You are on page 1of 7

cutting

edge. equity deRivatives

NO T

FO

Multi-asset options

have long been prominent among equity derivatives. The number of securities referenced in an equity payout may range from a few in the case of, say, a basket option on indexes to 50 for the case of a correlation swap on the components of the Stoxx 50 index. In addition, option payouts are generally written on baskets of securities that span all geographical areas. In multi-asset models, securities are modelled as correlated processes that are simultaneously observable. This is suitable for options whose underlying securities all belong to the same geographical area and whose deltas can be calculated and readjusted simultaneously, usually on a daily basis. However, exchanges throughout the world have asynchronous opening and closing times. For example, while European and American exchanges operating times have some overlap, European and Asian exchanges, and Asian and American exchanges, usually have no overlap at all. This precludes simultaneous observation of all securities in a given basket. Despite this, derivatives trading desks typically use the following methodology: n Use standard multi-asset models based on the assumption of continuously traded securities typically a multi-asset BlackScholes or local volatility model.

RE

Lorenzo Bergomi addresses the issue of pricing multi-asset options in the context of asynchronous markets. Using the criterion that the carry profit and loss (P&L) vanishes, he derives the expression of the correlation estimator for the asynchronous case. He studies its historical behaviour for the case of the Stoxx 50, S&P 500 and Nikkei indexes, and compares his estimator with popular heuristic estimators. Finally, he characterises practical situations whereby correlations larger than one are materialised as a P&L

76

Risk November 2010

PR
1

OD

Estimating correlations and volatilities in asynchronous markets

The academic literature has mostly focused on the case of correlated processes that exist continuously in time, yet are observed asynchronously. The issue is that of backing out from discrete asynchronous observations the latent correlations of the processes and characterising the variance and bias of the resulting estimators. In our context, the notion that there exist true correlations driving the motion of securities even while their respective markets are closed and which it is our task to uncover, may seem incongruous and unnecessarily ambitious. Rather, we are concerned with the issue of parameterising our model so that, on average, no money is made or lost as we risk-manage an option over its lifetime. This is the criterion we use in this work for deriving correlation and volatility estimators. Let us consider two stocks trading on asynchronous markets, as shown in figure 1. The Stoxx 50 daily closing times are denoted by ti1, ti, ti+1 .... They are equally spaced by D = ti+1 ti = 24 hours. The Nikkei closes 10.5 hours earlier than the Stoxx 50, at times ti1 d, ti d, ti+1 d .... Let us assume that valuation of the option is done at the close of the Stoxx 50 and that deltas are calculated and traded at the market closes of each security, using the last known values for both indexes. The options value is given by f(t, S1, S2), where S1 denotes the Stoxx 50 and S2 the Nikkei. S1,i and S2,i are the values of the Stoxx 50 and the Nikkei on their respective closes, for calendar day i: S1,i = S1(ti), S2,i = S2(ti d). Without loss of generality, let us assume that interest rates and repos are vanishing. Our P&L over [ti, ti+1] reads:

While for pricing we may not be free to stray from market implied values, we need historical volatilities and correlations as benchmarks, to ensure that the carry levels of our position are reasonable

UC

TI

Correlations in asynchronous markets

n When readjusting the delta-hedge on each security daily, calcu-

late the delta by using stale values for the securities that are not trading at the time of computation. n Likewise, daily valuation of the option is carried out using available closing quotes for securities that are live at the time of valuation and stale values for other ones. These stale values are usually the last known closing quotes of the respective securities. As deltas are calculated and traded at different times for different securities, using different information sets, a natural question arises: how should one estimate the correlation parameters needed for pricing and what is the nature of these correlations? Surprisingly, given its practical relevance and the size of correlation books in equity derivatives departments, this issue seems to have attracted little attention in the literature. Attempting to frame it into a multi-dimensional stochastic control problem with processes that exist only in asynchronous periodic time windows seems a daunting task. We take here a more practical route and answer the following question. Assume we use a standard continuous-time model for valuing and deltahedging an option on asynchronous stocks on a daily basis. Which volatility and correlation parameters should we use?1 We first derive the expression of the correlation estimator, then study its properties and its historical behaviour for the case of Stoxx 50, S&P 500 and Nikkei indexes and compare our estimator with commonly used heuristic estimators. We then study cases when it lies above one, and characterise practical situations whereby correlations larger than one may be materialised as a profit and loss (P&L).

