You are on page 1of 27

Journal of International Money and Finance 31 (2012) 1033–1059

Contents lists available at SciVerse ScienceDirect

Journal of International Money


and Finance
journal homepage: www.elsevier.com/locate/jimf

Exchange rate bubbles: Fundamental value estimation and


rational expectations test
Wilfredo L. Maldonado a,1, Octávio A.F. Tourinho b, *, Marcos Valli c, 2
a
Universidade Católica de Brasília, SGAN 916, Módulo B, Brasília, DF 70790-160, Brazil
b
Universidade do Estado do Rio de Janeiro, Rua São Francisco Xavier 524, Bloco F, Sala 8039, Rio de Janeiro, RJ 20550-013, Brazil
c
Banco Central do Brasil, SBS, Quadra 3 Bloco B, Brasília, DF 70074-900, Brazil

a b s t r a c t

JEL classification: We propose a model of periodically collapsing bubbles which


E32 extends the Van Norden (1996) model, and nests it, by considering
C22 a non-linear specification for the bubble size in the survival regime,
Keywords: and the endogenous determination of the level of the fundamental
Exchange rate bubbles value of the stochastic process. They allow us to test for rationality in
Regime-switching regression the formation of expectations, and remove the arbitrariness of
exogenously setting the level of the fundamental value. This general
model is applied to the exchange rate of the Brazilian real to the US
dollar from March 1999 to February 2011. The futures market
exchange rate is used as a proxy of its expected future value, and
three different structural models are considered for the determina-
tion of the fundamental value. The first two imply that the exchange
rate satisfies either purchasing power parity (PPP), or a modified
version of it. The third structural model is a version of the monetary
model of exchange rate determination, fitted to the period under
consideration. We obtain the maximum likelihood estimate of the
parameters of the models, explore the properties of the errors, test its
restricted versions, and compare the three specifications for the
fundamental. We find that the models we propose fit well the data,
and are useful in the heuristic interpretation of the exchange rate
movements of the period. Finally we select the structural models that
display the best performance, according to several criteria.
 2011 Elsevier Ltd. All rights reserved.

* Corresponding author. Tel.: þ55 21 2334 0172.


E-mail addresses: wilfredo@pos.ucb.br (W.L. Maldonado), octavio.tourinho@terra.com.br (O.A.F. Tourinho), marcos.valli@
bcb.gov.br (M. Valli).
1
Tel.: þ55 61 3448 7135.
2
Tel.: þ55 61 3414 2351.

0261-5606/$ – see front matter  2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jimonfin.2011.12.009
1034 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

1. Introduction

Brazil adopted a floating exchange rate regime in February of 1999, and the Brazilian real to US
dollar exchange rate has experienced wide fluctuations since then, in spite of the fact that the Brazilian
economy exhibited a relatively stable behavior throughout that period. This suggests that to charac-
terize the trajectory of that exchange rate one should use a formulation that is able to account for the
occurrence of large deviations, as well as for some mean-reversal behavior. Of course, it must also take
into account the asset nature of foreign exchange stocks, as well as its role in pricing tradable goods
relative to non-tradable ones. This can be done with a model that allows for the occurrence of bubbles.
It is assumed that in equilibrium the exchange rate is the solution of a structural system of equations
that has two components: the fundamental value and the bubble. Tirole (1982, 1985) and Blanchard
and Fisher (1989) present several simple macroeconomic models that produce bubbles. Some alter-
natives can be considered regarding the stochastic specification of the bubble. In Obstfeld and Rogoff
(1983) and Engsted (1993) the bubble only disappears when the system itself ceases to exist. Blanchard
(1979) and Blanchard and Watson (1982) consider a more flexible specification with an exogenous
probability of the bubble collapsing. Several tests for his type of bubble have been proposed in the
literature, and have been applied to study foreign exchange and stock markets, as indicated below. Diba
and Grossman (1988b) discuss the inception and survival of bubbles.
Evans (1986) proposes a nonparametric procedure that takes into account the possibility of
occurrence of more than one bubble, each covering part of the sample period, and finds evidence of
a negative bubble in the excess return to holding sterling rather than dollar assets during 1981–1984.
Meese (1986) considers the occurrence of bubbles in a monetary model of the exchange rate. He rejects
the hypothesis of inexistence of bubbles in the dollar to deutsche mark and in the dollar to pound
exchange rates over the period 1973–1982. West (1987) proposes a parametric approach based on the
Hausman specification test, and rejects the null hypothesis of no bubbles in the Standard and Poor’s
500 index (1871–1980) and the Dow Jones index (1928–1978). Diba and Grossman (1988a) propose
testing the stock price and the dividend series for the presence unit roots, and for cointegration
between them, to test for bubbles. They conclude that the empirical evidence is inconsistent with the
existence of an explosive rational bubble in prices.
Some of these models have a restrictive feature which may be undesirable in some contexts: they
assume that the bubble increases indefinitely until some exogenous structural change occurs. This
limitation can be removed by specifying a regime-switching model for the bubble, which can be
appropriate if agents contemplate the existence of two alternative dynamics of the bubble size, and
incorporate the possibility of a regime change in their expectations. In this formulation the bubble is
either: (i) collapsing, and its expected size is decreasing, or (ii) it survives, and its expected size is
increasing. Some examples of this approach, which is also followed in the model proposed here, are
discussed below.
Evans (1991) is the standard reference for a bubble that collapses periodically and switches its
growth regime, depending on its size. Whenever it exceeds a certain value, it enters a regime where
a stochastic variable with a Bernoulli distribution shifts its local behavior between two trajectories: it
can either grow faster, or can it collapse, reverting to a mean value.3 Van Norden and Schaller (1993)
provide an alternative specification of the periodically collapsing bubble, where the bubble regime is
a non-observable stochastic binary variable whose probability of occurrence is a function of the bubble
size. In addition to the different specification of the bubble, these models display another crucial
difference: in Evans (1991) the regime is observable, while in Van Norden and Schaller (1993) it is not.
Since the relation between the innovation and the bubble size is determined by the regime, the later
model allows for a more flexible interaction between them.
Van Norden and Schaller (1993) apply their model to the Toronto stock market index from January
1956 to November 1989, and find evidence of regime switches in stock market returns that are
influenced by apparent deviations from fundamentals. Van Norden (1996) used this model to model

3
Evans (1991) points out that, when applied to this type of bubble, the Diba and Grossman (1988b) unit root based test
would lead to incorrect conclusions regarding the presence of a bubble. However, Charemza and Deadman (1995) later rejoin
this discussion, and propose a unit root based test of explosive regime-switching bubbles.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1035

the dynamics of speculative bubbles in the exchange rate between the US dollar and three other major
currencies, and finds mixed evidence regarding the occurrence of regime-switching bubbles of that
type. Several other studies test for the presence of regime-switching bubbles in a variety of data sets.
For example, Funke et al. (1994) examine the inflation rate in Poland, Roche (2001) analyzes the price of
land in Iowa, and Brooks and Katsaris (2005) study the S&P 500, with this type of model.
A related body of literature employs switching models with Markovian state dependence
(Hamilton, 1994). This type of regime switching is not considered in the models mentioned above, and
substantially complicates the estimation because it requires the use of more than two states, as
indicated by Evans (1996).
The model proposed here generalizes the Van Norden (1996) formulation of two regimes for the
bubble component by considering the possibility of a non-linear relation between the innovation and
the bubble size in the survival regime. This allows us to test for rational expectations, using the
procedures proposed by White (1982, 1987) and Hamilton (1996). We also include the endogenous
estimation of the reference level of the fundamental value, which is exogenous in his model.
We apply our model to study the occurrence of regime-switching bubbles in the Brazilian real to US
dollar exchange rate on monthly data extending from March 1999 to February 2011. The chosen period
starts with the adoption of the floating exchange rate regime, and ends at the last month for which the
data was available when the empirical analysis reported here was performed. It is sufficiently long to
permit the identification of bubbles, but is not long enough to raise questions regarding the possibility
of significant structural shifts within the sample. It is reasonable to specify a regime-switching bubble
for the stochastic process of the Brazilian exchange rate in that period because it was determined in
a context where there existed a governmental authority (the Central Bank) which frequently inter-
vened in the market.
To define the fundamental value and the bubble component of the exchange rate we consider three
structural models. Models I and II are based on the application of the inter-temporal non-arbitrage
principle to the investment decision of firms engaged in international trade. They imply that the
fundamental exchange rate satisfies one of two dynamic pricing rules: it either maintains purchasing
power parity (PPP) between the Brazilian real and the US dollar (Model I), or it satisfies a modified
version of that principle, which accounts for the interest rate differential of Brazilian government bonds
traded domestically and abroad (Model II). Model II can be interpreted as adding the uncovered interest
rate parity hypothesis to the PPP hypothesis of Model I. In Model III the dynamic pricing rule of the
fundamental exchange rate is obtained from a suitably modified version of the monetary model of
exchange rate determination proposed in Meese (1986). It essentially adds a monetary equation to
Model II.
The evidence regarding the presence of regime-switching bubbles of the type we postulate is
different in each of the models indicated above. This is to be expected since, as widely pointed out in
the literature, there is generally a way in which bubbles can be reinterpreted in terms of the process
driving fundamentals.
The paper is divided into five sections. Section 2 describes the structural equation for the stochastic
process that has solutions with bubbles. It also presents the regime-switching model that will describe
the dynamics of the innovations to the process, and its relation to the bubble size. In Section 3, we
detail the empirical model, specifying the functions chosen to characterize the behavior of the bubble
component of the exchange rate, describe the data, discuss the estimation and testing procedures, and
present the results. In Section 4 we make a heuristic evaluation of the results, contrasting the several
models we test with the actual dynamics of the Brazilian exchange rate. In Section 5, we summarize the
main conclusions. Appendix 1 shows the derivation of the three structural equations used to calculate
the fundamental values for the exchange rate and Appendix 2 describes the formulae used for
statistical inference.

