)
k -5 -4 -3 -2 -1 0 1 2 3 4 5
275
Table IV Phillips-perron test for unit roots in stock index and stock index futures prices ZL and ZD are the Phillips-Perron statistics of the series in levels and in rst differences, respectively. For a model with intercept the MacKinnon critical values are -2.868 and -3.448 at the 1% and 5% levels, respectively. 22 -1.348 -18.880 23 -1.296 -18.725 Panel A: lpct 24 27 28 1.820 -1.898 -2.319 -16.390 -17.096 -17.964 Panel B: lpft 24 27 28 2.332 -0.601 -0.605 -20.300 -17.742 -16.754 29 -0.872 -16.080 30 0.147 -14.894
ZL ZD
ZL ZD
22 -1.186 -18.160
23 -1.236 -19.525
29 -0.484 -16.731
30 -0.393 -15.612
Table V Johansen cointegration test results for stock index and stock index futures prices The rst column shows the corresponding day and the number of observations and lags are in parenthesis. i (i = 1, 2) is the estimated value of the characteristic root (eigenvalue). The last column gives the statistic trace that tests the null hypothesis, which, versus a more general alternative, considers that the number of distinct cointegration vectors is lower or equal to r. Each day has an intercept in a cointegration equation. October 24th and 30th have intercept and deterministic trend. * , ** and *** denote signicance at the 1% , 5% and 10% level. Critical values of the trace statistic are obtained from Osterwald-Lenum (1992).
Day 22 (416,6) 23 (416,3) 24 (416,4) 27 (416,4) 28 (402,9) 29 (380,7) 30 (408,7) r r r r r r r r r r r r r r H0 =0 1 =0 1 =0 1 =0 1 =0 1 =0 1 =0 1 i 0.058 0.009 0.116 0.006 0.078 0.001 0.052 0.006 0.052 0.004 0.029 0.006 0.062 0.003 trace 28.153* 3.657 52.992* 2.301 33.943* 0.578 24.361** 2.286 22.556*** 1.600 13.229** 2.341 26.625* 1.082
276
Table VI Error corrections model Parameter estimates of the error correction model for the IBEX-35 spot and futures prices. 1 (2 ) give the coecient of the error correction term in a spot (future) equation and the t-statistic is in parentheses.
22 -0.152 (-4.221) 0.066 (1.464) 23 -0.299 (-6.960) -0.053 (-0.943) 24 -0.269 (-5.677) -0.039 (-0.756) 27 -0.150 (-4.562) -0.103 (-2.309) Asian crisis 28 29 -0.076 -0.057 (-4.519) (-3.201) -0.011 0.003 (-0.289) (0.084) 30 -0.158 (-5.008) -0.098 (-1.911)
1 t-stat. 2 t-stat.
Table VII Measures based on cross-market arbitrage The rst column shows all the measures. The rest of the columns give the results for the corresponding day. The second (third) row gives the maximum (minimum) value of p. The fourth (fth) row shows the maximum values of g and m. The sixth (seventh) row gives the number of minutes in which the contemporaneous spot price (Ct ) is lower (higher) than the theoretical spot price (Ct ). The eighth row shows the number of cash-and-carry (C.C.) arbitrage opportunities. The ninth row shows the number of reverse cash-and-carry (R.C.C.) arbitrage opportunities. The transaction costs have been computed for the corresponding day.
VARIABLE Maximum p Minimum p Maximum g Maximum m Ct > Ct Ct > Ct C.C. R.C.C. 22 1.002382 0.997282 0.001547 0.002718 148 268 0 0 23 1.002769 0.997058 0.001653 0.002942 100 316 0 0 24 1.002459 0.995792 0.002358 0.004208 53 363 0 1 27 1.003566 0.989183 0.005674 0.010817 202 214 0 8 Asian crisis 28 1.038149 0.970087 0.017762 0.036747 31 371 27 322 29 1.004643 0.986578 0.007035 0.013422 53 327 4 125 30 1.010057 0.995893 0.005316 0.009957 224 184 20 1
Table VIII Mean reversion in mispricing series Test of unit roots in mispricing series. The variable ADF represents the Augmented Dickey-Fuller test of a single unit root in the mispricing series. The critical value of ADF at 1% level is 3.449. is the mean reversion parameter of the mispricing process and its p-value appears in parentheses.
22 -8.278 -0.276 (0.000) 23 -4.645 -0.241 (0.000) 24 -8.717 -0.351 (0.000) 27 -1.785 -0.074 (0.075) Asian crisis 28 29 -0.920 -3.549 -0.011 -0.091 (0.358) (0.000) 30 -3.714 -0.146 (0.000)
ADF p-value
277
References
[1] Abhyankar, A.H. (1995): Return and volatility dynamics in the FT-SE 100 stock index futures markets, Journal of Futures Markets 15, No. 4, pages 457488. [2] Arshanapalli, B. and Doukas, J. (1997): The Linkages of S& P 500 Stock Index and S& P 500 Stock Index Futures Prices during October 1987, Journal of Economic and Business 49, pages 253266. [3] Balb as, A.; Longarela, I. R. and Pardo, A. (2000): Integration and Arbitrage in the Spanish Financial Markets: an Empirical Approach, The Journal of Futures Markets 20, no. 4, pages 32144. [4] Balb as, A. and Mu noz, M. J. (1998): Measuring the Degree of Fulllment of the Law of One Price. Applications to Financial Markets Integration, Investigaciones Econ omicas, Vol. XXII (2), pages 153177. [5] Bessembinder, H. (1992): Systematic Risk, Hedging, Pressure, and Risk Premiums in Futures Markets, Review of Financial Studies 5, No. 4, pages 637667. [6] Chan, K. (1992): A further analysis of the lead-lag relationship between the cash market and Stock Index Futures Market, Review of Financial Studies 5, No. 1, pages 123152. [7] Chen, Z. and Knez, P.J. (1995), Measurement of Market Integration and Arbitrage, Review of Financial Studies 8, No. 2, pages 56379. [8] Cornell, B. and French, K. (1983a): The pricing of stock index futures, Journal of Futures Markets 3, pages 114. [9] Cornell, B. and French, K. (1983b): Taxes and the pricing of stock index futures, Journal of Finance 38, No. 3, pages 675694. [10] Dickey, D.A. and Fuller, W.A. (1981): Likelihood ratio statistics for autoregressive time series with a unit root, Econometrica 49, pages 10571072. [11] De Jong, F. and Donders, M.W.M. (1998): Intraday lead-lag relationships between the futures, optins and stock market, Discussion Paper no. 108, Tilburg University. [12] Engle, R.F. and Granger, C.W. (1987): Cointegration and error correction representation, estimation, and testing, Econometrica 55, No. 2, pages 251276.
278
[13] Garbade, K.D. and Silber, W.L. (1983): Price movements and price discovery in futures and cash markets, Review of Economics and Statistics 65, No. 2, pages 289297. [14] Granger, C.W.J. (1969):Investigating causal relations by econometric models and cross-spectral methods, Econometrica 37, pages 424438. [15] Harris, L. (1989): The October 1987 S& P 500 stock-futures basis, Journal of Finance 44, No. 1, pages 7799. [16] Ingersoll, J.E. Jr. (1987): Theory of Decision Making, Ed. Rowman& Littleeld Publishers Inc., Maryland. [17] Johansen, S. (1988): Statistical analysis of cointegration vectors, Journal of Economic Dynamics and Control 12, pages 231254. [18] Johansen, S. (1991): Estimation and hypothesis testing of cointegrating vectors in Gaussian Vector autorregresive models, Econometrica 59, pages 15511580. [19] Johansen, S. and K. Juselius (1990): Maximum likelihood estimation and inference on cointegration with applications to the demand for money, Oxford Bulletin of economics and statistics 52, pages 169210. [20] Jokivuolle, E. (1995): Measuring true stock index value in the present of infrequent trading, Journal of Financial and Quantitative Analysis 30, No.3, pages 455464. [21] Kawaller, I.G.; Koch, P.D. and Koch, T.W. (1987): The temporal price relationship between S& P 500 futures and the S& P index, Journal of Finance 5, pages 13091330. [22] Kempf, A. and Korn, O. (1998): Trading System and Market Integration, Journal of Financial Intermediation 7, pages 220239. [23] Kleidon, A.W. (1992): Arbitrage, nontrading and Stale prices: october 1987, Journal of Business 65, No 4, pages 483507. [24] Kleidon, A.W. and Whaley, R.E. (1992): One market? stocks, futures and options during October 1987, Journal of Finance 47, No. 3, pages 851877. [25] Lee, C. and Mathur, I. (1999): Eciency Tests in the Spanish Futures Markets, Journal of Futures Markets 19, No. 1, pages 5978.
279
[26] Miller, M.H.; Muthuswamy, J. and Whaley, R. (1994): Mean reversion of Standard& Poors 500 index basis changes: arbitrage induced or statisticall illusion?, Journal of Finance 49, No. 2, pages 479513. [27] Modest, D. (1984): On the pricing of stock index futures, Journal of Portfolio Management 10, No. 4, pages 5157. [28] Modest, D. and Sundaresan, M. (1983): The relationship between spot and futures prices in Stock Index Futures Markets: some preliminary evidence, Journal of Futures Markets 3, No. 1, pages 1541. [29] Neal, R. (1996): Direct tests of Index Arbitrage Models, Journal of Financial and Quantitative Analysis 31, No. 4, pages 541562. [30] Ng, N. (1987): Detecting spot forecasts in futures prices using causality tests, Review of Futures Markets 2, pages 250267. [31] Osterwald-Lenum, M. (1992): A Note with Quantiles of the Asymptotic Distribution of the Maximum Likelihood Cointegration Rank Test Statistics, Oxford Bulletin of Economics and Statistics 54, pages 461472. [32] Pardo, A. (1998): Integraci on y arbitraje entre el mercado espa nol de renta variable y su mercado derivado, PhD Thesis, University of Valencia, Spain. [33] Protopapadakis, A. and Stoll, H.R. (1983): Spot and Futures Prices and the Law of One Price, Journal of Finance 38, No. 5, pages 143155. [34] Sims, C.A. (1972): Money, income and causality, American Economic Review 62, No. 4, pages 540552. [35] Stephan, J.A. and Whaley, R.E. (1990): Intraday price change and trading volume relations in the stock and stock option markets, Journal of Finance 45, pages 191220. [36] Stoll, H.R. and Whaley, R.E. (1990): The dynamics of stock index and stock index futures return, Journal of Financial and Quantitatie Analysis 25, No. 4, pages 441468. [37] Sutclie, C. (1997): Stock Index Futures: Theories and International Evidence, 2nd. ed., Cambridge, International Thomson Business Press. [38] Wang, G.H.K. and Yau, J. (1994): A time series approach to testing for market linkage: unit root and cointegration tests, Journal of Futures Markets 14, No. 4, pages 457474.
280
[39] Yadav, P.K. (1992): Studies on stock index futures pricing: a UK perspective, PhD Thesis, University of Straathclyde, UK, June. [40] Yadav, P.K.; Pope, F.P. and Paudyal, K. (1994): Threshold autoregressive modeling in nance: the price dierences of equivalent assets Mathematical Finance 4, No. 2, pages 205221.
Angel Pardo Departamento de Econom a Financiera y Matem atica Facultad de Ciencias Econ omicas y Empresariales Universidad de Valencia Campus Naranjos, Edicio Departamental Oriental 46071 Valencia e-mail: Angel.Pardo@uv.es
Alejandro Balb as Departamento de Econom a de la Empresa Universidad Carlos III de Madrid C/ Madrid, 126 28903 Getafe (Madrid) e-mail: balbas@emp.uc3m.es
Vicente Meneu Departamento de Econom a Financiera y Matem atica Facultad de Ciencias Econ omicas y Empresariales Universidad de Valencia Campus Naranjos, Edicio Departamental Oriental 46071 Valencia e-mail: Vicente.Meneu@uv.es
Crash Prediction: Science or Alchemy?

Common Points Between Earthquakes, Sand Piles and Markets
Gabriele Susinno1
The Market has its law. Were there no law, there could be no centre about which prices could revolve, and therefore, no market. R. N. Elliot, The Wave Principle (1938) Abstract Financial markets evolution is the macroscopic result of a huge number of human interactions. The term human is what separates social sciences and natural sciences. Indeed the unpredictability of human behaviour is often taken as the key factor for the impossibility to apply methods borrowed from physics to describe, explain, and predict market evolution. Despite this fact there are a number of market players struggling to nd the Holy Grail by identifying market ineciencies allowing them to beat the market, and some of them seem to have succeeded (at least for some time). From a more scientic viewpoint, a number of scientists have analysed markets phenomenology trying to deduce the hidden rules governing the macroscopic indicators of human interactions. Indeed, similarities between nancial markets and a hierarchical system of interacting particles may be an interesting way to enhance the synergy between social and physical sciences. From a physicist viewpoint one of the most interesting phenomena in nancial markets is the occurrence of crashes. What triggers market to start to act co-operatively in a short time window? Is it possible to identify precursors to such catastrophic events? Maybe concepts of scaling, universality and self-organised systems bring some new light to understand this phenomena.
Introduction
Most of the advanced models used in nance are based on basics assumptions that allow for the pure mathematical tractability. Albeit these models allow for a good estimation of fair derivative prices for liquid assets during stable periods, they become
1 Gabriele Susinno es Director de Capital Management Advisors. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de abril de 1999.
282
Gabriele Susinno
hardly applicable in turmoil. Instable periods are characterised by extreme market movement whose behaviour is not in the applicability domain of most models. Two major concepts are necessary to cope with large market uctuations: a correct estimation of eects induced by large uctuations in theoretical prices and hedging factors, and a qualitative understanding of causes producing market instabilities. Catastrophes on the derivatives market awaken the conscience about implicit risks taken in neglecting extreme uctuations. Indeed, extreme events appear to be more frequent than those predicted by the elementary diusion processes used in nancial modelling (e.g. geometric Brownian motion). Today the quality of information available from any kind of source allows for a critical view of the traditional Black-Scholes dogma. Pricing mistakes or inaccuracies, produced by unrealistic assumptions, may constitute a serious danger for market players. Risk against which they hardly nd a correct protection. Potters et al. 1998 [11] have shown, by studying in detail the market prices of options on liquid markets, that the market has empirically corrected the simple, but inadequate Black-Scholes formula to account for two important statistical features of asset uctuations: fat tails and correlation in the amplitudes of uctuations. These aspects, although not included in conventional pricing models, are reected in the price xed by the market as a whole. This means that models inadequacies are complemented, in average, by the feeling of market participants and in that sense liquid markets behave as rather ecient adaptive systems. Large uctuation eects become of utmost importance during instability periods and crashes. It is during those periods that most of the mathematical models show their instabilities and failures just because the fundamental assumptions are no longer satised. In this framework some fresh air brought by physicists may be benecial to nance. In particular crashes are of particular interest for physicists since they seem to share common points with complex systems. Indeed the challenge is to understand how simple non-linear behaviours acting repetitively may induce large co-operative behaviours. Large-scale patterns may arise from a series of repetitive interactions on many scales that lead to a progressive build-up of large-scale correlation and culminate with a crisis.
Learning Form Nature
In Physics a critical point indicates regimes of large-scale co-operative behaviour. Fundamentally it is a phenomenon resulting from the repeated interaction between microscopic elements of a system, which progressively construct a macroscopic selfsimilar state. A system in a critical state lies at the boundary between order and disorder, it exhibits strong correlation between his dierent elements, and an extreme susceptibility to external factors (a good textbook for an educated reader is reported on ref [6]). Classical examples of such systems are liquids and magnets,
283
where the system progressively orders under small external changes. A typical example of long range ordering, which we use frequently on our every day life (e.g. magnetic tape recording), is the ferromagnetism. Bulk ferromagnetic materials exhibit long range order which creates regions (domains) where magnetic moments of neighbouring atoms are locked into a rigid parallel order over a large number of atoms in spite of thermal agitation which tends do destroy any atomic level order. But for a given ferromagnetic material the long-range order abruptly disappears at a certain temperature, which is called the Curie temperature. The abrupt change from long range order to disorder may share common points whit market crashes. Another interesting experimental observation is that near the critical point systems have no characteristic length, which means that if a part of the system is magnied as large as the original it is impossible to distinguish the magnied part and the original. Near a market crash, individual investors tend to follow big investors, who in turn follow bigger investors in a coherent and co-operative way, a generalised euphoria as prelude of a nancial distress. In nature some systems naturally evolve towards a critical state, the so-called Self-Organised Critical (SOC) systems [1]. The expression of an underlying unstable dynamical critical point may be found in systems like sand-piles, earthquakes, forest res, etc. For instance, in sand-piles, starting from an undercritical slope with a steady supply of grains, the pile grows until the structure become unstable and an avalanche is initiated. In this way, the pile reaches a stationary critical state, characterised by a critical slope, in which additional grains of sand will fall o the pile via avalanches distributed in lifetime and size according to a power law (Fig. 1).
Figure 1: Number/Amplitude versus Size/frequency relation for sand piles and S&P500. Both systems exhibit a power-law behaviour characteristic of scale invariance.
284
Gabriele Susinno
Albeit the general conditions under which a physical system exhibits SOC are largely unknown, some facts have however been established: The large-scale evolution should obey a diusion process (like markets are supposed to behave) which satises a global conservation law. More generally a feedback mechanism must operate which attracts the dynamics to a critical state. The second point is one of the most neglected in mathematical modelling in nance since market and market players are considered as totally independent and not inuencing each other [5].
What about Markets?
It could be interesting to understand market dynamics and to identify crash precursors from a fundamental structure of the market evolution. In a general view nancial markets seem to exhibits a scale invariant behaviour [4]. If for a physicist, scale invariance means the absence of characteristic scale, from a business point of view, it means the existence of catastrophic risk, which can bankrupt a company (in the same way the critical slope of sand piles may produce catastrophic avalanches). It is rather tempting to identify the mechanism responsible for the scale invariance with Self-Organised Criticality [13] i.e. the expression of an underlying unstable dynamical critical point. In general co-operative behaviours of complex systems cannot be reduced to a decomposition of elementary causes. A crash emerges naturally as an intrinsic signature of the functioning of the market. Indeed there is a need to insert in market models the eect of a positive feedback interaction in which traders exchange information according to a hierarchical structure [14]. As a general comment, it is interesting to note the existing similarities between log-periodic structure observed in the market [14] and Elliot waves [12], an strongly rooted technique in the nancial analysis folklore. Crash statistics are very poor and the natural excitation, that any kind of evidence of a crash precursor can produce, should be always moderated by critical views. Albeit it is tempting to see nancial crashes as a critical phenomenon described by statistical mechanics where, in a particular situation, all the subparts of the system react cooperatively, there are only few experimental situations where these scenarios can be tested. In that sense, any empirical nding is not statistically signicant and could be even more dangerous than the crash itself if its conclusions are used to assess a strategy [7], based on an ex ante prediction. To use such empirical results, the lack of statistically relevant set of experimental congurations must be compensated by a a priori knowledge of the market conditions. Any conclusion on a market crash prediction, based on empirical results, must be handled as a bayesian inference on magnitude and time-scale on the next catastrophic event given the knowledge of the past.
285
Scaling and Criticality in Financial Markets
Which is the cause of the scaling laws and volatility time dependency observed in nancial Markets? Two main hypotheses may be made They have the same origin than those characterising physical systems with a large number of interacting units. They reect an exogenous cascade of information. Numerical simulation [3, 9, 10] tends to support the rst hypothesis. Suppose that a crash is a natural consequence of a slow build-up of a long-range correlation between market agents. In this case the crash will appear as a second order phase transition where the system abruptly moves from an ordered to a disordered state, very much like ferromagnetism at the Curie temperature. In this case the natural evolution of the system towards a critical state via a selfreinforcing imitation process can be quantied by the theory of critical phenomena. In this case the critical time, i.e. the time at which the crash occurs, plays the same role as the Curie temperature in ferromagnetism. In an ecient market, where the global available information is reected in current market price, it may be rational for traders to remain invested even if they fear to the risk of a crash. Indeed the remuneration for the risk of a crash can triggers a higher growth rate of assets in an inationary bubble. In this case market price reects the wish of buyers who hope for a price increase and the fear of sellers that it may fall. Again, as in the example of ferromagnetism, two forces regulate each trader: the opinions of his colleagues (which tend to order the system), and his own decisions, which tend to generate disorder. Moreover a global information is available to each trader. Taking forward the analogy with physics, each trader can be modelled as a two state unit for which buy corresponds to the state +1 and sell to the state 1. In this case the state si of a unit i with N(i) neighbours can be modelled as [13, 14]: si = sign I
j N (i )
sj + i i + G ,
where i describes the tendency for imitation, the idiosyncratic noise induced by the decision of unit i, i a random number from N (0, 1), and G an exogenous global inuence. By symmetry, in absence of a global inuence, the average state of the system is zero since units in the state +1 are balanced by units in the state 1. The presence of a global inuence (positive or negative) will move the system since units in one state will outnumber the others. The susceptibility of the system, i.e. the sensitivity of the average state of the system to an innitesimal change in
286
Gabriele Susinno
information will then be a direct measurement of the hazard rate2 . Indeed a critical point is characterised by the divergence of the susceptibility of the system, where innitesimal changes of the conditions induce the collapse. A common feature of systems near their critical state is that their susceptibility evolves as a power law, i.e.: |t tc | , where is called the critical exponent. In the vicinity of the crash the divergence of the correlation implies a power-law divergence of the hazard rate which is strictly related to the assets price evolution, and this until the crash occurs or the inationary bubble deates. An additional feature needs to be added. Indeed, for hierarchical systems the critical exponent may be a complex number, in this case the power-law divergence near the critical point starts to be decorated by log-periodic oscillations whose frequency diverges approaching the critical time. Considering market players organised in a hierarchical structure, organised by size and therefore inducing a dierent market impact may not be such an unrealistic assumption. An exponential evolution of market indexes modulated with log-periodic oscillations of increasing frequency may be a signature of the increasing probability of a future market crash, and the estimation of the critical time can be obtained by tting the model on observed data. Dynamical critical points exhibit a characteristic log-periodic signature. In this case, interesting results have been obtained by Sornette et al. [14] and Vandewalle et al. [15] in the analysis of the two major crashes of this century: October 1927 (Dow Jones) , October 1987 (S&P500), and October 1997. They found, using historical data before the crashes, a critical time (i.e. the market crash) in very good agreement with the real timing of the events. They suggest that this agreement reect the fundamental co-operative nature of the behaviour of stock markets.
Physics and Finance
Unfortunately it is impossible to perform experiments in the market, a scientist can only observe its evolution and guess from it the underlying mechanism which could be responsible for the observed pattern. In such an investigation game, the role of a physicist appears as a connector between all dierent and well-established disciplines in economical science, ranging from the actuarial approach to equivalent martingale theories. Indeed the instinct of a physicist is to model the observed phenomenology, to understand the fundamental principles, and to test for the validity of all the assumptions of existing models, constitute his principal asset in this eld.
2 Hazard rate dened as the probability per unit time that a crash, which has not already happened, will happen in the next instant.
287
Some people do have problems to really understand the interplay between physics and nance. It is so just because, in the mind of the majority of people, physics has always belonged to the abstract world of the deep secrets of the nature. How can the study of stars or subatomic particles be applied to nance? It probably cannot be applied directly, but all the technologies, ideas, and computational tools, which are used for the quest of the natures secrets, may be of some use in nance. But the most valuable contribution to the eld may be brought by the natural instinct of physicists to rationalize and then model the problems they tackle. In that respect, two complementary approaches appear. The observation of the market dynamics and all the properties driving the time evolution of traded quantities, allows for a guess on the underlying process followed by the observed phenomena. In the other hand it is of utmost importance to understand why we observe what we observe, what are, if any, the fundamental rules followed by the markets and producing the observed evolution. Indeed it is interesting to analyze the possible correspondence between thermodynamic models describing the evolution of a gas of particles and the market where particles are replaced by brains of market players. But if in thermodynamic the interactions are known and well dened, in nance they can evolve and be modied so that a model, which is valid today, will not be anymore so tomorrow. Market practitioners seem to solve this problem by continuously re-calibrating their models, which in turn is like adjusting the force strength on a day-by-day basis. Is there some intrinsic and unavoidable feature of the market itself or what we think to be the fundamental axioms are only the eect of more fundamental quantities? By understanding this question it will be possible to answer in a much more realistic, and maybe precise way to all VaR dreamers.
Conclusions
Since the beginning of this century, with the seminal work of Bachelier, nancial world has attracted a rising interest from a wide range of scientists. Applications of scientic methodologies in this eld have an increasing success and attract an increasing number of applied mathematicians. They performed an enormous analytical work contributing to the realisation of solid basis for the mathematical nance. Unfortunately there is a sizeable mismatch between the complexity, the mathematical sophistication of models and the rough description of the underlying object to which they apply. It is exactly here where physicists can bring their contribution. Indeed, quoting Richard Feynman, the physicist cannot understand the mathematicians care in solving an idealised physical problem. The physicist knows the real problem is much more complicated. It has already been simplied by intuition which discards the unimportant and often approximates the remainder. It is clear that physicists do not have the pretension to come into the nancial game bringing, with their methods, the absolute way of solving problems.
288
Gabriele Susinno
Quantitative skills, and their dierent approach to problems, can contribute to enhance the synergy in nancial research. Adapting the existing models and theories in such a way, they correctly reproduce the real world observations or, at least, give a coherent estimation of the risk implied in any nancial transactions. The basic principle allowing for a real progress in any scientic model is based on the following strategy: First you guess. Dont laugh, this is the most important step. Then you compute the consequences. Compare the consequences to experience. If it disagrees with experience, the guess is wrong. In this simple statement is the key to science. Richard P. Feynman One of the basic dreams of the growing community of nancial physicists or econophysicists is to achieve a general understanding of market movements and dynamics, by identifying the underlying elementary processes responsible for it. To do so they have in mind elementary questions, which often do not have an elementary answer. Why prices move? What determines the amplitude and time scale of those movements? How can the dierent processes inuencing demand and supply be identied and mathematically modeled? Since few years physicists from academies have started to collaborate between them to formulate an answer to many question left open in nancial modeling. There is no certitude to contribute with a brand new theory of everything to which physicists are aectionate. However a joint eort from the various scientic cultures involved in nancial research may lead to interesting results.
References
[1] Per Bak, How Nature Works : The Science of Self-Organized Criticality, Copernicus Books; ISBN: 0387947914. [2] P. Bouchaud et al., Risk, 9, no. 3 (March 1996). [3] S.H. Chen, T. Lux and M. Marchesi, Testing for Non-Linear Structure in an Articial Financial Market, www.ge.infm.it/~ecph/papers/marchesipapers/marchesipapers.html [4] R. Cont, Proceedings of the CNRS Workshop on Scale Invariance, Les Houches, 1997. [5] J. Cvitanic, in Mathematics in Derivative Securities, Cambridge University Press (1997) and references therein. [6] K. Huang, Statistical Mechanics, John Wiley & Sons. [7] L. Laloux et al., cond-mat/9804111.
289
[8] D. Lando, in Mathematics in Derivative Securities, Cambridge University Press (1997) and references therein. [9] T. Lux and M. Marchesi, A Micro-Simulation of Interactive Agents, www.ge.infm.it/~ecph/papers/marchesipapers/marchesipapers.html [10] T. Lux and M. Marchesi. Scaling and Criticality in a Stochastic Multi-Agent Model of a Financial Market, Nature 397 (1999), 498500. [11] M. Potters et al., Europhysics Letters, 41, no. 3 (February 1, 1998). [12] A. Prost and R. Prechter, Elliot Waves Principle, New Classic Library, 1985. R. Beckman, Elliot Wave Explained, Probus Publishing, 1992. [13] D. Sornette, in Physics of Complexity, Editions Frontieres. [14] D. Sornette et al., cond-mat/9510036. D. Sornette and Anders Johansen, condmat/9704127. [15] Vandewalle, N. et al., How the nancial crash of October 1997 could have been predicted, Eur. Phys. J. B 4, 139-141.
Gabriele Susinno Capital Management Advisors e-mail: susinno@cma-it.com
Una modelizaci on GARCH bivariante del futuro sobre el Ibex-35. Implicaciones para la cobertura del contado
Juan Angel Lafuente1
Resumen En este trabajo se discute la relaci on entre el ratio de inversi on optimo y el ratio de cobertura optimo cuando existe un mercado de futuros para el activo de renta variable cuyo valor de mercado se pretende inmunizar. Si bien a nivel te orico el ratio unitario constante ser a la soluci on optima tanto en el problema de inversi on como de cobertura, los supuestos requeridos son demasiado restrictivos. En la operativa real, ambos problemas tender an a tener soluciones diferentes. El trabajo propone tres metodolog as alternativas para la estimaci on de ratios de cobertura din amicos, evalu andose su efectividad en t erminos de reducci on de volatilidad respecto a la observada en el mercado de contado (posici on no cubierta). La simulaci on de una determinada estrategia sugiere que los ratios estimados con una metodolog a GARCH multivariante consigue una efectividad similar a la que proporcionan tanto el ratio unitario constante como los estimados a partir de modelos GARCH univariantes. Sin embargo, un menor n umero de contratos por t ermino medio es utilizado2 .
Introducci on
La excelente aceptaci on de los instrumentos derivados en el sistema nanciero espa nol ha incrementado considerablemente las posibilidades operativas con carteras de renta variable. En particular, desde Enero de 1992, existe la posibilidad de negociar contratos de futuros con el ndice Ibex 35 como activo subyacente, instrumento derivado especialmente relevante para efectuar coberturas sobre carteras de renta variable, es decir, inmunizar el valor de mercado de una cartera frente a las posibles uctuaciones de cotizaciones de los activos que la integran. Para ello, el valor del ratio de cobertura determina el n umero o ptimo de contratos de futuros que es necesario vender al llevar a cabo la operaci on.
1 Juan Angel Lafuente es profesor ayudante del Departamento de Finanzas y Contabilidad de la Universidad Jaume I de Castell on. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de mayo de 1999. 2 El autor desea agradecer a Alfonso Novales los valiosos comentarios y sugerencias recibidos. Cualquier error es de mi exclusiva responsabilidad.
292
Juan Angel Lafuente
Con objeto de proceder a la estimaci on del ratio de cobertura en los mercados de futuros nancieros, se propuso inicialmente en la literatura la utilizaci on de regresiones a partir de muestras de cambios en los precios o rendimientos, siendo las variables end ogena y ex ogena las correspondientes al mercado de contado y de futuros respectivamente (v ease, por ejemplo, Ederington (1979)). Los estudios recientes acerca de la estimaci on del ratio de cobertura o ptimo han se nalado fundamentalmente tres cuestiones relevantes acerca de la metodolog a inicialmente planteada: a) la existencia de una relaci on de equilibrio a largo plazo entre los precios del mercado de futuros y los precios del mercado del contado. Esta pauta debe ser incorporada al especicar el modelo econom etrico a partir del cual los ratios de cobertura ser an estimados; v ease por ejemplo Ghosh (1993), Lien y Luo (1993) o, m as recientemente, Lien (1996); b) evidencia emp rica acerca de la existencia de heterocedasticidad condicional autorregresiva para las series de rendimientos de los mercados. Si las distribuciones condicionales dieren de las distribuciones incondicionales, no tiene sentido plantear el estimar un ratio de cobertura constante, puesto que tanto la varianza condicional del mercado de futuros como la covarianza condicional entre los precios de ambos mercados ser an cambiantes en el tiempo, y c) la presencia de una relaci on causal entre ambos mercado, con unas determinada caracter sticas acerca de la interacci on entre rendimientos y volatilidad. La simplicaci on de estas relaciones din amicas intertemporales puede introducir sesgos en la estimaci on convencional de los ratios de cobertura (v ease, por ejemplo, Theobald y Yallup (1997)). En este trabajo se revisa la relaci on entre el ratio de cobertura de o ptimo y el ratio de inversi on o ptimo. A partir de un modelo de dos periodos se observa que la expresi on anal tica de ambos ratios es id entica si a) el mercado de derivados es eciente, es decir, el precio del contrato en un periodo es la proyecci on del precio del contrato en el periodo siguiente o b) la correlaci on entre el precio del mercado de futuros y el precio del mercado de contado es perfecta. En la pr actica ambos supuestos son excesivamente simplicadores, de forma que ambos ratios tender an a no ser iguales. En este sentido, incluso en el caso de que el activo que se pretende inmunizar sea una cartera r eplica del activo subyacente del contrato de futuros utilizado para la cobertura el ratio no deber a ser unitario. Partiendo de la base de que el ratio de inversi on y cobertura o ptimo ser an diferentes, el trabajo se centra en la estimaci on de este u ltimo cuando la cartera a cubrir es una r eplica del Ibex 35. Para ello se proponen tres alternativas metodol ogicas. Con objeto de comparar su potencialidad operativa a partir de simulaciones ex-post de operaciones de cobertura se calibra la ecacia en t erminos de reducci on de volatilidad respecto a la observada en el mercado de contado (posici on no cubierta), y tomando como marco de referencia el ratio miope (ratio unitario y constante). Los resultados ponen de maniesto que, si bien el ratio estimado teniendo en cuenta los tres aspectos se nalados anteriormente posee una ecacia similar a la obtenida mediante la aplicaci on del ratio unitario, un menor n umero de contratos por t ermino medio es requerido.
n GARCH bivariante del futuro sobre el Ibex-35 Una modelizacio
293
El trabajo est a estructurado como sigue. En la secci on 2 se presenta la expresi on te orica del ratio de cobertura a partir del modelo de dos periodos habitualmente utilizado en la literatura, as como su relaci on con el ratio que subyace al problema de inversi on o ptima. La secci on 3 muestra en detalle las diferentes alternativas metodol ogicas que se han considerado para la estimaci on del ratio de cobertura. En la secci on 4 se analiza la ecacia de la cobertura, comparando los resultados que se obtienen al aplicar el ratio estimado con las diferentes alternativas propuestas en la secci on 3. La secci on 5 resume las principales conclusiones obtenidas en el trabajo y se nala l neas de investigaci on futura.
El ratio de cobertura y el ratio de inversi on
Cuando se pretende efectuar una operaci on de cobertura el objetivo es inmunizar el valor de la cartera frente a las futuras uctuaciones de los precios de los activos que la integran. Para proceder a la operaci on de cobertura, una vez jado el horizonte de la misma, el elemento crucial es el ratio de cobertura, el cual determina el n umero de contratos de futuros que deber an ser vendidos. El fundamento de la operaci on radica en que las p erdidas de un mercado ser an compensadas con las ganancias en el otro mercado, dado que la posici on de contado es larga mientras que la correspondiente al mercado de derivados es corta. La cobertura puede ser utilizada con el objetivo de bloquear las p erdidas que est a experimentando una cartera, o alternativamente para garantizar los benecios obtenidos hasta una determinada fecha. La aplicaci on de ratios sesgados conlleva una p erdida de efectividad. Supongamos que para una cartera, la verdadera beta( ) es superior a uno, pero el operador la estima con un < 0 y se cubre con futuros frente a una posible ca sesgo negativo, es decir da de cotizaciones spot. En este caso el operador est a infracubierto: podemos resumir el resultado nal en un fen omeno dicot omico: a) si el mercado de contado experimenta un incremento de precios, los benecios de contado ser an superiores a las p erdidas en futuros, b) si el mercado se comporta a la baja las p erdidas de contado ser an superiores al benecio en la posici on de derivados. Por tanto, el operador no estar a ortogonalizando el valor de su cartera respecto al valor de la cartera del mercado. La expresi on del ratio de cobertura para un operador que pretendiera utilizar el contrato de futuros sobre el Ibex 35 para proteger una determinada cartera es: (Valor de la cartera) Beta . (Ibex 35) Multiplicador (1)
Seg un la anterior expresi on, en el caso de que la cartera sea una r eplica del Ibex 35, el ratio de cobertura deber a ser unitario, puesto que la beta de la cartera del mercado es la unidad. Este resultado conlleva impl citamente unos determinados supuestos. Veamos a continuaci on este aspecto.
294
Juan Angel Lafuente
2.1
El ratio de cobertura o ptimo. Minimizaci on de la volatilidad de la riqueza nal
Consideremos un agente que en el periodo t tiene una parte de su riqueza invertida en dos activos, uno libre de riesgo y otro con riesgo (renta variable) y pretende cubrir el valor de su cartera frente al comportamiento estoc astico del precio del activo con riesgo. Supongamos que no hay costes de transacci on, y que la toma de posiciones en el mercado de futuros no requiere ning un desembolso de riqueza (no existen dep ositos en garant a). El problema de optimizaci on al que se enfrenta este agente ser a: min {bt } s.t. V ar (Wt+1 /t ) V art (Wt+1 ) Wt+1 = (1 + r) [Wt pt qt ] + pt+1 qt + (ft+1 ft ) bt , (2)
astica) en el periodo t +1, Wt is la riqueza inicial, donde Wt+1 es la riqueza nal (estoc t es el conjunto de informaci on disponible al comienzo del periodo t, r es el tipo de inter es del activo libre de riesgo, pj (j = t, t + 1) hace referencia al precio del activo con riesgo en el periodo j , fj (j = t, t + 1) es el precio del contrato de futuros en el periodo j , qt es el n umero de activos con riesgo adquiridos al comienzo del periodo t y bt es el n umero de contratos de futuros vendidos en el periodo t para cubrir la posici on de contado. A partir de la restricci on presupuestaria, teniendo en cuenta que pt , ft t :
2 V art (Wt+1 ) = qt V art (pt+1 ) + b2 t V art (ft+1 ) + 2 qt bt Covt (pt+1 , ft+1 ) .
(3)
La condici on de primer orden del problema es: 2 bt V art (ft+1 ) + qt Covt (pt+1 , ft+1 ) = 0 . Por tanto, la soluci on o ptima viene dada por la siguiente expresi on: b t qt = Covt (pt+1 , ft+1 ) . V art (ft+1 ) (4)
El signo negativo indica que la operaci on de cobertura para una posici on larga en el mercado de contado requiere vender contratos de futuros (posici on corta en el mercado de derivados), mientras que el valor absoluto del ratio indica la proporci on entre ambas posiciones. La condici on de segundo orden corrobora que el candidato a o ptimo corresponde a un m nimo global del problema: 2 V art (Wt+1 ) = 2V art (ft+1 ) > 0 . b2 t
295
2.2
El ratio de inversi on o ptimo. Maximizaci on de la utilidad esperada de la riqueza nal
A continuaci on relacionemos el problema anterior con el de un agente que decide conjuntamente la inversi on en el activo con riesgo y la posici on en el mercado de derivados. Supongamos adem as que el inversor posee una funci on de utilidad U () von Neumann-Morgenstern, denida sobre el periodo nal t + 1, estrictamente creciente, estrictamente c oncava y dos veces diferenciable. El problema de optimizaci on al que se enfrenta el agente es max {qt , bt } s.t. E [U (Wt+1 ) /t ] Et U (Wt+1 ) Wt+1 = (1 + r) [Wt pt qt ] + pt+1 qt + (ft+1 ft ) bt . Et [U (Wt+1 ) [pt+1 (1 + r) pt ]] Et [U (Wt+1 ) (ft+1 ft )] = 0, = 0, (5)
Las condiciones de primer orden son:
donde Et es la esperanza condicional al conjunto de informaci on disponible en el periodo t. La condici on de segundo orden identica un m aximo global del problema (5), puesto que U (Wt+1 ) pt+1 (1 + r) pt ft+1 ft (pt+1 (1 + r) pt , ft+1 ft )
es siempre una matriz denida negativa. Bajo el supuesto de que la distribuci on condicional conjunta del vector (Wt+1 , pt+1 , ft+1 ) es Normal, a partir de las ecuaciones (2) y (3), aplicando el resultado3 obtenido en Rubinstein (1976), el valor o ptimo del ratio de inversi on es
qt b t
1 t+1/t
Et
U (Wt+1 ) U (Wt+1 )
p rt,t +1 f rt,t +1
(6)
donde t+1/t es la matriz de varianzas-covarianzas condicional del vector de precios del activo con riesgo y el contrato de futuros,
p rt,t +1 = Et (pt+1 ) (1 + r ) pt
f rt,t +1 = Et (ft+1 ) ft .
A partir de la ecuaci on (6) la expresi on del ratio de inversi on o ptimo puede reescribirse de la siguiente forma: bt qt
3 Sea
f p V art (pt+1 ) rt,t +1 Covt (ft+1 , pt+1 ) rt,t+1 p f V art (ft+1 ) rt,t +1 Covt (ft+1 , pt+1 ) rt,t+1
(7)
el vector aleatorio (X, Y ) N (, ) y sea g () una funci on diferenciable. Entonces dg (X ) Cov (X, Y ) . Cov [g (X ) , Y ] = E dX
296
Juan Angel Lafuente
2.3
Relaci on entre ambos problemas
Relacionemos a continuaci on las expresiones (4) y (7). Para ello, la expresi on (7) puede reescribirse de la siguiente forma: bt qt donde t,t+1 = y t+1/t = Covt (ft+1 , pt+1 ) [V art (ft+1 )] [V art (pt+1 )] .
Covt (ft+1 , pt+1 ) + t,t+1 , V art (ft+1 )
(8)
f 2 t+1/t 1 V art (pt+1 ) rt,t+1 p f [V art (ft+1 )] rt,t +1 Covt (ft+1 , pt+1 ) rt,t+1
(9)
A partir de (7) y (8) es inmediato observar que la discrepancia entre ambos ratios viene dada por (9). Por tanto ambos ratios son equivalentes si y solo si t,t+1 = 0, lo cual ocurrir a si:
f a) rt,t +1 = Et (ft+1 ) ft = 0, es decir Et (ft+1 ) = ft ,
b) o bien
2 t+1/t 1 V art (pt+1 ) = 0, lo que implica t+1/t = 1, puesto que
V art (pt+1 ) > 0. Interpretemos a continuaci on estos resultados. Si la correlaci on entre el precio del activo con riesgo y el del contrato de futuros es perfecta entonces la nube de puntos correspondiente a ambos precios debe estar sobre una l nea recta, es decir, existen valores y tal que pt = + ft , de forma que es posible formar una cartera de riesgo nulo a partir de la combinaci on lineal de ambos activos. En este caso, no solo t,t+1 = 0, sino que Covt (ft+1 , + ft+1 ) Covt (ft+1 , ft+1 ) Covt (ft+1 , pt+1 ) = = =. V art (ft+1 ) V art (ft+1 ) V art (ft+1 ) Por tanto, en el caso en que no existe riesgo de base, el ratio de cobertura podr a obtenerse a trav es de la estimaci on por m nimos cuadrados del par ametro4 en la anterior ecuaci on. Te oricamente todos los residuos deber an ser iguales a cero. Por otro lado, si el mercado de futuros es eciente, es decir, el precio de un periodo reeja toda la informaci on relevante, entonces la mejor predicci on del precio del contrato para el periodo siguiente es el precio cruzado del contrato en el periodo presente. Este es el caso en el que los precios del contrato de futuros evolucionan seg un un paseo aleatorio.
4 N otese que si, como es habitual, los precios presentan una ra z unitaria, entonces ser a necesario utilizar series de rendimientos o de cambios en los precios. En este u ltimo caso se estimar a la ecuaci on: st st1 = + (ft ft1 ) + t .
297
Por tanto, bajo alguno de los dos supuestos anteriores el ratio de inversi on o ptimo es asimismo un ratio de cobertura o ptimo. En caso contrario, la discrepancia entre ambos ratios depender a del comportamiento estoc astico del error de predicci on acerca del precio del contrato de futuros. Si los agentes utilizan de forma o ptima el conjunto cabr a esperar que su error de predicci on fuese cero en media, de forma que por t ermino medio el ratio de uno u otro problema deber an ser iguales en t erminos estad sticos. A partir de la ecuaci on (1) hemos se nalado anteriormente que el ratio de cobertura para una cartera r eplica del subyacente del contrato de derivados deber a ser unitario. A continuaci on razonamos que este resultado puede obtenerse a partir los supuestos a) y b) junto con la hip otesis de expectativas racionales. En efecto, bajo la hip otesis de expectativas racionales el precio del contrato de futuros deber a ser un estimador insesgado y o ptimo del valor del subyacente en la fecha del vencimiento del contrato. En el caso del modelo anterior de dos periodos ft = Et pt+1 . A partir de la relaci on lineal entre ft+1 y pt+1 , aplicando esperanzas condicionales al conjunto de informaci on disponible en t: Et pt+1 = + Et ft+1 = + Et ft+1 , es decir, ft = + Et ft+1 = + ft , donde la u ltima igualdad es consecuencia del supuesto de paseo aleatorio para la evoluci on de los precios del mercado de futuros. Por tanto, la estimaci on de la recta de regresi on pt = + ft deber a producir estimaciones no signicativas de la constante, a la vez que no deber a rechazarse la hip otesis nula de que la pendiente fuese unitaria5 . Sin embargo, el ratio unitario tambi en puede interpretarse como soluci on o ptima en el caso de que el precio del contrato de futuros no presente discrepancia alguna respecto a su valoraci on cost of carry, es decir bajo la ausencia te orica de oportunidades on este caso con un sencillo ejemplo en tiempo de arbitraje6 . Ilustremos a continuaci continuo. Consideremos que el subyacente evoluciona de acuerdo a un movimiento browniano geom etrico: (10) dpt = p pt dt + t pt dz1,t , on spot, p,t y s,t son la media y desviaci on t pica condicional de donde pt es la cotizaci la rentabilidad del mercado de contado, y dz1,t = 1t dt, siendo 1t i.i.d. N (0, 1), un proceso de Wiener. De acuerdo con la valoraci on costofcarry, bajo la cual se supone que los mercados son perfectos, el tipo de inter es del activo libre de riesgo es estable a corto plazo, y la cuant a de los dividendos de que ser an pagados por cada uno de los activos que integran el subyacente es conocida, el precio te orico de un contrato de futuros en el momento del tiempo t con vencimiento en T es igual al coste de oportunidad de mantener una cesta r eplica del activo subyacente desde la fecha de valoraci on hasta
5 N otese en este caso que, si los precios del mercado de contado y el mercado de futuros son integrados de orden uno, entonces existe una relaci on de equilibrio a largo plazo, siendo el vector de cointegraci on (1,-1), es decir, la base es estacionaria. 6 En la operativa real pueden existir discrepancias (mispricing ) sin que ello implique una oportunidad rentable de arbitraje, dada la existencia de costes de transacci on y diferenciales de precios bid-ask.
298
Juan Angel Lafuente
la fecha de vencimiento, es decir, ft,T = pt e(rd) (T t) , (11)
donde pt es el valor de mercado del subyacente en el periodo t y (r d) es el coste neto asociado a mantener una cesta r eplica hasta el vencimiento. A partir de las ecuaciones (10) y (11), aplicando el Lema de Ito, podemos obtener la ley de evoluci on para el precio del contrato de futuros: dft,T = f ft,T dt + t ft,T dz1,t , donde7 f = p (r d). En este caso, bajo la ausencia de oportunidades te oricas de arbitraje se tiene que: Et dpt Et dft,T = = p pt,T dt , p pt,T dt ,
2
(12)
(13) (14)
Et (dpt dft,T ) = Et f p ft,T pt (dt) + (p + f ) pt ft,T t dz1,t dt +