ON

P & L = f (ti+1,S1, i+1,S2, i+1 ) f (ti ,S1, i ,S2, i ) df + (ti ,S1, i ,S2, i ) (S1, i+1 S1, i ) dS1 df + (ti ,S1, i1,S2, i ) (S2, i+1 S2, i ) dS2

1 example of two non-overlapping asynchronous securities: the stoxx 50 and nikkei indexes
r1i1 Stoxx 50 ti1 r2i1 Nikkei ti1 ti r2i ti+1 r2i+1 r1i

df f (t i+1,S 1, i+1,S 2, i+1 ) f (t i ,S 1, i, S 2, i ) + dS ( S 1, i+1 S 1, i ) 1

FO

The other correction terms, representing the difference between df/dS2 evaluated at t and evaluated at t d contribute terms of higher order in D to the P&L. As f and its derivatives are now all evaluated with the same arguments (ti, S1,i, S2,i), we no longer carry them. Rewriting our P&L, we now get:

RE

df d2 f = (ti ,S1, i ,S2, i ) dS dS (ti ,S1, i ,S2, i ) (S1, i S1, i1 ) dS2 1 2

PR
(1)

df (ti ,S1, i1,S2, i ) dS2

OD

The first piece is the P&L of the option itself, calculated at the close of the Stoxx 50 using S1,i+1 and the stale value S2,i+1. The second piece is the P&L generated by the delta on the Stoxx 50, while the third piece is the P&L generated by the delta on the Nikkei. Note that the arguments of df/dS1 and df/dS2 differ. At the Stoxx 50 close, we use the closing value of the Nikkei on the same (calendar) day, that is, S2,i, while on the Nikkei close we use the closing value of the Stoxx 50 one day before, that is, S1,i1. Because the two deltas are not evaluated using the same arguments, their contribution will not exactly cancel the first-order term in an expansion of the options P&L in S1 and S2 , in contrast with the synchronous case. Let us expand the expression above at order one in D and order two in dS. This is consistent as, practically, variances and covariances of returns scale linearly with their time scales. To cancel the first-order term in dS2 in the options P&L, we need to rewrite the delta on S2 so that it involves the same arguments as f at time t:

2 df 1 2 d 2 f 2 2 d 2 f d2 f + S1 2 + 2 S2 2 + 1 2 S1S2 =0 dt 2 2 dS1dS2 S1 S2

We get the following expression for the P&L:


2 1 2 d 2 f S 2 P & L = S1, i 2 1+ 1 2 dS1 S1, i 2 1 2 d 2 f S S2, i 2 2+ 2 2 2 dS2 S2, i S 2 d f 1 S1+ S2+ S1, i S2, i + 1 2 dS1dS2 S1, i S1, i S2, i

UC
*2 1 =

contributions. In contrast with the case of synchronous securities, we get an extra cross-gamma term involving the product dS1dS2+. This expression for the P&L is general to back out a constant correlation parameter, let us make the assumption that f solves a BlackScholes equation with constant volatilities and correlation s1, s2, r:

TI
*2 = 2 1 1 S S S 1 + 1+ 2+ S 1, i S1, i S2, i
*2 = 2 1 2 r2i , * =

ON
ti ti+1

(2)

From this expression, we get volatility and correlation estimators s* , s* , r* that ensure that, at order one in D and two in dS, 1 2 the P&L vanishes on average:

NO T

df d f + (S 1, i S 1, i1 )(S 2, i+1 S 2, i ) dS 2 dS 1dS 2 = f (t i+1,S 1, i + S 1+ ,S 2i + S 2+ ) f (t i ,S 1, i ,S 2, i ) 2 df d f df S 1+ + S 1 S 2+ + dS 1 dS 2 dS 1dS 2

S 2 1+ S 1, i

S 2 2+ S 2, i

* *1 * = 2

where dS1+ = S1,i+1 S1,i, dS1 = S1,i S1,i1, dS2+ = S2,i+1 S2,i. At order one in D and order two in dS, we get:

where brackets denotes averages. Let us replace dS1/S1,i with dS1/ S1,i1 as this does not change the covariance at lowest order in D. We then get:
*2 1 =