2. The stochastic process with bubbles

In macro dynamic models it is common to describe the trajectory of a variable as depending on the
expectations of its future value, and on the values of other exogenous variables. The simplest linear
model with such a structure is described below, and is the basis of our model.
1036 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Let Yt be the stochastic process for the asset value, and assume the system that governs its price
evolves according to:
e
Yt ¼ aYtþ1 þ f ðXt Þ (1)

where t denotes the time period, Yt is the value of the endogenous variable, and Xt is a vector of
e
exogenous variables, Ytþ1 is the expected value of the endogenous variable for period t þ 1, a ˛ (0,1) is
a constant, and f is a well behaved function of the exogenous variables that reflects the equilibrium
conditions of our system, the absence of riskless arbitrage opportunities, and any rationality conditions
that are imposed on the behavior of the agents.
The rational expectations hypothesis requires that Ytþ1 e ¼ E½Ytþ1 jJt , where Jt is the s-algebra
generated by {Xt,Yt,Xt1,Yt1,.}. Henceforth, we will denote E½$jJt  as Et ½$. Therefore, the rational
expectations solution to (1) is a stochastic process (Yt)t0 that satisfies
Yt ¼ aEt ½Ytþ1  þ f ðXt Þ (2)
Any solution of (2) may be written as a sum of two components: the fundamental value, which only
depends on the exogenous variables and their expected values, and the bubble component, which is
equal to the present value of the expected long-run level of the endogenous variable. To prove this,
solve equation (2) forward in time to obtain the following decomposition of the endogenous variable:
X
þN
Yt ¼ ai Et ½f ðXtþi Þ þ lim aT Et ½YtþT  (3)
T/þN
i¼0

where the first term in the right hand side (r.h.s.) is the fundamental value, denoted Yt and defined by:
X
þN
Yt ¼ ai Et ½f ðXtþi Þ (4)
i¼0

Noting that the value of the exogenous variables at time t is known, and using the law of iterated
expectations, the fundamental value can be written as:
X
þN   
Yt ¼ ai Et ½f ðXtþi Þ þ f ðXt Þ ¼ aEt Ytþ1 þ f ðXt Þ (5)
i¼1

Letting Jt be the s-algebra generated by fXt ; Yt ; Xt1 ; Yt1  ; .g note that J 4J implies
t t
E½E½$jJt jJt  ¼ E½$jJt . Therefore, taking the expectation of (5) conditional on Jt shows that funda-
mental value Yt is also a solution of equation (2), since Yt ¼ aE½Ytþ1  jJ  þ f ðX Þ. This is denoted by
t t
calling it the fundamental solution associated with (Yt)t0.
The difference between the solution and the fundamental value is the bubble component. It is equal
to the last term in the r.h.s. of (3), assuming that the limit exists4:

bt ¼ Yt  Yt ¼ lim aT Et ½YtþT  (6)


T/þN

The dynamic equation for the bubble size can be obtained by substituting (2) and (5) in (6):

bt ¼ aEt ½btþ1  (7)


Note from (7) that the bubble size is governed by the same type of expectations mechanism present
in the asset price, but it is not affected by the exogenous variables. Therefore, it isolates the speculative
mechanism, which functions autonomously: its current size depends only on its expected size next
period. This dependence, however, is not unrestricted, since (7) is a linear function, and the parameter
a is exogenous and constant. Since it is assumed that a ˛ (0,1), the bubble size will be expected to
increase in the next period if the current bubble size is not zero.

4
Regarding a as a discount factor, the hypothesis that the limit exists is the conventional specification of the transversality
condition for the dynamic optimization of the agents’ behavior, and can be interpreted as stating that the present value of the
terminal asset stock is finite, when the horizon is indefinitely extended.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1037

Following Blanchard (1979) and Blanchard and Watson (1982), assume that in each period one of
two events occurs in the bubble dynamics: either collapse (C) or survival (S). These events will
determine the current regime of the bubble. However, since by hypothesis the regime is not observable,
only a probabilistic statement can be made about it. This can be formalized by letting (rt)t0 be the
stochastic process generated by the occurrence of those events, defined as rt ¼ 1 if S occurs, and rt ¼ 0 if
C occurs, in period t. Let qtþ1 be the probability of that the bubble survives in period t þ 1, given the
information available in period t: qtþ1 ¼ Pðrtþ1 ¼ 1jJt Þ.
The two assumptions below characterize the dynamics of the bubble and of the fundamental value,
specifying how they depend on the bubble’s regime. Loosely speaking, the first hypothesis says that the
size of the bubble itself summarizes all the information required to determine the endogenous aspects
of the bubble dynamics, and the second hypothesis makes precise the sense in which the fundamental
value is unaffected by the bubble’s regime.
Assumption 1. There exist two functions u:R / R and q:R / [0,1] that satisfy equations (8) and (9):
Et ½btþ1 jrtþ1 ¼ 0 ¼ uðbt Þ (8)

qtþ1 ¼ qðbt Þ (9)

Equation (8) asserts that if, in addition to the information in Jt , it is known that collapse will occur
in period t þ 1, then the expected bubble size only depends on the current bubble size. Equation (9)
implies that probability of survival of the bubble in period t þ 1 is independent of all the variables
in Jt , and depends only on bt.
The expected bubble size in the next period if it survives (equation (10)) can be calculated by
using (7)–(9) to substitute for the expectations terms in the following identity:
Et ½btþ1  ¼ qtþ1 Et ½btþ1 jrtþ1 ¼ 1 þ ð1  qtþ1 ÞEt ½btþ1 jrtþ1 ¼ 0.

bt uðbt Þ
Et ½btþ1 jrtþ1 ¼ 1 ¼  þ uðbt Þ (10)
aqðbt Þ qðbt Þ
As can be seen in equation (6), the bubble is not observable because the fundamental value is
unobservable. However it is feasible to infer from the data the size of the innovations to the asset value,
Rt defined by:

Rtþ1 ¼ Ytþ1  Et ½Ytþ1  (11)


The innovation is closely related to the bubble, as shown by equation (12), obtained by letting

Ytþ1 ¼ Ytþ1 þ btþ1 in equation (11) and using equation (7).
bt
Rtþ1 ¼ Rtþ1 þ btþ1  (12)
a
where Rtþ1 is the innovation in the fundamental solution, i.e. Rtþ1 ¼ Ytþ1
  .
 Et ½Ytþ1
The expected value of the innovation, conditioned on the bubble collapse and survival, is calculated
as:
   bt
Et ½Rtþ1 jrtþ1 ¼ 0 ¼ Et Rtþ1 rtþ1 ¼ 0 þ uðbt Þ  (13a)
a
  
   1 bt
Et ½Rtþ1 jrtþ1 ¼ 1 ¼ Et Rtþ1 rtþ1 ¼ 1 þ 1  uðbt Þ (13b)
qðbt Þ a

Assumption 2. The current forecast of the next period fundamental value is not affected by the
knowledge of regime of the bubble in the next period:
     
Et Ytþ1 rtþ1 ¼ Et Y  ; for all t (14)
tþ1

Noting that Assumption 2 implies that, conditional to either regime, the expected value of the
innovation in the fundamental solution is null, (13a) and (13b) can be written as follows:
1038 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

bt
Et ½Rtþ1 jrtþ1 ¼ 0 ¼ uðbt Þ  (15a)
a
  
1 bt
Et ½Rtþ1 jrtþ1 ¼ 1 ¼ 1  uðbt Þ (15b)
qðbt Þ a

3. The empirical model

To estimate equations (15a) and (15b), and recover the parameters that characterize the bubble
dynamics, it is necessary to further restrict the model, and specify functional forms for the functions q
and u.
Following Van Norden (1996), we use a logit function for the probability of survival:
h  i1
qðbÞ ¼ 1 þ exp bq0 þ bq2 b2 ; bq2 > 0 (16)

Note, however, that unlike his formulation, the exponent of the exponential function in (16) does
not include a linear term.5 The fact that the probability of survival is decreasing in the bubble size i.e.
ðdq=djbjÞ < 0, controls the explosive behavior of the bubble.
Noting that in (15a) and (15b) the expected value of the innovation is linear with respect to u, we
adopt a linear specification for the function u(bt). When used in (15a), it yields (17a). Expanding the
r.h.s. of (15b), and using (16) and (17a), yields the following: Et ½Rtþ1 jrtþ1 ¼ 1 ¼
expðbq0 Þexpðbq2 b2t ÞðbC0  bC1 bt Þ. Equation (17b) is a generalization of this expression, since it can be
obtained by setting bS0 ¼ bC0 expðbq0 Þ; bS1 ¼ bC1 expðbq0 Þ; bS00 ¼ 0 and bp2 ¼ bq2 in (17b). The
main advantage of this more general formulation is allowing us to test for rational expectations in the
futures market for foreign exchange.6