2 +Et ft,T t dz1,t dt + pt ft,T t (dz1,t ) 2 2 dt . = f p ft,T pt (dt) + pt ft,T t 2
(15)
Por tanto, la covarianza entre los cambios instant aneos en los precios:
2 Covt (dpt dft,T ) = Et (dpt dft,T ) [Et dpt ] [Et dft,T ] = pt ft,T t dt .
(16)
Por otro lado, V art (dpt ) = V art (dft,T ) =

2 2 Et [dpt Et dpt ] = Et [t pt dz1,t ] = t pt dt , 2 2
(17) dt . (18)
Et [dft,T Et dft,T ] = Et [t ft,T dz1,t ] =
2 t
2 ft,T
Teniendo en cuenta las expresiones (16) (17) y (18) podemos observar que la correlaci on condicional de los cambios instant aneos de los precios es igual a la incondicional, siendo perfecta y positiva: Covt (dft,T , dpt ) = [V art (dft,T )] [V art (dpt )]
2 pt ft,T t dt 2 (t
= 1. dt
(19)
p2 t
dt)
2 t
2 ft,T
7 N otese que a partir de la ecuaci on (11), tomando logaritmos neperianos, retardando la ecuaci on un periodo y restando las expresiones correspondientes a t y t + 1 se obtiene que
ln
pt+1 pt
= ln
ft +1 ft
+ (r d) .
299
Por tanto, existen valores 1 y 2 > 0 tales que dpt = 1 + 2 dft,T por lo que el valor absoluto del ratio de cobertura e inversi on o ptimo ser a8 : Covt (dft,T , dpt ) [V art (dft,T )] [V art (dpt )] = = Covt [dft , (1 + 2 dft,T )] [V art (dft,T )] [V art (1 + 2 dft,T )] 2 V art (dft,T ) = 1. 2 V ar (df [V art (dft,T )] [2 t t,T )]
(20)
Sin embargo, existe evidencia en la literatura no solo acerca de que los precios de los activos no evolucionan a lo largo del tiempo seg un en paseo aleatorio (v ease, por ejemplo, Lo y Mackinlay (1999)), sino tambi en sobre la existencia de sistem aticas desviaciones entre el precio t eorico de un contrato de futuros seg un la valoraci on cost-of -carry y el precio cruzado en el mercado (mispricing ) (v ease Mackinlay y Ramaswamy (1988), Lim (1992), Miller et al. (1994), Yadav and Pope (1990, 1994), y B uhler y Kempf (1995), entre otros). Bajo estas pautas de comportamiento la soluci on al problema de cobertura e inversi on o ptima ser a diferente, sino que el ratio de cobertura para una cartera r eplica del subyacente en el contrato de derivados no deber a ser unitario, recogiendo el hecho de que en los mercados reales existe un riesgo de base. Volveremos sobre esta cuesti on en la secci on 4.
Estimaci on del ratio de cobertura
En este secci on se proponen diferentes especicaciones alternativas para estimar el ratio de cobertura de un cartera r eplica del Ibex 35.
3.1
Datos
El periodo muestral analizado abarca desde el 20 de Diciembre de 1993 hasta el 20 de Diciembre de 1996, el cual ha sido seleccionado por tres razones: a) como ya hemos observado anteriormente, atendiendo al n umero de contratos negociados, el despegue del mercado de derivados se produce entre 1992 y 1994, de forma que el nivel de liquidez alcanzado a nales de 1993, que permanece estable hasta nales de 1996, permite utilizar datos de alta frecuencia; b) las especicaciones del contrato de futuros no han sido objeto de modicaci on en dicho periodo; al respecto, debemos tener en cuenta que a nales de Enero de 1997 el multiplicador del contrato de futuros pas o de 100 ptas. a 1.000 ptas. por punto b asico, y c) el periodo considerado, que comprende aproximadamente el 33% del tiempo de vida del mercado de futuros, si bien es homog eneo en cuanto al nivel de liquidez en el instrumento derivado recoge un comportamiento claramente diferenciado para el mercado de contado, tanto en
8 T engase en cuenta que si el diferencial hace referencia al incremento instant aneo entre t y t + 1 entonces, tendiendo en cuenta que toda variable contenida en el conjunto de informaci on disponible en t es no estoc astica, se tiene que Covt ft+1,T , pt+1 = Covt dft,T , dpt y V art ft+1,T = V art dft,T .
300
Juan Angel Lafuente
lo que concierne a la din amica de precios como en lo que respecta al nivel de riesgo (consid erese el comportamiento de la Bolsa Espa nola durante el a no 1996, con alzas generalizadas de cotizaciones y bajo nivel de volatilidad, con relaci on a los mayores cambios de tendencia de precios observados en los a nos 1994 y 1995). Por tanto, el ejercicio de simulaci on realizado de la secci on 4 pueden considerarse robusto frente al comportamiento general del mercado de contado. El contrato de futuros considerado en cada momento del tiempo es siempre el de pr oximo vencimiento, puesto que es el que sistem aticamente presenta un mayor nivel de liquidez y en consecuencia el u nico susceptible de ser analizado con datos de alta frecuencia. Las observaciones disponibles para el contrato de futuros sobre el Ibex 35 no est an equiespaciadas en el tiempo, dado que la frecuencia de negociaci on, si bien es muy alta, no es uniforme a lo largo del periodo de negociaci on diario. Por tanto, es necesario crear series de precios sobre un intervalo temporal jo, a la vez que emparejar para cada observaci on del futuro el valor del ndice Ibex 35 en el mismo instante de tiempo (id entico minuto). La m axima frecuencia de observaci on muestral utilizada es de 5 minutos. En este sentido, cada d a de negociaci on es dividido en intervalos de 5 minutos, comenzando, como ya hemos dicho, a las 11:00 horas. Posteriormente seleccionamos el primer cruce de operaci on de futuro dentro de cada intervalo, y asociamos a dicho precio el valor del Ibex 35 en el mismo minuto en que se llevo a cabo la operaci on en el mercado de derivados. De esta forma, a priori, deberemos disponer de 72 observaciones diarias, puesto que cada d a estamos considerando 6 horas de negociaci on; no obstante, para algunos d as es inevitable la p erdida de alguna observaci on como fruto de una escasa liquidez ocasional en el mercado de futuros, fundamentalmente entre las 14:30 y las 15:30 horas. La muestra disponible a partir de los 743 d as de negociaci on utilizados contiene contiene 51.929 observaciones9 . El objetivo es generar estimaciones del ratio, al menos de frecuencia diaria, con objeto de simular una determinada estrategia de cobertura y calibrar posteriormente su ecacia en funci on de cada una de las alternativas de estimaci on propuestas. Existe contundente evidencia emp rica en la literatura acerca de que las series de precios de los mercados de futuros sobre ndice burs atil y los de su correspondiente subyacente presentan una tendencia com un o relaci on de equilibrio a largo plazo; para el mercado espa nol v ease Blanco (1998), Lafuente (1998) y Pardo y Climent (2000), entre otros. En este sentido, en la l nea de lo propuesto en Engle y Granger (1987) la representaci on VAR bivariante para la din amica de las series de precios (integradas de orden uno) debe incluir el t ermino de correcci on de error. Esta aspecto ser a tenido en cuenta al proponer las diferentes especicaciones econom etricas que pasamos a describir en las siguientes subsecciones.
9 Por
tanto, en media, se han perdido aproximadamente dos observaciones diarias.
301
3.2
Modelos de correcci on de error con perturbaciones homoced asticas
Una primera aproximaci on para la estimaci on de ratios de cobertura diarios es la estimaci on de una recta de regresi on a partir de las series de cambios en los precios, pero incluyendo la presencia de una relaci on de equilibrio a largo plazo entre los precios de ambos mercados. Es el modelo inicialmente propuesto en la literatura (v ease Ederington (1979)) al que se ha incorporado la presencia de un mecanismo de correcci on de error. En este sentido, para cada d a de la muestra, se estima por m nimos cuadrados el siguiente modelo: ln St = + ln Ft + ut1 + t , (21)
ndice y Ft es el donde B es el operador de retardos, = 1 B , St es el valor del precio cruzado del contrato de futuros, ut1 = ln St1 1 2 ln Ft1 es el t ermino de correcci on de error, estimado en una primera etapa. La estimaci on del modelo anterior se lleva a cabo cada d a de negociaci on a partir de los datos cada 5 minutos disponibles en el mismo desde las 11:00 horas hasta las 17:00 horas. La pendiente estimada del modelo es utilizada como estimaci on del ratio de cobertura; si tenemos en cuenta que como estimaci on del ratio de cobertura. La anterior especicaci on tiene asociado un bajo coste computacional de estimaci on, pero por contrapartida presenta la restricci on de que la perturbaci on t es homoced astica. La gura 1 muestra la evoluci on din amica de los ratios de cobertura estimados mediante la anterior metodolog a. Sin embargo tambi en existe abundante evidencia emp rica en la literatura acerca del comportamiento heteroced astico de las series de rendimientos; para el mercado espa nol, los trabajos de Blanco (1998), Lafuente (1998), Arag o (2000) y Le on y Mora (1999), entre otros, ofrecen evidencia de esta pauta de comportamiento tanto para el mercado de futuros sobre el Ibex 35 como para el mercado de contado. Es por ello, que en las siguientes subsecciones se proponen estimaciones alternativas del ratio de cobertura a trav es de modelos de heterocedasticidad condicional autorregresiva.
3.3
Modelos de heterocedasticidad condicional autorregresiva
En esta secci on consideramos la estimaci on del ratio de cobertura a trav es de modelos de correcci on de error con perturbaciones heteroced asticas. En este caso, dado el mayor coste computacional asociado a la estimaci on utilizaremos datos de rendimientos horarios, una vez depurada la muestra de los rendimientos overnight 10 .
10 Con la muestra con datos cada cinco minutos, se ha seleccionado precios horarios comenzando con el de las 11:00 horas. A partir de las siete observaciones de precios disponibles entre las 11:00 y las 17:00 horas se han generado los rendimientos horarios a partir de la primera diferencia logar tmica, eliminando posteriormente los rendimientos overnight . Por tanto, una vez as depurada la muestra, se dispone de seis observaciones diarias de rendimientos horarios para cada mercado. Por otro lado, t engase en cuenta que cuanto mayor es la frecuencia de observaci on muestral m as leptoc urtica es la
302
Juan Angel Lafuente
3.3.1
Un modelo bivariante de correcci on de error
Sean rs,t y rf,t los rendimientos del mercado de contado y de futuros, respectivamente, es decir rs,t = ln St ln St1 , y rf,t = ln Ft ln Ft1 . La din amica propuesta para la evoluci on de los rendimientos es descrita por la siguientes ecuaciones: rs,t rf,t = 11 21 12 22 rs,t1 rf,t1 + s f (st1 (1 + 2 ft1 )) + s,t f,t , (22)
on condicional Gaussiana: t = siendo t , el vector de perturbaciones, con distribuci (s,t f,t ) | t1 N (0, t ), donde t1 es el conjunto de informaci on disponible en el periodo t 1 y t la matriz de varianzas-covarianza condicional del vector de rendimientos. El t ermino (ln St1 (1 + 2 ln Ft1 )) es el factor de correcci on de error. Los coecientes s y f miden la respuesta ante desviaciones del equilibrio a largo plazo en el mercado de contado y futuros respectivamente. Si el correspondiente coeciente es estad sticamente signicativo, pero peque no en valor absoluto, entonces los precios de mercado tienen una escasa capacidad para ajustarse y recuperar el equilibrio de largo plazo. A partir de la formulaci on habitual, la din amica de la matriz segundos momentos condicionales correspondiente a un modelo GARCH(p,q ) viene dada por la siguiente expresi on: vecht = vech + q (B ) vech (t t ) + p (B ) vecht , (23) donde B es el operador de retardos, q (B ) = 1 (B ) + 2 B 2 + + q (B q ) , p (B ) = 1 (B ) + 2 B 2 + + p (B p ) ,
t es el vector de innovaciones y vecht vech = =

2 2 s,t sf,t f,t 2 s 2 sf f
, .
vech (t t ) =
2 2 s,t s,t f,t f,t
Esta ecuaci on puede asimismo expresarse como un modelo VARMA (vectorial autoregressive moving average ). Consideremos el siguiente vector estoc astico trivariante: t = vech t t vecht . Sustituyendo la ecuaci on (24) en la ecuaci on (23) y reordenando: + p (B ) t , r (B ) vech t t = vech (25) (24)
distribuci on de frecuencias incondicional as como que los modelos de heterocedasticidad condicional tienen una capacidad limitada para capturar el exceso de curtosis en las series de rendimientos.
303
donde r (B ) = [I (p (B ) + q (B ))], siendo r = max{p, q } y p (B ) = [I p (B )] .
Esto es, una representaci on ARMA(r, p) vectorial. De acuerdo con la ecuaci on (25), la especicaci on utilizada en el trabajo se corresponde con una media m ovil pura11 : + 1 B + 2 B 6 + 3 B 12 + 4 B 18 t . vech t t = vech (26)
Si las ra ces del polinomio caracter stico caen fuera del c rculo unidad, la representaci on anterior puede expresarse como un modelo autorregresivo de orden innito estacionario, es decir, la anterior formulaci on puede considerarse un modelo m as parsimonioso para capturar una estructura de dependencia entre los residuos al cuadrado a lo largo de un periodo de tiempo muy amplio. Como se nalan Bollerslev et al. (1994) el potencial n umero de par ametros en una formulaci on GARCH multivariante es aplastante, y, en consecuencia, toda especicaci on u til, dado un objetivo de partida, debe reducir el n umero de par ametros. En este sentido, se han introducido las siguientes restricciones: a) las matrices 2 , 3 , y 4 son diagonales,
1 b) y si denotamos por 1 ij , el elemento (i, j ) en la matriz 1 , se impone que 12 = 1 1 1 1 21 = 22 = 23 = 32 = 0.
A pesar de las restricciones introducidas, exclusivamente con objeto de evitar problemas num ericos en la estimaci on, el modelo se caracteriza por las interacciones cruzadas entre la volatilidad de cada mercado quedan recogidas 1 a trav es de los elementos 1 31 y 13 de la matriz 1 , los elementos de la diagonal principal de las matrices j (j = 1, 2, 3) capturan el patr on estacional de comportamiento intrad a en forma de U en la volatilidad y la covarianza de ambos mercados, relacionando cada momento condicional con el observado en la misma hora en d as previos. Adem as el modelo es una formulaci on m as general que las propuestas habitualmente en la literatura para el an alisis de los principales mercados de futuros sobre ndice en el sentido de que no se impone un coeciente de correlaci on condicional entre los rendimientos de ambos mercados constante a lo largo del tiempo (v eanse Park y Switzer (1995), Iihara et al. (1996), Koutmos y Tucker (1996), Ackert y Racine (1999), entre otros).
11 Esta especicaci on responde a las caracter sticas de las funciones de autocorrelaci on simple de los rendimientos al cuadrado, as como de la pauta detectada en la funci on de correlaci on cruzada entre los rendimientos al cuadrado de ambos mercados. Una explicaci on detallada del comportamiento de dichas funciones puede encontrarse en Lafuente (2000).
304
Juan Angel Lafuente
El modelo ha sido estimado por m axima verosimilitud exacta con la toolboox E412 . en la ecuaci Para estimar, jamos los tres elementos del vector de constantes vech on (26) igual a los segundos momentos incondicionales estimados con la muestra global de rendimientos de ambos mercados. Por tanto, el algoritmo de optimizaci on (BFGS ) no tiene en cuenta dichos par ametros en la estimaci on13 . En conjunto, es necesario estimar 19 par ametros en el modelo bivariante propuesto. A partir de los segundos momentos condicionales de los rendimientos recuperamos la estimaci on del ratio de cobertura a partir en un momento del tiempo t a partir de la siguiente expresi on14 : pt1 Covt1 ( ln St , ln Ft,T ) . ft1,,T V art1 ( ln Ft,T ) (27)
La gura 2 muestra la evoluci on del ratio de cobertura estimado a partir del modelo multivariante. 3.3.2 Dos modelos univariantes de correcci on de error
Si bien la anterior especicaci on en la din amica de la matriz de varianzas-covarianzas condicional posee un alto grado de parsimonia la estimaci on del modelo implica un alto coste computacional, ya que a pesar de que las restricciones impuestas el n umero de par ametros a estimar es elevado. Es por ello que la convergencia en gradiente en el proceso de estimaci on resulta costosa, reejando el hecho de que la funci on de verosimilitud presenta una supercie excesivamente plana. Frente a la especicaci on multivariante, en esta secci on proponemos la estimaci on de un modelo univariante de correcci on de error con perturbaci on GARCH para las series de rendimientos horarios de cada mercado. A partir de la estimaci on de la volatilidad de cada mercado, recuperaremos la covarianza condicional entre los rendimientos de ambos mercados a trav es del coeciente de correlaci on incondicional
12 Una descripci on detallada de la toolbox, desarrollada en el Departamento de Econom a Cuantitativa (Universidad Complutense de Madrid) puede encontrarse en Terceiro et. al. (2000). 13 Como se muestra en Engle y Mezrich (1996) esta restricci on mejora notablemente el comportamiento num erico en el proceso de estimaci on. 14 T engase en cuenta que dada una variable X , entonces
Xt Xt1 Xt1
ln
Xt Xt1
si Xt Xt1 0. Esta puede considerarse una buena aproximaci on en el caso de rendimientos horarios. Por tanto se tiene que: Covt1 ln
pt , ln f ft p t 1 t1,T ft1,T ft 1 1 pt1 ft1,T 1 2 ft 1,T
Covt1 (pt pt1 ) , ft ft1,T V art1 ft ft1,T ,
V art1 ln
ft1,T Covt1 (pt pt1 ) , ft ft1,T pt1 V art1 ft ft1,T
puesto que ft1,T y pt1 pertenecen al conjunto de informaci on disponible en t 1.
305
entre los rendimientos, estimado con toda la muestra. Ahora, por tanto, no solo se impone una correlaci on constante, sino que adem as no se permite la interacci on en media y en varianza entre los dos mercados. Como contrapartida, no restringimos la especicaci on de la parte regular del polinomio a una media m ovil pura. El modelo estimado para cada mercado es el siguiente: rm,t = rm,t1 + (st1 (1 + 2 ft1 )) + m,t ,
2 0, m,t
(28)
donde mt | t1 N , y el sub ndice m = s, f denota el mercado de contado o futuros. Siguiendo la notaci on de la ecuaci on (26), la din amica de la varianza condicional viene dada por la siguiente expresi on ARMA univariante para la innovaci on al cuadrado: (1 1 B )2 mt = 1m + 1m t1 + 2m t6 + 3m t12 + 4m t18 , (29)
2 donde t = 2 nal abamos anmt m,t , y B es el operador de retardos. Como se teriormente, ahora la especicaci on es menos restrictiva al permitir un par ametro autorregresivo en la parte regular del polinomio caracter stico de 2 mt . La gura 3 muestra la evoluci on del ratio de cobertura a partir de la estimaci on por m axima verosimilitud de los dos modelos univariantes, del que obtenemos las volatilidades estimadas para los rendimientos del mercado de contado y futuros, utilizando en la estimaci on de la covarianza la correlaci on incondicional.
Simulaci on de operaciones de cobertura
Con objeto de analizar la aplicabilidad pr actica de los modelos propuestos, en esta secci on se llevan a cabo simulaciones de operaciones de cobertura en donde se aplica el ratio estimado a partir de las tres metodolog as propuestas. Realizamos una simulaci on para cada uno de los 36 contratos disponibles tomando como punto inicial de riqueza una cesta de Ibex 35 al cierre del d a posterior al vencimiento del contrato inmediatamente anterior en el tiempo. Suponemos que la revisi on de la posici on del mercado de derivados se realiza diariamente15 al cierre del mercado de contado (16:55 horas), as como que la cartera a cubrir es una cesta r eplica del Ibex 35. En este sentido, asimismo se lleva a cabo el ejercicio de simulaci on aplicando el ratio unitario de forma sistem atica. A partir de las series de rendimientos obtenidas correspondientes a cada una de las 36 sendas de valores de mercado de la posici on global se eval ua la ecacia de la cobertura de las cuatro alternativas comparando la volatilidad de la rentabilidad correspondiente a la posici on cubierta con la volatilidad del mercado (posici on no cubierta), es decir tomamos como indicador de efectividad de la cobertura el siguiente estad stico: 100 dt (posicio n cubierta) dt ( posicio n no cubierta) , dt ( posicio n no cubierta) (30)
15 Por supuesto, la frecuencia con la que en la pr actica deber a revisarse la posici on de derivados no tiene que ser tan alta. El objetivo de las simulaciones es conocer que habr a sucedido si el operador hubiese tenido que corregir la posici on corta del mercado de futuros con alta periodicidad.
306
Juan Angel Lafuente
donde dt denota la desviaci on t pica de la senda de rentabilidad. Si el anterior ratio es negativo, cuanto mayor sea su valor absoluto una mayor reducci on de volatilidad respecto a la observada en el mercado de contado se habr a alcanzado. Las tablas 1 a 3 presentan los ratios medios aplicados con cada contrato seg un la metodolog a de estimaci on utilizada, mientras que las tablas 4 a 6 muestran el valor del estad stico dado en (30) no solo para cada una de las especicaciones econom etricas utilizadas para la estimaci on de la volatilidad de cada mercado, sino tambi en para el caso en que el ratio de cobertura aplicado es sistem aticamente unitario. A partir de la informaci on contenida en las tablas 1 a 6 cabe destacar los siguientes aspectos: a) si bien por t ermino medio el ratio medio aplicado a partir de la estimaci on del modelo de correcci on de error con perturbaciones homoced asticas es sistem aticamente menor, la ecacia de la cobertura es irrelevante a lo largo a lo largo de los vencimientos comprendidos entre Diciembre de 1994 y Diciembre de 1996, como maniesta la existencia de una contribuci on de volatilidad por t ermino medio respecto a la observada en el mercado; b) los ratios estimados a partir de la familia de modelos de heterocedasticidad condicional autorregresiva son superiores en el caso de la metodolog a univariante, es decir requieren de un mayor n umero de contratos por t ermino medio. Sin embargo, la reducci on de volatilidad que se consigue aplicando los ratios estimados con el modelo bivariante es superior a la obtenida con los ratios estimados a partir de los modelos univariantes en 29 de los 36 contratos analizados. Por otro lado cabe destacar que si bien el ratio unitario tiene asociado la mayor ecacia en t erminos de cobertura, la discrepancia con el modelo bivariante es marginal por t ermino medio. Sin embargo, es el caso en el que mayor n umero de contratos es sistem aticamente requerido. En resumen, las coberturas realizadas a partir de los ratios estimados procedentes del modelo bivariante de correcci on de error con ruidos heteroced asticos no solo requiere un menor n umero de contratos por t ermino medio sino que tienen asociado un mayor grado de efectividad respecto a las obtenidas con los modelos univariantes, tanto hetero como homoced asticos. La reducci on de volatilidad conseguida con el modelo bivariante es muy similar a la que se deriva del ratio miope unitario. Sin embargo un menor n umero de contratos es utilizado por t ermino medio, en torno a un 25% inferior a lo largo de toda de la muestra.
307
Resumen y conclusiones
En este trabajo se deriva la expresi on del ratio de cobertura o ptimo a partir de un modelo de dos periodos en donde un agente cubre su posici on de contado ex ogena a trav es de un contrato de futuros con id entico activo subyacente con el objetivo de minimizar la varianza condicional de su riqueza nal. Asimismo el trabajo relaciona el problema de cobertura anterior con el problema de inversi on o ptima en donde el agente decide conjuntamente su posici on de contado y de futuros y cuyo objetivo es maximizar la utilizada esperada de la riqueza nal. El trabajo muestra que el ratio de inversi on o ptimo es asimismo un ratio de cobertura o ptimo cuando a) la correlaci on entre el valor del subyacente y el precio del contrato de futuros es perfecta o b) el precio del contrato de futuros evoluciona seg un un paseo aleatorio. Sin embargo, ambos supuestos son excesivamente simplicadores de la realidad operativa de los mercados de futuros, por lo que ambos ratios tender an a ser diferentes. Partiendo de la idea anterior el trabajo se centra en la estimaci on del ratio de cobertura para una cartera r eplica del Ibex 35. Si bien a nivel te orico la soluci on o ptima ser a el aplicar un ratio sistem aticamente unitario, en la pr actica, el supuesto de ausencia de discrepancias entre el precio cruzado del contrato de futuros y su valoraci on cost-of-carry es excesivamente restrictivo. En este sentido, el trabajo propone tres metodolog as alternativas para la estimaci on del ratio; partiendo de una especicaci on de correcci on de error en media se considera tanto perturbaciones heteroced asticas como homoced asticas. Finalmente el trabajo calibra el potencial operativo de los ratios estimados a partir de un ejercicio de simulaci on de una determinada estrategia de cobertura con revisi on diaria de la posici on de derivados. Los resultados sugieren que a) los modelos con perturbaciones homoced asticas, estimados diariamente, no son efectivos en cuanto a reducci on de volatilidad y b) los ratios estimados a partir del modelo GARCH bivariante consiguen una reducci on de volatilidad superior a la obtenida con los modelos GARCH univariantes y similar a la que se deriva de la aplicaci on del ratio miope unitario. Sin embargo, un menor n umero de contratos es utilizado. Este trabajo abre nuevas l neas de investigaci on futura. Entre otras, ser a interesante calibrar la robustez de los resultados obtenidos con los modelos de correcci on de error y perturbaciones heteroced asticas para diferentes horizontes de cobertura as como para horas alternativas de revisi on. Por otro lado, resulta asimismo de inter es estimar el ratio dentro de la muestra y efectuar las operaciones de cobertura teniendo en cuenta los costes de transacci on. En este caso la revisi on de la posici on en el mercado de derivados tendr a una frecuencia menor, lo que podr a dar lugar a resultados cualitativamente diferentes.
308
Juan Angel Lafuente
Ap endice 1: Tablas estad sticas