1 d 2 f 2 1 d 2 f 2 df P & L = S1+ + S2+ 2 2 dt 2 dS2 2 dS1 d 2 f d2 f + S1+S2+ S1S2+ dS1dS2 dS1dS2


The first two pieces of the P&L are the familiar theta and gamma

1 2 r1i ,

(r1i1 + r1i ) r2i


2 r1i 2 r2i

where r1i = (S1,i+1 S1,i)/S1 and likewise for r2i. i The volatility estimators are the usual ones. They involve daily returns sampled at the closing time of each asset, which is natural as delta-hedging is performed daily. The covariance, and hence the correlation parameter, also depends on daily returns and involves

risk-magazine.net

77

cutting edge. equity deRivatives

2 six-month eWMa running correlations: rS (blue), rA (pink), r* = rS + rA (green)


100 80 60 % 40 20 0 Dec 4, 99 100 80 60 % 40 20 0 Dec 4, 99 100 80 60 % 40 20 Stoxx 50/S&P 500

size effect of order 1/N where N is the sample size:


1 N 1 N 1 (r1i1 + r1i ) r2i 1 r1i (r2i + r2i+1 ) N N 1 = ( r10 r21 r1N r2 N+1 ) N

Nikkei/Stoxx 50

Dec 3, 01

Dec 3, 03 Dec 2, 05 Nikkei/S&P 500

NO T

0 Dec 4, 99

FO

Dec 3, 01

Dec 3, 03 Dec 2, 05

the product of a return of S2 and the sum of the two returns of S1 that straddle it (see figure 1). Had we chosen the closing times of the Nikkei for the options valuation, we would have obtained the following symmetrical correlation estimator, with two returns on the Nikkei straddling one return of the Stoxx 50:
* = r1i ( r2i + r2i+1 )
2 r1i 2 r2i

If the returns of S1 and S2 are time-homogeneous, r1i1r2i = r1ir2i+1 and the two correlation estimators are identical. In practice, averages are evaluated as averages over historical samples and the difference between the two covariance estimators is a finite

RE

78

Risk November 2010

PR

Dec 2, 07

Dec 1, 09

Dec 2, 07

Dec 1, 09

OD UC
2

Dec 3, 01

Dec 3, 03 Dec 2, 05

Dec 2, 07

Dec 1, 09

TI
* = S + A

We will not distinguish them in what follows. In conclusion, in asynchronous markets, we have two correlations denoted by rS (synchronous) and rA (asynchronous), whose estimators are: r1i r2i r1i r2i+1 S = , A = (3) 2 2 2 2 r1i r2i r1i r2i and derivatives should be priced in a constant correlation setting with r* given by:

Among all estimators proposed in the literature, r* resembles the Hayashi-Yoshida estimator.2 Their covariation estimator is simply expressed as the sum of all products of returns on S1 and S2 , provided they have non-zero overlap. In the above derivation, we calculate deltas using stale values for securities not trading at the time of computation. What if we had calculated deltas differently, for example, predicting the value of the security that is not trading by using data from the other security would we have priced correlation and hence our option differently? Consider a long position in an option deltahedged one way combined with a short position in the same option delta-hedged a different way: the resulting position is a pure investment strategy whose final payout is of the form Dt (St+ St), where Dt depends on information available at time t. The price of such a payout is zero, so the option prices in the two hedging schemes coincide. The expression for r* in equation (3) is thus general, and does not depend on the particular delta strategy used in the derivation. In the case of synchronous markets, rA involves the product of non-overlapping returns. While the realised covariance r1ir2i may differ from zero because of serial correlations, its implied value is zero, as it is the P&L of a pure delta strategy. This implies that rA = 0. We then recover the usual estimator rS . The volatility and correlation estimators s* , s* , r* ensure that 1 2 on average no money is made or lost as we delta-hedge an option. More precisely, if S2 d2f/dS2 , S2 d2f/dS2 , S1S2 d2f/(dS1dS2) are con1 1 2 2 stant, equation (2) shows that our estimators exactly quantify the P&L incurred. In practice, options have non-constant gammas: practitioners adjust volatility and correlation levels accordingly. We do not address this point, whose relevance extends to the usual synchronous case as well. While we call r* correlation, as this is the role it plays in pricing models and because it quantifies the realised cross-gamma P&L, it does not share the usual mathematical definition of correlation. An important issue, which we examine later, is whether r* [1, 1]. Before we analyse historical data, let us consider the theoretical situation in which processes for S1 and S2 exist continuously and could be hedged synchronously yet are observed asynchroT Hayashi and N Yoshida, 2005, On covariance estimation of non-synchronously observed diffusion processes, Bernouilli 11(2), pages 359379