Et ½Rtþ1 jrtþ1 ¼ 0 ¼ bC0 þ bC1 bt (17a)

 
Et ½Rtþ1 jrtþ1 ¼ 1 ¼ bS00 þ ðbS0 þ bS1 bt Þexp bp2 b2t (17b)

This derivation is exact and, unlike Van Norden’s (1996) formulation, does not require the log-
linearization of the system resulting from the combination of (15) and (16). Our empirical imple-
mentation nests his, and his model can be tested against ours by testing for bp2 ¼ 0. If bp2 s 0, it is also
possible to test the rational expectations hypothesis, a verification that is impossible in his model.7
System (17) is written in terms of the expected innovation. To obtain an empirical model, we
assume that, conditional on the regime that will prevail in t þ 1, the deviations of the innovation from
its expected value are independent, homoscedastic, serially uncorrelated, normally distributed
stochastic errors8:
 
C C
3 tþ1 ¼ Rtþ1  Et ½Rtþ1 jrtþ1 ¼ 0; 3t wN 0; s2C (18a)

 
S S
3 tþ1 ¼ Rtþ1  Et ½Rtþ1 jrtþ1 ¼ 1; 3 t wN 0; s2S (18b)

5
The linear term in the exponential for the logit distribution of the probability of survival appears in Van Norden (1996)
when he log-linearizes the bubble. In our approach we do not approximate the model so that term does not appear.
6
Note that what we are considering here is a rather specific failure of rational expectations that could possibly be described
as a form of limited rationality, because agents are only unable to fully recognize the nature of the dependence of the size of the
innovation if the bubble survives on probability of survival.
7
To preserve identification of the model and the ability to distinguish between bS0 and bS00, we always include the
hypothesis bS00 ¼ 0 when the test considers the possibility that bp2 ¼ 0.
8
The hypothesis of homoscedasticity and lack of serial autocorrelation of the errors are tested when the model is estimated.
The normality of the errors cannot be tested because the regime-dependent errors are not observable. Nevertheless, we test for
normality of the observed average error using the JB test, and note that it is rejected in all models, due observations that appear
to be outliers.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1039

These errors can be interpreted as noise, admitting that all the relevant information present in the
innovation is contained in its expected value. Further, in each regime, they are equal to the innovation
to Ytþ1, computed under the assumption that, in addition to Jt , the information that the regime will
prevail in t þ 1 is also available. To show this, use the definition Rtþ1 ¼ Ytþ1  Et[Ytþ1] in (18), and apply
the law of iterated expectations to obtain:

C
3 tþ1 ¼ Ytþ1  Et ½Ytþ1 jrtþ1 ¼ 0 (19a)

S
3 tþ1 ¼ Ytþ1  Et ½Ytþ1 jrtþ1 ¼ 1 (19b)

The average error 3 tþ1 is a convex combination of 3 Ctþ1 and 3 Stþ1 , and is equal to the innovation to Ytþ1,
as follows:

C S
3 tþ1 hð1  qðbt ÞÞ3 tþ1 þ qðbt Þ3 tþ1 ¼ Ytþ1  fð1  qðbt ÞÞEt ½Ytþ1 jrtþ1 ¼ 0 þ qðbt ÞEt ½Ytþ1 jrtþ1 ¼ 1g
¼ Ytþ1  Et ½Ytþ1 
(20)
This implies that Y is in fact an AR(1) process, where the average error is the innovation. This can be
shown writing the structural model of equation (2) as Et[Ytþ1] ¼ a1Yt  a1f(Xt), adding 3 tþ1 to both
sides of this equation, and using (20):

Ytþ1 ¼ a1 Yt  a1 f ðXt Þ þ 3 tþ1 (21)


The regime-switching equations are obtained by substituting for the expectation terms in (17) from
(18):
(
bC0 þ bC1 bt þ 3 Ctþ1 ;  
Prob ¼ 1  qðbt Þ ðaÞ
Rtþ1 ¼ (22)
bS00 þ ðbS0 þ bS1 bt Þexp bp2 b2t þ 3 Stþ1 ; Prob ¼ qðbt Þ ðbÞ

We obtain maximum likelihood estimates for the parameters in (22), and use the Likelihood Ratio
test to verify the following hypotheses.9

C Rational expectations hypothesis

To test if expectations regarding the innovations are rational we verify if the following conditions
are satisfied by the estimated parameters.10

bS00 ¼ 0; bp2 ¼ bq2 ; bS0 þ bC0 exp bq0 ¼ 0; bS1 þ bC1 exp bq0 ¼ 0 (23)

C Two linear regimes

The empirical implementation in Van Norden (1996) assumes that the innovation is a linear
function of the bubble size in both regimes. We test if the more general model proposed here better
supports our data by testing the conditions:
bp2 ¼ 0; bS00 ¼ 0 (24)

We also reproduce the following tests proposed by Van Norden (1996):

C Two linear regimes with the same slope

9
Goldfeld and Quant (1973) and Kiefer (1978) proved that maximum likelihood estimators for this type of model are
consistent and efficient. See Appendix 2 for the details on how the tests of the different hypothesis were implemented.
10
To test the rational expectations hypothesis we use (22a) and (22b), in the identity
Et ½Rtþ1  ¼ qðbt ÞEt ½Rtþ1 jrtþ1 ¼ 1 þ ð1  qðbt ÞÞEt ½Rtþ1 jrtþ1 ¼ 0, and test Et ½Rtþ1  ¼ 0; cbt .
1040 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Test if the impact in the innovation of changes in the bubble size is the same in both regimes.

bp2 ¼ 0; bS00 ¼ 0; bC1 ¼ bS1 (25)

C Linear regression model

Test if the single regime stochastic process fits the data better.
bp2 ¼ 0; bS00 ¼ 0; bC1 ¼ bS1 ; bC0 ¼ bS0 (26)

C Normal mixture model

Test if distribution of the innovations is a mixture of normal random variables that occur with
probabilities q and 1  q, as indicated in (22), with regime probabilities that are independent of the
bubble size.
bp2 ¼ 0; bS00 ¼ 0; bC1 ¼ bS1 ¼ 0; bq2 ¼ 0 (27)

C Restricted normal mixture model

Test a special case of the previous hypothesis: having just one regime that is independent of the
bubble.
bp2 ¼ 0; bS00 ¼ 0; bC1 ¼ bS1 ¼ 0; bq2 ¼ 0; bC0 ¼ bS0 (28)

3.1. Data

In this section we describe the data used to estimate de model described above for the Brazilian real
to US dollar exchange rate, report the values of the estimated parameters, and discuss the results of the
statistical inference analysis. We denote this particular time series St, to emphasize that this is an
application of the general model described in the previous section, where the time series was
denoted Yt.
For all variables we use monthly data from March 1999 to February 2011. The data on prices,
production and money stocks has been seasonally adjusted, but the exchange and interest rates have
been used in their original form. The sample period was constrained by the availability of the proxy
variable for the innovations, as discussed below.
For the spot exchange rate (St) we use “sell” price of the US dollar on the first day of the month, as
reported by the Central Bank of Brazil. We obtain the one month forward exchange rate (Ft) from data
of the Brazilian Mercantile Exchange database (BM&F of Brazil), and use it as a proxy for the expected
value of St, i.e. Et[Stþ1] ¼ Ft. Therefore, the innovation is: Rtþ1 ¼ Stþ1  Ft.
To calculate the bubble it is necessary to specify the fundamental value St . We consider three
structural models to determine it. Model I assumes only the PPP (Purchasing Power Parity) hypothesis.
Model II is an extension of Model I that takes into account the interest rate differential for the Brazilian
government debt in the domestic and international markets. Model III is an extension of the Meese
(1986) monetary model for the exchange rate. It essentially adds a monetary equation to Model II.
Appendix 1 justifies the use of Models I and II as the rational pricing relation if firms arbitrage their
investment opportunities in the country and abroad, and develops Model III in detail.
For Model I, St is determined by (29), where PDt and PFt are the domestic and foreign price levels.
PDt
St;I ¼ (29)
PF t
In this case the bubble size is given by (30), where PDI and PFI are domestic and foreign price
indexes, and k is a constant that must be estimated in the maximization of the likelihood function.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1041

PDIt
bt;I ¼ St  k (30)
PFI t
For the Model II, St is given by:

1 þ IF 0t PDtþ1
St;II ¼ Et (31)
ð1 þ IDt Þ PF tþ1
where ID and IF0 are the nominal interest rates observed in the domestic and foreign markets for
Brazilian government bonds, respectively. The expected value of the next period’s price ratio is taken to
be the realized price ratio, assuming that expectations are confirmed. In this case the bubble size is
given by (32), where PDI and PFI are as before.