Vencimiento Enero Febrero Marzo Abril Mayo Junio Julio Agosto Septiembre Octubre Noviembre Diciembre Total Vencimiento Enero Febrero Marzo Abril Mayo Junio Julio Agosto Septiembre Octubre Noviembre Diciembre Total Vencimiento Enero Febrero Marzo Abril Mayo Junio Julio Agosto Septiembre Octubre Noviembre Diciembre Total UNIV(a) 0.78 1.04 1.08 1.17 0.90 0.86 0.90 0.92 0.74 0.89 0.91 1.08 0.94 UNIV(a) 0.77 1.05 0.90 0.85 0.93 0.93 0.93 1.17 0.99 0.99 1.22 0.92 0.97 UNIV(a) 1.14 0.88 0.80 1.02 0.91 1.01 0.88 0.84 0.85 0.97 1.06 0.92 0.94 MULTIV(b) 0.78 0.78 0.76 0.73 0.71 0.78 0.69 0.76 0.73 0.71 0.75 0.79 0.75 MULTIV(b) 0.74 0.79 0.72 0.76 0.78 0.74 0.81 0.81 0.83 0.78 0.80 0.80 0.78 MULTIV(b) 0.79 0.80 0.76 0.81 0.83 0.82 0.79 0.81 0.79 0.80 0.80 0.81 0.80 MCE(c) 0.57 0.51 0.47 0.57 0.61 0.62 0.52 0.47 0.66 0.65 0.63 0.50 0.56 MCE(c) 0.73 0.38 0.73 0.63 0.62 0.71 0.53 0.45 0.41 0.54 0.42 0.54 0.56 MCE(c) 0.55 0.54 0.82 0.64 0.56 0.66 0.60 0.69 0.68 0.69 0.60 0.73 0.65
Tabla 1: Ratios de Cobertura Medios A no 1994
Notas:
(a)
Modelos de correcci on de error univariantes con perturbaciones GARCH (correlaci on incondicional igual a la incondicional). Modelo de correcci on de error bivariante con perturbaciones GARCH (correlaci on condicional no constante). Modelo de correcci on de error con perturbaciones homoced asticas.
(b)
( c)
309
Tabla 4. Reducci on de volatilidad respecto a la posici on no cubierta

Vencimiento Enero 94 Febrero 94 Marzo 94 Abril 94 Mayo 94 Junio 94 Julio 94 Agosto 94 Septiembre 94 Octubre 94 Noviembre 94 Diciembre 94 Total UNIV(a) 48% 49% 29% 17% 67% 57% 49% 60% 57% 54% 53% 48% 49% MULTIV(b) 58% 43% 75% 76% 68% 72% 40% 63% 72% 67% 78% 68% 65% MCE(c) 32% 18% 32% 46% 41% 60% 45% 35% 64% 57% 58% 43% 44% R=1(d) 57% 50% 85% 82% 83% 77% 56% 71% 83% 78% 75% 65% 72%
Notas:
(a)
Modelos de correcci on de error univariantes con perturbaciones GARCH (correlaci on incondicional igual a la incondicional). Modelo de correcci on de error bivariante con perturbaciones GARCH (correlaci on condicional no constante). Modelo de correcci on de error con perturbaciones homoced asticas. Ratio sistem aticamente unitario.
(b)
( c) (d)

Vencimiento Enero 95 Febrero 95 Marzo 95 Abril 95 Mayo 95 Junio 95 Julio 95 Agosto 95 Septiembre 95 Octubre 95 Noviembre 95 Diciembre 95 Total UNIV(a) 68% 7% 68% 71% 55% 67% 57% 55% 31% 77% 40% 58% 55% MULTIV(b) 65% 37% 62% 74% 68% 66% 71% 78% 72% 78% 80% 68% 68% MCE(c) -26%() -3% -31% -31% 7% -16% -12% 21% -85% -2% -17% -13% -21% R=1(d) 71% 25% 72% 78% 66% 71% 68% 80% 66% 87% 81% 75% 70%
Notas:
(a), (b), (c), (d) ( )
como en la tabla anterior.
Un signo negativo indica que la estrategia proporcion o un incremento de volatilidad respecto de la observada en el mercado de contado.
310
Juan Angel Lafuente

Vencimiento Enero 96 Febrero 96 Marzo 96 Abril 96 Mayo 96 Junio 96 Julio 96 Agosto 96 Septiembre 96 Octubre 96 Noviembre 96 Diciembre 96 Total UNIV(a) . 62% 72% 64% 75% 73% 71% 72% 72% 71% 69% 58% 76% 69% MULTIV(b) . 63% 71% 84% 79% 69% 74% 78% 79% 78% 76% 66% 81% 75% MCE(c) . -20%() -19% -27% -77% 0% -21% -11% -32% -51% -78% -7% -26% -31% R=1(d) 79% 7% 85% 82% 73% 79% 81% 85% 80% 81% 64% 85% 72%
Notas:
(a), (b), (c), (d) ()
como en las tablas anteriores.
En este caso la estrategia proporcion o un incremento de volatilidad respecto de la observada en el mercado de contado.
Ap endice 2: Gr acos
Figura 1: Ratios de Cobertura. Modelo de Correcci on de Error Homoced astico
311
Figura 2: Ratios de Cobertura. Modelo GARCH bivariante
Figura 3: Ratios de Cobertura. Modelo GARCH univariantes
312
Juan Angel Lafuente
Referencias
[1] Ackert, Lucy F. y M.D. Racine (1999), Stochastic trends and cointegration in the market for equities, Journal Of Economics And Business 51, no. 2, p aginas 133143. [2] Arag o, V. (2000), Vencimiento del contrato de futuros y ratio de cobertura de m nima varianza: Evidencia emp rica para diferentes horizontes temporales de inversi on., Revista Espa nola de Financiaci on y Contabilidad, en prensa. [3] Blanco, R. (1998), Transmisi on de informaci on y volatilidad ente el mercado de futuros sobre el ndice Ibex 35 y el mercado al contado, III Jornadas de Econom a Financiera, Volumen 1, p aginas 219289. [4] Bollerslev T., Engle R.E. y D.B. Nelson (1994), ARCH models, Handbook of Econometrics, Engle R.E. y D.L. McFadden (editores), Volumen 4. Elsevier, Amsterdam. [5] B uhler, W. y A. Kempf (1995), DAX index futures: mispricing and arbitrage in German markets, Journal of Futures Markets 7, p aginas 833859. [6] Engle, R.F. y C.W. Granger (1987), Cointegration and error correction: Representation, estimation and testing, Econometrica 55, p aginas 251276. [7] Ederington, L.H. (1979), The hedging performance of the new futures markets, The Journal of Finance 3, p aginas 157170. [8] Ghosh, A. (1993), Hedging with Stock Index Futures: Estimation and Forecasting with Error Correction Model, The Journal of Futures Markets 13 (7), p aginas 743752. [9] Engle, R.F. and C.W. Granger (1987), Cointegration and error correction: Representation, estimation and testing, Econometrica 55, p aginas 251276. [10] Engle, R.F. y J. Mezrich (1996), GARCH for groups, Risk 9 (8), p aginas 3640. [11] Iihara Y., Kato, K. and T. Tokunaga (1996), Intraday returns dynamics between cash and the futures markets in Japan, Journal of Futures Markets 16, p aginas 147162. [12] Koutmos G. and M. Tucker (1996), Temporal relationships and dynamics interactions between spot and futures stock markets, Journal of Futures Markets 16, p aginas 5569.
313
[13] Lafuente, J.A. (1998), Estrategias din amicas de cobertura en el mercado de futuros sobre el Ibex 35, III Jornadas de Econom a Financiera, Volumen 2, p aginas 85138. [14] Lafuente, J.A. (2000) Intraday return and volatility relationships between the Ibex 35 stock index and stock index futures markets, Working paper 00-02. Universidad Carlos III de Madrid. [15] Leon A. y J. Mora (1999), Modelling conditional heteroskedasticity: Application to the Ibex 35 stock-return index, Spanish Economic Review 1, no. 3, p aginas 215238. [16] Lien y Luo (1993), Estimating Multiperiod Hedge Ratios in Cointegrated Markets, The Journal of Futures Markets 13, no. 8, p aginas 909-920. [17] Lim, K. (1992), Arbitrage and price behavior of the Nikkei stock index futures, Journal of Futures Markets 12, p aginas 151161. [18] Lien, D.(1996), The Eect of the Cointegration Relationship on Futures Hedging: A Note, The Journal of Futures Markets 16, no. 7, p aginas 773780. [19] Lo, A. and A.C. Mackinlay (1999), A non-random walk down Wall street, Princeton Unversity Press. [20] Mackinlay, A.C. y C. Ramaswamy (1988), Index futures arbitrage and the behavior of stock index futures prices, Review of Financial Studies 1, p aginas 137158. [21] Miller, M.H., Muthuswamy, J. and R.E. Whaley (1994), Mean reversion of Standard & Poors 500 index basis changes: Arbitrage-induced or statistical illusion?, Journal of Finance 49, p aginas 479513. [22] Pardo, A. y F. Climent (2000) Relaciones Temporales entre el Contrato de Futuro sobre el Ibex 35 y su Activo Subyacente, Investigaciones Econ omicas 24 (1), p aginas 219235. [23] Park T.H. y L.N. Switzer (1995), Bivariate GARCH Estimation of the Optimal Hedge Ratios for Stock Index Futures: A Note, The Journal of Futures Markets 15, no. 1, p aginas 6167. [24] Rubinstein, M. (1976), The Valuation of Uncertain Income Streams and the Pricing of Options, Bell Journal of Economics 7, p aginas 407425. [25] Terceiro, J., Casals J.M., Jerez, M., Serrano G. and Sotoca (2000), A Matlab Toolbox for reliable time series modeling and forecasting in State-Space. ICAE working paper.
314
Juan Angel Lafuente
[26] Theobald, M. y P. Yallup (1997): Hedging Ratios and Cash/Futures Market Linkages, The Journal of Futures Markets 17, no. 1, p aginas 101115. [27] Yadav, P.K. and P.F. Pope (1990), Stock index futures arbitrage: International evidence, Journal of Futures Markets 10, p aginas 573603. [28] Yadav, P.K. and P.F. Pope (1994), Stock index futures mispricing: Prot opportunities or risk premia?, Journal of Banking and Finance 18, p aginas 921953.
Juan Angel Lafuente Departamento de Finanzas y Contabilidad Facultad de Ciencias Econ omicas y Empresariales Universidad Jaume I 12071 Castell on e-mail: lafuen@cofin.uji.es
Risk Assessment and Extreme Value Theory in Finance

rez1 Alberto Sua
Abstract The estimation and management of risk is an important and complex task with which market regulators and nancial institutions are faced. It has become apparent that accurate and reliable quantitative measures of risk are needed in order to avert, or at least minimize, the undesirable eects on a given portfolio of large uctuations in the conditions of the market. In order to accomplish this task, a number of mathematical tools are available. Classical risk analysis focuses on the estimation of Value at Risk, which is a percentile of the returns distribution at a given probability level for a given time horizon. Besides this standard measure, it is possible to carry out a more sophisticated risk analysis based on Extreme Value theory. A series of novel risk measures (Shortfall, MaxVaR, etc.) based in this theory have been introduced by a number of authors in order to address the insuciencies of VaR to properly characterize the response of a given portfolio to uctuations in market conditions. In order to illustrate the use of these tools, a comprehensive risk analysis for the IBEX35 stock index (Spain) is presented.
Introduction
The nancial portfolio of an institution comprises a number of products (bonds, assets, derivatives, and the like) that are openly traded in nancial markets. The prices of these products depend on the values of a number of risk factors: interest rates, asset prices, volatility, and the like. Market risk analysis consists in estimating the response of the portfolio to uctuations in the values of these risk factors. Portfolio managers may use this analysis to assess dierent strategies. In situations where the goal is to secure the value of a portfolio, risk analysis can be used to minimize the eects of unexpected uctuations in market conditions. Alternatively, a portfolio manager may attempt a competitive advantage through a calculated and carefully controlled exposure to risk. Market regulators, however, enforce restrictions to discourage an excessive or uncontrolled exposure to risk, which can lead to disruptions in the economic system. These restraints should be also supple enough not to have a negative eect on the functioning of the markets.
1 Alberto Su arez es Profesor Titular de la Escuela T ecnica Superior de Inform atica (Universidad Aut onoma de Madrid) y responsable de Computaci on Cient ca del Risklab Madrid. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de junio de 1999.
316
rez Alberto Sua
There are several elements that make the problem of risk assessment a dicult one. A partial list includes the complexity of the portfolio itself, the correlation structure between the dierent portfolio assets, and a synthesis of these inputs into a small set of numbers, such as VaR or expected Shortfall, that capture the information about the risk prole of the portfolio in a form that is both manageable and easy to communicate. Here, we assume statistically that the portfolio exhibits the same near future behavior as its recent past. The analysis focuses on a synthetic time-series composed of reconstructed values from the recent past of a hypothetical portfolio whose composition is xed and identical to that of the actual portfolio. The reconstruction process may involve some portfolio compression, where the portfolio is replaced by a few components that account for a large fraction of risk, replication, and other manipulations. With such a time-series, the most common measure of risk is the Value at Risk, VaR (see [1]). The objective is to estimate the worst trading losses of a xed portfolio over a period T , the time horizon, at a given probability level P . Assuming the time horizon to be one day, the histogram of daily returns is constructed from the historical data. This empirical distribution is then tted to a parametric model from which a percentile at level P is obtained. The usual assumption is that the daily returns are independent normally distributed variables. This procedure is quickly implemented, but has recently come under criticism (see, for instance [2, 3]). Indeed, there seems to be a consensus in the nancial community that current Value at Risk measures fail to capture some of the essential features of actual market risks. It is empirically observed that the tails of the distribution of returns on nancial products have more weight than what would be predicted by a t to a normal distribution. In particular, the assumption of normality is known to fail for large uctuations in the value of the portfolio [4]. These extreme uctuations are precisely the ones that a risk measure such as VaR is trying to capture. The failure to represent the behavior of the portfolio correctly in worst-case scenarios has prompted several suggestions for alternative, more sophisticated models, better able to deal with extreme events. One extension of the classical VaR methodology drops the assumption of normality and posits a dierent distribution with a positive kurtosis and possible skewness (e.g. hyperbolic distributions [5], mixture of Gaussians [4], stable distributions [6]). This provides a more accurate model of the behavior of the market. The parameters of these non-normal distributions can be determined by maximum-likelihood estimation. Finally, a measure of risk analogous to the normal VaR is extracted by calculation of percentiles. Extreme Value Theory [7, 3, 8] also provides a framework for the denition of other non-standard risk measures such as MaxVaR and Shortfall. In order to extract these parameters, the analysis focuses on the distribution of extreme events itself.
317
The paper is organized as follows: section 2 gives a general introduction to the problem and describes the computation of VaR with dierent parametric models (normal, mixture of normals, hyperbolic distribution) for the empirical distribution of portfolio returns. In section 3, risk measures are calculated by analyzing the tail of the distribution, assuming Pareto behavior. Section 4, focuses on the distribution of maxima of the time series. Finally, a summary of the results of a risk analysis for the IBEX35 data is presented.
IBEX35 ( from 4/1/93 to 23/12/97)
7500 7000 6500 6000 5500 5000 4500 4000 3500 3000 2500 0 200 400 600 800 1000 1200
Figure 1: Daily values for the IBEX35 stock index.
Unconditional risk measures
Assume we have a reconstructed series of the negative of the returns of a given portfolio St+1 St . (1) rt = St The quantities St are the values of the hypothetical portfolio with the same xed composition as the actual portfolio. We focus on the distribution of losses, since they are the important quantity in the problem of risk assessment. The negative sign is included so that losses appear to the right-hand side of the probability distributions. A histogram displaying the numbers of observations in non-overlapping segments covering the whole range of observed returns gives a graphical representation of the
318
rez Alberto Sua
Relative return level (%)
200
400
600 Time
800
1000
1200
Figure 2: Daily relative returns for the IBEX35 stock index (from 4/1/93 until 23/12/97). empirical probability density distribution for the relative losses (negative of the returns) for the portfolio (see gure 3). The usual risk measure, VaR (Value-at-Risk), is a percentile of the prot-loss distribution at a given probability level P (expressed as a percentage) for a given time horizon T . Intuitively, if the time horizon is one day, this quantity represents the minimum relative loss a portfolio will have in at least P of one hundred days, assuming the composition of the portfolio remains unaltered during that period. The value of this percentile can, in principle, be obtained from the empirical probability density distribution function (pdf). Due to the limited data available, this measure is not robust. In particular, the VaR estimations are sensitive to sample uctuations. This is especially true in the tail of the distribution, where extreme losses may occur. Furthermore no out-of-sample VaR estimates can be given by this procedure, which may become crucial for situations with small quantities of data, such as portfolios based on new markets or new economies products. A more robust estimate is obtainable if we assume a parametric form for the protloss distribution and nd the parameters by likelihood maximization. The value of VaR can be obtained as a percentile of the tted distribution. The problem of selecting a model from the data is a dicult one, with no general solution can be given, except if a market model can be explicitly formulated. In its absence, one can assume various parametric forms as suggested in the litera-
319
120
100
80
Count
60
40
20
0 6
0 Relative loss (%)
Figure 3: Daily relative losses for the IBEX35 stock index. ture. Parametric ts with a normal distribution, mixtures of normals, and hyperbolic distributions are included in the analysis.
2.1
Normal VaR
It is commonly assumed that portfolio returns behave as an asset in a Black-Scholes model, and that the relative returns are normally distributed. Figure 4 depicts the main results of a risk analysis with a time horizon of one day and a probability level of 99%. The empirical distribution and corresponding normal t appear on the left hand side of gure 4. The results of two dierent tests to assess the quality of the t (the Kolmogorov-Smirnov statistical test and the quantile-quantile plot) are presented on the right-hand side of the gure. The Kolmogorov-Smirnov statistical test [9] gives a measure of the likelihood of the hypothesis that a set of data is an empirical sampling of a given models pdf with certain values of the parameters. The test denes a sample statistic as the maximum distance between the empirical and the model cumulative probability distributions. An analytic form for the probability distribution of this statistic can be given if the parameters of the model distribution are independent of the data. Despite the fact that the parameters are obtained by maximum-likelihood optimization from the sample itself,we make use of this analytical form. Nonetheless, it is expected that low values of the likelihood of the hypothesis as measured through the Kolmogorov-Smirnov statistic indicate a low quality t. The quantile-quantile plot gives the empirical quantiles as a function of the model
320
rez Alberto Sua
100
80
60
6
40
4
Y Quantiles
20
2 0 2 4 6 4
0 X Quantiles
Figure 4: Risk measures for the IBEX35 using a normal model. The parameters of the t are = 0.0901, = 1.0249.
quantiles. It also provides a way to assess the quality of the parametric t by inspection. If the t is good, the quantile-quantile plot should be a straight line with slope one. Deviations from this behavior are easy to detect by inspection in this type of plot. Besides the value of VaR at a certain probability level P , it is possible to calculate the Shortfall for the data as Shortfall E [X | X > V aR(P )] . (2)
Shortfall is a measure of the average loss a portfolio would have, given that the loss is above a certain threshold. It has been proposed by several authors as an alternative measure of risk with the desired properties of subadditivity and coherence [10]. There are several indications that the normal model fails to reect the distribution of extreme events. The Gaussian t severely underestimates the magnitude of the tails. The low value of the test on the likelihood for the Kolmogorov-Smirnov (KS) statistic indicates that we should be wary of the quality of the t obtained. The quantile-quantile plot reveals that the main discrepancies occur at the tails of the distribution. It can also be seen that the VaR level predicted by the normal t is much lower than the VaR derived directly from the empirical distribution (labeled as sample VaR in the gures).
321
2.2
VaR with mixture models
Among the possible generalizations of the normal model for the portfolio returns, the Gaussian Mixture (GM) model is able to expand the range of describable phenomena while remaining close to the Gaussian paradigm. The mixture-of-normals model is exible enough to capture features commonly observed in actual nancial time series. An important example is the fact that extreme events in the nancial world occur more frequently than the classical normal models would predict. This leads to distribution functions of returns on a nancial asset that have tails with greater weight than normal. This has implications for the pricing of derivative products, for hedging and for the estimation of risk measures [11]. The parameters of the mixture models are estimated by maximization of a modied likelihood function with an Expectation Maximization algorithm [12]. Hull and White [11] introduced two new Greeks, and , to reect the sensitivity of a portfolio composed of derivatives to the kurtosis and skewness, respectively, in the distribution of returns. It also shows how a simple model based on a mixture of normals is sucient to capture this eect and correctly model the volatility smile implicit in foreign exchange rates and equity options. Let us examine in Figure 5 the IBEX35 data when the distribution of returns is modeled by a mixture of two Gaussians. The t is more plausible, as is indicated by the value close to one of KS statistic test, and by the quantile-quantile plot. The values of the model and the sample VaR are now close to each other. Given that we are obtaining within-sample values of VaR, this is an indication that the model produces estimates consistent with the observations. The mixture model is a simple yet exible extension of the normal model that allows one to model features that are observed in actual nancial data. In particular, it can be used to carry out a reliable and robust risk analysis, at least up to the sample edges. Note, however, that the decay of the tails of a mixture distribution is still Gaussian-like. It is unclear at this point whether the actual tail decay is algebraic or exponential. Consequently, out-of-sample extrapolations made with a mixture-ofGaussians model should be regarded with utmost caution. One should also try to limit the number of components that enter the mixture, lest some amount of overtting occur. Generally two or three components provide an appropriate t for the data.
2.3
VaR with hyperbolic models
Another proposal in the literature to account for heavy tails models the portfolio returns with a hyperbolic distribution [5, 13]. The hyperbolic family depends on four parameters PHyp (x; , , , ) = 2 2 2 K1 ( exp 2 2 ) 2 + (x )2 + (x ) , (3)
322
rez Alberto Sua
100
80
60
6
40
4 2
Y Quantiles
20
0 2 4 6 8 10
0 X Quantiles
10
Figure 5: Risk measures for the IBEX35 using a mixture of two normal distributions. The parameters of the t are p1 = 0.8988, 1 = 0.1052, 1 = 0.8934 for the rst normal in the mixture and p2 = 0.1012, 2 = 0.0438, 2 = 1.8053 for the second normal. where 0; > | | 0, and K1 (x) is a modied Bessel function of the third kind with index 1. The quantities and are the location and scale parameters, respectively. The values of and determine the shape of the distribution. Several other distributions (normal, symmetric and asymmetric Laplace, exponential, generalized inverse Gaussian) appear as limiting cases of this distribution. One advantage of the hyperbolic distribution is that it can be seen as a mixture of an innite number of Gaussians, where the weights are given by a generalized inverse Gaussian distribution [5]. Furthermore, the hyperbolic distribution appears as a stationary distribution of a continuous-time Markov process described by a particular stochastic dierential equation [5]. This observation justies its possible use as a model to price options. Figure 6 displays the t obtained with IBEX35 data when the distribution of returns is modeled by a hyperbolic distribution. The resulting t is quite reasonable, except possibly at the tails, which still decay exponentially.
Modeling the tails: Shortfall
In estimating risk measures we are interested in events in the tails. In contrast to the parametric approach given in the previous section, where one attempts to model the full pdf of the portfolio uctuations, one can focus on nding an appropriate model for the tail of the distribution alone [7]. Assuming that the tail behaves regularly,
323
100 90 80 70 60 50 40 30 20 10 0 5 0 5
= 1.84159 = 0.0651013 = 1.02963 = 0.157995
6 4 2 0 2 4 6 8 10 5 0 5 10
Figure 6: Risk measures for the IBEX35 using a hyperbolic distribution.
the probability density distribution should behave asymptotically like a Generalized Pareto distribution Ptail (X ) = 1 1+ (X u)+
(1+ 1 )
X > u.
(4)
In the literature on extreme events, this method is known as the Peaks Over Threshold (POT) method [7]. Note that the predicted decay is algebraic except if the tail index is close to zero, when it corresponds to exponential decay. The parameter u should be chosen large enough to produce the asymptotic behavior predicted in equation (4), but as small enough to produce adequate data in the tail to yield reliable estimates for the remaining parameters and . The mean excess above a threshold for a random variable is dened to be the conditional expectation of that variable less the threshold for values of the variable above the threshold. For an empirical realization with N data samples, the mean excess is E [X | X > ] = 1 NX>
N
(Xj )+ ,
j =1
(5)
where NX> is the number of instances above the threshold in the sample. If for X > u the distribution of this variable can be approximated as a Generalized Pareto, equation (4), the mean excess should be approximately linear in the threshold. The
324
rez Alberto Sua
analytic formula for the mean excess for a Generalized Pareto pdf is E [X | X > ] = + ( u). 1 1 (6)
To select the parameter u of the Generalized Pareto (GP) distribution, the user should inspect the mean excess plot and identify regions for which the plot is approximately linear. Then the threshold u is chosen as small as possible, but within the linear region. For the IBEX35 data, we present the results with dierent choices for the threshold. Figure 7 displays the results with u = 1, corresponding to the at region ( 0, exponential decay) in the mean-excess plot. Even though the data are scarce, it is
200
150
1.5
100
50
1
0.5 0 1 2 3
Figure 7: Risk measures for the IBEX35 data using a Generalized Pareto Fit u = 1. The t parameters are = 0.6397, = 0.0141. possible to identify a separate linear region for values X > 2. The corresponding Pareto t for a threshold u = 2.25 is displayed in Figure 8. In contrast to the previous choice of the threshold, which predicts an exponential decay, this choice predicts algebraic decay. This question is of particular importance in calculating the shortfall, whose value is very sensitive to the type of decay for edge-of-sample probability levels (around and above 99.9%). It is possible that the true asymptotic decay is algebraic, as is predicted when focusing on points beyond the threshold u = 2.25. One could also conclude that the decay is exponential but the algebraic-like behavior observed for u 2.25 comes from sample uctuations. With the available data it is dicult to decide which hypothesis is closer to reality.
325
200
150
1.8
100
1.6 1.4
50
1.2 1
0.8 0.6 0.4 0 1 2 3
Figure 8: Risk measures for the IBEX35 data using a Generalized Pareto Fit u = 2.25. The t parameters are = 0.3830, = 0.3864.
Modeling extreme events: MaxVaR
The evolution of the value of a portfolio in response to large uctuations in the risk factors is one of the important elements in risk management. It is thus useful to develop models for distributions of extremal events [7]. Under certain relatively weak conditions (identical distribution of events, independence of the maxima), it can be shown that the maxima of a time series follow asymptotically a Generalized Extreme Value (GEV) distribution for the probability density P (x) = 1 1 + (x ) (x )
(1+ 1 )
exp 1 + (x )
, (7)
1+
> 0.
Parameter is the tail index introduced in the previous section and should be similar to the one estimated from the Generalized Pareto t. The analysis proceeds as follows: We choose a time lapse and partition the data into non-overlapping segments of the selected length. A new series is formed using the maximum values for the losses within each segment. The box size is chosen by compromising between suciently large boxes, so that the independence and asymptotic conditions are met, and having sucient data to produce reliable estimates of the GEV parameters.
326
rez Alberto Sua
35 30 25 20
(10%) = 0.280176
15 10 5
6 5 4 3 2 1 0 0 200 400 600
Figure 9: Risk measures for the weekly maxima of IBEX35 data. The parameters of the GEV distribution tted are = 0.6084, = 0.6179, = 0.0324. Let us consider the MaxVaR analysis for the IBEX35 data. We focus on weekly maxima (Figure 9) and biweekly maxima (Figure 10). The novel elements in Figures 9, 10 are the Hill plot and the risk analysis based on maxima. The Hill plot permits the estimation of the parameter when it is positive. The points in the time series {X1 , X2 , . . . , XT }, are ordered ascendingly XT ;T XT 1;T . . . X2;T X1;T . The Hill plot gives the value of the Hill estimator for the tail index (H ) = 1 k
k
(8)
log
j =1
X j ;T X k ;T
(9)
as a function of the integer k = 1, 2 . . . T . The point selected in the graphic by a cross is the 10% Hill estimator (k10% = 0.1 T ). We observe that this estimate is close to the tail index found by a t to a GP with a threshold u = 2.25, which indicates an algebraic decay in the distribution tails. However, the maximum-likelihood estimates for the distributions of weekly and biweekly maxima predict a tail index close to 0 (exponential decay of the tails), in agreement with the Pareto t with a threshold u = 1. In Figs 9, 10 we have also selected two probability levels Pl = 5% and Pu = 95%. By calculating the corresponding percentiles of the GEV distribution, one obtains a range where the corresponding maxima can be found with a certain
327
35 30 25 20
(10%) = 0.280176
15 10 5
6 5 4 3 2 1 0 0 200 400 600
Figure 10: Risk measures for the biweekly maxima of IBEX35 data. The parameters of the GEV distribution tted are = 0.9666, = 0.5864, = 0.0774. probability [M axl , M axu ]. The expected value for the maxima conditioned on being in the range derived is also reported in Figures 9, 10 E [M ax|M axl M ax M axu ] . (10)
Summary of results for IBEX35
The main results obtained for Value-at-Risk and shortfall for the IBEX35 data are summarized in tables 1 and 2. Table 1 displays Value at Risk and table 2 Shortfall for dierent probability levels and with dierent parametric ts. The rows are labeled according to the model used to compute the values: empirical distribution VaR (sample), normal (GM1), mixture models (GM2, GM3, mixtures with 2 and 3 Gaussians, respectively), hyperbolic distribution (HYP), and Generalize Pareto distributions (GP, with the chosen threshold in parentheses). The second column displays the result of the Kolmogorov Smirnov statistical test (KS test). The remaining columns are labeled according to the probability level at which the risk measures are computed. We observe that within-sample risk measures (95%) are insensitive to the model selected. For probability level P = 99%, all parametric ts predict essentially the same risk measures, except for the normal one, which severely underestimates both VaR and Shortfall. This tendency becomes marked for higher probability levels. Beyond this probability level, data are scarce and sample-derived measures should not be trusted.
328
rez Alberto Sua
For edge-of-sample probability levels P = 99.9 % and beyond, whether the asymptotic decay is algebraic or exponential is of some consequence. Shortfall is especially sensitive to this issue. We observe that the model with a mixture of two Gaussians and the Pareto t with u = 1 predict fairly consistent risk measures. The hyperbolic model, with fast decaying exponential tails displays excellent agreement for the body of the distribution but predicts thinner tails than observed. Much higher values for the risk measures are predicted by the Pareto t with u = 2.25, which shows slow algebraic decay. The model t with three Gaussians lies somewhere between. Table 1: Comparison between dierent measures of VaR KS test Sample GM1 GM2 GM3 HYP GP (u = 1) GP (u = 2.25) 0.12 0.88 0.99 0.99 0.78 0.96 95% 1.58 1.60 1.54 1.58 1.59 1.57 * 99% 2.48 2.29 2.53 2.46 2.57 2.63 2.45 99.9% 4.34 3.08 4.25 4.41 3.91 4.18 4.16 99.99% 5.11 3.72 5.63 6.86 5.30 5.78 8.34
Table 2: Comparison between dierent measures of Shortfall KS test Sample GM1 GM2 GM3 HYP GP (u = 1) GP (u = 2.25) 0.12 0.88 0.99 0.99 0.78 0.96 95% 2.19 2.02 2.16 2.17 2.20 2.23 * 99% 3.16 2.64 3.27 3.21 3.13 3.30 3.21 99.9% 4.81 3.36 4.86 5.51 4.49 4.87 5.98 99.99% 5.22 3.97 6.13 7.69 5.80 6.50 12.79
329
References
[1] P. Jorion. Value at Risk: The new Benchmark for Controlling Market Risk. McGraw-Hill, New York, 1997. [2] P. Artzner, F. Delbaen, J.-M. Eber, and D. Heath. Thinking coherently. Risk 10 (1997), no. 11, pp. 6871. [3] P. Embrechts, S. Resnick, and G. Samorodnitsky. Living on the edge. Risk (January 1998), pp. 96100. [4] J. Hull and A. White. Value at risk when daily changes in market variables are not normally distributed. Journal of Derivatives 5 (1998), no. 3, pp. 919. [5] E. Eberlein and U. Keller. Hyperbolic distributions in nance. Bernoulli 1 (1995), pp. 281299. [6] J. H. McCulloh. Financial Applications of Stable Distributions, in Statistical Methods in Finance, Handbook of Statistics Vol. 14. G.S. Maddala and C. R. Rao eds. North Holland, NY, 1996. [7] P. Embrechts, C. Kluplelberg, and T. Mikosch. Modelling Extremal Events for Insurance and Finance. Springer-Verlag, Berlin, 1997. [8] J. Danielsson, P. Hartmann, and C. de Vries. The cost of conservatism. Risk (January 1998), pp. 101103. [9] W. Press, W. T. Teukolsky, S. A. Vetterling, and B. Flannery. Numerical Recipes in C: The Art of Scientic Computing. Cambridge Univ. Press, Cambridge, 1993. [10] P. Artzner, Delbaen F., J.-M. Eber, and D. Heath. Coherent measures of risk. Mathematical Finance 9 (1999), no. 3, pp. 203228. [11] J. Hull and A. White. Evaluating the impact of kurtosis and skewness on derivative prices. NetExposure 3 (1997), pp. 8190. [12] J. D. Hamilton. A quasi-bayesian approach to estimating parameters for mixtures of normal distributions. Journal of Business and Economic Statistics 9 (1991), no. 1, pp. 2739. [13] U. Kuchler, K. Neumann, M. Soerensen, and A. Streller. Stock returns and hyperbolic distributions. Sonderforschungsbereich 373, Humboldt-Universtat zu Berlin, (23): 118, 1994.
Alberto Su arez ETS de Inform atica Universidad Aut onoma de Madrid Ciudad Universitaria de Cantoblanco 28049-Madrid e-mail: alberto.suarez@ii.uam.es
Valoraci on y medici on de riesgo de pasivos bancarios sin vencimiento denido