ON

3 six-month sliding eWMa estimates of rS (blue) and rA (pink) for the stoxx 50/s&P 500 pair, calculated normally (left) and after exchanging the two time series (right)
100 80 60 40 20 0 20 Dec 4, 99 % % Stoxx 50/S&P 500 normal 100 80 60 40 20 0 20 Dec 4, 99 Stoxx 50/S&P 500 reversed

Dec 3, 01

Dec 3, 03

Dec 2, 05

Dec 2, 07

Dec 1, 09

UC
Dec 3, 01 Dec 3, 03

TI
Dec 2, 05 Dec 2, 07 Dec 1, 09 risk-magazine.net

A =

NO T

t+Dd where we have used the fact that td s2 dt = tt+Ds2 dt. We recover 2 2 the correlation estimator for continuous processes: our asynchronous estimator has no bias when applied to situations where synchronous hedging is possible.

Historical correlations for the Stoxx 50, S&P 500 and Nikkei indexes

Figure 2 shows six-month exponentially weighted moving average (EWMA) correlations for the three pairs Stoxx 50/S&P 500, Nikkei/Stoxx 50 and Nikkei/S&P 500, from December 1999 to December 2009: rS (blue), rA (pink) and the pricing correlation r* = rS + rA (green). rA is not small and the graph for the Stoxx 50/S&P 500 pair features instances when r* goes above 100%. This graph also suggests that rS and rA fluctuate antithetically, with their sum being more stable. Let us use the toy model with deterministic volatilities and correlation, with rS and rA given by expression (4). Imagine a

FO

Summing rS and rA , we get:


* =

1 t+ ds 1 2 t t+ 2 1 t+ 1 ds t 2 ds 2 t

1 t+ ds 1 2 t+ t+ 2 1 t+ 1 ds t 2 ds 2 t

RE

S =

1 t+ ds 1 2 t t+ 2 1 t+ 2 ds 1 ds t 2 t

PR
(4)
3

nously. The argument above implies that correlation estimators in both situations will coincide. Let us verify this explicitly with a simple example. n Continuous processes with asynchronous observations. Suppose that S1 and S2 have lognormal dynamics with deterministic time-dependent volatilities s1, s2 and correlation r, where s1(t), s2(t), r(t) are periodic functions with period D = 24 hours. Figure 1 yields the following expressions for rS and rA:

constant instantaneous correlation r and given levels of daily Dd integrated variances Ds1(s)2 ds and d s2(s)2 ds. If most of the 0 variance is realised during the same calendar day, rS will be large and rA small. For example, let us make the simple assumption that s1(s) = s1l(s), s2(s) = s2l(s), where s1 and s2 are constant, and l is such that (1/D) Dl2(s)ds = 1, for example assum0 ing that l(s) reproduces the distribution of the intra-day quadratic variation of the S&P 500, and also determines that of the Stoxx 50. We then get:

OD
S =

1 2 1 0 ( s ) ds, A = 2 ( s ) ds While the sum rS + rA = r stays fixed, rS, rA vary in opposite ways depending on the shape of l. Notice how, as we move from the Stoxx 50/S&P 500 pair through the Nikkei/Stoxx 50 pair, to the Nikkei/S&P 500 pair, the hierarchy between rS, rA is reversed with rA > rS for the Nikkei/S&P 500 pair this is in fact natural if we assume, going back to the discussion above, that most of the covariance between the Nikkei and the S&P 500 is realised during the [open, close] interval of the S&P 500, which technically occurs during the following calendar trading day in Japan. While it is noisy, the signal for rA for the Stoxx 50/S&P 500 case embodies real information. This can be assessed by calculating rS, rA using the same historical data, but by reversing the order of the closing times, thus pretending that the S&P 500 closes before the Stoxx 50. The two corresponding graphs are shown in figure 3. In contrast to the left-hand graph, the righthand graph shows that the curve for rA , which now quantifies the correlation of returns that do not overlap at all, hovers around zero.3 n Comparison with other heuristic estimators. Trading desks long ago realised empirically that using rS for pricing underestimates correlation. A standard fix consists of calculating correlations using returns over several trading days, typically three or five, instead of daily returns. How do these estimates differ from r*? Before addressing this question, we consider the particular case of correlation swaps. Correlation swaps are financial instru-