1 þ IF 0t PDItþ1
bt;II ¼ St  k (32)
ð1 þ IDt Þ PFI tþ1
In Models I and II the domestic price index (PDI) is the Producer Price Index (IPA-OG), as suggested
by the development in Appendix 1.11 We consider two versions (A and B) of the above models, each
corresponding to one of two alternatives for the foreign price index (PFI). For models of type A, it is the
Producer Price Index for the USA (US-PPI).12 For models type B, we take a trade-weighted index of the
producer price index of the more significant partners of Brazil in international trade in the sample
period.13 In Model III we used the same price indexes are as in Models IA and IIA, for comparability.
For ID we consider the domestic fixed interest rate for 30 days implied by traded contracts for
swapping fixed nominal interest rate bonds for indexed bonds whose return is post-fixed on the basis
of the realized very short term interest rate. For IF0 we use a proxy: the interest rate paid by the Bra-
zilian Government on its foreign debt, calculated as the sum of the interest rate of the U.S. Treasury Bills
(IF) and the average “spread” on Brazilian foreign debt, denoted IS. For the latter we use the monthly
equivalent rate of the interest rate spread of Brazilian debt over US treasuries. It is measured by the
Emerging Markets Bond Index (EMBI) calculated by JP Morgan as the average interest rate differential
between Brazilian Bonds and US treasuries of same duration, weighted by their debt volume.14 All
interest rates are specified in decimals, not in basis points, i.e. 1% per month is represented as 0.01. This
fact is important when comparing our estimates for the model coefficients with those in other
empirical studies.
Model III is presented in detail in Appendix 1, but its implications for the fundamental rate can be
summarized by pointing out that, given an initial value for the exchange rate (S0), the fundamental
exchange rate of the monetary model can be calculated from the following equations:

Dxt hðDmt  a1 Dyt þ a2 DISt Þ (33)

Dxt ¼ cDxt1 þ dt (34)


gc
Dst ¼ Dxt þ ðDxt  Dxt1 Þ (35)
1  gc

St ¼ exp st (36)

where st is the log of the spot exchange rate (St), ft is the log of the forward exchange rate (Ft), mt and yt
are the logs of the ratio of the domestic aggregate to the foreign aggregates for the nominal money
stock ðm ¼ lnðMD=MFÞÞ and for industrial production ðy ¼ lnðYD=YFÞÞ. The parameters a1 > 0 and
a2 > 0 are, respectively, the income elasticity and the interest semi-elasticity of real money balances,

11
The variable IPA-OG is the “Índice de Preços no Atacado – Oferta Global” calculated by Instituto Brasileiro de Geografia e
Estatística (IBGE), obtained from IPEAdata: http://www.ipea.gov.br.
12
As reported by the Federal Reserve Bank of St. Louis.
13
Specifically, we choose twenty countries among those with greater trade volume (imports plus exports) with Brazil. They
are responsible for 76% of the total trade with Brazil. It is available from the authors, on request.
14
This proxy variable for IF has a weakness: the duration of the long term instruments it considers is much longer than that of
short term government debt used in measuring ID. However, no other better measure was readily available.
1042 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

and g ¼ a2/(1 þ a2) is an auxiliary parameter. The market fundamentals process, denoted Dxt, is defined
by (33), and its autocorrelation is jcj < 1. Sample determination of the trajectory for the fundamental
exchange rate in this model is done as follows.
First, the variables of the model, m  p, y, ID  IF, IS, s  p, and s  f are tested for the presence of
a unit root, using the procedure suggested by Doldado et al. (1990) to address the issue of inclusion of
a constant and a time trend in the test equations, as described in Enders (1995). The number of lags
used in the tests is chosen by the Akaike information criterion.15 This test, at the 5% significance level, is
unable to reject the presence of a unit root with zero drift in m  p, y, IS and s  p, and barely rejects it
in ID  IF. However, the presence of a unit root in ID  IF is not rejected at the 10% significance level, and
inspection of the graph of that series suggests the presence of a unit root if the possibility of a “break” in
the time series is taken into account. This is confirmed by applying the Perron (1997) tests, which are
unable to reject the presence of a unit root at the 5% confidence level.16 Finally, the innovation in the log
of the exchange rate s  f is classified as stationary by using the methodology above. Nevertheless,
examination of the graph of the series indicates the presence of significant noise superimposed on
a stochastic trend. To handle this, the series was subjected to the Hodrick and Prescott (1980, 1997)
filter, and it was verified that the presence of unit root cannot be rejected in the filtered series.
Therefore, all the variables listed in the beginning of this paragraph are classified as I(1).
Second, the parameters of the real money balances demand equation are obtained by estimating the
cointegrating vector of m  p, y, and ID  IF, by regressing the first variable on the other two, via the
Stock and Watson (1993) dynamic OLS approach, using 2 lags. The estimated values are significantly
different from zero (standard errors in parenthesis): a1 ¼ 0.95895 (0.47277) and a2 ¼ 2.70749
(1.07285).17 These are, approximately, double and half (respectively) the values found in Ball (2001).
However, considering the variability of estimates documented in that survey and in Stock and Watson
(1993), they appear to be within the range of reasonable values for the US in the period considered in
our sample. The estimates for Brazil, obtained using the methodology in Tourinho (1995) are similar to
the values above. The auxiliary parameter g ¼ 0.73028 is calculated using its definition.
Third, the two other structural hypotheses of the monetary model are verified. The UIP hypothesis,
extended to include a country risk spread variable, is tested by estimating the cointegrating vector
between ID  IF, s  f, and IS. It is obtained by regressing the first variable on the other two, via the Stock
and Watson (1993) dynamic OLS approach, using 1 lag. The estimated values are, respectively: 1.323089
(0.917568) and 0.933017 (0.212349) (standard errors in parenthesis). The estimated coefficients are
close to 1, and the expanded UIP hypothesis is not rejected. The importance of including the interest rate
“spread”, which reflects the country risk and the foreign capital supply conditions, in the UIP is evident in
its relatively small standard error in the estimated equation. The PPP hypothesis, that s  p is a random
walk, is confirmed by the presence of a unit root in that series, already discussed above.
Fourth, the fundamental market process Dxt is calculated from (33), using the sample values of
Dmt ; Dyt and DISt . The parameter c ¼ 0.46646 is obtained by estimating (34) in first differences by OLS.
Fifth, using the values of the parameters discussed above and the time series of the market
fundamentals process, equation (35) can be used to calculate Dst . Using an initial value s0 we can use
equation (37) to obtain the following fundamental exchange rate:
! !
X
t X
t
St ¼ expðs0 Þ  exp Dsi ¼ k exp Dsi (37)
i¼1 i¼1
So for Model III the bubble is:
!
X
t
bt;III ¼ St  k exp Dsi (38)
i¼1

In all three models the estimation of k jointly with the parameters of the bubble process (equations
(30), (32) and (38), respectively), allows the endogenous estimation of the fundamental value of the
exchange rate. In Van Norden (1996) the value of k is calibrated in such a way that the average value of

15
Unit root testing was done using the URAUTO procedure of RATS 7.0 (Estima, 2007).
16
The results of the PERRON97 procedure or RATS 7.0 (Estima, 2007) are in Appendix B.
17
Estimation was performed using the SWDOLS standard procedure of RATS 7.0 (Estima, 2007).
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1043

the bubble over the sample period is null.18 Here we estimate it simultaneously with all the other
parameters within the likelihood maximization procedure.
In the next sections we describe the estimation and the tests of the five models we consider: IA, IB,
IIA, IIB, and III.

3.2. Estimation

Table 1 shows the maximum likelihood estimates of the parameters of the regime-switching
empirical equation (22) that relates the innovation to the bubble size, for each model, for their
unrestricted and restricted versions. The first two columns indicate the model and the version.
The third column contains the maximum value of the log-likelihood in each case.19 These are used
to test the hypothesis previously described with the likelihood ratio test. Looking at the unrestricted
version of the models, in the first line in each section of Table 1, it can be seen that for this metric the
ranking of the models is as follows: IA, IIA, IB, IIB, III. Therefore, model I, where the fundamental
exchange rate is specified by the pure PPP relation, dominates model II which uses an interest rate
adjusted PPP rule for that purpose. The worse performance is that of model III, where the fundamental
rate is determined by the monetary model.
This ranking is strengthened by noting that the best models are also the more parsimonious. The
variant A of models I and II dominates the variant B, i.e. the models that adopt as a foreign price index the
Producer Price Index for the USA (US-PPI) dominates those that use trade-weighted index of the
producer price index of the more significant partners of Brazil in international trade in the sample period.
The other columns of Table 1 show the values of the estimated parameters. Column 4 refers to
parameter k, which scales the reference trajectory of the fundamental exchange rate, which are
comparable between models.20 It varies from 0.51 to 1.44 across models and versions. The smallest
values are produced by model IB, and the largest values are produced by the restricted versions of
model III. Examining the unrestricted versions of the models, it can be seen that the variant B of models
I and II produces smaller estimated values for k. Generally speaking, the base trajectory of the
fundamental value has to be scaled down by a larger proportion whenever the weighted foreign price
index is used. However, for the same variant of the two models, the estimated k is similar, as can be
seen, for example, by comparing its value for models IA and IIA, equal to 0.87 and 0.9 respectively.
It is also interesting to note that the estimate of k in the models restricted to satisfying the rational
expectations hypothesis is significantly smaller than those obtained in the unrestricted models (an
average decrease of 30.8%). This means that the fundamental rate is smaller under rational expecta-
tions, and that the likelihood of occurrence of negative bubbles is much smaller in those restricted
models than in the unrestricted model.
Columns 5 and 6 show the estimated values of the parameters of the logit function q(b) for the
probability of the bubble collapsing in the next period (equation (16)). The plot of that function for the
unrestricted version of the models IIB and III is displayed in Fig. 1. For model III it shows that the
probability of survival is very close to one whenever jbj  1:2, i.e. until a large bubble develops. After
that occurs, and while the bubble does not burst, that probability drops steeply and becomes null for jbj
 3, signaling an imminent regime change. This situation was observed in 2002. For Model IIB the
probability of survival is at most 0.57, and declines steadily as the bubble size increases, becoming null
for jbj  2.
The columns 7–11 of Table 1 show the estimated values of the parameters for linear part of the
functions that relate the innovation to the bubble size in the two regimes (see equations (22)). The
parameters with the “C” subscript (bC0 and bC1) refer to the collapse regime, while those with the “S”
subscript (bS0 and bS1) refer to the survival regime. The parameters of the non-linear part of the
function for the survival regime (bp2 and bS00) are displayed in the last two columns of the table.