s M. Tarriba Unger1 Jesu
Resumen En el contexto de la gesti on del riesgo de mercado de activos y pasivos en el balance de bancos comerciales, se presenta el problema de valuar y medir la sensibilidad ante movimientos de las tasas de inter es de los instrumentos de captaci on que no tienen un vencimiento contractual denido. Como ejemplo se puede citar el caso de cuentas de ahorro y de cheques, en las que el cliente tiene la opci on de retirar el dinero cuando as lo desee, opci on por la que normalmente paga un costo de oportunidad al recibir rendimientos por debajo de los ofrecidos en el mercado de dinero. Existen diversas propuestas metodol ogicas, enfocadas principalmente al sistema bancario norteamericano, que buscan sustentar sobre bases m as s olidas el tratamiento de los pasivos sin vencimiento denido. Los modelos parten de ajustes econom etricos para describir la dependencia del nivel de los saldos con las tasas de inter es de mercado y el rendimiento de las cuentas, que se complementan con formalismos desarrollados en el contexto de la valuaci on de derivados de tasa de inter es. En este documento se presenta el formalismo y resultados te oricos de un modelo que ilustra el tipo de acercamientos encontrados en la literatura sobre el problema de estimar la duraci on efectiva de pasivos sin vencimientos. El objetivo es dar un idea general de la riqueza de comportamiento a que dan lugar este tipo de modelos, con todas las simplicaciones de la realidad que presuponen.
Introducci on
En el contexto de la gesti on del riesgo de mercado de activos y pasivos en el balance de bancos comerciales, se presenta el problema de valuar y medir la sensibilidad ante movimientos de las tasas de inter es de los instrumentos de captaci on que no tienen un vencimiento contractual denido. Como ejemplo se puede citar el caso de cuentas de ahorro y de cheques, en las que el cliente tiene la opci on de retirar el dinero cuando as lo desee, opci on por la que normalmente paga un costo de oportunidad al recibir rendimientos por debajo de los ofrecidos en el mercado de dinero. Esta
1 Jes us M. Tarriba Unger trabaja en el area de Riesgo de Mercado Global del Banco Santander Central Hispano. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de septiembre de 1999.
332
s M. Tarriba Jesu
prima representa un contribuci on importante al margen nanciero de los bancos comerciales, ajustado desde luego por los costos derivados del servicio de las cuentas. Asimismo, el banco se reserva la opci on de reducir los rendimientos e inclusive aplicar comisiones de manejo de cuenta en el caso en que movimientos a la baja en el nivel general de las tasa de inter es reduzca el margen de intermediaci on. Debido a la dicultad de modelar y cuanticar estos efectos, es pr actica com un o bien ignorar total o parcialmente este tipo de pasivos al agregar la estructura de vencimientos del balance, o bien asignarles de forma arbitraria una estructura de plazos sin un an alisis que lo sustente (ver p.e. [12]). Sin embargo, existen diversas propuestas metodol ogicas enfocadas principalmente al sistema bancario norteamericano y que buscan sustentar sobre bases m as s olidas el tratamiento de los pasivos sin vencimiento denido, [1], [2], [5], [7], [8], [10], [11]. Todos los acercamientos parten de modelar dos aspectos esenciales al problema: (a) Para el caso de cuenta remuneradas, la relaci on entre las tasas de referencia y la tasa o costo que los bancos comerciales pagan a los clientes. (b) El comportamiento de los vol umenes o saldos de las cuentas en el tiempo y su relaci on con el comportamiento de las tasas de inter es de mercado de referencia y el rendimiento pagado por las cuentas. Los modelos combinan ajustes econom etricos de las variables arriba mencionadas con formalismos desarrollados en el contexto de la valuaci on de derivados de tasa de inter es, en los cuales el valor presente de un instrumento se calcula promediando el valor presente de los ujos de efectivo futuros bajo cierta distribuci on de probabilidad para la evoluci on estoc astica de las estructura de los tipos de inter es. Los par ametros de la distribuci on de probabilidad usada se deben calibrar para que sea consistente con el la estructura de tipos y sus volatilidades a la fecha de valuaci on, aplicando argumentos de no-arbitraje, [6], [9]. En este documento se presenta el formalismo y resultados te oricos de un modelo que ilustra el tipo de acercamientos encontrados en la literatura sobre el problema de estimar la duraci on efectiva de pasivos sin vencimientos. El objetivo es dar un idea general de la riqueza de comportamiento a que dan lugar este tipo de modelos, con todas las substanciales simplicaciones de la realidad que presuponen. En la pr actica, las dicultades reales se encuentran al tratar de ajustar los modelos a datos emp ricos, generalmente de corto alcance hist orico y contaminados por m ultiples factores que el modelo no considera, lo que convierte el an alisis en una labor casi artesanal de tratamiento caso por caso que se debe complementar con supuestos econ omicos razonables. En la secci on 2 se describe una versi on simplicada de balance a partir del cual se construye el modelo. En la secci on 3 se describe el tratamiento est atico del problema, el cual se sostica en la secci on 4 donde se presenta la versi on din amica. Finalmente en la secci on 5 se presentan los resultados de un ejemplo ilustrativo y se comentan los resultados.
n y riesgo de pasivos bancarios sin vencimiento definido Valoracio
333
Modelo de balance
En la tabla 1 se representa de forma muy esquem atica el balance de un banco comercial. El banco capta recursos nancieros a diversos plazos y costos que se colocan en el lado de los activos seg un las l neas de negocio que gestione. Los activos generan un rendimiento bruto normalmente por arriba del costo nanciero incurrido en nanciarlos. A la diferencia entre el rendimiento de los activos y el costo nanciero de los pasivos es lo que se conoce como el margen nanciero del banco. Este margen nanciero, neto de costos operativos y comisiones, representa una fuente importante a los ingresos de un banco comercial. En particular, una parte de los recursos pueden ser canalizados a otorgar prestamos a empresas y particulares, asumiendo el riesgo de cr edito aparejado y por lo que el banco cobra una sobretasa sobre el nivel de los tipos libres de riesgo, por los que se entiende el rendimiento pagado por deuda gubernamental o pr estamos colateralizados al 100%. ACTIVOS Mercado interbancario Cartera Crediticia Inversiones Varias Activos Fijos PASIVOS Mercado interbancario Dep ositos Clientes Plazo Fijo Dep ositos Clientes sin vencimiento denido Deuda Largo Plazo Capital
Tabla 1: Esquema de hoja de balance de banca comercial. Otra fuente importante de margen nanciero, relevante para los nes del presente an alisis, proviene de la captaci on de recursos por abajo del tipo libre de riesgo, t picamente a trav es de dep ositos de clientes, que pueden ir desde instrumentos a plazo hasta cuentas de cheques y de ahorro sin vencimiento denido. Como nuestro inter es es precisamente estimar de alguna forma la contribuci on de este grupo de pasivos al riesgo de tasa de inter es del balance, se construye un balance simplicado que se ilustra en la siguiente tabla: ACTIVOS Mercado interbancario PASIVOS Dep ositos Clientes sin vencimiento denido
Tabla 2: Esquema de hoja de balance simplicada. En el modelo simplicado de banco, al tiempo t = 0 se recibe un ingreso D0 en una cuenta de ahorro o de cheques. El banco invierte al corto plazo el dinero en el mercado interbancario, y hace lo propio con cualquier nuevo ingreso que reciba en el futuro,
334
s M. Tarriba Jesu
retir andolo cuando se requiera liquidar a los clientes. Los nuevos ingresos/retiros y el pago de intereses en las cuentas y en el marcado de dinero se efect ua en intervalos regulares de tiempo t. Denotamos por rt el rendimiento interbancario al tiempo t, st el rendimiento o costo de las cuentas y Dt el saldo de las mismas al principio del periodo (t, t + t). Entonces, los ujos de caja al tiempo t son2 : Activo (Mercado de Dinero): Pasivo (Cuentas de Ahorro) : Neto: Cat Cpt Ct = (Dt Dt1 ) + Dt1 rt1 t = (Dt Dt1 ) Dt1 st1 t = Dt1 (rt1 st1 ) t
En las dos primera ecuaciones, el primer t ermino corresponde al ujo proveniente del incremento (o decremento) neto en el saldo de las cuentas. Como el banco invierte (retira) de la cuenta de los activos cualquier ingreso (egreso) realizado en la cuenta de los pasivos, el ujo neto debido al cambio en los saldos se cancela. Por tanto, el ujo generado en cualquier intervalo futuro depende s olo del nivel de los saldos al principio del intervalo y del diferencial entre la tasa interbancaria y el costo de las cuentas. En el balance simplicado de banco, la u nica fuente de ingresos es el margen nanciero positivo neto de costo operativo, cuyo valor presente determina el valor te orico del balance.
Tratamiento est atico
Sup ongase primero que en fecha de valuaci on (t = 0) tanto los tipos de inter es de corto plazo como el costo de las cuentas y los saldos se conocen de antemano para todo tiempo futuro. En este caso, los ujos del balance simplicado descrito arriba est an a su vez predeterminados y en el l mite continuo su valor presente est a dado por
V =
Dt (rt st ) eZt t dt ,
(1)
donde Zt es el rendimiento anualizado a t = 0 de un bono cup on cero que vence al plazo t. Por argumentos de arbitrajes, a tasas deterministas las tasas cup on cero y de corto plazo futuras deben cumplir la relaci on: eZt t =
t 0
ers s ds .
(2)
A continuaci on se examinan varios casos l mites, suponiendo por simplicidad una estructura de tasas de inter es planas, i.e., Zt = rt = t para toda t.
2 Si existen requerimientos de encaje o reservas exigidos por las autoridades supervisoras, s olo una porci on de los dep ositos podr an ser invertidos a tasa interbancaria, en cuyo caso el t ermino del rendimiento se multiplica por la fracci on restante, p.e., 0.9 si el encaje es del 10%.
335
3.1
s < r constante, independiente de r y saldo constante D
En este caso los pasivos se tratan como una perpetuidad. El valor del balance esta dado por s D [r s] ert = D 1 V = . (3) r 0 El valor del pasivo es, simplemente, s (4) VP = D , r y por tanto su sensibilidad 1% y duraci on modicada est an dadas por: 1% TP = 0.01 = s dVP = 0.01 D 2 , dr r (5) (6)
1 1 dVP = . VP dr r
Para tasas del 6% se estar a hablando de una duraci on modicada de casi 17 a nos. Sin embargo, para nes de la gesti on de riesgo de mercado, el plazo relevante es aqu el al que se deben invertir los activos si se quiere tener un balance inmunizado ante peque nos desplazamientos en el nivel de las tasas de inter es. Para obtener este plazo, sup ongase que el banco invierte el dinero del dep osito a cierto plazo T en un instrumento que genera rendimientos a tasa r en composici on continua, de forma que el valor inicial de los activos est a a la par y la duraci on modicada del activo es precisamente T . El valor del balance de nuevo est a dado por la ecuaci on (3), pero en este caso la sensibilidad 1% de valor de balance es: s 1% (T ) = 0.01 D0 T 2 . (7) r Por tanto, el plazo al que el banco debe invertir el dep osito para obtener un balance neutral a peque nos movimientos en la tasa es: s s s 1 = TP . = (8) 2 r r r r De lo anterior se deduce que la duraci on modicada que deben tener los activos para neutralizar el riesgo de tasa de inter es no es igual a la duraci on de la perpetuidad, sino a esta u ltima escalada por la raz on entre el costo de la cuenta y la tasa interbancaria. Por ejemplo, con niveles de tasas de 6%, si una cuenta se puede considerar como una perpetuidad con costo jo del 1%, el plazo en que habr a de invertir los activos para cancelar el riesgo de tasa de inter es resulta del orden de (.01/.06)/(.06) = 2.8. Esto puede sonar parad ojico, en la medida en que existe la inclinaci on a pensar que el tener nanciamiento barato al muy largo plazo implica que se puede invertir al muy largo plazo sin incurrir en riesgo de tasa de inter es. Esto es cierto en cuanto al riesgo de variaciones en margen futuro, pero no en cuanto al valor presente de dicho margen, que a nal de cuentas determina la contribuci on al valor econ omico del balance de las cuentas sin vencimiento. T1% =0 =
336
s M. Tarriba Jesu
3.2
s ota con r de forma que s = k r, para cierta constante k < 1, y saldo constante D
En este caso, el valor del banco esta dado por: V =

0
D0 [r k r] er t dt = D0 [1 k ] ,
(9)
asicos y con V p = k D0 independiente de r, por lo que la sensibilidad puntos b duraci on en este caso son cero tanto en el pasivo como en el activo.
3.3
s constante, saldos dependientes de las tasas de inter es
Ahora consid erese el caso en que el saldo de los dep ositos depende de la tasa r de acuerdo a la ecuaci on D0 w r . (10) La constante se puede interpretar como la elasticidad de los saldos de las cuentas ante movimientos en el nivel de la tasa de inter es y es tal que, para cambios peque nos en r, se cumple que r D0 . (11) D0 r La ecuaci on (10) es la relaci on de equilibrio en el largo plazo que se asume en varios de los modelos analizados en la literatura, sobre los que se volver a m as adelante. Por ahora baste para ilustrar la contribuci on a la sensibilidad 1% que proviene de la elasticidad de los saldos respecto al nivel de las tasas de inter es. En efecto, para el caso de costo constante, la sensibilidad total del valor del balance est a dada por la expresi on D0 s s 1% = 0.01 . (12) 1 D0 T 2 r r r Asumiendo que se cumple la ecuaci on (10), D0 = w r1 = D0 , (13) r r y se obtiene, sustituyendo en la ecuaci on (12), s s 1% = 0.01 D0 , (14) 1 T 2 r r r de donde inmediato obtener el plazo efectivo de inversi on de los activos tal que la sensibilidad 1% del balance se anule: s s s s T1% =0 = (15) 1 + 2 = 1 + TP . r r r r r En la ecuaci on anterior, adem as de la correcci on sobre la duraci on de la perpetuidad obtenida anteriormente, se tiene un t ermino adicional que proviene de la constante que determina la dependencia del nivel de los saldos respecto al nivel de las tasa de inter es. Como en general s < r, entonces el efecto de < 0 es reducir el plazo efectivo. Con r = 6%, s = 1% y = 0.1 se tiene T1% = 1.4, aproximadamente.
337
3.4
F ormula general
Las expresiones anteriores se puede generalizar de forma directa para el caso en que el tanto el volumen de las cuentas como el costo dependen de forma arbitraria de r, aunque r en s sigue siendo tratada como variable determinista. En este caso, las expresiones para el valor siguen siendo igual a (1), pero se tienen contribuciones adicionales en la sensibilidad. Por ejemplo, la sensibilidad 1% de los pasivos se convierte en 1% = 0.01 d s(r) s dD(r) D(r) ds(r) D(r)s + + . D(r) = 0.01 dr r r dr r dr r2 (16)
En la medida en que las observaciones emp ricas de un conjunto de cuentas en particular se puedan adecuar a alguno de los casos sencillos considerados, el tratamiento de las cuentas se simplica enormemente.
Tratamiento din amico
Como ya se mencion o, determinar la relaci on del volumen con las tasas de inter es es un problema complejo, que por lo general involucra asumir una forma funcional con cierto n umero de par ametros libres, que se ajustan para tratar de reproducir los datos emp ricos lo mejor posible. A continuaci on se describe los formalismos desarrollados en las referencias [7] y [11], que ilustran la idea b asica detr as del tipo de modelos usados.
4.1
Modelos para el volumen: cuentas no remuneradas
Se parte de suponer que, a un valor dado de todos los dem as factores que inciden en el comportamiento de los vol umenes, existe un nivel de equilibrio Deq proporcional a cierta potencia de la tasa de inter es r, de forma que
eq Dt = kt rt ,
(17)
que se expresa alternativamente en forma logar tmica:

eq ln (Dt ) = ln (kt ) + ln (rt ) .
(18)
Dado que r y los dem as factores incluidos en la funci on k (t) est an en constante movimiento, D(t) en general se encontrar a fuera de su nivel de equilibrio y por tanto su din amica al corto plazo estar a dominada por la tendencia a regresar a dicho nivel. Esto se expresa en la ecuaci on siguiente:
eq ln (Dt+1 ) ln (Dt ) = (ln (Dt ) ln (Dt )) ,
(19)
para cierta > 0. De acuerdo a la ecuaci on (19), si el nivel de los saldos est a por arriba de su valor de equilibrio Deq , la evoluci on temporal tender a a revertir a Deq .
338
s M. Tarriba Jesu
Introduciendo (19) en (18) y deniendo los par ametros 0 1 Xt = = = 1 , = (1 0 ) , ln (kt ) , (20) (21) (22)
se obtiene la siguiente ecuaci on lineal en los logaritmos de los vol umenes y la tasa: ln (Dt+1 ) = 0 ln (Dt ) + 1 ln (rt ) + Xt , (23)
donde se incluyen en Xt todas las variables no relacionadas con las tasas de inter es que pudiesen inuir en el nivel de los saldos, tales como factores estacionales, tendencia, etc. Los par ametros relevantes en la ecuaci on (23) son 0 y 1 , que gobiernan la dependencia de Dt respecto a la tasa de inter es y su velocidad de reversi on al equilibrio. En el corto plazo, la elasticidad de los saldos a movimientos en las tasas est a dada por la constante 1 , mientras que, en el largo plazo, lo est a por la constante denida en la ecuaci on (17) y que se puede obtener a partir de 0 y 1 como = 1 . 1 0 (24)
4.2
Modelo para el volumen: cuentas remuneradas
Se procede en las mismas l neas que en la argumentaci on anterior, con la diferencia de que, en este caso, el nivel de saldos de equilibrio se asumen dados por una expresi on de la forma: eq Dt = F (rt , st ) ; (25) i.e., el saldo de equilibrio depende adicionalmente del rendimiento st que paga la cuenta. Por ejemplo, si se asume que la relaci on es de la forma3
eq 1 2 Dt = kt rt st ,
(26)
se llega, procediendo como en el caso anterior, a la siguiente ecuaci on: ln (Dt+1 ) = 0 ln (Dt ) + 1 ln (rt ) + 2 ln (st ) + Xt . (27)
Una vez especicado el modelo para determinar la evoluci on de los saldos dada la evoluci on de las tasas de inter es y el rendimiento de las cuentas, se procede a modelar la relaci on funcional entre dichas variables.
3 La relaci on funcional supuesta entre saldos y tasa varia seg un el autor. Puede incluir m as o menos t erminos de retraso (lags ) y puede o no ser lineal en los logaritmos de las diferentes variables.
339
4.3
Modelo para comportamiento del costo de las cuentas como funci on de las tasas interbancarias
Para modelar el comportamiento del costo de las cuentas en la simulaci on, seguimos la l nea de argumentaci on de las referencias [5] y [8]. Se asume que existe un nivel de rendimiento de las cuentas de equilibrio, esto es, el nivel al que tender a el costo a un nivel de tasa de mercado con todas las dem as variables constantes. Este rendimiento de equilibrio se asume compuesto de una parte ja y una otante: s t = a1 rt + a0 . (28)
En un mercado con competencia perfecta y donde los clientes act uan ecientemente, a1 estar a dado por la proporci on de reservas requeridas a los bancos, y a0 por los costos no nancieros de la gesti on de las cuentas. En la pr actica, efectos de monopolio y de comportamiento de la base clientes se traducir a en valores diferentes para estas constantes4 . Al igual que en el caso de los saldos, la din amica de corto plazo del rendimiento de las cuentas est a gobernada por la tendencia a regresar a su valor de equilibrio. Sin embargo, una diferencia importante es que en este caso se introduce la posibilidad de una asimetr a entre la velocidad de reversi on seg un los rendimientos est en por arriba o por abajo de su valor de equilibrio, por lo que en la ecuaci on din amica se introducen dos constantes seg un el caso:
st+1 st = + (st s t ) (st st ) (st st ) (1 (st st )) + Xt , (29)
as variables que pudieran impactar el nivel de los donde Xt representa todas las dem costos y (x) es la funci on escal on de Heaviside, denida como: (st s t) = 1, 0, si st s t, si st < s t. (30)
De acuerdo a la ecuaci on (30), si el costo nanciero se encuentra por arriba de su valor de equilibrio, revierte a este con velocidad + ; pero si se encuentra por debajo, lo hace con velocidad . Esta asimetr a se observa en la pr actica y se debe a la disparidad con la que reaccionan ante los movimientos de tasas los tenedores de las cuentas y los bancos. Este comportamiento se ilustra en las gr acas 1 y 2. En ellas se aprecia el efecto de disminuir el valor del par ametro + de un valor de 0.2 a 0.05. Cuanto menor sea el valor del par ametro, m as lenta es la respuesta de la variable costo, pareci endose m as las cuentas remuneradas a un instrumento de tasa ja. En ambas simulaciones se utiliza el mismo par ametro para reversi on, = 0.5, que lo hace revertir r apidamente a su valor de equilibro en escenarios de tasas a la baja.
4 Ver
p.e. la referencia [1].
340
s M. Tarriba Jesu
Gr aca 1: Simulaci on tasa y rendimiento de cuentas: + = 0.2, = 0.5
Gr aca 2: Simulaci on tasa y rendimiento de cuentas: + = 0.05, = 0.5
341
Como en el caso de los saldos, para proceder hay que introducir la relaci on de equilibrio en la ecuaci on din amica (30) y aplicar alg un algoritmo de ajuste para estimar los par ametros a partir de la serie hist orica que se quiera analizar. Dado el valor del rendimiento de las cuentas de acuerdo al modelo, el costo total estar a dado por la expresi on: ct = st + cN F , (31) donde se introduce la constante cN F , que representa el costo no nanciero de la variable, que incluir a todos los costes variables (i.e., por unidad de saldo en cuenta) asociados al manejo de las cuentas no incluidos en el rendimiento que pagan.
4.4
Modelo para la evoluci on de las tasas de inter es
De acuerdo a la teor a moderna de valuaci on de activos nancieros5 , la forma apropiada de valuar instrumentos de renta ja con pagos contingentes es calcular el valor esperado de todos los ujos futuros, utilizando una distribuci on de probabilidad neutral al riesgo consistente con la estructura de tasa de inter es y volatilidad prevalecientes. Denotando por E () el operador de valor esperado bajo dicha probabilidad, la expresi on general para el valor de nuestro banco simplicado es
V =E
D(rt ) (rt ct ) e
t 0
rk dk
dt
(32)
El t ermino dentro del operador de valor esperado representa el valor presente de todos los ujos futuros del banco para cada posible trayectoria de la tasa de corto plazo rt . El operador de valor esperado indica que para obtener el valor del banco se tiene que tomar un promedio de los ujos descontados sobre todas las trayectorias futuras. En un esquema de simulaci on Monte Carlo (que se adopta en el presente trabajo), las trayectorias se generan a partir de asumir un modelo estoc astico para las tasas de inter es. Un modelo de uso com un por su sencillez es el de Cox-Ingersoll-Ross (CIR), en el cual los cambios en la tasas de inter es en tiempos peque nos est an gobernados por la expresi on dr = k (rm r) dt + r1/2 dZ , (33) donde dZ es una variable estoc astica que sigue una distribuci on normal con media cero y varianza dt. Los par ametros k , y rm se estiman a partir de la estructura de tasas de mercado a fecha de valuaci on y su serie hist orica.
5 Ver,
p.e., la referencia [9].
342
s M. Tarriba Jesu
4.5
Simulaci on Monte Carlo
Una vez determinados los par ametros para controlan la din amica de las tasas de inter es y de los saldos en el tiempo, se procede a realizar la simulaci on Monte Carlo, donde se busca aproximar num ericamente la esperanza del valor presente de las cuentas. Como plazo de la simulaci on se toman usualmente 30 a nos, para un total de 360 meses, siendo un mes el intervalo discreto jado para la evoluci on de las variables. Sin embargo, en los resultados presentados en la siguiente secci on se utiliz o un horizonte de 1000 meses, que equivale para todo n pr actico a suponer un balance de alcance perpetuo. Para cada periodo se generan las trayectorias de tasas mediante la ecuaci on ri+1 = ri + k (rm ri ) 1 1 1/ 2 + ri N (1, 0) , 12 12 (34)
donde N (1, 0) representa n umeros aleatorios generados de acuerdo a una distribuci on normal est andar. Para cada posible trayectoria j , se genera el comportamiento de los saldos de acuerdo a la ecuaci on (23) o (27), seg un las cuentas sean o no remuneradas, los costos nancieros de acuerdo a la ecuaci on (29) y se calculan los ujos correspondientes, que a su vez son descontados sobre la misma trayectoria seguida por las tasas. La contribuci on al valor presente bajo la trayectoria j - esima est a dada por
1000
V0j
=
t=1
1 Dt,j (rt,j st,j cN F ) , 12 (1 + r1,j /12) (1 + r2,j /12) (1 + rt,j /12)
(35)
y el valor te orico del banco es V0 = 1 N

N
V0j ,
j =1
(36)
donde N es el n umero de escenarios Monte Carlo generados. Para calcular la sensibilidad a puntos b asicos del balance, se desplazan los escenarios un punto b asico y se recalcula el valor esperado dado por la ecuaci on (36). Dada la sensibilidad 1 punto b asico se puede derivar la duraci on efectiva de los activos en los que hay que invertir el saldo de los dep ositos para obtener la misma sensibilidad estimada: Tef 104 D0 N
N j =1
V1jpb V0j .
(37)
343
Ejemplo
A continuaci on se presentan resultados obtenidos a partir de un ejemplo sencillo, que da una idea del efecto de variar tres de los principales par ametros del modelo. Las tasas de inter es a corto plazo r se jan en un 6% y la remuneraci on inicial de las cuentas s en 3%, y se asume que en equilibrio s = 0.5 r. La velocidad de reversi on a la media en el modelo CIR se ja en 0.3 y se considera el caso en que el nivel de los saldos en equilibrio depende s olo de las tasas de inter es interbancarias, jando en 0.5 el par ametro 0 en la ecuaci on (23). Se examinan los siguientes casos: (a) Asumiendo un rendimiento de las cuentas estrictamente jo (i.e., + = = 0 en la ecuaci on (29)), se calcula el plazo efectivo variando la elasticidad de corto plazo de los saldos frente a movimientos a la tasa de inter es (la constante 1 en la ecuaci on (23)). (b) Lo mismo que en el caso (a), pero considerando la opci on del banco de revisar los costos a la baja cuando las tasas bajan tanto que el margen nanciero tiende a cero. En el ejemplo mostrado se asume simplemente que la revisi on se lleva a cabo en caso de que el margen se reduzca a cero, de forma que si r < 3%, entonces s = r. Esto implica que en la regi on r < 3% las cuentas se comportan como un instrumento a tasa otante. (c) Asumiendo saldo constante, independiente de la tasa de inter es (1 = 0), se calcula el plazo efectivo variando la velocidad de ajuste entre tasa y el rendimiento de las cuentas, las constantes + y de la ecuaci on (29), que para nes del ejemplo se asumen iguales. (d) Lo mismo que en el caso (c), pero considerando la opci on del banco de revisar los costos a la baja cuando las tasas bajan tanto que el margen nanciero tiende a cero, opci on que se trata como en el caso (b). En la gr aca 3 se presentan los resultados para el caso (a), variando la constante 1 entre 0 y -0.18 para volatilidades CIR entre 0 y 6%. Para el caso de volatilidades analizadas entre 0 y 4% CIR (del orden de 100 puntos b asicos en t erminos absolutos y 16% en relativos para el ejemplo considerado), el efecto del car acter estoc astico de las tasas de inter es es m nimo. El comportamiento de las duraci on efectiva como funci on de la elasticidad negativa de los saldos sigue un patr on mon otono decreciente, como era de esperarse de la ecuaci on (15) derivada en el an alisis est atico. Para volatilidades mayores se incrementa el plazo efectivo, pero el orden de magnitud y comportamiento cualitativo se mantiene. El aumento en la duraci on para volatilidades altas conforme aumenta la elasticidad negativa de los saldos se puede entender considerando que su efecto es reducir la contribuci on relativa de los escenarios de tasas a la alza respecto al l mite de volatilidad cero, aumentando la contribuci on de los escenarios a la baja. Como estos u ltimos descuentan a tasas menores, se traducen en duraciones efectivas mayores, todo lo dem as constante.
344
s M. Tarriba Jesu
En la gr aca 4 se presentan los resultados correspondientes al caso (b), introduciendo en el modelo la opci on del banco de revisar a la baja los rendimientos cuando el margen se hace negativo. De nuevo, para volatilidades moderadas las cosas cambian poco, pero ahora para volatilidades entre 4 y 6% se observa una reducci on dr astica en la duraci on efectiva conforme la volatilidad aumenta, llegando a tomar inclusive valores negativos. El efecto de las volatilidades altas es poblar de escenarios la zona en que se activa la opci on de banco, en la que las cuentas se comportan fundamentalmente como un instrumento a tasa otante, i.e., de duraci on cero. Para volatilidades sucientemente altas, el efecto de la opci on de banco es tal que invierte el signo de la sensibilidad, lo que se traduce en duraciones efectivas negativas para el balance simplicado. En la gr aca 5 se presentan los resultados para el caso (c), a saldo constante, variando la velocidad de ajuste anualizada de la tasa y el rendimiento de las cuentas entre 0 y 0.45, sin considerar la opci on de banco. En este caso el efecto de la volatilidad es nulo, observ andose el mismo patr on decreciente conforme el instrumento se asemeja m as a un instrumento de tasa otante. La duraci on efectiva pasa de su valor de rendimiento jo (8.33 a nos aproximadamente) a un valor alrededor de 1 a no para 12 + = 12 = 0.45. Esto se puede entender si se piensa que una velocidad de reversi on anualizada del orden de 0.5 implica un tiempo medio de repreciaci on del rendimiento del orden de 2 a nos. Como adicionalmente se tiene que en este caso el pasivo est a aproximadamente a 50% de su valor par, el plazo efectivo al que se deben invertir los saldos para neutralizar la sensibilidad del balance simplicado se sit ua alrededor de 1 a no. En la gr aca 6 se presentan los resultados para el caso (d), a saldo constante, variando la velocidad de ajuste anualizada de la tasa y el rendimiento de las cuentas entre 0 y 0.45, pero ahora incluyendo la opci on del banco. De nuevo, el efecto de la opcionalidad modica signicativamente la duraci on efectiva para volatilidades altas, pero en contraste con el caso anterior, las duraciones para las diferentes volatilidades convergen conforme la velocidad de ajuste crece. Esto es de esperar, dado que conforme la relaci on entre el rendimiento y la tasa de inter es se parece m as al de un otante, el efecto de la opcionalidad disminuye.
345
Gr aca 3: Plazo efectivo vs elasticidad de corto plazo para diferentes volatilidades CIR. Caso sin opci on del banco
Gr aca 4: Plazo efectivo vs elasticidad de corto plazo para diferentes volatilidades CIR. Caso con opci on del banco
346
s M. Tarriba Jesu
Gr aca 5: Plazo efectivo vs velocidad de ajuste tasa-rendimiento para diferentes volatilidades CIR. Caso sin opci on del banco
Gr aca 6: Plazo efectivo vs velocidad de ajuste tasa-rendimiento para diferentes volatilidades CIR. Caso con opci on del banco
347
Referencias
[1] Hutchison, D., Pennachi, G., Measuring Rents and Interest Rate Risk in Imperfect Financial Markets: The Case of Retail Bank Deposits. Journal of Financial and Quantitative Analysis 31, no. 3, September 1996, P. 339. [2] Jarrow, R., van Deventer, D., Disease or Cure. Risk 9, no 2, Feb 1996. [3] Jarrow, R., van Deventer, D., The arbitrage-free valuation and hedging of demand deposits and credit card loans. Journal of Banking & Finance 22 (1998) 249-271. [4] Judge, G. et al., The Theory and practice of econometrics. Wiley series in Probability and Mathematical Statistics. John Wiley, 1980. [5] Moore, G. R. et. al., Modeling the Disaggregated Demands for M2 and M1: The U.S. Experience in the 1980s. Board of Governors of the Federal Reserve System, 1990. [6] Nefti, Salhi N., Mathematics of Financial Derivatives. Academic Press 1996 [7] OBrien, James M., Estimating the Values and Interest Rate Risk of InterestBearing Transaction Deposits. Board of Governors of the Federal Reserve System, 2000. [8] OBrien, J., Orphanides, A., Small, D., Estimating the interest Rate Sensitivity of Liquid Retail Deposits Values. Proceeding of the Bank Structure and Competition 1994 conference. p. 400-435. Federal Reserve Bank of Chicago. [9] Rebonato, R., Interest-Rate Option Models, 2nd ed. Wiley 1996. [10] Sanyal, Ani, The Core of the Matter (parts I and II). Balance Sheet -Auntum 1997 and Winter 1997-98. [11] Selvaggio, R. D., Using the OAS Methodology to value and hedge commercial bank retail demand deposits premiums, in The Hand Book of Asset/Liability Management. Compiled by F. Fabozzi, Irwin 1996. [12] Uyemura, D., van Deventer, D., Financial Risk Management in Banking pp. 188., Irwin 1993.
Jes us M. Tarriba Unger Banco Santander Central Hispano Paseo de la Castellana 75, p. 6 28046-Madrid e-mail: jtarriba.madrid@sinvest.es
An alisis cr tico del mercado el ectrico espa nol