A Martian returning from a trip on planet Earth with historical closing quotes of the major exchanges would thus be able to determine the direction of rotation of the Earth

ON
79

cutting edge. equity deRivatives

4 example of daily returns of asynchronous securities

S A

Stoxx 50/S&P 500 100 80 60 % 40 20 0 1 100 80 60 % 40 20 0 3 5 7 Nikkei/Stoxx 50

NO T

FO R

5 Nikkei/S&P 500

100 80 60

40 20 0 1 3 5 7 9

RE
7 9

80

Risk November 2010

PR
9
4

OD UC
n = S +

5 n-day correlations as a function of n estimated using n-day returns (dark blue circles) or expression (5) (light blue circles), compared with r* = rS + rA (pink line), on the historical sample dec 1, 2004dec 1, 2009

ments whereby one party pays to the other party the average of the realised pair-wise correlations of a basket of securities, in exchange for a fixed strike. In term sheets of correlation swaps, correlation is expressed as the usual correlation estimator, evaluated using log-returns spanning n days: n = 1 and n = 3 are standard values. Measuring correlation using n-day returns and pricing correlation swaps are thus related issues. n Correlation swaps. Consider a correlation swap written on n-day returns and concentrate on the realised correlation of a single pair, which is the ratio of the covariance of n-day returns of two securities to the product of their standard deviations. Figure 4 shows that the n-day covariance involves n r1i r 2i terms (blue rungs) and n1 r1i r 2i1 terms (pink rungs). Thus an n-day correlation swap should be priced with correlation rn given by:

n 1 (5) A n For the n = 1 case, we have r1 = rS , while, for n = 3, r3 = rS + 2/3rA . In correlation swaps, covariance is expressed as Siln(S1,i+n /S1,i) ln(S2,i+n /S2,i). It is in fact a string of very short maturity payouts whose period n days is comparable with the delta-hedging period. This introduces some differences with respect to the derivation of r* in the first section and will cause rn to differ from r*. Consider first the case n = 1: the covariance payout has no delta and its value is simply r1i r 2i1. The resulting correlation is rS . Let us now turn to the general case n 1. The approach developed in the first section of this article holds, but for the first hedging interval. At time t d, when calculating the delta on S 2 we do not get the extra correction term in equation (1) as information prior to t, and especially S1,i1, is not being used at all. Thus, over the first daily interval [t, t + D], only rS is materialised as cross-gamma P&L, whereas over the remaining intervals the standard reasoning applies and the full correlation rS + rA is materialised. With respect to the pricing correlation for standard options, rS + rA , we get a correction given by (1/n)(rS (rS + rA)) = rA /n. The important conclusion is that correlation swaps should be priced with a different correlation than that of options.4 Figure 2 shows that this correction can be quite substantial. See figure 5 for the average differences over a five-year historical sample. n Correlation estimators with n-day returns. How do correlation estimators based on n-day returns compare with r*? First note, glancing again at figure 4, that if returns in our sample do not exhibit any serial correlation, the correlation of n-day returns can be expressed as a function of rS and rA using formula (5), that is, using estimators involving one-day returns only. It will then coincide with the correlation swap correlation. In practice, correlation calculated from n-day returns directly may be slightly different because of intermittent serial correlation. Figure 5 displays both estimators, along with r*. As expected, n-day correlations measured using n-day returns are very close to those calculated using equation (5). Only for the Nikkei/S&P 500 pair do we notice a bias, with the correlation computed using n-day returns lying above rn. This could be due

This is in addition to other effects such as volatility of volatility, which would make the implied correlations of covariance swaps, correlation swaps and basket options different, even for synchronous securities