18
More precisely, the parameter which is calibrated in Van Norden (1996) is the analog of k in the log-linearization of the
model for equation (25).
19
The normalization factor of the likelihood function was omitted in the calculations; therefore the log-likelihood is not
necessarily negative. Since the sample is the same for all models, that statistic is comparable across them.
20
The base trajectory of the exchange rate starts at approximately R$1.2/US$ in march 1999.
1044
Table 1
Maximum likelihood estimates of the parameters.

Restrictions Scaled k Transition Regime S Regime C Rational expectation Constant


log-likelihood
bq0 bq2 bS0 bS1 sS bC0 bC1 sC bp2 bS00
Model IA Unrestricted model 148.66 0.87 (0.77) 1.43 0.06 (0.02) 0.050 0.01 (0.06) 0.106 1.64 (0.12)

W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059
Rational expectations model 143.77 0.60 (0.60) 0.89 (0.04) 0.04 0.038 0.07 (0.07) 0.111 0.89 –
Two linear regimes (Van Norden) 141.45 0.55 (11.56) 3.10 0.01 (0.04) 0.077 0.90 (0.37) 0.320 – –
Two linear regimes w/ same slope 140.95 0.77 (10.23) 4.81 (0.01) (0.02) 0.078 0.11 (0.02) 0.339 – –
Linear regression model 140.30 0.85 (9.62) 5.64 (0.02) (0.02) 0.078 (0.02) (0.02) 0.354 – –
Normal mixture model 138.01 0.85 2.52 – 0.14 – 0.358 (0.03) – 0.072 – –
Restricted normal mixture model 136.98 0.85 (2.53) – (0.03) – 0.072 (0.03) – 0.398 – –
Model IB Unrestricted model 146.79 0.76 0.07 1.40 0.25 (0.11) 0.034 0.01 (0.06) 0.097 1.15 (0.30)
Rational expectations model 144.73 0.51 0.84 0.56 (0.09) 0.09 0.008 0.04 (0.04) 0.096 0.56 –
Two linear regimes (Van Norden) 139.65 0.92 (6.04) 5.64 (0.02) (0.02) 0.076 0.24 (0.15) 0.321 – –
Two linear regimes w/ same slope 139.45 0.90 (6.16) 5.29 (0.02) (0.02) 0.077 0.09 (0.02) 0.327 – –
Linear regression model 139.12 0.59 (9.51) 3.14 0.01 (0.04) 0.077 0.01 (0.04) 0.357 – –
Normal mixture model 138.01 0.59 2.52 – 0.14 – 0.358 (0.03) – 0.072 – –
Restricted normal mixture model 136.98 0.59 (2.53) – (0.03) – 0.072 (0.03) – 0.398 – –
Model IIA Unrestricted model 148.65 0.90 (1.15) 1.92 0.02 (0.01) 0.050 0.02 (0.07) 0.115 2.56 (0.07)
Rational expectations model 144.75 0.60 (0.67) 0.94 (0.04) 0.03 0.036 0.07 (0.07) 0.110 0.94 –
Two linear regimes (Van Norden) 141.50 0.69 (11.71) 4.45 (0.00) (0.03) 0.078 0.82 (0.38) 0.323 – –
Two linear regimes w/ same slope 140.97 0.82 (10.30) 5.73 (0.01) (0.02) 0.078 0.10 (0.02) 0.337 – –
Linear regression model 140.33 0.87 (9.89) 6.49 (0.02) (0.02) 0.078 (0.02) (0.02) 0.352 – –
Normal mixture model 138.01 0.87 2.52 – 0.14 – 0.358 (0.03) – 0.072 – –
Restricted normal mixture model 136.98 0.87 (2.53) – (0.03) – 0.072 (0.03) – 0.398 – –
Model IIB Unrestricted model 145.08 0.72 (0.30) 1.37 0.01 (0.00) 0.037 0.02 (0.06) 0.106 2.49 (0.04)
Rational expectations model 143.62 0.55 (0.43) 0.91 (0.04) 0.04 0.033 0.07 (0.07) 0.106 0.91 –
Two linear regimes (Van Norden) 138.62 0.93 (7.43) 7.75 (0.02) (0.03) 0.077 0.34 (0.24) 0.322 – –
Two linear regimes w/ same slope 138.31 0.97 (5.99) 7.13 (0.02) (0.02) 0.076 0.09 (0.02) 0.321 – –
Linear regression model 137.76 0.63 (8.73) 3.11 0.01 (0.04) 0.077 0.01 (0.04) 0.358 – –
Normal mixture model 136.38 0.63 (2.52) – (0.03) – 0.072 0.14 – 0.359 – –
Restricted normal mixture model 135.35 0.63 (2.53) – (0.03) – 0.072 (0.03) – 0.399 – –
Model III Unrestricted model 143.10 0.84 (5.07) 1.16 0.08 (0.02) 0.077 2.36 (1.38) 0.224 (0.09) (0.09)
Rational expectations model 129.96 0.56 1.92 0.72 0.26 (0.11) 0.189 (0.04) 0.02 0.079 0.72 –
Two linear regimes (Van Norden) 143.08 0.83 (5.08) 1.15 (0.01) (0.02) 0.077 2.37 (1.38) 0.225 – –
Two linear regimes w/ same slope 138.85 1.44 2.51 0.15 0.15 (0.01) 0.364 (0.03) (0.01) 0.072 – –
Linear regression model 137.78 1.39 2.59 0.02 (0.03) (0.01) 0.403 (0.03) (0.01) 0.072 – –
Normal mixture model 138.01 1.32 2.52 – 0.14 – 0.358 (0.03) – 0.072 – –
Restricted normal mixture model 136.98 1.32 2.53 – (0.03) – 0.398 (0.03) – 0.072 – –
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1045

PROBABILITY OF SURVIVAL OF THE BUBBLE


q
1.0
0.8
0.6 Model III

0.4
Model IIB
0.2
b
3 2 1 1 2 3

Fig. 1. Probability of survival of the bubble.

Examining the coefficients for the linear part of the unrestricted models, it can be seen that their
magnitude is generally small, indicating a weak linear dependence of the innovation on the bubble size,
for both regimes. Model III an exception in this respect, because for it bC0 and bC1 are one order o
magnitude larger than in the other models. Broadly speaking, this occurs because in models I and II the
non-linear part of the function for the survival regime plays the most important part, as can be seen in the
large relative values for bp2 in the next-to-last column, while in model III that parameter is small. In other
words, the function for the survival regime in model III is approximately linear. This is confirmed by
comparing the lines for the “Unrestricted” and the “Two linear regimes” for model III in Table 1, and noting
that the estimated parameters are very similar and that their log-likelihood is also almost identical.
The estimated standard deviation of the errors in the two regimes (sS and sC), displayed in columns
9 and 12, shows some interesting patterns, which can be seen by looking at their values for the
unrestricted models. They are summarized in Table 2, where the models are listed in increasing order of
the standard deviation of the errors.
This ranking is different from the one obtained earlier by comparing the log-likelihood value, since
this metric indicates that the variant “B” of models I and II fits the data better, offering evidence in favor
of the superiority of the trade-weighted foreign price index. The dominance of model I over II is
maintained and, as previously found, the fit of model III to the data on the innovations on the exchange
rate is inferior to that of the other models. The information on the standard deviation of the errors,
however, provides some quantitative indication on its performance relative to model IB, since the
estimated standard deviation of its errors is about 2.2 times larger, in both regimes.

3.3. Tests

Table 3 shows the results of the likelihood ratio test of the several hypotheses discussed earlier, for
all five models. For Models I and II, the last column shows that all of them can be rejected at the usual
5% of confidence level, except the rational expectations hypothesis. Due to the results of these tests, in
what follows we test misspecification and make inferences on the parameters only for the unrestricted
version of these 4 models.
For Model III, all hypotheses, except Van Norden’s (1996) specification of two linear regimes, can be
rejected at the 10% confidence level. The rejection of rational expectations in Model III may be related
to the fact that its fundamental exchange rate is more informative than in the other models, a fact that
may have elicited the following important feature of this application on our bubble model: since we

Table 2
Scaled log-likelihood of the models.

Model Scaled log-likelihood sS sC


IB 146.79 0.034 0.097
IIB 145.08 0.037 0.106
IA 148.66 0.050 0.106
IIA 148.65 0.050 0.115
III 143.10 0.077 0.224
1046 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Table 3
Likelihood ratio tests.