Rafael de Benito1
Resumen La liberalizaci on del mercado el ectrico iniciada en espa na el 1 de enero de 1998 culminar a el 1 de enero de 2003 con la posibilidad de que todos los clientes podamos elegir suministrador. Desde hace ya casi cuatro a nos, los precios de la electricidad en el mercado mayorista no est an intervenidos y surge un nuevo riesgo que las empresas el ectricas deber an gestionar. En este art culo se resumen los hechos m as relevantes ocasionados por las aprobaci on de la Ley 54/97 de 27 de noviembre del Sector El ectrico, incidiendo especialmente en el funcionamiento del mercado de contado gestionado por la empresa OMEL. Asimismo, se hace especial hincapi e en el desarrollo de instrumentos nancieros para le gesti on del riesgo de precio. Por u ltimo, se realiza una breve menci on a alguno de los enfoques que han tratado de explicar el comportamiento del precio de la electricidad, el cual registro altos niveles de volatilidad motivado, principalmente, por la imposibilidad de poder almacenar esta mercader a.
Introducci on
La Ley 54/97 de 27 de noviembre representa la piedra angular del proceso de liberalizaci on del mercado el ectrico espa nol. Una de las consecuencias m as importantes de la nueva regulaci on es la creaci on de un mercado de contado de la energ a el ectrica, en el que oferta y demanda presentan sus ofertas de venta y adquisici on de electricidad para el d a siguiente. En el nuevo marco legal se separan las diferentes actividades de generaci on, comercializaci on, transporte y distribuci on. Las dos primeras pasan a ser actividades liberalizadas sujetas a las fuerzas del mercado, mientras que las otras dos conservan su status de reguladas. No todos los clientes tienen la posibilidad de acceder al nuevo mercado libre de electricidad, sino que deben registrar un m nimo de consumo anual de energ a el ectrica. Estos son los llamados clientes cualicados. Estos clientes pueden adquirir la electricidad, bien acudiendo directamente al pool (mercado organizado de contado), bien contratando con un generador, bien a trav es de un comercializador o bien manteni endose con la tarifa. Estas modalidades son, de momento, compatibles entre s .
1 Rafael de Benito es Director de la Divisi on de Energ a de FC&M y responsable del desarrollo del mercado de futuros de electricidad de la Sociedad Rectora del Mercado de Futuros y Opciones sobre C tricos. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de octubre de 1999.
350
Rafael de Benito
Los umbrales de consumo para acceder a la categor a de cualicado se han ido reduciendo paulatinamente de la manera en que se muestra en la tabla adjunta. Fecha 1998 01/01/99 01/04/99 01/07/99 01/10/99 01/07/2000 Umbral de consumo anual 15 GWh 5 GWh 3 GWh 2 GWh 1 GWh 1 kV (nivel de tensi on)
El n umero de clientes elegibles el 1 de octubre de 1999 se estima en unos 7.850, cuyo consumo anual ascender a a unos 72.500 GWh (alrededor del 47,5% del total espa nol). El valor monetario de este consumo se situar a en el entorno de los 621.795 millones de pesetas (3.737 millones de euros). Un nuevo criterio de elegibilidad para adquirir la condici on de cliente cualicado se estableci o con el objeto de que los consumidores conectados a la red a una tensi on de 1 kV puedan serlo desde el 1 de julio de 2000. Se estima que la cifra de clientes cualicados se sit ua en unos 68.000, con lo que el grado de apertura traer a consigo un escenario de alta competitividad en el a mbito de la comercializaci on. En enero del 2007 est a previsto que todos los consumidores puedan libremente elegir suministrador, si bien desde el Ministerio de Industria y Energ a se ha apuntado recientemente que esa fecha podr a adelantarse en tres a nos como ya ha ocurrido con los umbrales previos, que en alg un caso se ha adelantado en casi 5 a nos. Esta fecha ha sido modicada por un Real Decreto posterior al cual se hace referencia en un punto posterior. Por lo que se reere a la nueva gura de los comercializadores, estos pueden vender electricidad a los clientes cualicados debiendo adquirirla previamente en el pool. Son, por lo tanto, una gura que, al igual que productores y clientes cualicados que acuden al pool, est a expuesta al riesgo de precios.
Funcionamiento, organizaci on y caracter sticas del Mercado Mayorista de Electricidad
El mercado mayorista de electricidad est a gestionado por la Compa n a Operadora del Mercado, OMEL (a partir de ahora pool ). En dicho mercado se vende y adquiere la energ a el ectrica para el d a siguiente. Esto signica, como es com un en todos los mercados de electricidad, que el contado equivale a que la transacci on debe tener lugar en D + 1. Las transacciones cuya perfecci on tiene lugar en un plazo superior son consideradas operaciones a plazo y, en todo caso, en la actualidad OMEL no gestiona un mercado cuyo horizonte temporal rebase el d a siguiente o day ahead, como se indica en lengua inglesa.
lisis cr ol Ana tico del mercado el ectrico espan
351
Seg un el art culo 2 del R.D. 2019/97 por el que se organiza y regula el mercado de producci on de energ a el ectrica, El Mercado de Producci on de energ a el ectrica es el integrado por el conjunto de transacciones comerciales de compra y venta de energ a y de otros servicios relacionados con el suministro de energ a el ectrica. El Mercado de Producci on de energ a el ectrica se estructura en Mercado diario y Mercado de servicios complementarios, integr andose tambi en en el los contratos bilaterales f sicos. Adicionalmente, existir a un mercado de ajustes de programaci on de car acter intradiario, que se denomina Mercado Intradiario. Los agentes que act uan o que pueden actuar en el mercado de electricidad son los que a continuaci on se enumeran: - Productores. - Distribuidores (mientras subsista la tarifa regulada). - Comercializadores. - Agentes externos (compradores y vendedores). - Consumidores cualicados. A grandes rasgos, la denici on de cada uno de estos cinco grupos es la siguiente: - Los productores son las empresas de generaci on que vierten electricidad a la red. Pueden vender su energ a al pool o rmar contratos bilaterales f sicos con clientes cualicados. - Los distribuidores son los responsables de las redes de media y baja tensi on. Para aquellos consumidores sujetos a tarifa regulada (la mayor a) el distribuidor hace tambi en las veces de comercializador, pues es quien factura por la energ a consumida. - Los comercializadores son una gura nueva creada por la Ley El ectrica de 1997, y son quienes pueden vender a los clientes cualicados. No han de ser necesariamente titualres de la red de distribuci on que llega hasta el contador del cliente nal. Con la normativa actual, la u nica manera en que pueden adquirir energ a el ectrica es compr andola en el pool. - Los agentes externos son aquellas empresas que importan a Espa na energ a el ectrica desde fuera de nuestras fronteras (vendedores) o que la exportan desde Espa na (compradores). - Por u ltimo, los consumidores cualicados son aqu ellos que pueden acceder al mercado liberalizado, comprando a un comercializador (la pr actica totalidad opta por esta posibilidad), acudiendo directamente al pool o acordando un contrato bilateral f sico con un productor.
352
Rafael de Benito
El modelo de organizaci on del mercado de contado de electricidad se asienta sobre los siguientes principios: Basado en sistema de ofertas de venta y adquisici on. El precio se ja por medio de un sistema de subasta en la que se aplica un principio de precio marginal. Per odo de programaci on: una hora. Horizonte de programaci on: el d a siguiente. Cierre recepci on de ofertas para D + 1: 10:00 h d a D. Resultados para D + 1: 11:00 h d a D. Respecto al per odo de programaci on, esto signica que para cada d a en que se realiza una sesi on de mercado se calculan 24 precios, uno por cada una de las 24 horas del d a. Seg un el art culo 8 del R.D. 2019/1997, Los agentes del mercado podr an presentar ofertas de venta de energ a el ectrica para cada per odo de programaci on, en el horario que se establezca en las normas de funcionamiento del mercado. Estas ofertas tendr an car acter de compromiso en rme una vez superado el plazo de admisi on establecido. Las ofertas de venta de energ a el ectrica deber an incluir, al menos, el precio y cantidad ofertada, la identicaci on del agente que las realiza y la unidad de producci on a que se reere. Por otro lado, el art culo 9 de ese mismo R.D. reza: Los agentes del mercado podr an presentar ofertas de adquisici on de energ a el ectrica para cada per odo de programaci on, en el horario que se establezca en las normas de funcionamiento del mercado. Estas ofertas tendr an car acter de compromiso en rme una vez superado el plazo de admisi on establecido. Las ofertas de adquisici on de energ a el ectrica deber an incluir la cantidad de energ a demandada, la identicaci on del agente y el per odo de programaci on a que se reere la oferta. Estas ofertas de adquisici on podr an incluir tambi en precio de la energ a demandada. En este caso, si la oferta no resultara casada, de acuerdo con el procedimiento previsto en el art culo 10, no se producir a suministro de energ a el ectrica.
353
En la Figura 1 se ilustra el proceso inicial que se realiza en el mercado diario.
Figura 1: Proceso inicial que se realiza en el mercado diario En esta primera fase, cuya responsabilidad en la gesti on corresponde al operador del mercado, deben ser tenidos en cuenta los contratos bilaterales f sicos. Los contratos bilaterales f sicos vienen regulados por los art culos 19 y 20 del citado R.D.. Seg un el art culo 19, Los consumidores cualicados o los agentes externos podr an formalizar con productores o agentes externos contratos bilaterales f sicos de suministro de energ a el ectrica. [. . . ] Las unidades de producci on que estuvieran afectas al cumplimiento de estos contratos quedar an exceptuadas de la obligaci on de presentar ofertas en el Mercado de Producci on, por la parte de su energ a generada vinculada al cumplimiento del contrato.. Es decir, la energ a el ectrica se puede negociar en el ambito mayorista entre los agentes indicados, por lo que la obligaci on de ofertar la energ a al pool lo es para aquella energ a no afecta al cumplimiento de contratos bilaterales f sicos. La cuesti on de los contratos bilaterales f sicos ser a de nuevo abordada m as adelante. Una vez son conocidas por el operador del mercado las unidades y cantidad de energ a comprometidas en los contratos bilaterales f sicos, y habiendo realizado previamente la casaci on de oferta y demanda al pool, se publica el programa base de funcionamiento. El siguiente paso consiste en la resoluci on de las denominadas restricciones t ecnicas por parte del operador del sistema, tal y como aparece en la gura 2.
354
Rafael de Benito
Figura 2: Incorporaci on de las restricciones t ecnicas al mercado de contado Seg un el art culo 12 del R.D 2019/97, El Programa Diario Base ser a comunicado por el operador del mercado a los agentes del mercado y al operador del sistema, quien a la vista del mismo, determinar a las restricciones t ecnicas que pudieran afectar a su ejecuci on, as como las necesidades de servicios complementarios a que diera lugar. A los efectos del presente Real Decreto, se entender a por restricci on t ecnica cualquier limitaci on derivada de la situaci on de la red de transporte o del sistema para que el suministro de energ a el ectrica pueda realizarse en las condiciones de seguridad, calidad y abilidad que se determinen reglamentariamente y a trav es de los procedimientos de operaci on. Para solventar las restricciones t ecnicas, el operador del sistema acordar a con el operador del mercado la retirada de la casaci on de las ofertas de venta que sean precisas y la entrada de otras ofertas presentadas en dicha sesi on, respetando el orden de precedencia econ omica. Las unidades de producci on que hubieran de entrar en funcionamiento como consecuencia de las citadas restricciones t ecnicas, percibir an por su energ a la retribuci on que corresponda por la oferta que hubieran presentado para aquellos per odos de programaci on en que funcionen. El programa resultante de la incorporaci on de estas nuevas transacciones y del resultado del Mercado de servicios complementarios a que se reere el art culo 14 se denominar a Programa Diario Viable y ser a comunicado por el operador del sistema al operador del mercado y a los agentes del mercado. . . Respecto de los servicios complementarios, el art culo 13 del R.D. mencionado, Se entiende por servicios complementarios aquellos que resultan necesarios para asegurar el suministro de energ a el ectrica en las condiciones de calidad, abilidad y seguridad necesarias. La asignaci on de servicios complementarios no obligatorios se rige tambi en por unos principios de mercado, pero no son objeto de an alisis en este art culo (ver gura 3).
355
En cualquier caso, los servicios complementarios representan un componente m as en la composici on del precio nal de la electricidad.
Figura 3: Incorporaci on de los servicios complementarios al mercado de contado Asimismo, existe un mercado denominado intradiario, del cual se celebran seis sesiones al d a con diferentes horizontes de programaci on. Seg un el art culo 15 del R.D. 2019/97, El Mercado intradiario tiene por objeto atender los ajustes que en la oferta y demanda de energ a se puedan producir con posterioridad a haberse jado el Programa Diario Viable. La tabla 1 muestra la distribuci on de horarios de las diferentes sesiones de mercado intradiario, as como sus diferentes horizontes de programaci on.
Sesi on 1 a 16:00 17:45 18:30 18:45 19:20 19:35 Sesi on 2 a 21:00 21:45 22:30 22:45 23:10 23:20 Sesi on 3 a 01:00 01:45 02:30 02:45 03:10 03:20 Sesi on 4 a 04:00 04:45 05:30 05:45 0:10 06:20 Sesi on 5 a 08:00 08:45 09:30 09:45 10:10 10:20 Sesi on 6 a 12:00 12:45 13:30 13:45 14:10 14:20
Apertura de sesi on Cierre de sesi on Casaci on Recepci on de desgloses An alisis de restricciones Recuadre por restricciones. Publicaci on PHF Horizonte de programaci on (periodos horarios)
28 horas (21-24)
24 horas (1-24)
17 horas (8-24)
17 horas (9-24)
13 horas (12-24)
9 horas (16-24)
Tabla 1: Los mercados intradiarios
356
Rafael de Benito
Funcionamiento del mercado diario
De los diferentes mercados que componen el mercado de producci on, sin duda, el m as relevante de todos ellos es el mercado diario, pues representa el porcentaje mayoritario del precio nal de la electricidad. En ella intervienen los agentes que han sido mencionados en los primeros apartados. Las ofertas econ omicas de venta de energ a el ectrica que los vendedores presenten al operador del mercado pueden ser simples o incorporar condiciones complejas en raz on de su contenido. Las ofertas simples son ofertas econ omicas de venta de energ a que los vendedores presentan para cada periodo horario y unidad de producci on de la que sean titulares con expresi on de un precio y de una cantidad de energ a. Las ofertas que incorporan condiciones complejas de venta son aquellas que, cumpliendo con los requisitos exigidos para las ofertas simples, incorporan adem as todas, algunas o alguna de las condiciones t ecnicas o econ omicas siguientes: - Condici on de indivisibilidad. - Gradiente de carga. - Ingresos m nimos. - Parada programada. La condici on de indivisibilidad permite jar en el primer tramo de cada hora un valor m nimo de funcionamiento. Este valor solo puede ser dividido por la aplicaci on de los gradientes de carga declarados por el mismo agente, o por aplicaci on de reglas de reparto en caso de ser el precio distinto de cero. El gradiente de carga permite establecer la diferencia m axima entre la potencia inicio de hora y la potencia nal de hora de la unidad de producci on, lo que limita la energ a m axima a casar en funci on de la casaci on de la hora anterior y la siguiente, para evitar cambios bruscos en las unidades de producci on que no pueden, t ecnicamente, seguir las mismas. La condici on de ingresos m nimos permite la realizaci on de ofertas en todas las horas, pero respetando que la unidad de producci on no participe en el resultado de la casaci on del d a si no obtiene para el conjunto de su producci on en el d a un ingreso superior a una cantidad ja, establecida en pesetas o euros, m as una remuneraci on variable establecida en pesetas o c entimos de euro por cada kWh casado. La condici on de parada programada permite que si la unidad de producci on ha sido retirada de la casaci on por no cumplir la condici on de ingresos m nimos solicitada, realice una parada programada en un tiempo m aximo de tres horas, evitando parar desde su programa en la u ltima hora del d a anterior a cero en la primera hora del d a siguiente, mediante la aceptaci on del primer tramo de las tres primeras horas de su oferta como ofertas simples, con la u nica condici on de que la energ a ofertada sea decreciente en cada hora.
357
El operador del mercado realizar a la casaci on de las ofertas econ omicas de compra y venta de energ a el ectrica (recibidas antes de las 10:00 horas del d a) por medio del m etodo de casaci on simple o compleja, seg un concurran ofertas simples o que existan ofertas que incorporen condiciones complejas. El m etodo de casaci on simple es aqu el que obtiene de manera independiente el precio marginal, as como el volumen de energ a el ectrica que se acepta para cada unidad de producci on y adquisici on en cada periodo horario de programaci on. El m etodo de casaci on compleja obtiene el resultado de la casaci on a partir del m etodo de casaci on simple, al que se a naden las condiciones de indivisibilidad y gradiente de carga, obteni endose la casaci on simple condicionada. Mediante un proceso iterativo se ejecutan varias casaciones simples condicionadas hasta que todas las unidades de oferta casadas cumplen la condici on de ingresos m nimos as como de parada programada, siendo esta soluci on la primera soluci on nal provisional, obtenida considerando una capacidad ilimitada en las interconexiones internacionales. Mediante un proceso iterativo se obtiene la primera soluci on nal denitiva que respeta la capacidad m axima de interconexi on internacional, considerando tanto las ofertas realizadas al mercado diario, como las ejecuciones de contratos bilaterales f sicos con afectaci on expresa a dichas interconexiones. El precio en cada periodo horario ser a igual al precio del u ltimo tramo de la oferta de venta de la u ltima unidad de producci on cuya aceptaci on haya sido necesaria para atender la demanda que haya resultado casada. Como resultado de la casaci on, el operador del mercado obtiene el Resultado de la Casaci on, entendiendo por el mismo la programaci on de entrada en la red establecida por el operador del mercado a partir de la casaci on de las ofertas de venta y adquisici on de energ a el ectrica y en el que se determina, para cada periodo horario de un mismo horizonte diario, el volumen de energ a el ectrica que se requiere que se produzca para cubrir la demanda de dicha energ a el ectrica. Desde el punto de vista de la formaci on del precio, debe incidirse en el hecho de que se basa en un sistema marginalista y que es la u ltima unidad de producci on que entra en la casaci on la que determina el precio, es decir, la m as ineciente de todas las que son casadas. Asimismo, el volumen con el que se marca el precio marginal para cada per odo de programaci on puede ser m nimo, no siendo representativo de la oferta en t erminos generales, si bien todos los oferentes se ven beneciados de un precio m as elevados, pues todos cobran el mismo precio. En la gura 4 se muestra el resultado de la casaci on correspondiente al 22 de junio de 2000.
358
Rafael de Benito
Figura 4: 22 de junio de 2000
La Garant a de Potencia
La garant a de potencia es otro de los componentes del precio nal de la electricidad. Seg un el art culo 24 del R.D. 2019/97, La retribuci on por garant a de potencia tiene por objeto proporcionar una se nal econ omica para la permanencia e instalaci on de capacidad de generaci on en el sistema el ectrico, con el objeto de conseguir un nivel de garant a de suministro adecuado. Las unidades de producci on obligadas a presentar ofertas econ omicas de venta (seg un el art culo 23 del R.D. 6/2000) tendr an derecho a obtener la retribuci on por garant a de potencia, siempre que hayan acreditado un funcionamiento de 480 horas equivalentes a plena carga durante el u ltimo a no. Por el contrario, los agentes externos en sus importaciones, los generadores en r egimen especial y los vendedores en los contratos bilaterales f sicos no tienen derecho de cobro por este concepto. La cuant a mensual por garant a de potencia asciende a 1,15 Pta/KWh por la demanda anual en barras de central de los clientes nales nacionales. La demanda en barras de central signica que no tienen en cuanta las p erdidas que se producen debido a la transmisi on de la electricidad por las diversas redes de transporte y distribuci on. Los comercializadores, consumidores cualicados y agentes externos efectuar an un pago por garant a de potencia producto de la energ a adquirida en cada hora por el precio correspondiente a la misma. Los precios horarios para cada uno de los periodos aparecen en la tabla 2.
359
Tabla 2: Garant a de potencia
Componentes del precio nal de la Electricidad
A efectos de liquidaci on, el precio de la energ a el ectrica a pagar por el comprador y a percibir por el vendedor incorporar a: - Precio de casaci on del mercado diario. - Coste resultante del proceso de soluci on de restricciones t ecnicas. - Coste del mercado de servicios complementarios. - Precio de casaci on del mercado intradiario. - Coste de la garant a de potencia. - Coste de los procesos de operaci on t ecnica del sistema. - Exceso/d ecit de los contratos internacionales a largo plazo. Debe se nalarse que, con relaci on a los servicios complementarios, estos ser an abonados por aquellos vendedores y compradores de electricidad que hayan provocado la necesidad de su utilizaci on. Por otro lado, el precio de casaci on del mercado intradiario ser au nicamente relevante para quienes hayan intervenido en el mismo.
360
Rafael de Benito
Los Costes de Transici on a la Competencia (CTC)
La regulaci on inicial de los CTCs aparec a en la disposici on transitoria sexta de la Ley 54/97 del Sector El ectrico. En dicha disposici on transitoria se establece: Se reconoce la existencia de unos costes de transici on al r egimen de mercado competitivo, previsto en la presente Ley, de las sociedades titulares de instalaciones de producci on de energ a el ectrica, que a 31 de diciembre de 1997 estuvieran incluidas en el a mbito de aplicaci on del Real Decreto 1538/1987, de 11 de diciembre. . . Durante un plazo m aximo de diez a nos desde la entrada en vigor de la presente Ley, el Gobierno podr a establecer anualmente el importe m aximo de esta retribuci on ja con la distribuci on que corresponda. . . Los costes que se deriven de esta retribuci on ser an repercutidos a todos los consumidores de energ a el ectrica como costes permanentes del sistema, en los t erminos que reglamentariamente se establezcan y su importe base global, en valor a 31 de diciembre de 1997, nunca podr a superar 1.988.561 millones de pesetas. . . Si el coste medio de generaci on a que se reere el art culo 16.1 de la presente Ley a lo largo del per odo transitorio, resultara en media anual superior a 6 pesetas por kWh, este exceso se deducir a del citado valor actual. El coste medio de generaci on denido en el art culo 16.1 comprende los siguientes conceptos: precio marginal, coste de las p erdidas en el transporte, garant a de potencia y servicios complementarios. Los controvertidos CTCs representan un factor que afecta de diversos modos al mercado de electricidad. Por un lado, act uan como un instrumento de cobertura natural en un entorno de precios bajos y, por otro, dicha cobertura act ua como una barrera de entrada en el mercado, puesto que las empresas propietarias de centrales de nueva construcci on no ser an acreedoras del cobro por este concepto. La resoluci on de la pol emica generada por la posible titulizaci on de los CTCs y su aprobaci on nal por parte del Comisario de la Competencia, Mario Monti, implicar a el esclarecimiento de un escenario de mercado que hasta ahora se ve distorsionado por este factor.
Los Contratos Bilaterales F sicos
Como ya ha sido indicado, los art culos 19 y 20 del R.D. 2019/97 regulan este tipo de contratos, siendo los consumidores cualicados o los agentes externos quienes pueden formalizar con productores o agentes externos este tipo de contratos de suministro de energ a el ectrica. Lo primero que llama la atenci on es que los comercializadores no pueden rubricar esta modalidad de contratos. Esta circunstancia es especialmente relevante para los comercializadores independientes, quienes se ven abocados a acudir al pool como tambi en han de hacerlo los comercializadores que pertenecen a las compa n as el ectricas que poseen activos de generaci on para adquirir la energ a el ectrica que posteriormente vender an a los consumidores nales.
361
Impacto de las medidas liberalizadoras del Real Decreto Ley 6/2000 de 23 de junio
Con el objetivo de incrementar el nivel de competencia en diferentes sectores de la econom a espa nola, el Gobierno aprob o el 23 de junio de 2000 un paquete de medidas que, entre otros, afectan al sector el ectrico (art culos 16 a 28, ambos inclusive). Cabe destacar los siguientes aspectos de la nueva norma: Establecimiento de unos l mites para el crecimiento de las empresas con mayor cuota de mercado, que afecta principalmente a Endesa e Iberdrola, si bien no se obliga a desinvertir en activos de generaci on. Reducci on de la cuant a por garant a de potencia a 0,80 Ptas./kWh. Adelanto de la apertura total del mercado al a no 2003. Reducci on de requisitos para el ejercicio de la condici on de cliente cualicado. Incorpora nuevas posibilidades de actuaci on a los comercializadores. A continuaci on se realiza una explicaci on m as detallada con relaci on a los tres u ltimos puntos.
8.1
Apertura total del mercado el 1 de enero de 2003
Todos los consumidores de electricidad tendr an la posibilidad de elegir suministrador, por lo que se espera un incremento en el nivel de competencia. Las experiencias del Reino Unido y de Alemania hacen prever que los precios puedan bajar si bien, dada la reducci on de la tarifa regulada desde el inicio del proceso liberalizador (en el RD 6/2000 se determina que en los pr oximos tres a nos la tarifa debe bajar un 9% ), el margen para bajadas ulteriores se reduce signicativamente. Un problema que se ha presentado en el Reino Unido a los clientes dom esticos ha sido el de la doble facturaci on, tras haber cambiado de suministrador. Este tipo de problemas burocr aticos y administrativos, junto con el pol emico asunto de los contadores, podr a surgir una vez se liberalice en su plenitud el mercado. En cualquier caso, la integraci on de servicios a trav es de un u nico proveedor (electricidad, gas, incluso agua y tel efono) suele ser uno de los rasgos fundamentales de mercados liberalizados y con un cierto grado de madurez, por lo que no ser a raro observarlo en Espa na dentro de unos a nos.
8.2
Ejercicio de la condici on de cliente cualicado
El art culo 20 de RD 6/2000 a nade un nuevo art culo, el 21 bis, al RD 2019/1997, por el que se regula el mercado de producci on de energ a el ectrica. Hasta ahora, los consumidores que deseasen ejercer su condici on de cualicados deb an inscribirse en el
362
Rafael de Benito
Registro Administrativo de Distribuidores, Comercializadores y Clientes Cualicados. Tras la aprobaci on de este RD, ya no es necesario, bastando para acreditar su derecho a ser consumidor cualicado la certicaci on sobre la caracter stica de su consumo o nivel de tensi on de suministro expedido por la empresa o empresas distribuidoras con las que est e conectado, tal y como reza el art culo arriba indicado. Por el contrario, aquellos consumidores cualicados que quieran intervenir directamente en el mercado (comprando en OMEL), s deber an registrarse en el mencionado Registro. Los distribuidores deber an remitir gratuitamente a todos aqu ellos de sus clientes, si lo solicitan, que vayan a adquirir la condici on de cualicados en 15 d as (de momento no se prev e la adquisici on de dicha categor a a nuevos clientes hasta 2003) de un certicado en el que har an constar su nivel de tensi on de suministro, identicaci on de la tarifa correspondiente, potencia o potencias contratadas, el consumo mensual de los dos a nos anteriores y el detalle de su facturaci on de dicho per odo. Con estas medidas, se pretende facilitar y agilizar el que, principalmente, las pymes puedan recibir ofertas de los diferentes comercializadores, pudiendo estos acceder a cierta informaci on t ecnica e hist orica del cliente en cuesti on, y posibilitando as la formulaci on de una oferta m as adecuada.
8.3
Nuevas posibilidades para los comercializadores
De acuerdo con el art culo 21 del RD 6/2000, los comercializadores ven ampliado su campo de actuaci on, poniendo a su alcance nuevas f ormulas contractuales que les permitir an tener un mayor papel en el mercado, sobre todo en lo que se reere a los comercializadores independientes que no poseen capacidad de generaci on. Hasta ahora, los comercilizadores s olo pod an adquirir energ a el ectrica acudiendo a OMEL. A partir de ahora, tienen otras opciones. En primer lugar, los comercializadores de energ a el ectrica podr an realizar contratos de adquisici on de energ a el ectrica con empresas autorizadas a la venta de energ a el ectrica en pa ses de la Uni on Europea o terceros pa ses, as como con productores nacionales de electricidad en r egimen especial y a partir del 1 de enero de 2003 o cuando todos los consumidores tengan la condici on de cualicados, tambi en con productores nacionales en r egimen ordinario. Dicha energ a podr a venderse a los consumidores cualicados o integrarse en los mercados diarios o intradiarios existentes. Si bien es un paso adelante respecto a la actual regulaci on, como bien se se nala en la Nota T ecnica no 67 de Analistas Financieros Internacionales, los contratos de suministro con entidades de fuera de Espa na est an condicionadas por las limitaciones en la capacidad de interconexi on el ectrica con Francia. Dicha capacidad de interconexi on es de 1.100 MW de entrada y 1.000 MW de salida. Por otro lado, los productores en R egimen Especial perciben en media una prima de tres pesetas sobre los precios
lisis cr ol Ana tico del mercado el ectrico espan del pool y su potencia est a limitada a 50 MW. En estas condiciones es dudoso que estos productores vayan a actuar de manera signicativa en la cobertura de las necesidades de energ a de los comercializadores.
363
En el supuesto de que comercializadores independientes puedan adquirir algunas cantidades de energ a el ectrica por alguna de las v as que se les abren ahora, no s olo podr an revenderla a clientes nales, sino que tienen la posibilidad de venderla en el pool, bien tanto en el mercado diario como en el intradiario, con lo que pueden generar un efecto modulador que ser a tanto mayor cuanta mayor sea la energ a el ectrica que puedan negociar. Mucho m as importante, pero sin consecuencias hasta el a no 2003, es el hecho de que los comercializadores podr an comprar directamente la energ a el ectrica a los productores espa noles en r egimen ordinario. Esta circunstancia posibiltar a que los propietarios de las centrales, cuya construcci on est a prevista que nalice aproximadamente para la primera mitad de ese a no, puedan negociar acuerdos de venta con comercializadores por la totalidad o parte de la energ a que produzcan. En Espa na est a proyectada la construcci on de m as de 4.000 MW de potencia correspondientes a merchant plants, es decir, empresas que poseen capacidad de generaci on pero no poseen comercializaci on (ver gura 5). Este escenario, propicio para la realizaci on de contratos estructurados vinculando el precio de la electricidad con el precio de suministro del gas (combustible de todas estas nuevas centrales y materia prima esencial en la liberalizaci on de los sectores el ectricos de toda Europa), aportar a un alto grado de dinamismo al mercado el ectrico espa nol.
Figura 5: Merchant Plants proyectadas
364
Rafael de Benito
Los Mercados de derivados sobre la Electricidad
Con la liberalizaci on de los mercados el ectricos en muchos pa ses del hemisferio occidental, la electricidad se ha convertido en una commodity m as, y como tal, ha sido negociada a trav es de distintos tipos de contratos derivados. La electricidad registra una caracter stica que le dota de una singularidad especial: no es almacenable. Este rasgo de la energ a el ectrica condiciona la manera en que estos contratos deben ser liquidados una vez se alcanza su vencimiento, con independencia de que dicha liquidaci on tenga lugar mediante una entrega f sica o tenga lugar por diferencias contra un precio o ndice de referencia. T omese como ejemplo los contratos de futuros de electricidad negociados en el Nord Pool noruego. Los diferentes contratos cotizados lo son sobre la base del per odo de entrega denido. Cuanto m as lejano en el tiempo, el per odo de entrega es mayor. Cuanto m as cercano en el tiempo, dicho per odo es menor, pero no se puede concebir un contrato de electricidad sin vincularlo con el per odo de entrega.
Figura 6: Cotizaci on de contratos de futuros en el Nord Pool La gura 6 muestra la forma en que se cotizan los contratos de futuros de electricidad en el Nord Pool. A como m nimo un a no antes del vencimiento, se negocian las ua por debajo del a no, la estaci on estaciones2 . Una vez el plazo al vencimiento se sit se divide en varios bloques de cuatro semanas que previamente compon an la estaci on. Al situarse a cuatro semanas al vencimiento, el bloque se divide en cuatro semanas
2 Debe signicarse que los contratos de estaci on fueron deslistados en octubre de 1999 en el mercado de futuros, permaneciendo en el mercado de forwards, mucho m as l quido para este tipo de contratos. En cualquier caso, se entiende procedente incluir en la explicaci on esta gura pues ha estado vigente varios a nos (sigue est andolo en el mercado de forwards ), y facilita la comprensi on del mecanismo de funcionamiento de los contratos de futuros que nos ata nen.
365
individuales. Cuando un per odo se divide en otros m as peque nos, estos u ltimos pueden negociarse de forma individual y separadamente. Cabe la posibilidad de negociar los d as de forma individual, si bien el per odo de liquidaci on es la semana. De qu e forma se liquida? Nord Pool gestiona tanto el mercado spot como el mercado de futuros. El precio del mercado spot es p ublico y transparente. Conocido el precio de cierre del futuro para una semana en cuesti on (precio que se conoce al concluir la sesi on de mercado correspondiente al viernes previo al inicio de dicha semana), se liquida contra el precio medio diario del mercado spot. Al ser el mercado spot del Nord Pool un mercado horario, como en el caso espa nol, el precio medio se calcula como la media aritm etica de los 24 precios horarios publicados por Nord Pool. La citada liquidaci on del precio de futuro contra el precio medio del mercado de contado tiene lugar de forma diaria durante la semana de entrega. Si el precio medio spot de un d a de la semana de entrega se sit ua por encima del precio de futuros, la liquidaci on es favorable al comprador dicho d a, y ser a favorable al vendedor en el supuesto contrario, tal y como se ilustra en la gura 7.
Figura 7: Liquidaci on del precio del futuro Carece de todo sentido el cotizar la electricidad a plazo para ser entregada (f sica o nocionalemente) en un momento concreto, dada la imposibilidad material de almacenarla. Por ello, la entrega (o si se quiere, la liquidaci on a vencimiento) debe de ir vinculada, inevitablemente, a un per odo de entrega. Cuesti on diferente resulta el hecho de si la liquidaci on a vencimiento del instrumento derivado en cuesti on se liquida por diferencias o por entrega f sica. La existencia de un mercado centralizado sucientemente representativo o la existencia de ndices de precios aceptados por los agentes que participan en el mercado, posibilita la liquidaci on por diferencias. La ausencia de estos factores determinar a una liquidaci on por entrega. De aquellos pa ses donde en la actualidad est an abiertos a negociaci on contratos de futuros sobre energ a el ectrica (Noruega, Estados Unidos, Australia y Nueva Zelanda),
366
Rafael de Benito
tan s olo el pa s norteamericano liquida sus contratos a vencimiento por entrega. Ello es as por la inexistencia de un precio de referencia contra el que liquidar el precio de futuro. En cualquier caso, el procedimiento de entrega correspondiente a los contratos norteamericanos tiene una duraci on de un mes. La tendencia actual en pa ses como Inglaterra o Alemania en cuanto a la estructura del mercado es la de separarse del formato del mercado de subasta (a modo de pool ) y adoptar la negociaci on continua en bloques, siendo los mercados horarios de subasta una parte marginal o meramente inexistente. Este hecho provocar a que, si se desea negociar contratos de derivados que se liquiden a vencimiento por diferencias, habr a de dise narse unos ndices bajo unos principios diferentes aplicados a los mercados de subasta. En tanto en cuanto estos ndices no sean conformados, los contratos tendr an que ser liquidados a vencimiento por entrega.
10
Desarrollo de contratos derivados sobre Electricidad en el mercado espa nol
El mercado el ectrico espa nol en lo que se reere a contratos derivados OTC ha registrado un crecimiento lento, con ciertos altibajos en su camino, pero que parece haber encontrado unas bases s olidas que auguran un futuro con un grado de liquidez m as que aceptable. Tras dar sus primeros pasos serios a principios a principios de 1999, a nales del a no 2000 se puede armar que es un mercado que posee unos cimientos rmes pero que necesita elevar su liquidez, profundidad y volumen negociado. En un principio, el mercado OTC se desarroll o a partir de la actividad de cuatro operadores: SKS, Synergia, Enron y Electrabel. La liquidez de este mercado era ciertamente escasa, pero a un as los mismos agentes entendieron necesaria la presencia de agentes intermediarios (brokers ). La aparici on de un broker (HOB) agiliz o levemente la negociaci on. A lo largo del u ltimo trimestre de 1999 fueron entrando nuevos agentes (Sempra, Uni on Fenosa, etc.), haciendo m as atractivo el mercado. El err atico comportamiento de los precios en el mercado spot a lo largo del primer trimestre de 2000 lleg o a provocar una ralentizaci on de la negociaci on en el mercado OTC que lo dej o con una nula actividad durante unos d as. Este dr astico descenso en la actividad vino provocado por una falta de conanza en el mercado de contado por parte de los agentes m as activos en el mercado de derivados. Los contratos OTC negociados en Espa na cubren diferentes per odos: semanas, meses, restos de meses, trimestres, semestres, a no y otros. La negociaci on se ha venido centrando en per odos m as largos o m as cortos en funci on del per odo del a no, habiendo habido momentos donde la pr actica totalidad de operaciones se centraba en la semana siguiente o como mucho la posterior a la siguiente, mientras que en otros per odos los traders se han centrado casi con exclusividad en la negociaci on de meses.
367
Por otro lado, el perl de los contratos (es decir, qu e horas quedan englobadas por estos) se centra en la actualidad en carga base (las 24 horas del d a) y picos (desde la hora 9 a la hora 24, ambas inclusive). Otros perles, tales como las noches, los valles o los nes de semana han ido perdiendo inter es. El nominal de los contratos, aparte del per odo y del perl, depende de la potencia pactada. En el mercado espa nol, no se suele pactar m as all a de una potencia de 20 MW, siendo 10 MW la cantidad m as habitual. El tipo de contrato que mayoritariamente se he negociado ha sido el contrato por diferencias (CPD). En dicho contrato se ja un precio para un determinado per odo y con un determinado perl. Si el precio medio del mercado spot durante dicho per odo y con el perl especicado es superior al precio acordado a trav es del contrato, la parte vendedora abona la diferencia a la compradora. El montante econ omico depende, asimismo, de la potencia contratada. A continuaci on se ilustra con un ejemplo como se calcula la liquidaci on de una operaci on. El trader A (comprador) pacta con el trader B (vendedor) un contrato de horas pico, 20 MW para la semana n umero 47 del a no 2000 a un precio de 6,500 Pta/kWh. El nominal del contrato asciende a 20 MW x 80 horas = 1.600 MWh. Ello signica que a cada uno de los cinco d as comprendidos en el contrato le corresponden 320 MWh (320.000 kWh). Los precios medios de las horas pico (de lunes a viernes, siempre que sean laborables, y de las 8:00:01 de la ma nana hasta las 12:00 de la noche) para los d as comprendidos en el contrato son los siguientes (en Pta/kWh): Lunes: 6,564 Martes: 6,140 Mi ercoles: 6,987 Jueves: 7,215 Viernes: 6,841
D a Lunes: Martes: Mi ercoles: Jueves: Viernes: Total Precio medio spot 6,564 6,14 6,987 7,215 6,841 Precio pactado 6,500 6,500 6,500 6,500 6,500 Diferencia 0,064 -0,36 0,487 0,715 0,341 Liquidaci on 20.480 -115.200 155.840 228.800 109.120 399.040
En este ejemplo, la liquidaci on ha favorecido al comprador, puesto que el precio medio del mercado spot correspondiente a las horas comprendidas en el contrato ha sido superior en su conjunto al precio pactado.
368
Rafael de Benito
A lo largo de los meses en que el mercado OTC ha estado en funcionamiento, se han realizado otro tipo de operaciones, tales como opciones o spreads, si bien la mayor a de las transacciones han sido CPD. Como ejemplo real de la situaci on del mercado OTC, a continuaci on se muestra la tabla que a este respecto la empresa Platts publica diariamente en su publicaci on especializada European Power Daily (gura 8).
Figura 8: Datos de Platts sobre el mercado OTC en Espa na
11
La evoluci on del mercado spot en gr acos
Figura 9: Evoluci on del precio medio diario spot en 1998
369
Figura 10: Evoluci on del precio medio diario spot en 1999
Figura 11: Evoluci on del precio medio diario spot en 2000 (hasta agosto)
370
Rafael de Benito
12
Diferentes an alisis del comportamiento de los precios en los Mercados de Electricidad
Existen diversas opiniones al respecto de c omo explicar el comportamiento de los precios en los mercados el ectricos y de c omo calcular la volatilidad de aqu ellos. Por ello, parece adecuado citar dos autores que han tratado estos temas, como son Dragana Pilipovic y David Shimko. La primera es la m axima responsable de Sava Risk Management, empresa consultora para cuestiones de gesti on de riesgo enfocada al sector energ etico. Es autora del libro Energy Risk. Pilipovic ha centrado parte de su trabajo en elaborar un modelo de estimaci on de precios que recogen el fen omeno de reversi on a la media. El efecto de la reversi on a la media es an alogo a lo que sucede cuando hacemos sonar una cuerda de guitarra o pegamos un empuj on a un columpio: tanto la cuerda como el columpio tienden a un punto de central de equilibrio, si bien cada uno lo hace a una velocidad diferente. Con la reversi on a la media tiene lugar algo parad ojico, y es que no puede determinarse la velocidad de esta reversi on si la variable en cuesti on (precio, tipo de inter es), no se aparta de su punto de equilibrio por cualquier motivo y retorna de nuevo al mismo. En los mercados energ eticos, la reversi on a la media responde a razones de diferente ndole: viene principalmente determinado por la presencia de acontecimientos y de c omo la oferta reacciona ante los mismos o de con qu e velocidad estos acontecimientos desaparecen. Hay ejemplos muy claros al respecto: el verano pasado, en el medio oeste americano, el precio del MWh lleg o a alcanzar la incre ble cifra de $ 7.500, cuando los niveles normales se sit uan en torno a los $ 35-40/MWh. Esta subida vino motivada por varias razones: elevadas temperaturas provocando un uso masivo de aparatos de aire acondicionado, paradas de algunas centrales y problemas en la red de transporte. En cuanto estas circunstancias desaparecieron, o se corrigieron signicativamente, los precios volvieron r apidamente a sus niveles habituales. Pilipovic (1998) realiza una aportaci on muy interesante en consonancia con los trabajos posteriores de Smith & Schwartz (1999). En este modelo de dos factores, el primero coincide con el precio spot y el segundo factor es el precio de equilibrio a largo plazo. Se asume que el precio spot revierte hacia el precio de equilibrio. Este es un modelo muy utilizado en los mercados energ eticos. El modelo Pilipovic es el siguiente: dSt dLt = = (Lt St ) dt + St dzt , Lt dt + Lt dwt .
Por un lado, este modelo representa una variaci on algebraica del modelo de Gibson
lisis cr ol Ana tico del mercado el ectrico espan & Schwartz (1990) del precio spot y convenience yield : dSt St dyt = ( yt ) dt + 1 dwt , = k ( yt ) dt + 2 dut ,
371
siendo una constante, e yt el convenience yield, siendo este componente el que revierte hacia un nivel de equilibrio, capturando este modelo un comportamiento aleatorio de la forma de la curva forward (contango o backwardation ), pero sigue contemplando a la commodity subyacente en cuesti on como activo de inversi on y al convenience yield como si fuesen dividendos. El segundo factor, el precio a largo plazo Lt , es lognormal. A partir de este punto de partida, Pilipovic propone su modelo para el c alculo del precio futuro:
corto plazo largo plazo
subyacente Ft,T
L e (T t)/365 , contribuci on estacionalidad.
(S L) e(T t)/365 +
subyacente Ft,T +
estacional = Ft,T
Por u ltimo, para acometer el factor de la estacionalidad, Pilipovic propone 3 modelos diferentes que puedan capturarla. Modelo del coseno:
subyacente estacioanlidad Ft,T = Ft,T + a cos (2 (T ta )) + sa cos (4 (T tsa )) .
Modelo exponencial:
subyacente estacioanlidad Ft,T = Ft,T + a ea (T ta ) + sa esa (T tsa ) .
2 2
Modelo potencia de N :
subyacente estacioanlidad = Ft,T + Ft,T
a sa + . 1 + a (T ta )N 1 + sa (T tsa )N
Respecto a los grados de libertad, el modelo del coseno permite una correcta calibraci on de la altura (del pico debido a la estacionalidad) y de la fecha central. Los modelos exponencial y de la potencia en esima permiten, a su vez, calibrar no s olo altura y fecha central sino tambi en la anchura (la duraci on del per odo donde se registra la desviaci on respecto del precio de equilibrio debido a la estacionalidad). Respecto a la representaci on gr aca de los picos, el modelo de la potencia en esima da lugar a guras planas. El modelo exponencial es el que permite un mejor ajuste. En denitiva, la naturaleza de los mercados energ eticos y su implicaci on en el comportamiento de los precios provoca la necesidad de unos modelos matem aticos
372
Rafael de Benito
que expliquen dichos comportamientos. El ejemplo extremo es la electricidad, donde la velocidad de reversi on a la media puede llegar a ser muy elevada dando lugar a elevadas tasas de volatilidad, si bien hay quien arma que esta magnitud no es tan alta como pueda parecer, aportando criterios alternativos para su c alculo. Este es el caso de David Shimko, de Bankerstrust. David C. Shimko, m aximo responsable de la empresa perteneciente al grupo Bankerstrust-Deutsche Bank, Covar, plantea junto a Jean-Paul St.Germain en su art culo Substandard deviation, una metodolog a de c alculo de la volatilidad distinta a la tradicional y generalmente aceptada. Seg un Shimko, cuando se calcula la volatilidad de la manera enunciada arriba, se est a asumiendo que: 1. Los rendimientos de los precios presentan una media y volatilidad constantes; 2. Los cambios en los precios son independientes de los cambios anteriores. Esto u ltimo signica que los precios siguen un random walk o paseo aleatorio (el logaritmo de los cambios en los precios no hace otra cosa que convertir unos rendimientos peri odicos en un rendimiento compuesto continuo, evitando as la posibilidad de los precios negativos). Estos puntos de partida asumidos como ciertos se aproximan a la realidad cuando se trata de la renta variable, aunque ciertamente hay innumerables excepciones. No es tan able en el mercado de bonos, en el cual la volatilidad se reduce conforme el instrumento de deuda se aproxima al vencimiento. Y desde luego, la segunda de las armaciones es la que menos se cumple, especialmente en lo que se reere a tipos de inter es, divisas y mercader as. Por qu e los precios no describen un random walk ? Si el precio de la electricidad se dispara a $ 100, ciertamente no va a mantenerse en esos niveles mucho tiempo; es obvio para todos que el precio revertir a r apidamente a niveles m as razonables. En estad stica, este comportamiento es conocido como reversi on a la media. Seg un lo presenta Shimko en su art culo, la media de los cambios en los precios o la esperanza de los cambios en los precios ser a negativa cuando los estos registren valores elevados, y positiva cuando sean bajos. A trav es de una ecuaci on, puede ser expresado de la siguiente forma Pt+1 Pt = k (m Pt ) + s et , que puede parecer complicada al principio. Tan s olo establece que el cambio de precio proyectado (Pt+1 Pt ) es proporcional en promedio a la diferencia entre el precio actual y el precio medio a largo plazo m. El t ermino k indica la rapidez de la reversi on a la media, ya que, cuanto mayor sea dicho t ermino, m as r apido retornar a el precio hacia su valor medio a largo plazo. No resulta dif cil aplicar esta ecuaci on. Para ello, los autores sugieren una peque na ayuda:
373
- Proceder a realizar una regresi on entre los rendimientos de los precios y el precio anterior correspondiente. - La velocidad de la reversi on a la media corresponde a la pendiente estimada, con el signo contrario. - La constante estimada, al dividirlo por la velocidad de la reversi on a la media da como resultado la media a largo plazo. - La desviaci on est andar de los residuos es la volatilidad de los cambios de los precios (en d olares) despu es de ajustar a una media variable. La aportaci on de Pilipovoc y Shimko resulta muy interesante. Conforme los mercados liberalizados adquieran un mayor grado de madurez, la actividad de trading de electricidad va a adquirir una posici on preponderante dentro de las empresas involucradas en la generaci on, compra y venta de electricidad. Esa mayor madurez puede perfectamente implicar que los modelos que son aplicables ahora no lo sean dentro de un tiempo, puesto que la evoluci on de los precios responda m as a factores que ahora tienen poca importancia. Por ello, en los a nos venideros, no ser a extra no observar la publicaci on de nuevos modelos que intenten modelizar el comportamiento de los precios bajo otros principios.
Referencias
[1] P agina web de OMEL, www.omel.com. [2] Notas t ecnicas de AFI sobre el mercado el ectrico. [3] Pilipovic, Dragana, European Power Daily Managing Energy Risk, 1998.
Rafael de Benito Divisi on de Energ a FC&M Libreros, 24 46002-Valencia (Spain) e-mail: energia.rdbenito@fcym.com
Uso de las martingalas en la valoraci on de activos derivados