TI

ON

RE

Realised correlations above one

to some serial correlation between return [i, i + 1] on the S&P 500 with return [i + 2, i + 3] on the Nikkei. Notice how r3 and r5 commonly used by practitioners underestimate r*, especially for the Nikkei/Stoxx 50 and Nikkei/ S&P 500 pairs about eight correlation points for r5. Figure 5 highlights how sizeable the difference between correlation swap correlation (n = 1, 3) and option correlation can be. For example, for n = 3: nine points for the Stoxx 50/S&P 500 pair, 12 points for the Nikkei/Stoxx 500 pair and 17 points for the Nikkei/S&P 500 pair. This is much larger than the typical bid/offer spread for these pairs. n The S&P 500 and Stoxx 50 as synchronous securities. European and American markets operating hours do have some overlap. We can thus either: delta-hedge asynchronously on the S&P 500 future at 4pm New York time and on the Stoxx 50 future at 5:30pm Paris time; or delta-hedge simultaneously both futures, say at 4pm Paris time, while both markets are open. In the first case, we should use r* as pricing correlation, while in the second case, the standard correlation estimator for synchronous securities, rS , should be used. In the first section of this article, we showed that both implied correlations are equal, for pricing purposes. The argument relies on the fact that the price of a pure delta strategy is zero, which has the consequence that the price of a product of non-overlapping returns vanishes: implied serial correlations vanish. In practice, though small, realised serial correlations of returns will not vanish and may cause realised values of both estimators to differ. Figure 6 shows both correlations, calculated over the interval December 1, 2005December 1, 2009. While the asynchronous correlation is more noisy, which is expected as it involves an extra term in the covariance estimator, both correlations track each other well on average. As figure 2 shows for the case of the S&P 500/Stoxx 50 pair, r* can go above one. Another example is shown in figure 7 with the three-month running correlation between the Citigroup (New York) and Royal Bank of Scotland (RBS) (London) stocks. Are instances when r* > 1 an artefact of the estimator itself or do they have financial significance? n Theoretical bounds for r*. Consider a sample of successive returns of two asynchronous securities S1, S2 as shown in figure 4. The two ladder uprights represent the sequence of daily returns of S1 and S2 . In practice, one calculates covariances of returns by taking time averages, for example, r1ir2i = (1/N) SNr1jr2j, but let us here imagine that an infinite number of such 1 samples have been drawn, and that covariances are calculated as ensemble averages. Let us assume that returns have no serial correlation, so that the only non-vanishing correlations are symbolised by the pink (rA) and blue (rS) rungs. This is the case, for example, if returns are generated with the process used previously in the example of continuous processes observed asynchronously. How large can rS + rA be? The correlation matrix of returns is positive by construction. Which constraints on rS , rA does this positivity impose? Let us assume that both ladder uprights consist of N segments, with periodic boundary conditions we will later take the limit N and let us look for eigenvectors of the full correlation matrix. Using an ansatz familiar to solid-state physicists, let us assume that eigenvectors have components eji on the ith return of the higher upright and components aeji for the ith return of the

6 three-month eWMa correlation of stoxx 50 and s&P 500 futures


120 100 80 60 40 20 0 Dec 2, 05 %

FO R

PR

NO T

OD UC
120 100 80 60 40 20 0 Dec 1, 04 Dec 1, 05

Dec 2, 06

TI
Dec 2, 07

Note: evaluated asynchronously using the 4pm quote for the S&P 500 and the 5:30pm quote of the Stoxx 50 (light blue), and synchronously using simultaneous quotes for both futures at 4pm Paris time (dark blue)

7 three-month eWMa correlations for the RBs (London) and citigroup (new york) stocks: rS (blue), rA (pink), r* (green)

Dec 1, 06 Dec 1, 07 Nov 30, 08 Nov 30, 09

lower upright, where j = 1, and let us denote by l the associated eigenvalue. Each return is correlated with itself and two returns of the other security. The equations expressing that l is an eigenvalue read:
S + + e j A = S +1+ e j A =

and yield the following expression for l:

ON
Dec 1, 08 Dec 1, 09

Periodic boundary conditions impose that = (2n)/N, where n = 0, ... , 1. We get the 2N eigenvalues of the global 2N 2N correlation matrix. Let us now take the limit N . The eigenvalues of the full correlation matrix are then given by expression (6) where [0, 2]. As the correlation matrix is positive, l 0. (rS + rA cos )2 + r2 sin 2 is extremal for = A 0, . For these values, l = 1 |rS rA|. This yields the following conditions on rS , rA:
risk-magazine.net