Null hypothesis Restriction Unrestricted Restricted LR statistics p-Value (%)


rank (n) Log-L Log-L Chi2(n)
Model IA Rational expectations model 5.00 148.66 143.77 9.78 8.17%
Two linear regimes (Van Norden) 2.00 148.66 141.45 14.43 0.07%
Two linear regimes w/ same slope 3.00 148.66 140.95 15.42 0.15%
Linear regression model 5.00 148.66 140.30 16.73 0.51%
Normal mixture model 5.00 148.66 138.01 21.30 0.07%
Restricted normal mixture model 7.00 148.66 136.98 23.37 0.15%
Model IB Rational expectations model 5.00 146.79 144.73 4.12 53.18%
Two linear regimes (Van Norden) 2.00 146.79 139.65 14.29 0.08%
Two linear regimes w/ same slope 3.00 146.79 139.45 14.68 0.21%
Linear regression model 5.00 146.79 139.12 15.34 0.90%
Normal mixture model 5.00 146.79 138.01 17.56 0.36%
Restricted normal mixture model 7.00 146.79 136.98 19.63 0.64%
Model IIA Rational expectations model 5.00 148.65 144.75 7.79 16.80%
Two linear regimes (Van Norden) 2.00 148.65 141.50 14.30 0.08%
Two linear regimes w/ same slope 3.00 148.65 140.97 15.35 0.15%
Linear regression model 5.00 148.65 140.33 16.63 0.53%
Normal mixture model 5.00 148.65 138.01 21.27 0.07%
Restricted normal mixture model 7.00 148.65 136.98 23.34 0.15%
Model IIB Rational expectations model 5.00 145.08 143.62 2.93 71.11%
Two linear regimes (Van Norden) 2.00 145.08 138.62 12.93 0.16%
Two linear regimes w/ same slope 3.00 145.08 138.31 13.55 0.36%
Linear regression model 5.00 145.08 137.76 14.64 1.20%
Normal mixture model 5.00 145.08 136.38 17.41 0.38%
Restricted normal mixture model 7.00 145.08 135.35 19.47 0.68%
Model III Rational expectations model 5.00 143.10 129.96 26.27 0.01%
Two linear regimes (Van Norden) 2.00 143.10 143.08 0.03 98.62%
Two linear regimes w/ same slope 3.00 143.10 138.85 8.49 3.70%
Linear regression model 5.00 143.10 137.78 10.63 5.93%
Normal mixture model 5.00 143.10 138.01 10.17 7.06%
Restricted normal mixture model 7.00 143.10 136.98 12.24 9.30%

measure the innovations by a proxy variable produced in a derivatives market, we in fact test the joint
hypothesis of rational expectations in the spot and in the derivative markets for foreign exchange.
Therefore, the rejection of RE could be due to imperfections in it, or in the modeling of the relation
between the spot and futures market.
Table 4 displays the results of the significance test for the coefficients of equation (22) for the
unrestricted models.21 Greater significance (smaller p-value) of the coefficients indicates larger like-
lihood that the bubble dynamics is relevant in explaining the innovations to the exchange rate. Models
IA and IIB have sharply defined bubble dynamics, since four of the five parameters are significant.
Models IB, IIA and III have only one significant parameter, at the conventional 5% level. Model III has
two significant parameters at the less stringent 10% significance level.
The relevance of one of our extensions to Van Norden’s model, represented by including a non-
linear term in the bubble dynamics in the survival regime (represented by finding bp2 significantly
different from zero), is confirmed in Models IA and IIB. In the others we cannot exclude the possibility
that bp2 ¼ 0.
Following Van Norden (1996), we used the White (1987) method to explore the possibility of
misspecification of the error terms, testing for autocorrelation, heteroscedasticity, and for Markovian
state dependence in the two regimes (3 Ct and 3 St ). We also test the joint hypothesis that the errors satisfy
all these hypotheses, and refer to it as a test that the errors are “well behaved”.
To test for presence of an AR(1) error process, we have to detect serial correlation in the derivatives
of the likelihood function with respect to bS00 and bC0 in the survival and collapse regimes. Analogously,

21
We use the Quasi-Maximum Likelihood method (Hamilton, 1994) to estimate the standard errors.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1047

Table 4
Unrestricted models – estimates and tests of the coefficients.

Null hypothesis bq2 ¼ 0 bS1 ¼ 0 bC1 ¼ 0 bp2 ¼ 0 bS00 ¼ 0


Model IA Estimate 1.43 (0.02) (0.06) 1.64 (0.12)
Std. dev. 0.56 0.03 0.02 0.47 0.04
p-Value (%) 1.03% 47.55% 0.16% 0.05% 0.65%
Model IB Estimate 1.40 (0.11) (0.06) 1.15 (0.30)
Std. dev. 4.54 0.26 0.18 10.24 0.01
p-Value (%) 75.87% 66.93% 75.36% 91.09% 0.00%
Model IIA Estimate 1.92 (0.01) (0.07) 2.56 (0.07)
Std. dev. 3.44 3.07 0.03 2.02 0.28
p-Value (%) 57.66% 99.75% 2.36% 20.51% 80.94%
Model IIB Estimate 1.37 (0.00) (0.06) 2.49 (0.04)
Std. dev. 0.49 0.00 0.02 0.78 0.01
p-Value (%) 0.55% 55.90% 0.98% 0.14% 0.25%
Model III Estimate 1.16 (0.02) (1.38) (0.09) (0.09)
Std. dev. 0.63 0.04 0.33 0.92 0.34
p-Value (%) 6.53% 70.33% 0.00% 92.41% 78.98%

to test for ARCH(1) error process in each regime, we test for serial correlation between the derivatives
of that function with respect to sS and sC. To test if a state-dependent Markov-switching regime is more
appropriate to the data, we investigate if there exists first order correlation in the series defined by the
derivatives with respect to bq0.22 In these tests we adopt a confidence level of 1%, because of the size
distortion of finite samples.
The results of the specification tests are shown in Table 5. The p-values in the last column indicate
that: (i) there is evidence of AR(1) errors in the collapse regime for all models except IB, and in the
survival regime only for models IB and IIB, (ii) in both regimes there is no evidence of heteroscedastic
errors in any model, (iii) the Markov property can be rejected for model IB, (iv) all models, except IA, fail
the joint specification test.
The results of this section can be summarized as follows. Model IA is the best model because it
achieves the largest likelihood in the unrestricted model, rejects all the restricted models and accepts
rational expectation, displays sharply defined bubble dynamics, and passes the joint specification tests
on the errors. Model IIB comes in second in the overall ranking because it displays the second-lowest
standard deviation of the errors, in spite of placing third in terms of the value of the log-likelihood of
the unrestricted model, also rejects all the restricted models, accepts rational expectation, and displays
sharply defined bubble dynamics.

4. Heuristic evaluation of the results

The fit of models has been compared in the last section in terms of the estimated equation (22),
which relates the innovation to the bubble in the two regimes. However, there is another dimension
over which the models be contrasted: the extent to which they track the level of the exchange rate. To
discuss this, Figs. 2–6 display the sample values of the fundamental value of the exchange rate ðSt Þ, the
bubble (given by St  St ), and the probability of collapse (1  q(bt)) for each model.
Since the fundamental rate is determined endogenously in our approach, it is instructive to compare
its value to spot exchange rate at the end of the sample, which is equal to 1.66 in February of 2011. The
evidence regarding the presence or absence of a speculative bubble in recent data is important because
it may have policy implications.
The variant “A” of models I and II, that use the US-PPI as the foreign price, display a trajectory for the
fundamental value which implies a negative bubble of size approximately equal to 0.6 at that date. The

22
Recall that the model here displays is a particular regime-switching behavior, where the transition probability does not
depend on the state (equation (9), of hypothesis 1). We actually test for indications of Markovian dependence in the dynamics
of the states.
1048 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Table 5
Specification tests.

Specification Degrees of freedom (n) Value of statistic chi2(n) p-Value


Model IA AR(1) in the error for the survival regime (S) 1 0.81 36.68%
AR(1) in the error for the collapse regime (C) 1 8.44 0.37%
ARCH(1) in the error for the survival regime (S) 1 0.05 81.87%
ARCH(1) in the error for the collapse regime (C) 1 2.69 10.12%
Markovian effect 1 4.25 3.94%
Joint test 5 9.87 7.91%
Model IB AR(1) in the error for the survival regime (S) 1 10.06 0.15%
AR(1) in the error for the collapse regime (C) 1 5.11 2.38%
ARCH(1) in the error for the survival regime (S) 1 0.33 56.79%
ARCH(1) in the error for the collapse regime (C) 1 2.52 11.27%
Markovian effect 1 6.74 0.94%
Joint test 5 19.57 0.15%
Model IIA AR(1) in the error for the survival regime (S) 1 1.16 28.16%
AR(1) in the error for the collapse regime (C) 1 10.44 0.12%
ARCH(1) in the error for the survival regime (S) 1 1.04 30.79%
ARCH(1) in the error for the collapse regime (C) 1 2.27 13.16%
Markovian effect 1 0.93 33.44%
Joint test 5 14.50 1.27%
Model IIB AR(1) in the error for the survival regime (S) 1 11.44 0.07%
AR(1) in the error for the collapse regime (C) 1 7.17 0.74%
ARCH(1) in the error for the survival regime (S) 1 0.96 32.65%
ARCH(1) in the error for the collapse regime (C) 1 2.68 10.19%
Markovian effect 1 4.55 3.29%
Joint test 5 20.11 0.12%
Model IIB AR(1) in the error for the survival regime (S) 1 0.75 38.67%
AR(1) in the error for the collapse regime (C) 1 26.13 0.00%
ARCH(1) in the error for the survival regime (S) 1 2.64 10.40%
ARCH(1) in the error for the collapse regime (C) 1 0.23 63.05%
Markovian effect 1 0.81 36.77%
Joint test 5 31.12 0.00%