Eloy Fontecha1
Resumen En el presente art culo se muestra c omo la t ecnica de cambio de numerario puede ser empleada en la valoraci on de un n umero considerable de activos derivados. Dado un activo derivado cualquiera, su valoraci on se podr a llevar a cabo utilizando como numerario el precio de cualquier activo con pagos positivos. Sin embargo, teniendo en cuenta caracter sticas propias del derivado como la expresi on de sus pagos en t erminos de su activo subyacente y la evoluci on estoc astica de este, podemos encontrar un numerario cuya elecci on como tal facilita enormemente la valoraci on. En el presente art culo se explica la t ecnica de valoraci on de activos derivados relacion andola con la elecci on de un numerario. Posteriormente y para distintos activos derivados se muestra la elecci on del numerario adecuado y se procede a su valoraci on.
Introducci on
En el presente documento se va a explicar la metodolog a empleada en la valoraci on de activos derivados. En general un activo derivado es aqu el cuyo pay-o depende del precio de otro activo subyacente sobre el que est a emitido. La t ecnica de valoraci on trivial consistir a en asumir una cierta medida de probabilidad y, bajo la misma, un proceso estoc astico para la evoluci on del precio del activo subyacente. Despu es, descontar el valor esperado bajo esa medida del precio futuro del activo derivado a un tipo de inter es que incorpore una prima de riesgo. El problema de esta t ecnica reside en la discrecionalidad de la medida y de la prima de riesgo que podr amos emplear para descontar. Para eliminar esa discrecionalidad lo que se hace es elegir una medida para la que el descuento se pueda hacer a trav es del precio conocido objetivamente de un activo que elegiremos como numerario. Dicho de otra manera, elegiremos una medida para la que el precio de cualquier activo expresado en unidades del numerario tenga media cero.
1 Eloy Fontecha es Responsable del Departamento de Nuevos Productos en el Area de Mercados de BBVA. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de noviembre de 1999.
376
Eloy Fontecha
En los apartados que siguen vamos a ver que esa medida existe para cualquier numerario que pudi eramos elegir, y dicha existencia la mostraremos tanto en tiempo discreto como en tiempo continuo. Finalmente mostraremos unos ejemplos en los que vemos c omo emplear esta t ecnica y c omo llevar a cabo valoraciones mediante cambios de una medida a otra.
B usqueda de martingalas en la valoraci on de activos. Elecci on de numerario. Tiempo discreto
Denotemos por j el ndice de estado correspondiente al instante T , j = 1, . . . , J . Supongamos que un activo A paga ujos en T y denotemos por A0 Aj
j P0
su precio en t = 0, el ujo que paga en T si ocurre el estado j .
Sea P j un activo que paga 1 en T si ocurre j y 0 en caso contrario. Esto es, es el precio Arrow Debreu del estado j del momento T , si j = j y 0 en caso contrario. Pjj
=1
La ausencia de arbitraje implica que

J
A0 =
j =1
j Aj P0 .
Sea N otro activo que elegimos como numerario; entonces, A0 = N0 donde

N wj = Pj J j =1
Aj j P = N0 0 Nj N0
J j =1
Aj Nj Pj = Nj N0
J j =1
Aj N w , Nj j
para cada j = 1, . . . , J . Nj = 1, N0
Se tiene que
N wj = j j
Pj
de manera que A0 = EN N0 A N .
Esto es, existe una medida wN bajo la que A/N es una martingala y esa medida es independiente del activo A: depende s olo del numerario. Restricci on para que N se pueda considerar numerario: ha de ser Nj > 0 para todo j .
n de activos derivados Uso de martingalas en la valoracio
377
Cambio de numerario. Cambio de medida. Derivada de Randon-Nikodym. Tiempo discreto

A0 = EM M0 A M
Sea otro numerario M con Mj > 0 para todo j . Igualmente, .
Ahora:
M wj
= Pj =
Mj N N0 Mj = wj , M0 Nj M0
wM wN
N0 M . N M0
wM /wN es la derivada de Radon Nikodym entre las dos medidas, que permite pasar del valor esperado calculado a partir de una de ellas al calculado a partir de la otra: en general, wM EM (X ) = EN X . wN En particular, podemos emplear esto en la valoraci on de activos para pasar de un numerario a otro: A0 = M0 EM A M = M0 EN A wM M wN = N0 EN A N .
Valoraci on de activos
Sea un activo A y sea F un activo derivado tal que FT = f (AT ) y cuyo precio en 0 queremos calcular. Elegimos un numerario N y la medida wN bajo la que A/N es martingala. Por lo ya visto, tambi en F/N es martingala: F0 = N0 EN f (AT ) NT . (1)
C alculo del hedge : teorema de representaci on de martingalas. Cartera r eplica
Teorema 1 (de representaci on de martingalas) Si suponemos el intervalo temporal dividido mediante una partici on de ndice i, dada una cierta medida y dos martingalas X e Y , existe un proceso i previsible (el valor que toma en i lo conocemos en i 1) tal que i+1 Xi . YT Y0 =
i
378
Eloy Fontecha
Tenemos que Z = A/N es martingala, y queremos replicar FT , que no tiene por qu e ser martingala. Pero FT Et = EN t NT s lo es, por las esperanzas condicionadas sucesivas. Por lo tanto existe i previsible tal que: ET E 0 Ei =
i
i+1 Zi ,
= i+1 Zi .
Construyamos una cartera formada en i por i+1 i+1 = Ei i+1 Zi unidades de A, unidades de N .
En i, el valor de la cartera es i+1 Ai + i+1 Ni = Ei Ni . En vencimiento (T ), el valor de la cartera es FT . Para que sea una r eplica exacta de F bastar a que sea autonanciada. Pero i+1 Ai+1 + i+1 Ni+1 = = = = i+1 Zi+1 Ni+1 + Ni+1 Ei i+1 Zi Ni+1 Ni+1 i+1 (Zi+1 Zi ) + Ni+1 Ei Ni+1 (Ei+1 Ei ) + Ni+1 Ei = Ni+1 Ei+1 i+2 Ai+1 + i+2 Ni+1 .
De aqu extraemos varias conclusiones: Tenemos una cartera que replica el activo derivado construida con A y N (el subyacente y el numerario). El valor en cualquier momento del derivado ha de ser el de la cartera: Ft = Et Nt = Nt EN En particular, F0 = EN FT NT . FT t NT .
Necesariamente ha de ser Ft /Nt = Et una martingala. Tambi en podemos escribir que la variaci on de la cartera y por tanto del derivado es (Ei Ni ) = Fi = i+1 Ai + i+1 Ni .
379
Tiempo continuo
En general asumiremos que el proceso estoc astico seguido por el precio de un activo S (t) es lognormal y viene dado por la ecuaci on diferencial estoc astica dS = S dt + S dw , S con w(t) un proceso de Wiener bajo la medida bajo la que se da esa representaci on. Primero necesitamos ser capaces de construir matingalas: Teorema 2 (Girsanov) Si w(t) es un proceso de Wiener bajo una cierta medida de probabilidad P y (t) es un proceso previsible, entonces existe una medida Q equivalente a P tal que: (i) El proceso w (t) dado por dw = dw + (t) dt es de Wiener bajo Q. (ii) La derivada de Randon-Nikodym entre ambas medidas en T , con la informaci on disponible en 0, viene dada por dQ (T ) = exp dP
T 0
(t) dw(t)
1 2
T 0
2 (t) dt
Construcci on de martingalas. Elijamos un activo N como numerario y consideremos un activo gen erico S . Supongamos que ambos siguen procesos lognormales dados por dS S dN N = S dt + S dwS , = N dt + N dwN .
Sea Z el proceso dado por Z = S/N . Por el lema de It o tenemos que dZ Z = = Z dt + Z dwZ = dN (dN )2 dS dS dN + 2 S N N S N
2 S N + N N S SN dt + S dwS N dwN .
Si suponemos que hay un u nico factor de riesgo, esto es, un u nico proceso de Wiener dw(t) com un para todos los activos, entonces SN = 1 y podemos escribir: dZ Z =
2 S N + N N S SN dt + (S N ) dw
S N N dt + dw S N = (S N ) ( (t) dt + dw) = (S N ) dw . = (S N )
380
Eloy Fontecha
Obs ervese que, eligiendo (t) de esa manera y aplicando el teorema de Girsanov, ya hemos encontrado una medida Q tal que S/N sea martingala para cualquier activo S . S olo falta probar que la medida Q, o equivalentemente (t), es independiente de S. Si eligi eramos como numerario el activo B (t), la cuenta corriente, denido por B (0) = 1 , el proceso seguido por Z = S/B ser a dZ Z = = (S r) dt + S dw = S S r dt + dw S dB = r(t) dt , B
S ( (t) dt + dw) = S dw .
Obs ervese que cuando el numerario es B , la independencia de Q respecto del activo S considerado es equivalente a que el precio de mercado del riesgo , (t), sea el mismo para todos los activos. Adem as, esta condici on es suciente para que cuando el numerario es un activo N cualquiera la medida Q sea independiente del activo S , pues S r N r = S N = (t) = S N N r N = N . S N N
De esta manera, tenemos la herramienta para, dado un numerario cualquiera N , encontrar una medida bajo la que S/N es martingala para cualquier S . S olo hace falta que el precio de mercado del riesgo sea el mismo para cualquier activo S . Valoraci on de activos. Elegido el numerario, ya podemos aplicar la f ormula de valoraci on de cualquier derivado F cuyo pay-o nal viene dado por una funci on f del precio nal de un activo S : FT Ft = f (ST ) , = Nt EN FT t NT = N0 EN f (ST ) t NT .
Si tuvi eramos dos posibles numerario N y M , podr amos cambiar de medida entre ambos de la siguiente manera: Ft EN XT t = Nt EN = Mt M E Nt FT t NT = Mt EM FT t MT ,
XT NT t MT
Para construir el hedge usaremos la versi on en tiempo continuo del teorema de representaci on martingalas:
381
Teorema 3 (de representaci on de martingalas) Dada una cierta medida Q y dos martingalas X e Y , existe un proceso (t) previsible tal que dY (t) = (t) dX (t) . Tenemos que Z = S/N es martingala, y queremos replicar FT , que no tiene por qu e ser martingala. Pero Et = EN (FT /NT | t) s lo es, por las esperanzas condicionadas sucesivas. Por lo tanto existe (t) previsible tal que dE (t) = (t) dZ (t) . Construyamos una cartera formada en t por: (t) (t) = E (t) (t) Z (t) unidades de S , unidades de N .
En t el valor de la cartera es V (t) = (t) S (t)+(t) N (t) = E (t)N (t). En vencimiento (T ), el valor de la cartera es FT . Para que sea una r eplica exacta de F bastar a que sea autonanciada. Pero por el lema de It o tenemos que dV (t) = dE (t) N (t) + E (t) dN (t) + dE (t) dN (t) . Por el teorema de representaci on, dV (t) = (t) dZ (t) N (t) + ((t) + (t) Z (t)) dN (t) + (t) dZ (t) dN (t) = (t) (dZ (t) N (t) + Z (t) dN (t) + dZ (t) dN (t)) + (t) dN (t) . Y por el lema de Ito de nuevo, dV (t) = (t) d(Z (t) N (t)) + (t) dN (t) = (t) dS (t) + (t) dN (t) , luego la cartera es autonanciada. De aqu extraemos varias conclusiones: Tenemos una cartera autonanciada que replica el activo derivado construida con A y N (el subyacente y el numerario). El valor en cualquier momento del derivado ha de ser el de la cartera Ft = Et Nt = Nt EN En particular, F0 = N0 EN FT NT . FT t NT .
Necesariamente ha de ser Ft /Nt = Et una martingala.
382
Eloy Fontecha
Tambi en podemos escribir la ecuaci on diferencial estoc astica seguida por la cartera, y por tanto por el activo derivado, de la siguiente manera: d(E (t) N (t)) = dF (t) = (t) dS (t) + (t) dN (t) . Si tenemos en cuenta los procesos seguidos por S y N y aplicamos el lema de It oa F (t) = F (S (t), t), tendremos: dF (S (t), t) = = Por otro lado, dF (S (t), t) = (t) dS (t) + (t) dN (t) = (S S (t) + N N (t)) dt + (S S (t) + N N (t)) dw . F F 1 2F dS + dt + (dS )2 S t 2 S 2 S S 1 2 2 2F F F + S S + S 2 S 2 t dt + S S F dw = F dt + F dw S
Por tanto, S S (t) + N N (t) = S S (t) + N N (t) = esto es, (t) = (t) = F N F N , S S N S S N F S S (t) . N N F , F ,
De esta manera, usando la f ormula de valoraci on obtenemos F (S (t), t). De aqu deducimos F y F . Hecho esto, podemos usar esas f ormulas para construir la cartera que exactamente replica el derivado que estamos valorando. Obs ervese que si tomamos como numerario el activo cuenta corriente B , tendremos: F F = , (t) = S S (t) + N N (t) = F S S S = S S (t) = F (t) = F S S (t) . rB F = (t) S + (t) B = Fr = (t) S r + F S S (t) = rS 1 2 2 2F F F + S . S + S 2 S 2 t
383
Esto es,
F 1 2 2 2F F + S F r = 0, S + 2 S 2 S t que es la ecuaci on diferencial en derivadas parciales que dene el precio del activo derivado en funci on del del activo subyacente. rS
P erdida y recuperaci on de martingalas. Recuperaci on del drift. Ejemplos
Ejemplo 1. Modelo binomial. Cambio de medida

Sean dos activos S y T y un activo derivado F cuyos precios en t = 0 son S (0) y T (0). Supongamos que s olo hay un momento futuro t = 1, en el que se pueden dar s olo dos estados U y D; en ellos, los respectivos activos toman valores S (U ) S (D) = uS S (0) = dS S (0) T (U ) T (D) = uS T (0) = dS T (0) F (U ) F (D)
Para replicar el valor de F en t = 1, construimos una cartera V formada por unidades de S y unidades de T . Si esta cartera es r eplica de F , deber a cumplir que F (U ) = uS S (0) + uT T (0) , dS S (0) + dT T (0) , S (0) + T (0) .
F (D) = F (0) = Operando ah obtenemos: = = F (0) =
F (U ) dT F (D) uT , S (0) (uS dT dS uT ) F (D) uS F (U ) dS , T (0) (uS dT dS uT ) F (U ) dT dS uS u T + F (D) . uS dT dS uT uS dT dS uT
Esa misma expresi on se puede escribir de dos maneras: Manera 1 Asumimos S como numerario. F (0) S (0) = = F (U ) uS dT uS dS F (D) dS uS dS uT + S (U ) uS dT dS uT S (D) uS dT dS uT F (U ) F (D) pS + (1 pS ) . S (U ) S (D)
384
Eloy Fontecha
Se demuestra que, bajo ausencia de arbitraje, se ha de tener que 0 pS 1, y por tanto se puede considerar como una probabilidad bajo la cual se cumple que F (0) = ES S (0) F S .
Adem as, dicha probabilidad la podr amos haber obtenido sin m as que exigir que T /S fuera martingala. T (0) = ES S (0) T S = pS dT T (0) uT T (0) + (1 pS ) uS S (0) dS S (0) = pS = uS dT uS dS . uS dT dS uT
Manera 2 Asumimos T como numerario. F (0) T (0) = = F (U ) uT dT uT dS F (D) dT uS dT uT + T (U ) uS dT dS uT T (D) uS dT dS uT F (U ) F (D) pT + (1 pT ) . T (U ) T (D)
Se demuestra que, bajo ausencia de arbitraje, se ha de tener que 0 pT 1 y por tanto se puede considerar como una probabilidad bajo la cual se cumple que F (0) = ET T (0) F T .
Adem as, dicha probabilidad la podr amos haber obtenido sin m as que exigir que S/T fuera martingala. S (0) = ET T (0) S T = pT dS S (0) uS S (0) + (1 pT ) uT T (0) dT T (0) = pT = uT dT uT dS . uS dT dS uT
Ejemplo 2. Pre-f ormula de Black y Scholes

Supongamos que tenemos un activo F cuyo precio es una martingala bajo una cierta medida P y viene dado por la siguiente ecuaci on: dF = F dw , F donde w(t) es un proceso Weiner bajo P . Vamos a calcular unos valores esperados que aparecer an regularmente al valorar opciones de compra o venta. Dado un momento futuro T , queremos calcular, con la informaci on en t = 0, EP (F (T ) K )+ y EP (K F (T ))+ .
385
Por el lema de It o, es f acil ver que F (T ) = F (0) e0.5 F +F w(T ) , siguiendo w(T ) una distribuci on N (0, T ) bajo P . Esos valores esperados se podr an calcular evaluando una integral. Sin embargo, nosotros los vamos a calcular usando el Teorema de Girsanov: EP (F (T ) K )+ = = = Pero EP 1{F (T )>K } F (T ) = EP 1{F (T )>K } F (0) eF /2+F w(T ) = EQ 1{F (T )>K } F (0) eF /2+F w(T ) donde Q es otra medida cualquiera. Si la elegimos de forma que
2 Q = eF /2+F w(T ) , P 2 2 2
EP 1{F (T )>K } (F (T ) K ) EP 1{F (T )>K } F (T ) K EP 1F (T )>K EP 1{F (T )>K } F (T ) K P (F (T ) > K ) .
P Q
tendremos EP 1{F (T )>K } F (T ) = EQ 1{F (T )>K } F (0) = F (0) Q(F (T ) > K ) . Por el teorema de Girsanov, eligiendo (t) = F , tendremos que w (t) = w(t) F t es un proceso de Wiener bajo la medida Q, esto es, w (t) es N (0, t) bajo Q. Por lo tanto, EP (F (T ) K )+ = F (0) Q (F (T ) > K ) K P (F (T ) > K ) . Si llamamos ln d1 = tenemos: P (F (T ) > K ) = = Q (F (T ) > K ) = = P (F (0) eF /2 T +F w(T ) > K ) P (F (0) eF /2 T +F
2 2 2
F (0) 2 + F T K 2 F T
ln y d2 =
F (0) 2 F T K 2 , F T
> K ) = P ( > d2 ) = N (d2 ) ,
Q(F (0) eF /2 T +F (w (T )+F T ) > K ) Q(F (0) eF /2 T +F

2
> K ) = Q( > d1 ) = N (d1 ) .
386
Eloy Fontecha
Finalmente, EP (F (T ) K )+ = = = An alogamente, EP (K F (T ))+ = K P (F (T ) < K ) F (0) Q (F (T ) < K ) = K N (d2 ) F (0) N (d1 ) = BSput (F (0), K, T, F , r = 0, q = 0) . F (0) Q (F (T ) > K ) K P (F (T ) > K ) F (0) N (d1 ) K N (d2 ) BScall (F (0), K, T, F , r = 0, q = 0) .
Ejemplo 3. Activo subyacente: tipo de inter es de corto plazo

Sea t = 0 el momento de valoraci on y sean T1 y T2 dos momentos futuros, siendo T1 < T2 . es de dep osito que se ja en t para un Denotemos por R(t, T1 ) el tipo de inter dep osito que va de t a T1 . Denotemos por F (t, T1 , T2 ) el tipo FRA que se ja en t para el per odo que va de T1 a T2 . Y llamemos P (t, T1 ) al tipo de descuento que se ja en t para T1 . Es claro que F (t, T1 , T2 ) = 1 T2 T 1 P (t, T1 ) 1 P (t, T2 ) = 1 T2 T 1 P (t, T1 ) P (t, T2 ) P (t, T2 ) .
Eligiendo P (t, T2 ) como numerario, existe una medida P bajo la cual F (t, T1 , T2 ) (alguna vez lo denotaremos por F sin m as) es martingala. Si asumimos lognormalidad para el proceso seguido por F , dF = F dw , F tendremos que F (T, T1 , T2 ) = F (0, T1 , T2 ) = F (0, T1 , T2 ) eF /2 T +F
2
EP (F (T, T1 , T2 )) = EP (F (T1 , T1 , T2 )) = EP (R(T1 , T2 )) ,
con w(T ) siguiendo una distribuci on N (0, T ) bajo P . Valoremos una serie de activos V (t): Dep osito a tipo variable, que paga en T2 una cantidad igual a R(T1 , T2 ). Elegimos como numerario P (t, T2 ): V (0) = = P (0, T2 ) EP V (T2 ) P (T2 , T2 )
P (0, T2 ) EP (R(T1 , T2 )) = P (0, T2 )F (0, T1 , T2 ) .
387
Cap, que paga en T2 una cantidad igual a (R(T1 , T2 ) K )+ . Elegimos como numerario P (t, T2 ): V (0) = P (0, T2 ) EP V (T2 ) P (T2 , T2 )
+
= P (0, T2 ) EP (F (T1 , T1 , T2 ) K )
= P (0, T2 ) BScall (F (0, T1 , T2 ), K, T1 , F , r = 0, q = 0) . Cualquier pay-o pagadero en T2 , que paga en T2 una cantidad igual a f (R(T1 , T2 )). Elegimos como numerario P (t, T2 ): V (0) = = = P (0, T2 )EP V (T2 ) P (T2 , T2 )
+
= P (0, T2 ) EP (f (F (T1 , T1 , T2 )))

2
P (0, T2 ) EP f F (0, T1 , T2 ) eF /2 T1 +F P (0, T2 )

2
T1 T1
2 1 e /2 d . 2
f F (0, T1 , T2 ) eF /2 T1 +F
Ejemplo 4. Activo subyacente: tipo de inter es de corto plazo distorsionado