= 1

(S + A cos)

+ 2 sin 2 A

(6)

81

cutting edge. equity deRivatives

8 Paths of the two stocks used in the back-test of the basket option
120

80 Paris stock Tokyo stock 0 50 100 150 200

60

9 Final P&L of a short delta-hedged basket option (green) and initial price of the option (blue) as a function of correlation
8 6 P&L/price (%) 4 2 0 Final P&L Initial option price

2 4 60 80

The first condition shows that r* = rS + rA is restricted to the interval [1, 1]. r* may exceed one only if serial correlations are present. This argument strictly holds for covariances calculated as ensemble averages and very long samples. In practice, the sample has finite length and covariances are calculated as time averages. However, the practical consequence of this is that instances when r* exceeds one are not an artefact of our estimator and may instead indicate periods when returns are serially correlated. How will this affect a trading desk? As expression (2) for the daily P&L shows, it will materialise as cross-gamma P&L a correlation larger than one. Even though it may have priced-in the correct realised volatility for both securities and used an implied correlation of 100%, a trading desk with a short correlation position would have lost money something not possible in the synchronous case. We now illustrate this with a simple example. n An example with a basket option. Imagine we have sold a sixmonth basket option on two stocks, one trading, say, in Tokyo, the other in Paris. We delta-hedge the Japanese stock at the market close in Tokyo and the French stock at the close in Paris. The payout is given by:

NO T

FO

1 S + A 1 1 S A 1

100 Correlation (%)

120

140

1 2 1 + 2 + 21 2 2 2 Nothing prevents us from using values of r larger than one for risk-managing this basket option. To illustrate the effect we would like to highlight, imagine that, over the life of the option, S1 and S2 have trending behaviour, with both stocks moving by, say, +1% daily. The effect of serial correlation is then maximal and we get a realised value of r* = 200%. We now consider a stylised case that is somewhat more realistic, with intermittent trending behaviour. Figure 8 shows the path of the two basket components over the lifetime of the option. They have a short-term trend and hence their returns will be serially correlated with no appreciable longterm drift. The realised volatilities over the options lifetime are, respectively: 21.8% for S1 and 23.6% for S2 . The realised correlations are rS = 63.3% and rA = 57.6%, which gives a realised value for correlation r* = 121%. We now assume that we have sold the basket option and delta-hedge it until maturity, using a Black-Scholes model, with zero interest rates. We use the exact realised volatilities for implied volatilities s1 and s2 . Our final P&L is very plainly the initial premium minus the final payout plus the sum of delta P&Ls from the daily rehedges on both securities, which has the form SDi(Si+1 Si). It is shown in figure 9 as a function of the correlation level r used for calculating prices and deltas, along with the initial price of the basket option. First note that, as we would expect since the two stocks have no long-term trend, the curves representing the initial price charged for the option and the final P&L are roughly parallel. Then notice that the final P&L vanishes for r around 125% close to r* = 121% and that, had we initially priced the option with r = 100% and the exact realised volatilities, we would have lost money. =
Conclusion

RE

82

Risk November 2010

PR

OD UC

In this article, we show it is possible to price and risk-manage derivatives in the standard synchronous framework, in situations where underlying securities trade on asynchronous markets, provided one uses a special correlation estimator, which we derive using the criterion that the carry P&L vanishes on average. Our estimator quantifies the correlation that is materialised as the cross-gamma P&L of a delta-hedged option position. We highlight the difference between the fair correlation of options and that of correlation swaps and show that, in an asynchronous context, serial correlation of returns may push realised correlation above one. A short-correlation option position will lose money even though one has used implied volatilities that match future realised volatilities and a 100% correlation parameter. n
Lorenzo Bergomi is head of quantitative research in the global markets division at socit gnrale. He would like to thank members of his team for useful discussions and suggestions. email: lorenzo.bergomi@sgcib.com

TI

ON

100

S S + 1,T 2,T 1 S S 1, 0 2, 0 We choose the geometric over the usual arithmetic average as we have closed-form formulas for prices and deltas in the BlackScholes model. The geometric average of S1 and S2 is lognormal, with a volatility given by:

You might also like