Note: Bold values indicate rejection of the tested hypothesis at the 1% significance level.

variant “B” of those models, that use a trade-weighted foreign price index, yields very small values for
the end-of-sample bubble. The bubble in Model III is virtually also null at that date. Therefore, the
models with the richer description of the fundamental exchange rate suggest that no significant
change of exchange rate policy is needed, since no major distortions (produced by rational bubbles) are
detected recently.
In a general sense, as can be seen in Fig. 6, the evolution of the fundamental exchange rate in model
III more closely follows that of the spot exchange rate, i.e., there exists a higher correlation between the
fundaments and the observations. This implies a linear adjustment between the innovation and the
bubble. This has been observed earlier, in the discussion of Table 1, when it was noted that the esti-
mated value of bp2 is not significant, and in that of Table 3, when it was pointed out that the “two linear
regimes” hypothesis cannot be rejected.
Comparison of the models with respect to the general behavior of the bubble indicates that all the
models are very similar in the first part of the sample. The bubble is positive and increasing until the
end of 2002, and then decreases until it becomes null, several years later. For models IA, IIA and III this
occurs in the third trimester of 2007, and one year later for models IB and IIB.
The speculative bubble “centered” in the end of 2002 is easily correlated with the political and
economic events of that period, and can be associated with the uncertainty with respect to the result of
the presidential elections in that year and the policies that would be implemented in case the leftist
leaning candidate won the election. Lula da Silva, of the Laborers’ Party was elected, but the worst fears
of the agents, which had driven the spot exchange rate to almost 4.00 at the end of 2002, did not
materialize. The policies he implemented were essentially orthodox, especially the monetary policy
and, the speculative bubble shrank steadily until it disappeared six years later.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1049

Fig. 2. Model IA: spot and fundamental rates versus collapse probability.

After the bubble size becomes null toward the later part of the sample, however, its behavior differs
across models. In models IA and IIA two negative bubbles develop. One starts in third trimester of 2007
and lasts for one year, and the other starts in the second trimester of 2009, and lasts to the end of the
sample. In models IB and IIB only one positive bubble develops after the third trimester of 2008. In
model III one negative bubble starts growing in the third trimester of 2007 (like in models IA and IIA)

Fig. 3. Model IB: spot and fundamental rates versus collapse probability.
1050 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Fig. 4. Model IIA: spot and fundamental rates versus collapse probability.

followed by a positive bubble after the third trimester of 2008 (as in models IIA and IIB) that lasts until
the end of the sample.
The collapse probability is also displayed in Figs. 2–6. The general shape of its trajectory is similar in
both variants of models I and II, and is consistent with the description above. At the peak of the “big”
speculative bubble (at the end of 2002), the collapse probability reaches 100%, and then decreases to
about 40%, toward the end of the sample. In model III, however, the collapse probability has a very
different behavior. It is null almost always, except for a period that extends from the first trimester of

Fig. 5. Model IIB: spot and fundamental rates and collapse probability.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1051

Fig. 6. Model III: spot and fundamental rates and collapse probability.

2001 to the first trimester of 2005, during the 2002 speculative bubble. It briefly reaches 50%, but
generally remains below 20%. When compared to the bubble of models I and II, the bubble in model III
is much more persistent. This was shown earlier in the discussion of Fig. 1, when it was pointed out that
the probability of survival of the bubble in model III is close to one for jbjh1:2.
As can be seen in (11), the innovation is the gain of an investor that detains a long position in
forward contracts for the exchange rate. The expectation of a positive innovation associated to the
occurrence of a positive bubble in period t þ 1 implies a positive gain for the agent. Therefore, rational

Fig. 7. Average error relative to fundamental (absolute value).


1052 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Fig. 8. Average error relative to 1 þ bubble (absolute value).

expectations implies Et[Rtþ1] ¼ 0. The expected innovation in the exchange rate can be calculated by
the following expression, and tested:
Et ½Rtþ1  ¼ qðbt ÞEt ½Rtþ1 jS þ ð1  qðbt ÞÞEt ½Rtþ1 jC (39)
As seen in the previous section, rational expectations hypothesis is only rejected in model III.
The average error over the sample can be calculated by the identity in equation (20), with the
sample errors given by equations (19a) and (19b). In Figs. 7 and 8 they are compared, in absolute value,
to S* (the fundamental exchange rate) and to 1 þ jbj for models IIB and III, respectively.
Fig. 7 shows that the error is generally less than 10% of the fundamental, except during the 2002–
2003 bubble. Relative to 1 þ jbj, the average error is also around 10% most of the time, except during the
2002 and 2008 bubbles. These two graphs suggest that the fit of the model is relatively good.

5. Conclusions

In this paper we propose a model of periodically collapsing bubbles which extends the Van Norden
(1996) model, and nests it, by considering a non-linear specification for the bubble size in the survival
regime, and the endogenous determination of the level of the fundamental exchange rate. These
generalizations allow us to test for rationality in the formation expectations regarding the level of the
stochastic process, which was impossible in the original model, and remove the arbitrariness of
exogenously setting a reference level of the fundamental value, which is another feature of it. We feel
that these are important improvements.
This general model is applied to the exchange rate of the Brazilian real to the US dollar from March
1999 to February 2011, a period characterized by very large variations in its level and its rate of growth.
The possibility that these may have been caused by speculative bubbles that may have occurred in that
period is considered. Accordingly, the spot exchange rate is modeled as the sum of a fundamental value
and a bubble component.
The futures exchange rate is used as a proxy of its expected future value, and three different
structural models, of increasing generality, are considered for the determination of the fundamental
value. The first two (models I and II) are based on the application of the inter-temporal non-arbitrage
principle to the investment decision of firms engaged in international trade, and imply that the
exchange rate satisfies either purchasing power parity (PPP), or a modified version of it that accounts
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1053

for the difference in the interest rate of Brazilian government bonds traded domestically and abroad.
Two variants of models I and II are tested, varying the foreign price index. Models IA and IIA use the PPI,
and models IB and IIB use the trade-weighted foreign producer’s prices. The third structural model
(model III) is a suitably modified version of the monetary model of exchange rate determination. The
Meese (1986) formulation is extended to take into account the interest rate spread that characterizes
the “country risk”, as measured by the EMBI rate. The model is calibrated to the period of analysis.
The empirical implementation adequately models the dynamics of the Brazilian foreign exchange in
the period we consider. All five models are tested along several dimensions, and the overall best
models are deemed to be IA and IIB in their unrestricted forms. For all models, except the monetary
exchange rate model (model III), we reject Van Norden’s empirical implementation, but are unable to
reject rational expectations, at the 5% significance level.
In summary, we were able to construct a model for the Brazilian foreign exchange rate for the period after
it was allowed to float that contemplates bubble dynamics and describes well the dynamics of that rate.

Acknowledgments

The authors thank Søren Johansen, Simon Van Norden, and an anonymous referee of this journal,
for extensive comments on earlier versions, but retain full responsibility of any remaining errors.
Maldonado’s work was supported, in part, by CNPq (Brazil) though grants 304844/2009-8 and 401461/
2009-2, and acknowledges that the final revision was made during his academic visit to the Australian
National University. Valli’s work was supported, in part, by a scholarship granted by CAPES (Brazil) for
graduate studies. The reasoning, interpretation and conclusions presented here do not necessarily
reflect that of the institutions to which the authors are affiliated.

Appendix 1. The determination of the fundamental exchange rate


In this appendix we present three models that suggest the formulation to be used to calculate the
fundamental solution of the stochastic process proposed in the text for the exchange rate. They are
based on asset approach to the determination of the exchange rate, the principle of absence of riskless
arbitrage opportunities in the international markets for goods and capital, uncovered interest rate
parity, and the determination of the demand for monetary aggregates.

Model I: The PPP pricing relation

Suppose that the balance of trade is in equilibrium in (at least) two periods, and that it involves
a strictly positive flow of tradable goods. Then an agent which is a net importer must be indifferent
between importing a given good in period t or in period t þ 1. If he is neutral to risk, then the profit of
doing it immediately must be equal to the discounted expected future profit of doing it next period:
  
PDt 1 PDtþ1 e e
 St ¼ Stþ1 (A1.1)
PF t 1þr PF tþ1
where St is the exchange rate, quoted as the price of foreign exchange in domestic currency, PDt and PFt
are the domestic and the foreign prices of the representative tradable good, and r is the riskless interest
rate. Note that PDt/PFt is the revenue in domestic currency per unit of foreign exchange invested in the
import operation, while St is the cost. We assume for simplicity there are no import tariffs or other costs
of importing.
Assuming rational expectations, equation (A1.1) can be written as:

PDt PDtþ1
St ¼ aEt ½Stþ1  þ  aEt (A1.2)
PF t PF tþ1
where a ¼ (1 þ r)1 is the discount factor. Noting that equation (A1.2) is analogous to equation (2) in the
main text, it follows that f is equal to the last 2 terms in the r.h.s.
Recalling the definition of the fundamental value of the process from equation (4), the fundamental
exchange rate process is given by:
1054 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059


P
þN P
T PDtþi PDtþiþ1
St ¼ ai Et ½f ðXtþi Þ ¼ lim ai Et  aEtþi
i¼0 i¼0 PF tþi PF tþiþ1
 (A1.3)
PDt PDTþtþ1
St ¼  lim aTþ1 Et
PF t T/þN PF Tþtþ1
which is just the Purchasing Power Parity (PPP) value for the exchange rate minus the present value of
the expected long-run PPP value. We can discard this last term, assuming that the present value of the
domestic currency in terms of the foreign currency tends to zero:

PDTþtþ1
lim aT Et ¼ 0 (A1.4)
T/þN PF Tþtþ1
This is a familiar transversality condition in dynamic models: in essence it implies that the relative
price will not increase faster than the discount factor, as the terminal time increases without bound.
Using (A1.4) in (A1.3), we obtain the PPP pricing equation.