Supongamos ahora que el activo subyacente es el mismo que antes, s olo que no genera un ujo en T2 sino en T1 (In Arrears ). Obs ervese que bajo el numerario P (t, T1 ), F (t, T1 , T2 ) ya no es martingala. En general, vamos a tener dos opciones para resolver estos problemas de p erdida de las martingalas: Soluci on 1: mantener como numerario P (t, T2 ), de tal manera que F (t, T1 , T2 ) siga siendo martingala. De esta manera aparecer an m as activos cuya volatilidad y correlaci on con F deberemos conocer. Soluci on 2: cambiar de numerario P (t, T2 ), de tal manera que no aparezcan m as activos cuya volatilidad y correlaci on con F debamos conocer. En este caso, F (t, T1 , T2 ) ya no ser a martingala y por lo tanto deberemos calcular su drift. Ve amoslo con los ejemplos anteriores: sea el activo derivado V (t) que paga en T1 una cantidad dada por f (R(T1 , T2 )).
388
Eloy Fontecha
n 1 Elegimos como numerario P (t, T2 ): Solucio V (0) = P (0, T2 ) EP V (T2 ) P (T2 , T2 )
= P (0, T2 ) EP (f (F (T1 , T1 , T2 )) (1 + F (T1 , T1 , T2 ) (T2 T1 ))) = P (0, T2 ) EP (g (F (T1 , T1 , T2 ))) = P (0, T2 ) EP g F (0, T1 , T2 ) eF /2 T1 +F
+
2
T1
2 1 , e /2 d . 2
= P (0, T2 )
g F (0, T1 , T2 ) eF /2 T1 +F
T1
n 2 Elegimos como numerario P (t, T1 ). Sea Q la medida bajo la que los Solucio precios de los activos relativos al numerario son martingalas. V (0) = P (0, T1 ) EQ V (T1 ) P (T1 , T1 ) = P (0, T1 ) EQ (f (F (T1 , T1 , T2 ))) .
Ahora tenemos que calcular el drift de F bajo esa medida. Para ello elegimos otro activo derivado dependiente de F (T1 , T1 , T2 ) cuyo valor s sepamos calcular. Elijamos un contrato forward con expiraci on en T1 de un bono cup on cero que expira en T2 . Su precio spot es P (t, T2 ). Su precio forward en T1 ser a P (t, T1 , T2 ) = 1 P (t, T2 ) = . P (t, T1 ) 1 + F (t, T1 , T2 ) (T2 T1 )
Dadas esas expresiones, es claro que P (t, T1 , T2 ) es martingala bajo Q. Supongamos que el proceso seguido por F (t, T1 , T2 ) bajo Q es dF F F (T, T1 , T2 ) = F dt + F dw , = F (0, T1 , T2 ) eF T eF /2 T +F
2
con w(T ) siguiendo una distribuci on N (0, T ) bajo Q. Obs ervese que, conocido F , podemos proceder igual que en el caso est andar sin m as que sustituir F (0, T1 , T2 ) por el tipo forward ajustado de convexidad, F (0, T1 , T2 ) eF T1 . Ahora bien, P (t, T1 , T2 ) = EQ (P (T1 , T2 )) = EQ = EQ 1 1 + F (T1 , T1 , T2 ) (T2 T1 ) 1 1 + F (0, T1 , T2 ) eF T1 eF /2 T1 +F
2
T1
(T2 T1 )
389
S olo resta calcular el valor de F que hace cierta esa igualdad. Otra forma de calcular el drift de F ser a: F (0, T1 , T2 ) eF T1 = EQ (F (T1 , T1 , T2 )) P (0, T2 ) P P (T1 , T1 ) E = F (T1 , T1 , T2 ) P (0, T1 ) P (T1 , T2 ) P (0, T2 ) P E (F (T1 , T1 , T2 ) (1 + F (T1 , T1 , T2 ) (T2 T1 ))) , = P (0, T1 )
expresi on que se puede calcular, puesto que F (t; T1 , T2 ) es martingala bajo P .
Ejemplo 5. Activo subyacente: tipo de inter es de corto plazo distorsionado

Supongamos ahora que el activo subyacente es el mismo que antes, s olo que no genera un ujo ni en T2 ni en T1 , sino en T3 > T2 . Obs ervese que, bajo el numerario P (t, T3 ), F (t, T1 , T2 ) ya no es martingala. n 1 Elegimos como numerario P (t, T2 ) y la medida por tanto P . Solucio V (0) = P (0, T2 ) EP V (T2 ) P (T2 , T2 ) = P (0, T2 ) EP f (F (T1 , T1 , T2 )) (1 + F (T2 , T2 , T3 ) (T3 T2 )) .
Bajo P sabemos que F (t, T1 , T2 ) es martingala, pero no lo es F (t, T2 , T3 ). Sin embargo, ya sabemos calcular el ajuste de convexidad que deber amos aplicar a F (t, T2 , T3 ) bajo esa medida.. Esto es, bajo la medida P podemos calcular esa integral asumiendo los procesos: dF (t, T1 , T2 ) F (t, T1 , T2 ) dF (t, T2 , T3 ) F (t, T2 , T3 ) = = F (t,T1 ,T2 ) dwF (t,T1 ,T2 ) , F (t,T2 ,T3 ) dt + F (t,T2 ,T3 ) dwF (t,T2 ,T3 ) ,
con (wF (t,T1 ,T2 ) , wF (t,T2 ,T3 ) ) un proceso de Wiener bivariante con dos volatilidades y una correlaci on. n 2 Elegimos como numerario P (t, T3 ) y Q la medida asociada: Solucio V (0) = P (0, T3 ) EQ V (T3 ) P (T3 , T3 ) = P (0, T3 ) EQ (f (F (T1 , T1 , T2 ))) .
Ahora tenemos que calcular el drift de F bajo esa medida. Para ello elegimos otro activo derivado dependiente de F (T1 , T1 , T2 ) cuyo valor s sepamos calcular.
390
Eloy Fontecha
Elijamos un contrato consistente en invertir una unidad en T1 en un bono cup on cero hasta T2 y desde T2 reinvertirlo todo hasta T3 . El valor de este activo en t = 0 es P (0, T1 ) y su valor en T3 ser a (1 + F (T1 , T1 , T2 )(T2 T1 )) (1 + F (T2 , T2 , T3 )(T3 T2 )) . Por lo tanto, P (0, T1 ) P (0, T3 ) = EQ (1 + F (T1 , T1 , T2 ) (T2 T1 )) (1 + F (T2 , T2 , T3 ) (T3 T2 )) P (T3 , T3 )
= EQ ((1 + F (T1 , T1 , T2 ) (T2 T1 )) (1 + F (T2 , T2 , T3 ) (T3 T2 ))) , asumiendo dF (t, T1 , T2 ) F (t, T1 , T2 ) dF (t, T2 , T3 ) F (t, T2 , T3 ) = F (t,T1 ,T2 ) dt + F (t,T1 ,T2 ) dwF (t,T1 ,T2 ) , = F (t,T2 ,T3 ) dwF (t,T2 ,T3 ) ,
dado que F (t, T2 , T3 ) es martingala bajo Q y siendo (wF (t,T1 ,T2 ) , wF (t,T2 ,T3 ) ) un proceso de Wiener bivariante con dos volatilidades y una correlaci on. Ahora deber amos calcular el valor del drift de F (t, T1 , T2 ) que hace cierta esa igualdad dadas las volatilidades y correlaci on cotizadas. Otra forma de calcular el drift de F ser a: F (0, T1 , T2 ) eF T1 = = EQ (F (T1 , T1 , T2 )) = P (0, T2 ) P E P (0, T3 ) P (0, T2 ) P P (T1 , T3 ) E F (T1 , T1 , T2 ) P (0, T3 ) P (T1 , T2 ) (1 + F (T1 , T1 , T2 ) (T2 T1 )) F (T1 , T1 , T2 ) , (1 + F (T1 , T1 , T3 ) (T3 T1 ))
expresi on que se puede calcular puesto que F (t, T1 , T2 ) es martingala bajo P y el drift de F (t, T1 , T3 ) bajo P ya lo sabemos calcular del ejemplo anterior.
Ejemplo 6. Activo subyacente: activo cualquiera

Sea t = 0 el momento de valoraci on, y sea T un momento futuro. Denotemos por S (t) el precio spot que vemos en t, por F (t, T ) el precio forward que se ja en t para entrega en T y por P (t, T ) al tipo de descuento que se ja en t para T . Es claro que F (t, T ) = S (t) . P (t, T )
Eligiendo P (t, T ) como numerario, existe una medida P bajo la cual F (t, T ) (alguna vez lo denotaremos por F sin m as) es martingala. Si asumimos lognormalidad para el proceso seguido por F , dF = F dw , F
391
tendremos que F (t, T ) F (0, T ) = F (0, T ) eF /2 T +F

P P
2
= E (F (T, T )) = E (S (T )) ,
con w(T ) siguiendo una distribuci on N (0, T ) bajo P . De la misma manera que en el ejemplo 3, ya sabr amos valorar cualquier activo derivado consistente en un ujo dependiente de S (T ) y que se paga en T . Supongamos que el derivado paga en T una cantidad igual a f (S (T )). Elegimos como numerario P (t, T ): V (0) = P (0, T ) EP V (T ) P (T, T ) = P (0, T ) EP (f (F (T, T )))
2
= P (0, T ) EP f (F (0, T ) eF /2 T +F
+
T T
2 1 e /2 d . 2
= P (0, T )
f F (0, T ) eF /2 T +F
En el supuesto de que el pay-o fuera el de una call o una put, esa integral la reducir amos a las expresiones ya conocidas.
Ejemplo 7. Activo subyacente: activo cualquiera distorsionado

Supongamos ahora que el activo subyacente es el mismo que antes, s olo que no genera un ujo en T sino en T1 > T . Obs ervese que, bajo el numerario P (t, T1 ), F (t, T ) ya no es martingala. n 1 Elegimos como numerario P (t, T ) y la medida por tanto P : Solucio V (0) = P (0, T ) EP V (T ) P (T, T ) = P (0, T ) EP f (F (T, T )) (1 + F (T, T, T1 ) (T1 T )) .
Bajo P sabemos que F (t, T ) es martingala, pero no lo es F (t, T, T1 ). Sin embargo ya sabemos calcular el ajuste de convexidad que deber amos aplicar a F (t, T, T1 ) bajo esa medida. Esto es, bajo la medida P podemos calcular esa integral asumiendo los procesos: dF (t, T ) F (t, T ) dF (t, T, T1 ) F (t, T, T1 ) = = F (t,T ) dwF (t,T ) , F (t,T,T1 ) dt + F (t,T,T1 ) dwF (t,T,T1 ) ,
con (wF (t,T ) , wF (t,T,T1 ) ) proceso de Wiener bivariante con dos volatilidades y una correlaci on.
392
Eloy Fontecha
n 2 Elegimos como numerario P (t, T1 ) y Q la medida asociada: Solucio V (0) = P (0, T1 ) EQ V (T1 ) P (T1 , T1 ) = P (0, T1 ) EQ (f (F (T, T ))) .
Ahora tenemos que calcular el drift de F bajo esa medida. Para ello elegimos otro activo derivado dependiente de F (T, T ) cuyo valor s sepamos calcular. Elijamos un contrato consistente en comprar forward con expiraci on en T una unidad de activo y el benecio o p erdida experimentado en T reinvertirlo o nanciarlo hasta T1 . El valor de este contrato en t es 0 y su valor en T1 ser a (F (T, T ) F (t, T )) (1 + F (T, T, T1 )(T1 T )) . Por lo tanto, 0 = EQ (F (T, T ) F (t, T )) (1 + F (T, T, T1 ) (T1 T )) P (T1 , T1 )
= EQ ((F (T, T ) F (t, T )) (1 + F (T, T, T1 ) (T1 T ))) , asumiendo dF (t, T ) F (t, T ) dF (t, T, T1 ) F (t, T, T1 ) = = F (t,T ) dt + F (t,T ) dwF (t,T ) , F (t,T,T1 ) dwF (t,T,T1 ) ,
dado que F (t, T, T1 ) es martingala bajo Q y siendo (wF (t,T ) , wF (t,T,T1 ) ) proceso de Wiener bivariante con dos volatilidades y una correlaci on. Ahora deber amos calcular el valor del drift de F (t, T ) que hace cierta esa igualdad dadas las volatilidades y correlaci on cotizadas. Para calcular el drift de F tambi en podr amos haber hecho lo siguiente: EQ (F (T, T )) = = F (0, T ) eF T = P (0, T ) P E P (0, T1 ) P (T, T1 ) F (T, T ) P (T, T )
P (0, T ) P E (P (T, T1 ) F (T, T )) , P (0, T1 )
expresi on que se puede calcular en t erminos de las volatilidades y correlaci on entre P (t, T, T1 ) y F (t, T ) ambos martingalas bajo la medida P .
Ejemplo 8. Activo subyacente: swap

Sea t = 0 el momento de valoraci on, y sea T un momento futuro. Denotemos por S (t) el tipo que vemos en t para un swap que comienza en t0 > T y paga cupones en fechas t1 , t2 , . . . , tn con sendas bases entre fechas: b1 , b2 , . . . , bn ;
393
por P (t, ti ) al tipo de descuento que se ja en t para ti , y por 01(t) el precio de un activo formado por tantas unidades del bono cup on cero de expiraci on ti como indica la base bi , y esto para cada i desde 1 a n. Esto es,
n
01(t) =
i=1
P (t, ti ) bi .
Es claro que S (t) =
P (t, t0 ) P (t, tn ) . 01(t)
Eligiendo 01(t) como numerario, existe una medida P bajo la cual S (t) (alguna vez lo denotaremos por S sin m as) es martingala. Si asumimos lognormalidad para el proceso seguido por S , dS = S dw , S tendremos que S (T ) = S (0) eS /2 T +S S (0) = E (S (T )) , con w(T ) siguiendo una distribuci on N (0, T ) bajo P . De la misma manera que en el ejemplo 3, ya sabr amos valorar cualquier activo derivado consistente en un conjunto de ujos que se pagan en las mismas fechas t1 , . . . , tn y que se obtienen multiplicando una cierta funci on de S (T ) por las correspondientes bases. Esto es, supongamos que el derivado paga en ti una cantidad igual a f (S (T )) bi . Elegimos como numerario 01(t): V (0) = = = 01(0) EP V (T ) 01(T ) = 01(0) EP f (S (T )) 01(T ) 01(T )
2 2
01(0) EP (f (S (T ))) = 01(0) EP f (S (0) eS /2 T +S

+
01(0)
f (S (0) eS /2 T +S
2 1 ) e /2 d . 2
En el supuesto de que el pay-o fuera el de una call o una put, esa integral la reducir amos a las expresiones ya conocidas. Si por ejemplo es un payers swaption de fecha de expiraci on T , tendr amos V (0) = 01(0) EP V (T ) 01(T ) = 01(0) EP (S (T ) K )+ 01(T ) 01(T )
= 01(0) EP (S (T ) K )+ .
394
Eloy Fontecha
Ejemplo 9. Activo subyacente: swap distorsionado. CMS

Supongamos ahora que el activo subyacente es el mismo que antes, s olo que no genera una serie de ujos en t1 , . . . , tn de la forma mencionada en el ejemplo anterior, sino que genera un ujo u nico S (T ) en T . Obs ervese que bajo el numerario P (t, T ), S (t) ya no es martingala. Elegimos como numerario P (t, T ). Sea Q la medida asociada: V (0) = P (0, T ) EQ V (T ) P (T, T ) = P (0, T ) EQ (f (S (T ))) .
Ahora tenemos que calcular el drift de S (t) bajo esa medida. Para ello consideramos como activo un bono que paga en las mismas fechas que el swap a un tipo igual al que se ja para el swap en t = 0, esto es S (0). El precio de este contrato es obviamente igual a cero. Por tanto, 0 = P (0, T ) EQ V (T ) P (T, T ) = P (0, T ) EQ (01(T ) (S (t) S (T )) .
Bajo esa medida, S (t) tendr a un drift que es lo que queremos calcular. Por otra parte, 01(t) tampoco es martingala bajo esa medida. De hecho es una suma de tipos de descuento, que tendr an una matriz de varianzas y covarianzas y cuyo drift se debe obtener mediante t ecnicas similares a las ya vistas para activos.
Ejemplo 10. Activo subyacente: tipo cup on cero compuesto

Supongamos que queremos valorar en t = 0 un derivado que paga en T1 un ujo que es funci on del tipo impl cito, R(T, T1 ), en el precio en T , P (T, T1 ), de un bono cup on cero de vencimiento T1 . Asumimos que el tipo impl cito es cup on cero, esto es, P (T, T1 ) = 1 . (1 + R(T, T1 ))(T1 T )
Si consideramos el precio forward en t con entrega en T que vemos para ese bono: P (t, T, T1 ) = 1 . (1 + F (t, T, T1 ))(T1 T )
Para calcular el precio del derivado tomamos como numerario P (t, T ) y la medida asociada P : V (0) = P (0, T ) EP f (F (T, T, T1 )) P (T, T1 ) P (T, T )
= P (0, T ) EP (f (F (T, T, T1 )) P (T, T1 )) .
395
Pero F (t, T, T1 ) no es martingala bajo P . Sin embargo P (t, T, T1 ) s lo es: P (t, T, T1 ) = EP (P (T, T, T1 )) = EP 1 (1 + F (T, T, T1 ))
(T1 T )
De esa expresi on deducimos el drift de F (t, T, T1 ) bajo la medida P .
Referencias
[1] Baxter, M., Rennie A. (1996), Financial Calculus. An Introduction to Derivative Pricing. Cambridge University Press, Cambridge. [2] Musiela, M., Rutkowski, M. (1997), Martingale Methods in Financial Modelling. Springer-Verlag.
Eloy Fontecha Fern andez BBVA V a de los Poblados, s/n 28036-Madrid e-mail: efontecha@grupobbva.com
Modelos multivariantes de valoraci on de bonos convertibles con riesgo de cr edito

s P Jesu erez Colino1
Resumen El bono convertible es la estructura nanciera m as compleja, vers atil y l quida, de cuantos se conocen en los mercados en la actualidad. Este tipo de bonos agrupa subyacentes de diferente naturaleza: por un lado la evoluci on de la acci on, la evoluci on de los tipos de inter es libre de riesgo, el spread de cr edito del propio emisor y una prima de liquidez. El trabajo que se presenta es tanto una descripci on de los conceptos b asicos y componentes de un bono convertible, como de los diferentes modelos publicados, desde los m as sencillos modelos univariantes, hasta los modelos multivariantes que intentan introducir la prima de riesgo de cr edito del emisor, que es realmente, el centro de este trabajo2 .
1
1.1
Introducci on a los bonos convertibles

Denici on
El bono convertible es cualquier activo de renta ja (bonos, acciones preferenciales, etc.) que sea o pueda ser, en cualquier momento o en un determinado momento, transformados o convertidos en un n umero de acciones determinadas, o en su valor monetario, o en opciones sobre dichas acciones que permitan al tenedor del bono la compra de un n umero de t tulos a un determinado precio, en un determinado momento o durante un periodo de tiempo (continuo o discreto). Pero esta denici on de convertible tambi en incluye la posibilidad de convertibles sint eticos, los cuales pueden ser creados como una combinaci on de activos, cuya cesta da lugar a otro activo con las dos principales caracter sticas de un convertible: un ujo de renta ja, y un derecho de compra sobre un titulo de renta variable o derecho de rentabilidad.
1 Jes us P erez Colino era, en el momento de realizar este trabajo, el Director del Departamento de Product Research del Area de Mercado de Capitales de Caja Madrid. En la actualidad es gerente de Accenture en el Area de Financial Services. Esta charla se imparti o en la sesi on del Seminario MEFF-UAM de diciembre de 1999. 2 El autor desea agradecer expresamente el apoyo y soporte prestado por Carlos Contreras G omez, entonces Director de Mercado de Capitales de Caja Madrid, sin el cual este trabajo, y muchos otros de mi carrera, no hubiesen sido posibles.
398
s P Jesu erez Colino
1.2
Conceptos b asicos
Los diferentes conceptos que se deber an manejar para poder valorar dichos activos son los siguientes:
Par Value : valor nominal del bono. Issue Value : valor al cual el bono es emitido (normalmente a la par). Redemption Price : valor al que el bono no convertido a vencimiento ser a amortizado. Conversion ratio : n umero de acciones que recibir a el inversor en caso de conversi on, por cada bono.
n Modelos multivariantes de valoracio Parity or Conversion Value : ratio de conversi on por el valor de la acci on.
399
Conversion Price : precio al cual la acci on debe cotizar para que tenga lugar la conversi on. Conversion Premium : porcentaje de diferencia entre la paridad y el Conversion Price. Dividend yield : tasa de dividendos. Current yield : tasa de rentabilidad anual del bono convertible (cup on/precio bono convertible). Yield advantage : diferencia entre el Current yield y el Dividend yield. Especialmente interesantes son las siguientes caracter sticas: Conversion Privilege . El privilegio de conversi on se describe normalmente en t erminos de precio o ratio de conversi on, es decir, como el n umero de acciones que se obtienen de la conversi on de un bono de nominal 1000$. Habitualmente, en el momento de emisi on el precio de conversi on suele situarse entre un 15 y un 30% del precio de mercado de la acci on. Coupon . Es l ogico que el cup on que se paga en este tipo de emisiones est e entre 300400 bps por debajo del bono corporativo de id entico riesgo (en media, dependiendo del plazo y estructura). Maturity . Inicialmente, los bonos convertibles se emit an con vencimientos de veinte a nos o m as, pero, tras los fuertes crecimientos burs atiles de la d ecada de los 80, el vencimiento medio se ha reducido por debajo de los diez a nos. Sin embargo, todav a es posible ver Convertible preferreds bonds perpetuos (con call y/o put ). Call Protection . Opciones de amortizaci on anticipada por parte del emisor no aparecieron en los bonos convertible hasta inicios de los 80, cuando numerosas emisiones se estaban amortizando anticipadamente en periodos menores al a no e incluso los seis meses, con el resultado de que el inversor se ve a forzado a la conversi on, perdiendo el cup on corrido hasta el momento de la conversi on. Actualmente, existen dos tipos de protecciones contra las amortizaciones anticipadas, hard call protection, que proh be las amortizaciones anticipadas bajo determinadas circunstancias, y las Provisional or soft call protection, que limita las posibilidades de amortizaci on a que se den determinadas circunstancias en el precio del subyacente.
400
1.3
Qu e da valor a un bono convertible?
A un bono convertible le dan valor las siguientes caracter sticas: + Cup on alto, emisi on bajo par, baja prima de conversi on. + Largo periodo hasta la primera call del principal por parte del emisor. + Protecci on contra el efecto diluci on que genera la propia conversi on sobre el valor de las acciones. + Compensaci on ante sucesos extraordinarios como el reparto de dividendos especiales. + Cl ausula del Inversor despistado, que permite amortizaciones al precio m aximo de la acci on. + Inmediatamente convertible y de forma continua. + Deuda senior. + Protecci on ante cambio en el control de la compa n a. + Protecci on ante posibilidad de quiebra o default. + Protecci on ante el impago de dividendos/cupones. De igual forma, le quitar a valor a una estructura convertible, Cup on bajo o cero, emisi on a la par (o sobre par), alta prima de conversi on. Callability inmediatamente posterior al momento de la emisi on. Protecci on b asica, como a stocks splits. No existencia de compensaciones a los inversores por bonus o excesos de dividendos. El inversor deber a recordar en todo momento la opci on de conversi on. Convertibilidad limitada en tiempo o valor del subyacente, incluyendo la alternativa del cash. Subordinada. Sin referencias al cambio de control en la compa n a. Sin protecci on ante quiebra. Sin primas de pago en siguientes cupones por impago del anterior.
n Modelos multivariantes de valoracio
401
1.4
Tipolog a de bonos convertibles
LYONs (Liquid Yield Option Notes ). Bonos convertibles introducidos por ML en 1985, como bonos cup on cero convertibles en cualquier momento de la vida del bono, valorados normalmente entre el 20 y el 40% de la par para dar una YTM 47% (largo plazo), normalmente con hard/soft put options, y cinco a nos de call protection, e interesantes ventajas scales para el emisor (deduce el cup on corrido). TOPrS (Trust Originted Preferred Securities ). Bono convertible de preferentes que paga trimestralmente dividendos (in arrears ), de vencimiento a largo plazo, con call protection de 35 a nos, y emitido por una sociedad intermediadora (Trust ) que compra deuda subordinada del mismo plazo e id entico riesgo y coletariza la emisi on del bono convertible. Esta estructura apareci o fundamentalmente para aprovechar las ventajas scales que ofrecen determinados estados americanos a las trust. Convertible MIPS (Monthly Income Preferred Securities ). Son introducidos por primera vez en 1994, y emitidos especialmente por una SPS (Special Purpose Subsidiary ), la cual no es m as que una sociedad (normalmente limitada) que emite y vende el MIPS, comprando deuda de id entico vencimiento al emisor. Vencimiento 30 a nos con cinco a nos de call protection, prima de conversi on del 2025% y paga dividendos in arrears. Excelentes ventajas scales y crediticias. PERCS (Preferred Equity Redemption Cumulative Stock ). Bonos convertibles emitidos a tres a nos, con un cap de apreciaci on del subyacente al 3035%, emitido
402
normalmente a la par, sin protecci on a la baja, paga cupones por encima de mercado, con hard call, y s olo convertible a vencimiento (lo que genera problemas de cobertura serios seg un nos aproximamos al vencimiento). Otras estructuras similares son los MCPDPS, TARGETS, YES, CHIPS, ELKS, EYES, PERQS, o YEELDS. PRIDES (Prederred Redeemable Increased Dividend Equity Securities ). Bono convertible introducido por primera vez en 1993, normalmente construido como acci on preferente con opci on a intercambio por un numero de acciones normales a un ratio de conversi on variable en funci on del valor del subyacente, a cambio de un dividendo por encima del de mercado, y con vencimiento de aproximadamente 3 - 4 a nos, tres a nos de call protection, prima de conversi on entre 20 y 30%, que paga dividendos trimestralmente (normalmente in advance ). Este tipo de estructuras no ofrece protecci on ante ca das, y sin embargo tampoco ofrecen todo el valor de la subida del activo. Otras estructuras similares son los ACES, DECS, MARCS, PEPS, o SAILS. Reset PRIDES. Acci on preferente con opci on de conversi on a acciones normales, con ratios de conversi on inversamente proporcionales al valor del subyacente, proporcionando ante subidas del subyacente revalorizaciones por debajo de mercado, y si cae, p erdidas inferiores. Normalmente emitido a tres a nos a la par frente al subyacente, paga dividendos trimestrales con rentabilidades entre 3 - 4 p.p. por encima del rendimiento de la acci on normal. Feline PRIDES. Fundamentalmente consiste en la compra a futuro de acciones (a unas determinados ratios de conversi on) con un colateral, que normalmente puede intercambiarse entre Corporate Debt o Treasury (verdaderamente se convierte en una unidad m as b asica denominada Income PRIDE, que a su vez se puede separar en el secundario en un Growth PRIDES (contrato forward combinado con Treasury ) y un TOPrS). Es una estructura compleja pero que permite ser personalizada a trav es de combinaciones de sus partes b asicas (l quidas) en el secundario. Enhanced PRIDES. Consisten en bonos libres de riesgo que pagan el cup on del Treasury mas un cup on de un Yield enhanced payments con un contrato forward de compra de stock a ratios entre 0.83 y 1 dependiendo del precio de mercado. Con esta estructura se elimina en parte el riesgo de cr edito del emisor. Flex Caps . Acci on preferente con opci on de conversi on a common stock que participa del 100% de la revalorizaci on de la acci on hasta un determinado nivel o cap (al 50%) 3 a nos de vida, callable, y ofrece normalmente entre 3 y 4 p.p. por encima del rendimiento de la acci on normal. STRYPES (STRuctured Yield Products Exchangeable for Stock ). Estructura (convertible sint etico) emitida por un broker o trust, que compra a futuro acciones comunes de una compa n a y los emite en un estructurado. El inversor acepta el riesgo broker o utiliza bonos Treasury como colateral, y el vendedor del futuro suele tener la opci on de pagar a vencimiento con cash o equity.
403
1.5
Historia
Este tipo de activos tuvo su origen en el mandatory convertibles emitido por primera vez en abril de 1982 por la compa n a norteamericana Manufactures Hannover, como bono a 10 a nos que a su vencimiento pagaba principal en forma de acciones de la propia compa n a. Este tipo de emisi on r apidamente se hizo popular en Estados Unidos, y ya en 1984 el 40% del total de emisiones lanzadas por las 50 mayores compa n as norteamericanas ten an la forma de mandatory convertibles. Despu es de una r apida expansi on en el mercado americano, este producto se fue transformando y desarrollando (como las Equity Commiment Note de JPMorgan), hasta que en enero de 1986 llega al Euromercado, con una emisi on de Thomson-Brant International, garantizada por Thomson SA., de 50 millones de d olares en equity contract notes con un cup on del 8%, y opci on de vencimiento anticipado para el tenedor del bono (puttable americano) desde septiembre de 1986, de 6,194 acciones por cada 1.000$ del principal. Variantes europeas de los mandatory convertible fueron los bonos Balladur, emitidos por el Gobierno franc es en 1991, que daban derecho preferencial en la participaci on de los procesos de privatizaci on de compa n as francesas; o los emitidos por Swedish Steel en 1992, bonos cup on cero con un warrant sobre acciones de su propia compa n a. Desarrollos posteriores de convertibles en el mercado americano fue el Debt Exchangeable for Common Stock (DECS), emitido por primera vez por American Express en Octubre de 1993, como bonos a tres a nos con cup on jo del 6,25%, que a vencimiento devuelve el principal en acciones (o su valor) de Frist Data Corporation, del siguiente modo: si el precio de la acci on se situaba por debajo de 36,75$ (precio al
404
que cotizaban las acciones en el momento de la emisi on del bono) el emisor recibir a 1 acci on, mientras que si cotizara por encima de 44,875$ obtendr a el 0.82 de una acci on, ajust andose dicho ratio en el tramo intermedio en el break even del inversor (cap sobre el precio de la acci on), por lo que dicha estructura s olo permite recoger un 22% de la revalorizaci on de la acci on, mientras asume todo el riesgo de ca da de precio del stock. En la estructura del DECS el activo es un bono, mientras que en un PECS (Preferred Exchangeable for Common Stocks ) el activo es una acci on preferente solo delimitada, generalmente, por un cap. Otro ejemplo interesante fue la emisi on, en abril de 1995, de 75 millones de d olares en DECS sint eticos llamados STEP (Securities Tied to Equity Performance ) emitidos por JP Morgan con el objeto de liberarse de una posici on excesivamente larga en acciones de la compa n a americana America Online. Existen numerosas variaciones sobre las estructuras DECS. En julio de 1995 Merrill Lynch construyo un Structured Yield Product Exchangeable for Stocks (STRYPES) que transcurridos 3 a nos se convert an en el 10% de MGIC Investment Corporation, lo que permiti o al cliente de Merrill Lynch deshacerse de la mitad de la cartera sobre dicha compa n a. Una emisi on de similares caracter sticas fue la de Houghton Miin usando una estructura Stock Appreciation Income Linked Securities (SAILS) para la venta de 119 millones de d olares de acciones de Inso Corporation. Otra interesante estructura de STRYPES fue la emisi on de la compa n a Browning Ferris International, en junio de 1995. En este caso, el emisor recib a U.S. Treasury Bonds, con un valor id entico al precio de la oferta, m as la obligaci on de compra por parte del inversor de acciones de BFI al vencimiento del bono, donde principal m as cup on ser a reinvertidos en acciones de BFI. Luego, lo que hace en realidad BFI es una venta forward de acciones, aunque el n umero de acciones emitidas puede variar, como en la estructura de los DECS, es decir, y para este caso concreto, si en tres a nos el precio del stock est a por debajo del precio de oferta, entonces el inversor recibe 1 acci on, y si est a por encima recibe el 0.8% de la acci on. Un nuevo tipo de bono convertible aparece en noviembre de 1993, denomin andose Convertible Exchangeable y permit a al inversor sustituir el bono bien por acciones preferentes o bien por otro bono que pagar a el mismo tipo que dichas acciones preferentes. Un caso interesante de un tipo similar de activos fue la emisi on de Barclays Bank en ese mismo a no de 300 millones de d olares de lo que se llam o Convertible Capital Notes. La estructura dise nada permit a a Barclays emitir deuda perpetua al 8%, libre de retenciones impositivas en Estados Unidos, las cuales pod an convertirse en cualquier momento en acciones preferentes. Una t ecnica similar fue utilizada en la emisi on de 1993 de Bankers Trust, de 150 millones de d olares en Prefered Purchase Units. Esta emisi on consisti o en la emisi on de deuda subordinada a 40 a nos que paga cup on al 7,625% y callable al quinto a no. El inversor dispon a del derecho de convertir el bono en una acci on preferente al mismo
405
tipo, Bankers Trust, sin embargo, dispon a de la opci on de previa comunicaci on 60 d as antes reducir el cup on un 1,5% en cualquier momento, lo cual forzar a la conversi on a acciones preferentes. Una nueva estructura interesante, Fixed/Floating Convertibles, aparece en 1993, cuando BOT Cayman Finance Ltd., entidad subsidiaria del Bank of Tokyo, emiti o cincuenta mil millones de yenes en bonos subordinados perpetuos que pagaban un cup on del 4,25% hasta marzo del 2003, y despu es Libor yen + 1,8, un tipo lo sucientemente bajo como para forzar al inversor a convertir el bono en acciones. Los bonos de cup on cero convertibles aparecen por primera vez en abril de 1985, como LYONS (Liquid Yield Options Notes ), caracterizados por ser bonos de largo plazo (de 15 a 20 a nos) emitidos al descuento y con opci on para el inversor de convertibilidad cada cinco a nos de acuerdo con unos ratios de convertibilidad similares a los de un DECS. Sin embargo, Merrill Lynch en noviembre de 1994 se sale de la norma, creando lo que luego se denomin o Exchangeable LYONS en una emisi on para Microsoft de bonos cup on cero con vencimiento en 4,6 a nos, precio de emisi on cercana a la par y pago de prima al vencimiento, con convertibilidad. Este tipo de bono convertible ha recibido un especial tratamiento te orico, por McConnell y Schwartz (1986), ofreciendo una soluci on anal tica como ecuaci on diferencial estoc astica con unas determinadas condiciones de contorno. En 1994 aparecen bonos convertibles, con ajuste anual del precio o ratio de conversi on, emitidos por Royce Value Trust con el nombre de ICONS (Investment Company Convertible Notes ). Fueron emitidos 60 millones de d olares en bonos a 10 a nos convertibles en acciones de Royce. En julio de 1996 aparece emitido en el mercado de convertibles el primer PARCKS (Puttable and Redeemable Convertible Knockout Securities ) en una emisi on del Banco de Bangkok, con valor de 350 millones de d olares. Estos son bonos convertibles en acciones subyacentes del emisor, pero que a los cinco a nos pueden ser amortizados anticipadamente a petici on del inversor a un precio entre el 123,5 y el 127,5 por ciento, si el precio de la acci on es menor que un predeterminado nivel, situado un 7% por encima del nivel que nos dar a un precio del bono del 130%. Lo que obtenemos es un bono convertible puttable con opci on knockout. Otra sorprendente variaci on de bono convertible, sobre todo por su exito, apareci o a principios de 1997, emitido por Indian Petrochemicals Corporations por una cuant a de 150 millones de d olares de Credit Enhanced Debt Indexed to Stock (CREDITS), un bono convertible cuyo principal fue garantizado por una carta de cr edito del Bank of America, lo que le permiti o nanciarse hasta 155 p.b. por debajo del tesoro americano. Se conocen como Convertibles sint eticos aquellos bonos que a fecha de vencimiento devuelven el principal en dinero (no activo), pero con una rentabilidad ligada al valor de las acciones de determinada compa n a. La primera emisi on de este tipo de bono apareci o en junio de 1991, emitido por la compa n a francesa Lafarge Copp e, a la que le sigieron otras empresas francesas como Carr efour, Louis Vuitton, Moet
406
Hennessy, y Pernod Ricard. El emisor consigue nanciarse por debajo del PIBOR sin sufrir el efecto diluci on que sufren el valor de las acciones de la compa n a en el momento de la convertibilidad del bono por acciones. Por u ltimo, como curiosidad, destacar el Risk Insurance Convertible, una estructura de bono convertible realmente extra na, emitida por la compa n a de seguros Winterthur a nales de 1997, como un bono a 3 a nos en francos suizos convertible en 5 acciones de Winterthur. La caracter stica llamativa de esta emisi on es que paga un cup on anual del 2,25% siempre y cuando no ocurra una cat astrofe atmosf erica, tormentas de granizo, que da ne en un solo d a m as de 6.000 veh culos, en cuyo caso, no pagar a cup on.
1.6
En denitiva, por qu e existen los bonos convertibles?
Razones por el lado del emisor: Una ampliaci on de capital implica como media una reducci on de entre un 3 y un 8% en la cotizaci on de la acci on (Mikkelson y Parch 1986, y Smith 1986) frente al 2% de media como efecto a la emisi on de un convertible, y adem as el efecto de diluci on es postergado en el tiempo hasta la conversi on. Un bono convertible se emite con cupones inferiores a los de la deuda corporativa de mismo riesgo. Los bonos convertible ofrecen ventajas scales a los emisores frente a la simple ampliaci on de capital. La progresiva conversi on de una emisi on de bono convertible mejora los ratios de endeudamiento de la compa n a. Y razones por el lado del inversor: Ratios de rentabilidad/riesgo hist oricos muy superiores a los ratios de la deuda y renta variable. Los bonos convertible permiten participar de los mercados de renta ja y de renta variable, diversicando el riesgo. Los inversores en bono convertible experimentan en una gran parte toda la subida de valor de la renta variable, aunque normalmente los bonos convertible se emitan a una paridad mayor que la de valor del mercado.
407
2
2.1
Evoluci on de las diferentes estructuras de bonos convertibles y modelos de valoraci on

Valoraci on de bonos convertibles con tipos deterministas: modelo de un factor
Para introducir las ideas que se utilizan en la valoraci on de los bonos convertibles, empezamos con el modelo m as sencillo, el que asume que los tipos de inter es son deterministas y deja el valor de un bono convertible de vencimiento T , dependiendo solo del subyacente y del tiempo a vencimiento. En estas condiciones, el valor del bono convertible ser a: V = V (S, t) . Si construimos una cartera o portfolio compuesta por un bono convertible y S , repitiendo el an alisis de Black-Scholes, nos encontramos que la variaci on en el valor de la cartera es: d = V 1 V 2V dt + dS + 2 S 2 dt dS , t S 2 S 2
donde, si tomamos el valor de la delta como = V , S
nos queda una cartera libre de riesgo cuya rentabilidad deber a ser la del tipo de inter es libre de riesgo (condici on de no-arbitraje), que es expresada a trav es de la desigualdad de Black-Scholes, V 1 V 2V dt + (rS D(S, t)) dS + 2 S 2 dt r V 0 , t S 2 S 2 y unas ciertas condiciones, nales y de contorno, que se describen a continuaci on: V (S, T ) = 1 , = V S, t+ V S, t c c +K, V nS , V (S, t) nS , V (0, t) = e r(T t) + donde (1) implica que a vencimiento el valor del bono ser a 1, y justo antes del vencimiento ser a max(n S, 1). K er (tc t) ,
S
(1) (2) (3) (4) (5)
408 (2) es el salto del valor en fecha de pago del cup on K .
(3) implica que el valor del bono convertible ser a siempre mayor o igual al valor de la conversi on (por no-arbitraje). (4), donde el valor del bono convertible tiende al valor de conversi on a medida que crece el valor del stock, y si este es sucientemente grande, el bono convertible se convierte en us sustituto perfecto del stock. (5): y si el subyacente vale cero, el bono convertible pasa a ser valorado como un bono. Algunas complejidades a nadidas podr an ser: El ratio de conversi on puede variar con el tiempo y/o con el valor del subyacente: n = n(S, t). Pueden restringirse los periodos de conversi on, o hacerlos intermitentes, o incluso establecer convertibilidad europea. Pueden introducirse opciones de amortizaci on anticipada (call/put ).
2.2
Brennan - Schwartz (1977)
Siguiendo la misma notaci on que la del art culo de M. J. Brennan y E. S. Schwartz (1977), denimos: S (t) V (V, t) CP (t) B (V, t) D(t) q (t) Nc N0 I : valor de mercado de la acci on del emisor, : valor de mercado del bono convertible, con valor a la par 1000, : valor de la opci on de vencimiento anticipado en el momento t, : valor del bono de la misma compa n a no convertible, : pago de dividendo de la acci on del emisor, : n umero de acciones por los cuales se convierte el bono convertible, : n umero de bonos convertibles, : n umero de acciones antes de la conversi on, : pago de cup on en cada fecha de pago,
C (V, T ) : valor de conversi on, donde q (t) S (t) = z (t) S (t) , N0 + Nc q (t) q (t) . N0 + Nc q (t)
C (V, T ) = z (t) =
409
Durante la vida del propio bono de convertible, la condici on de arbitraje que debe de cumplirse es: V (V, t) C (V, t) . El valor de mercado del bono convertible ha de ser mayor que el valor de conversi on, dado que en el momento que el valor de convertir fuese mayor que el propio bono, la conversi on en acciones ser a masiva. Por otro lado, la caracter stica de amortizaci on anticipada introduce una nueva condici on sobre el valor del bono convertible, dado que el valor del bono si se amortizara ser a: V IC (S, t) = max [CP (t), C (t)] . El valor del bono convertible en caso de amortizaci on anticipada ser a el mayor entre el valor al que se amortizar a o el propio valor de mercado de la conversi on en ese mismo momento. Evidentemente, el emisor amortizar a anticipadamente cuando el precio de amortizaci on sea menor que el del propio mercado, y todav a no haya entrado en el breakeven de conversi on. Adem as, si llamamos t = t al momento en el que se produce la amortizaci on anticipada, ocurre que V (S, t ) = C (S, t ) si C (S, t ) CP (t ). Brennan y Schwartz proponen un modelo de un solo factor donde el valor de mercado el bono convertible (W = W (V, t)) depende tan solo del tiempo (t) y de la evoluci on de la acci on (S = S (t, z )), la cual sigue la siguiente evoluci on estoc astica: dS = dt + dz , S donde dz es un proceso de Gauss-Wiener. Repitiendo el an alisis de Black y Scholes, el valor del bono convertible debe satisfacer la siguiente ecuaci on diferencial (ecuaci on diferencial de Black y Scholes), 1 2 2 V 2 2W V 2 +rV W V rV = 0,
410
acompa nada de las consiguientes ecuaciones de contorno: Nc V (S, T ) V (Smin , t) V (S, t) V (S, t) V (S, t) V (S, t)
S
= > = = =
S, 0, B (S, t) + z (t) S , C (S, t) = z (t) S , z (T ) S , z (T ) S 1000 , 1000 , 1000 Nc S 1000/z (T ) , S/Nc , S 1000 Nc , CP (t) , z (t) , max V (S D, t+ ), z (t ) S , V (S I, t+ ) + i , min V (S I, t ) + i, CP (t ) ,
+
(6) (7) (8) (9) (10) (11) (12) (13) (14) (15)
lim
V (S, t) S V (S, t )
V (S, t ) = V (S, t ) =
(6): el valor total de los bonos ha de ser inferior al valor de la compa n a. (7): el bono convertible no tiene valor en el caso de que el valor de la acci on caiga a partir de determinado nivel (Merton). (8), siendo z (t) el valor m aximo de z (h), para h = (t, T ), por lo que el valor del bono convertible es menor que un bono m as el valor del m aximo n umero de acciones en que se puede convertir. (9): el valor del bono convertible ser a siempre mayor que el valor de la conversi on. (10). Condici on del valor a vencimiento: a vencimiento el bono convertible da el valor en acciones si este es mayor que su nominal, (11). Opci on de amortizaci on anticipada que tiene el emisor, que amortiza siempre en el caso que el valor del bono convertible supere al valor de amortizaci on o call price. (12). Condici on para un valor sucientemente grande de la acci on. (13). Condici on que se cumple en cada fecha de pago de dividendos y pago de cupones en el instante anterior t y en el instante inmediatamente posterior t+ . (14). La u ltima condici on es para el pago de cupones cuando coincida con el momento de la amortizaci on anticipada.
411
2.3
Valoraci on de los bonos convertibles con tipos estoc asticos: modelo de dos factores
Dado que la vida de un bono convertible es normalmente mayor que la vida de una opci on sobre el stock, es mucho m as precisa la valoraci on bajo el supuesto de tipos estoc asticos, por lo que el valor del bono convertible V = V (S, r, t) queda como dS dr E [dX1 dX2 ] = S dt + S dX1 , = u(r, t) dt + w(r, t) dX2 , = dt .
Desarrollando por Taylor, con los cambios propios del c alculo estoc astico, obtenemos
2 = dt , dX1 2 dX2 = dt ,
dX1 dX2 = dt 2V 2V 2V + w2 + 2S w 2 S Sr r2 dt .
dV =
V V 1 V dt + dS + dr + t S r 2
2 S 2
Y si construimos una cartera de valores compuesta por el bono convertible de vencimiento T1 , con 2 bonos de cup on cero con vencimiento T2 y 1 de acciones, obtenemos una cartera = V 2 Z 1 S 2 1 = =
V r V S Z r
Y como en anteriores casos, tomando para 1 y 2 los valores adecuados, eliminamos la parte estoc astica de la cartera y obtenemos 1 1 2 2V V V 2V V 2V + 2 S 2 + w + (u w) rV = 0, + S w +rS t 2 S 2 S r 2 r2 S r donde (r, S, t) es nuevamente la prima de riesgo que asigna el mercado al tipo de inter es, la misma que para cualquier bono que no dependa de otro activo.
3
3.1
Modelos de bonos convertibles con riesgo de cr edito

Modelos de valoraci on del riesgo de cr edito
Aunque la valoraci on de activos con riesgo de cr edito a trav es de metodolog a en tiempo continuo tuvo unos inicios prometedores con las propuestas iniciales de Black y Scholes (1973) y Merton (1974), ha sido un a rea que ha tenido que esperar a los u ltimos desarrollos de modelizaci on de la curva de tipos por Vasicek (1977) o Heath, Jarrow y Morton (1987), entre otros, para comenzar a tener una base matem atica y cient ca m as consistente. S olo recientemente la modelizaci on del riesgo de cr edito
412
ha recibido una renovada atenci on por parte tanto de los acad emicos como de los profesionales de los mercados nancieros. En la actualidad existen dos enfoques en los modelos de medici on de riesgo de cr edito. El primer enfoque parte de los primeros trabajos de Black y Scholes (1973), donde un activo con riesgo de cr edito llevaba incorporado un derivado sobre el valor de la compa n a o rma emisora del bono, de forma que pod a ser valorado de acuerdo con la teor a del precio de las opciones. En estos modelos el valor de la compa n a o rma sigue un proceso de difusi on y la quiebra o default es valorado como un tiempo de parada que se produce cuando el valor de la rma alcanza un determinado punto o frontera. Gracias a la continuidad de los procesos modelados, el momento de parada que dene el default es predecible. El payo en caso de default es normalmente una tasa de recuperaci on del principal que depende del proceso de liquidaci on de la compa n a. Referencia
Merton (1977) Shimko et al. (1993) Leland (1994) Longsta y Schwartz (1995) Das (1995) Anderson et al. (1996) Sa aRequejo y Santa Clara (1997) Briys y de Varenne (1997) Zhou (1997)
Denici on de default
V F al vencimiento T V F en T , con tipos estoc asticos, correlacionada con V V K con K end ogena o ex ogenamente determinado V F entre 0 y T con tipos de inter es estoc asticos, y cuya din amica se correlaciona con V V F entre 0 y T valoraci on como compound option, y Ho y Lee como modelo de tipos de inter es El proceso de default no es solo un proceso ex ogeno sino que tambi en intervienen comportamientos de las partes V K entre 0 y T donde K sigue un proceso de difusi on correlacionado con los procesos del short rate y V V F P (t, T ) entre 0 y T con tipos estoc asticos correlacionados con V V F entre 0 y T donde dV /V sigue un proceso de jumpdiusion
Tabla 1. Enfoque estructural: especicaciones te oricas. En la tabla anterior, seguimos la siguiente notaci on: V , valor de la rma o compa n a. T , vencimiento de la deuda. F , valor de mercado de la deuda de la compa n a. P (t, T ), precio en t de 1 u.m. libre de riesgo en el momento T (bono cup on cero). Un segundo enfoque es aqu el en que el momento de default es modelizado directamente como un momento de parada de una determinada intensidad. Este es el enfoque seguido por Artzner y Delbaen (1992, 1994) Jarrow y Turnbull (1995) Lando(1994, 1998) Jarrow, Lando y Turnbull (1997), Madan y Unal (1998), Flesaker et al. (1994), Due y Singleton (1997, 1999), Due, Schroder y Skiadas (1994) Due y Huang (1996), y Due (1994). La principal diferencia entre los diferentes modelos esta en la forma en que obtienen la tasa de recuperaci on.

Referencia Jarrow y Turnbull (1991, 1995)
413
Madan y Unal (1993, 1998)
Longsta y Schwartz (1995) Das y Tufano (1996) Jarrow et al. (1997)
Due y Lando (1997) Due y Singleton (1999)
Denici on de default De forma discreta, existe una probabilidad de default (martingala) para cada intervalo discreto de tiempo. En continuo, la estructura a termino de las probabilidades de default vienen exponencialmente distribuidas con un par ametro de intensidad resultante de un proceso de salto. Modeliza tanto la probabilidad de default como la tasa de recuperaci on o severidad de la p erdida, bajo el supuesto que ambas variables est an expl citamente valoradas en los mercados nancieros. Modelo de reversi on a la media de para los spreads de cr edito Modelo similar a Jarrow et al. (1995) pero con tasas de recuperaci on estoc asticas Primer modelo que expl citamente incorpora la informaci on de las matrices de transici on entre rating en el m etodo de valoraci on El proceso de default esta condicionado a determinados datos internos de la propia compa n a El activo con riesgo es valorado con una curva de descuento, modelizada sobre el shortrate, con riesgo.
Tabla 2: Enfoque de probabilista o de intensidad: especicaciones te oricas
3.2
Modelo de Brennan y Schwartz (1980)
El segundo modelo de Brennan y Schwartz es un modelo de valoraci on de bonos convertibles con dos subyacentes (tipos de inter es y valor de la acci on) de tipo estructuralista en cuanto a la forma de introducir el riesgo de cr edito en la valoraci on del convertible. Denimos como valor de la compa n a o rma (V ) a la suma de los valores de mercado de los tres activos emitidos por esta: bono convertible, acciones y deuda. Valor de la rma, antes (T ) y despu es (T + ) del momento de la conversi on: F VT F VT + = = NB BT + NCB C BT + NS ST , NB BT + + (NS + N ) ST + .
Modelizamos la evoluci on de los tipos a trav es de la SDE de Vasicek (1977) sobre el short rate, y por otro lado las variaciones del valor de la rma: drt dF Vt E [dX1 dX2 ] = a (b rt ) dt + dX1 , = [F V F V Q(F V, t)] dt + F Vt F V dX2 , = dt ,
414
donde los par ametros tradicionales de Vasicek son conocidos, y donde F V es la rentabilidad de la compa n a, y Q(F V, t) es la suma de retribuciones de capital de la rma, es decir cupones de deuda, del bono convertible y dividendos, de forma que Q (F Vt , t) = IB + ICB + D(F V, t) . Bajo los citados supuestos, Brennan y Schwartz demuestran que el valor del bono convertible satisface la PDE:
2 1 2 1 2 2 2V 2V V 2 V + F + r + F V + F V V FV t 2 F V 2 F V r 2 r2 V V (r F V Q(F V, r)) + (a (b r) r ) rV + cF = 0. F V r Y las siguientes condiciones: q V (F Vt , rt , t) (F Vt NB B (F V, r, t)) , (N0 + N )
V (F Vt , rt , t) V (F Vt , rt , t) Adem as, V (F VT , rT , T ) es
CP (t) , = kF si F V = NB B (: 0) + k F VCB .
q (F Vt NB B (F V, r, t)) q (F Vt NB B (F V, r, t)) si FV N0 + N N0 + N q (F Vt NB B (F V, r, t)) F V NB B (: 0) F si FV N0 + N Nc F V NB B (: 0) F V NB B (: 0) si F V 0 N Nc c 0 si F V < NB B (: 0) Condiciones que pese a su complejidad, son muy similares a las descritas en su anterior trabajo, descritas en la subsecci on 2.2 del presente estudio.
3.3
Modelo multivariante con jump diusion
El presente modelo parte de la modelizaci on del valor del bono convertible que depende del tiempo, subyacente y del grado o status de default, con Y = 0 si no se ha producido default e Y = 1 en el momento de default, de forma que ahora es V (S, t, Y ), donde Y sigue un proceso de Poisson de la forma: P Y = 1 en (t + dt) Y = 0 en t P Y = 0 en (t + dt) Y = 0 en t P Y = 1 en (t + dt) Y = 1 en t = dt , = 1 dt , = 1.
n Modelos multivariantes de valoracio Suponiendo la evoluci on del subyacente dS = S dt + S dX1 + (J 1) S dY , y el valor de la cartera = V (S, r, Y, t) 1 S (r, Y, t) ,
415
aplicando It o, y la cobertura por delta, obtenemos la ecuaci on diferencial de la cartera cubierta s olo con respecto al proceso browniano, y no al de Poisson Y : d = dt K D V V 2 2 2 V + + S S t 2 S 2 + dY [V (S, t + dt, Y + dy ) V (S, t, Y )] ,
donde K es cup on del bono convertible y D el dividendo del subyacente. Por tanto, la contribuci on de los procesos de Poisson no puede ser eliminada a trav es de la tradicional cobertura delta (deber amos elegir entre cubrir u nicamente la difusi on peque na o minimizar las variaciones del valor de la cartera con la difusi on completa), y entramos en mercados incompletos; pero, sin embargo, podemos suponer que nos encontramos en probabilidad de riesgo-neutro si se cumple como media que E [d] = r E [] dt . Es decir, que a pesar de los saltos, el mercado descuenta impl citamente, y seguimos encontr andonos en la probabilidad riesgoneutral. Adem as, suponiendo que el valor del bono convertible en el caso de default es una fracci on V (S, t, 0), obtendr amos que el valor esperado de la parte estoc astica no cubierta de la cartera ser a E [dY (V (S, t + dt, Y + dY ) V (S, t, Y ))] E [dY ] V (S, t, 1) dt V (S, t) ( 1) , = = dt V (S, t, 0) .
La evoluci on de la cartera de convertible y acci on, en aplicaci on de la condici on de no arbitraje, nos quedar a como: r E [] dt = r V S V S dt .
Poniendo todo esto en conjunto, nos queda una versi on riesgoajustada de Black Scholes: K + (rS D) V 1 V 2V + + 2 S 2 (r + (1 )) V = 0 . S t 2 S 2
416
De igual forma, si consideramos un modelo de evoluci on sobre tipos de inter es, el resultado es similar: dS dr E [dX1 dX2 ] = S dt + S dX1 + (J 1) S dY , = u(r, t) dt + w(r, t) dX2 , = dt .
La cartera autonanciada del bono convertible contra sus subyacentes: = V (S, r, Y, t) 2 Z (r, t) 1 S (r, Y, t) , 1 1 2V V 2V 2V + 2 S 2 + w2 + S w 2 t 2 S S r 2 r2 V V + (u w ) (r + (1 )) V rS S r + = 0.
3.4
Modelo multivariante de Goldman Sachs
Goldman Sachs (1993) incorpora el riesgo de cr edito incrementando los spread en funci on de lo out-of-the-money en que se encuentre el bono sobre la acci on. Es decir, utiliza un tipo de descuento que est a en funci on de la probabilidad de conversi on. El modelo evoluciona el stock con un a rbol binomial de CRR, y en cada nodo nal se valora el bono convertible, que ser a igual al valor m aximo entre la conversi on del bono, o el propio valor nominal + u ltimo cup on del propio bono; y utilizando la inducci on backward t pico del a rbol: Vn,i = max m Si , p Vn+1,i+1 ern+1,i+1 t + (1 p) Vn+1,i ern+1,i t , donde m es el ratio de conversi on y p es la probabilidad de evoluci on del precio de la acci on, n el estado e i el momento del tiempo medido en t. G-S introduce el citado efecto modelizando la evoluci on de los tipos de descuento en funci on de la probabilidad de conversi on qn,i de cada nodo, de forma que si en T el comportamiento optimo es la conversi on por parte del inversor, dicha probabilidad es igual a uno, y cero si no lo es; por lo que partiendo desde el vencimiento, obtenemos qn,i = p qn+1,i+1 + (1 p) qn+1,i , estimaci on backward de las probabilidades de conversi on, en funci on de las probabilidades de evoluci on del stock. El ajuste del riesgo de cr edito en la funci on de descuento se calcula por la mixtura de tipos entre tipo libre de riesgo y el tipo con riesgo corporativo, ponderada por las probabilidades de conversi on, de forma que: rn,i = qn,i r + (1 qn,i ) (r + k ) .
417
Luego si la probabilidad de conversi on es igual a 1, la tasa de descuento es la tasa libre de riesgo, y si la probabilidad de conversi on es 0, la tasa de descuento es la tasa libre de riesgo mas un spread de cr edito.
3.5
Modelo propuesto
Es evidente que necesitamos una nueva metodolog a que permita estimar el valor del bono convertible de acuerdo con una visi on m as integral del riesgo de cr edito, de acuerdo con las tasas de recuperaci on y las probabilidades o pseudo-probabilidades de default que cotizan impl citamente los dos activos en los que se basa el bono convertible: los precios de la acci on y la deuda corporativa. Asimismo, consideramos necesario que cualquier metodolog a de valoraci on debe tener en cuenta dos factores importantes: la posibilidad de que aparezcan burbujas especulativas que distorsionen la percepci on del riesgo de cr edito, y la prima de iliquidez de la deuda corporativa. Desde el punto de vista tradicional de B-S, una cobertura delta del bono convertible con acci on y bono corporativo y/o bono libre de riesgo no es posible, dada la distorsi on en el market price of risk que ocasiona el riesgo de cr edito. Lo que buscamos es, al menos, dar una valoraci on de riesgo de cr edito al bono convertible consistente con sus dos subyacentes. Necesitamos una nueva modelizaci on de los diferentes activos tradeables que intervienen en la replicaci on del bono convertible, teniendo especialmente en cuenta el riesgo de cr edito y su prima: 1. Bono cup on cero libre de riesgo: sea p(t, T ) el precio de 1 euro en t, a pagar en T , y que matem aticamente podemos expresarlo a trav es de los tipos forward instant aneos impl citos en la curva de tipos cup on cero hoy, y la cuenta bancaria, expresada en t erminos de short rate : p(t, T ) = exp
T t t
f (t, u) du
B (t) = exp
r(u) du
2. Bono cup on cero de riesgo corporativo: sea v (t, T : i) el precio de X euros en t que deber a pagar en T el emisor xxx del bono convertible, con un grado de seniority i, y que matem aticamente podemos expresarlo a trav es de los tipos cup on cero R de un determinado riesgo de cr edito:
Q v (t, T : i) = Et exp T t
Ru du
X ,
i Rt = rt + ht L: t + lt ,
con ht la probabilidad condicional en t de default en el momento [t, t + 1] suponiendo que en t todav a no se ha dado, Lt la fracci on de p erdida si se produce default (1-tasa de recuperaci on) y lt la prima de liquidez en el momento t.
418
Donde adem as podemos aplicar el teorema de Singleton y Due (1999) por el que
Q v (t, T : i) = Et exp T t
Ru du
}
Q X = Et exp
T t
ru du
Z ,
= X 1{T <T
+ X 1{T T
X = (1 Lt ) v (t , T : i) .
3. Acci on subyacente: supongamos que tratamos el equity o acci on como el t tulo de deuda con peor seniority, que paga cupones deterministas llamados dividendos en los momentos 1, 2, . . . , hasta T , fecha de liquidaci on de la compa n a, a menos que haga default antes de T , de forma que S (t) T S (t) + Dj v (t, j : e) dj si t < , (t) =
t
e S ( ) +
Dj v ( , j : e) dj
si t = .
El valor actual de liquidaci on de la acci on, bajo probabilidad riesgo neutra y en funci on del propio riesgo de cr edito, por el teorema de Singelton y Due, ser a S (t) =
Q Et
[S (T )]
Q Et
exp
T t
e [ru + hu L: u ] du
Q = Et [S (T )] v (t, T : e)
Q = Et exp
T t
e [ru + hu L: u ] du
S (T ) ,
donde Et es la esperanza, bajo probabilidad riesgo neutral, condicionada a que el default no se haya producido antes o en el momento t, y ht es la seudoprobabilidad de default bajo probabilidad riesgo neutra. Si le aplicamos una hip otesis de independencia en t, entre las distribuciones de los procesos de S (T ), el short rate ru y la seudoprobabilidad de default hu para u en [t, T ], obtenemos
Q Q :e S (t) = Et [S (T )] Et exp T t [ru + hu Lu ] du
Q = Et [S (T )] v (t, T : e) .
no, y suponemos que la tasa de dividendos Para el desarrollo emp rico, T es 1 a continua de t + 1 es la de t; asimismo a la esperanza del valor de liquidaci on le imponemos que
Q Et [S (T )] = exp { (t, T )}
= (t ) exp {0 + 1 r(t) + 2 [log( (t)) log ( (t ))] + 3 M (t)} El riesgo de liquidez es un factor importante que afecta a la valoraci on y diculta la replicaci on, si deseamos cubrir el bono convertible. El introducir el riesgo de
419
liquidez de los activos es consistente con la condici on de no arbitraje, pero introduce incompletitud en los mercados. El factor de liquidez, al igual que en la teor a de la valoraci on de derivados sobre commodities, la introducimos a trav es del llamado convenience yield, de forma que el bono corporativo con riesgo de liquidez queda denido como vL (t, T : i) = ei (t,T ) v (t, T : i) , Si existen excesos de demanda, perjudica deshacer posiciones cortas. Si se dan excesos de oferta, perjudicando posiciones largas, (t, T ) = 0 + 1 vol0 + 2 volt , vol0
donde vol es el volumen negociado en el momento t, y 0 , 1 y 2 son constantes. Por tanto, deberemos estimar la probabilidad de default y la tasa de recuperaci on, que a su vez dependen de t, y del vector de las variables de estado observables X (t) (tipo spot y/o ndices de mercado) tal que h(t) i = h (t, X (t)) , = i (t, X (t)) .
Y debemos estimar, para cada t, usando regresiones secci on cruzada y series de tiempo i (t, T ) = (t, T ) = Inputs: 1. Serie de bonos cup on cero corporativos vL (t, T : i) para diferentes t, T y diferentes : i, y bonos libre de riesgo P (t, T ). 2. Precios observados de la acci on, a fecha T y los dividendos D1 , . . . , DT . 3. Y las variables observables de estado X (t). Con estas observaciones estimamos el siguiente sistemas de ecuaciones: vL (t, T : i) = vL (t, T : i, r(t), (t, X (t)), i (t, X (t)), i (t, T, X (t))) (t) = exp { (t, T , X (t))} v (t, T : e, (t, X (t))) +
t T
i (t, T, X (t)) , (t, T , X (t)) .
Dj v (t, j : e) dj
420
Referencias
[1] Black, F. y Scholes, M., The pricing of options and corporate liabilities. Journal of Political Economy 81 (1973), 63754. [2] Brennan, M.J. y Schwartz, E.S., Convertible bonds: valuation and optimal strategies for call and conversion. Journal of Finance 32 (1977). [3] Brennan,M.J. y Schwartz, E.S., Analyzing convertible bonds. Journal of Financial and Quantitative Analysis 15 (1980), 907-929. [4] Briys, E. y de Varenne, F., Valuing risky xed rate debt: An extension. Journal of Financial and Quantitative Analysis 32 (junio 1997), 239-248. [5] Clewlow y Strickland, Implementing derivatives models. Wiley Series in Financial Engineering (1998). [6] Das, S.R. y Sundaram, R. K., A direct approach to arbitrage-free pricing of credit derivatives. Working Paper 6635, National Bureau of Economic Research, Julio 1998. [7] Due, D. and Lando, D., Term structures of credit spreads with incomplete accounting information. Working paper, Graduate School of Business, Stanford University, Junio 1997. [8] Due, D., Forward rate curves with default risk. Working paper, Graduate School of Business, Stanford University, Diciembre 1994. [9] Due, D. y Kan, R., A yield-factor model of interest rates.Mathematical Finance 6 (1996), 379406. [10] Due, D. y Singleton, K. J., An econometric model of the term structure of interest rate swap yields. The Journal of Finance 52 (1997), 12871322. [11] Darrell Due, D. y Singleton, K. J., Modeling term structures of default able bonds. The Review of Financial Studies 12 (1999), 687720. [12] Geske, R., The valuation of corporate liabilities as compound options. Journal of Financial and Quantitative Analysis 12 (1977), 541552. [13] Goldman Sachs, Valuing Convertible Bonds as Derivatives. Quantitative Strategies Research Notes, Nov. 94. [14] Heath, D., Jarrow, R., Morton, A., Bond pricing and the term structure of interest rates: A new methodology for contingent claims valuation. Econometrica 60 (1992), 77105. [15] Jarrow, R., Madan, D., Option pricing using the term structure of interest rates to hedge systematic discontinuities in asset returns. Mathematical Finance 5 (1995), 311336.
421
[16] Jarrow, R., Lando, D., Turnbull, S.M., A markov model for the term structure of credit risk spreads. The Review of Financial Studies, 10 (1997), 481523. [17] Jarrow, R. A., Turnbull, S. M., Pricing derivatives on nancial securities subject to credit risk. Journal of Finance 50 (1995), 5385. [18] Johnson, N. L., Kotz, S., Continuous Univariate Distributions, volume 2. Houghton Miin Co., New York, 1970. [19] Lando, D., Three Essays on Contingent Claims Pricing. PhD thesis, Graduate School of Management, Cornell University, 1994. [20] Smith, G. D., Numerical Solution of Partial Dierential Equations: Finite Difference Methods. Oxford Applied Mathematics and Computing Science Series 1985. [21] Longsta y Schwartz, Interest rate volatility and the term structure: A two factor general equilibrium model. The Journal of Finance 47 (1992), 125982. [22] Longsta, F. A., Schwartz, E. S., The pricing of credit risk derivatives. Journal of Fixed Income 5 (1995), 614. [23] Longsta, F. A., Schwartz, E. S., A simple approach to valuing risky xed and oating rate debt Journal of Finance 50 (1995), 789 - 819. [24] Longsta, F. A., Schwartz, E. S., A simple approach to valuing risky xed and oating rate debt. Working Paper 2293, Anderson Graduate School of Management, Univ. of California, Los Angeles, octubre 1993. revisado noviembre 1994. [25] Madan, D., Unal, H., Pricing the risks of default. Review of Derivatives Research 2 (1998), 121160. [26] Merton, R. C., On the pricing of corporate debt: The risk structure of interest rates. Journal of Finance 29 (1974), 449470. [27] Merton, R. C., Option pricing when underlying stock returns are discontinuous. Journal of Financial Economics 3 (1976), 125144. [28] Merton, R. C., On the pricing of contingent claims and the Modigliani-Miller theorem. Journal of Financial Economics 5 (1977), 241249. [29] Zhou, Ch., A jump-diusion approach to modeling credit risk and valuing default able securities. Finance and Economics Discussion Paper Series 1997/15, Board of Governors of the Federal Reserve System, Marzo 1997. Jes us P erez Colino Accenture Torre Picasso 28020-Madrid, Spain e-mail: jesus.perez.colino@accenture.com

Volumen 02

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Volumen 02

Uploaded by

Copyright:

Available Formats

Seminario de Matem atica Financiera MEFF-UAM

Volumen 2. A nos 1998 y 1999

Remedios Romeo Directora General Instituto MEFF

Santiago Carrillo Men endez Jos e Luis Fern andez P erez

Cobertura en tiempo discreto de una cartera de opciones

El ratio de cobertura de m nima varianza

Cobertura en tiempo discreto de una cartera de opciones

K er (T tt) N De igual forma, y en tercer lugar,

2 2 var(Dt ) = E (Dt +t ) (E (Dt+t )) .

y A1 , A2 , B1 y B2 son constantes, se obtiene que

Resultados emp ricos

Ejemplo num erico

Cobertura en tiempo discreto de una cartera de opciones

On the Relevance of Modeling Volatility for Pricing Purposes

On the Relevance of Modeling Volatility for Pricing Purposes

The Bond Pricing Equation

more details, see Moreno (1996).

On the Relevance of Modeling Volatility for Pricing Purposes

A2 ( ) = B ( ) = C ( ) = with q1 = k1 + b1 , q2 = k2 + d2 , Proof: See Appendix.

2 2 1 = (k1 1 a1 )/q1 , s = 1 1 /(2q1 ),

lim P (s, L, ) = lim P (s, L, ) = lim P (s, L, ) = 0 ,

On the Relevance of Modeling Volatility for Pricing Purposes

On the Relevance of Modeling Volatility for Pricing Purposes

On the Relevance of Modeling Volatility for Pricing Purposes

On the Relevance of Modeling Volatility for Pricing Purposes

forecasts were also computed. Results are available upon request.

On the Relevance of Modeling Volatility for Pricing Purposes

On the Relevance of Modeling Volatility for Pricing Purposes

On the Relevance of Modeling Volatility for Pricing Purposes

and a little algebra leads to C ( ) = 2 (exp{ } 1) . (q2 + ) exp{ } + ( q2 ) A2 ( ) = 0. A2 ( ) (33)

Tables and Figures

st st1 = a1 + b1 st1 + s t, Lt Lt1 = a2 + b2 Lt1 + L t , Lt Lt1 = a3 + b3 Lt1 + L t ,

2 s t IID (0, 1 ) 2 L t IID (0, 2 Lt1 ) 2 L t IID (0, 3 )

k 0.01544 (3.0756) 0.01042 (3.6285) 0.00728 (2.3988)

-0.1347 103 (-0.021) 0.103532 (33.2787) 0.100574 (20.881)

0.003467 0.070703 0.001159

Table II. Averages of Pure Cross-Sectional Regressions

1 eq1 2(exp{ } 1) , C ( ) = q1 (q2 + ) exp{ } + ( q2 ) 1 = (k1 1 a1 )/q1 , 2 = k2 2 /q2 ,

Table III. Averages of Market Prices of Risk

1 (Spread) 2 (Long rate)

1 (Spread) 3 (Long rate)

1994 -0.7678 (-2.07) [18.0933] -1.7900 (-29.76) [44.0095]

1995 -5.7401 (-54.31) [7.5040] 2.8856 (1.75) [20.7386]

On the Relevance of Modeling Volatility for Pricing Purposes

Table IV. Within-Sample Pricing Error Measures. 1991-1994

Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year

MAPE 0.000106 0.000321 0.001239 0.007631 0.021114 0.053265 0.070541 0.256526

RMSPE 0.000421 0.001519 0.005413 0.028254 0.075686 0.162536 0.159114 0.467754

Table V. Within-Sample Pricing Error Measures. 1992

Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year

MAPE 0.000035 0.000124 0.000533 0.003838 0.011595 0.034493 0.050313 0.181938

RMSPE 0.000068 0.000230 0.000940 0.006357 0.019969 0.057827 0.072872 0.281741

Table VI. Within-Sample Pricing Error Measures for 1-year Bonds

Period 1991:I 1991:II 1992:I 1992:II 1993:I 1993:II 1994:I 1994:II

ME -0.000138 0.000386 -0.000062 -0.000022 0.000586 0.000480 0.000753 -0.000096

RMSPE 0.079741 0.121425 0.048048 0.066439 0.364635 0.149408 0.154830 0.048524

Table VII. Comparison of One-Step-Ahead Forecasts. 1995

Maturity 7-day 15-day 1-month 3-month 6-month 1-year 3-year 5-year

MAPE 0.001648 0.003242 0.006781 0.018363 0.033589 0.063822 0.180321 0.357493

RMSPE 0.002410 0.004738 0.009874 0.026064 0.045608 0.084462 0.245460 0.473140

On the Relevance of Modeling Volatility for Pricing Purposes

Table VIII. Comparison of One-Step-Ahead Forecasts. 1-year Bonds

Table IX. Comparison of Five-Step-Ahead Forecasts. 1995