PDt
St ¼ (A1.5)
PF t

Model II: Modified PPP pricing relation

In this model we will consider as exogenous variables not only the price levels in the country and
abroad, but also the respective nominal interest rates. The arbitrage argument is similar to the one
employed in the previous model.
Suppose that in a period t an importer is evaluating an operation to be undertaken either in period
t þ 1 or t þ 2. He must be indifferent between these two dates.
If he decides to do it in period t þ 1, then the present value of the expected profit for each unit of
foreign currency is given by:
  
1 þ IF 0t PDtþ1 e
St (A1.6)
1 þ IDt PF tþ1
where IDt and IF 0t are the nominal interest rates observed in the domestic and foreign markets for
Brazilian government bonds, respectively, and the other variables are as before. The reasoning is the
same as the one employed to derive the expression in the l.h.s of (A1.1), adjusted for the fact that when
the import operation is conducted in period t þ 1, when the capital of the importer would have been
increased by the financial gain of purchasing a foreign bond with a maturity of 1 period yielding
ð1 þ IF 0t Þ in period t þ 1.
Analogously, if the importer delays the operation for one period, the present value of his expected
profit per unit of foreign exchange will be given by:
  e
1 þ IF 0tþ1 PDtþ2
Setþ1 (A1.7)
1 þ IDtþ1 PF tþ2
Equilibrium in international trade with strictly positive quantities of imports in t þ 1 and t þ 2 will
imply that the expected profit in both strategies be the same. Therefore the expressions in (A1.6) and
(A1.7) must be equal under rational expectations:
      
1 þ IF 0t PDtþ1 1 þ IF 0tþ1 PDtþ2
Et  St ¼ aEt  Stþ1 (A1.8)
1 þ IDt PF tþ1 1 þ IDtþ1 PF tþ2
Rearranging terms,
   "  #
1 þ IF 0t PDtþ1 1 þ IF 0tþ1 PDtþ2
St ¼ aEt ½Stþ1  þ Et  aEt (A1.9)
1 þ IDt PF tþ1 1 þ IDtþ1 PF tþ2
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1055

Equation (A1.9) has the same form as equation (2) in the main text, and the function f is equal to the
last two terms in the r.h.s. Again the fundamental solution can be found using equation (4):
X
St ¼ ai Et ½f ðXtþi Þ (A1.10)

     
1 þ IF 0t PDtþ1 1 þ IF 0t PDTþtþ1
St ¼ Et  lim aTþ1 Et (A1.11)
1 þ IDt PF tþ1 T/þN 1 þ IDt PF Tþtþ1
Therefore, the fundamental value for the exchange rate is the expected future value of the PPP
adjusted by the spread of the nominal interest rate, minus a residual term which represents the present
revenue value of investing one unit of foreign currency in period t during T þ t periods. As in the
previous model, we will assume that this residual vanishes as T goes to infinity:
  
1 þ IF 0Tþt PDTþtþ1
lim aT Et ¼ 0 (A1.12)
T/þN 1 þ IDTþt PF Tþtþ1
Using (A1.11) and (A1.12) we conclude that the fundamental value for the exchange rate is given by
the PPP adjusted for the interest rate differential:
  
1 þ IF 0t PDtþ1
St ¼ Et (A1.13)
1 þ IDt PF tþ1

Model III: The monetary model of exchange rate determination

In this model we consider a modified version of the Meese (1986) monetary model of exchange rate
determination. As in his model, the variables are specified in logs of the ratio of the domestic aggregate
to the foreign aggregate, for the money stock (MD and MF), prices (PD and PF), and income (YD and YF).
The exchange rate (S) is also specified in logs.

mt ¼ lnðMD=MFÞ; pt ¼ lnðPD=PFÞ; yt ¼ lnðYD=YFÞ; st ¼ lnðSt Þ (A1.14)


The domestic and foreign interest rates (ID and IF) are the rates paid by the country and abroad on
government debt.
The model has three equations which are specified below. The deterministic terms (a constant and
seasonal dummies) are suppressed in the theoretical model specification, to simplify the analysis, but
are included in the empirical model.
Assume a log-linear demand for real money balances which accounts for transactions and specu-
lative motives for holding money:
mt  pt ¼ a1 yt  a2 ðIDt  IF t Þ (A1.15)
where a1 > 0 is the income elasticity and a2 > 0 is the interest semi-elasticity of real money balances.
Following customary practice in specifying the monetary model of exchange rates, a disturbance term
is omitted from the money equation.
An extended uncovered interest parity relation is assumed, where the difference between domestic
(ID) and foreign (IF) interest rates is equal to the expected devaluation plus the interest rate “spread”
between foreign-denominated domestic bonds and foreign bonds of similar maturity, denoted IS.

IDt  IF t ¼ Et ½stþ1  st  þ ISt (A1.16)


The interest rate spread (IS) is a measure of the risk premium assigned by the market to foreign
investments in the country, and is often denominated “country risk”. It reflects not only risks associated
with the performance of the domestic economy, but also the supply of capital in the international
capital markets. This variable is very important in explaining the interest rate differential between the
Brazil and the U.S. in the sample considered in the empirical analysis of this paper. The interest rate on
Brazilian government bonds denominated in US dollar used in Model II is, therefore, IF 0t ¼ IFt þ ISt .
1056 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Lastly, assume that the deviations from PPP are a random walk.
st  pt ¼ ut ; ut ¼ ut1 þ 3 t (A1.17)
where 3 t is white noise, and ut is the real exchange rate.
Substituting the interest rate differential in (A1.15) for its value in (A1.16) yields:

mt  pt ¼ a1 yt  a2 ðEt ½stþ1  st  þ ISt Þ


Using (A1.17) to substitute for pt in the above equation,
mt  st ¼ a1 yt  a2 ðEt ½stþ1   st Þ  a2 ISt  ut
Collecting terms,
st ¼ gEt ½stþ1  þ ð1  gÞðmt  a1 yt þ a2 ISt Þ þ ð1  gÞut (A1.18)
where g ¼ a2 =ð1 þ a2 Þ 0 ð1  gÞ ¼ 1=ð1 þ a2 Þ. This is the structural equation for the dynamics of the
log of the exchange rate.
Following Meese (1986), assuming that st follows a borderline stationary process, we rely on first
difference of (A1.18) for empirical applications:
Dst ¼ ð1  gÞðDmt  a1 Dyt þ a2 DISt Þ þ gðEt ½stþ1   Et1 ½st Þ þ ð1  gÞ3 t (A1.19)
For notational simplicity, define the market fundamental process Dxt variables:

Dxt hðDmt  a1 Dyt þ a2 DISt Þ ¼ cDxt1 þ dt ; jcj < 1 (A1.20)


Equation (A1.19) can be solved forward to yield:
X
T 1 TX
1
Dst ¼ ð1  gÞ gs ½Et ½xtþs   Et1 ½xt1þs  þ gT ½Et ½xtþT   Et1 ½xtþT1  þ ð1  gÞ gs 3 tþs
s¼0 s¼0

If the transversality condition holds


lim gT ½Et ½xtþT   Et1 ½xtþT1  ¼ 0 (A1.21)
T/N

then the unique, no bubbles solution to (A1.19) is:


X
N
Dst ¼ ð1  gÞ gs ½Et ½xtþs   Et1 ½xt1þs  þ 3 t (A1.22)
s¼0

From (A1.20) the optimal prediction formula for xtþs is:


s
X
Et ½xtþs  ¼ xt þ ck Dxk (A1.23)
k¼1

Then, the rational expectations solution to (A1.19) can be shown to be:


gc
Dst ¼ Dxt þ ðDxt  Dxt1 Þ þ 3 t (A1.24)
1  gc
Recalling that the error 3 t is white noise, the changes to the fundamental value of the log of the
exchange rate can be calculated as:
gc
Dst ¼ Dxt þ ðDxt  Dxt1 Þ (A1.25)
1  gc
Appendix 2. The statistics for tests of the model
This section describes formally the statistics used to test the several models presented in the text.

C Likelihood ratio test

The Likelihood Ratio test (LR test) is one of the most popular approaches for testing hypotheses on
parameter restrictions, when they are estimated by the maximum likelihood method. Typically, the
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1057

hypotheses are restrictions on the parameters of the model, which are represented by a linear system
of equations with these parameters. The LR statistics is defined by the expression in (A2.1):
h  i
2 Lð^qÞ  L ~q (A2.1)

where ^qð~qÞ is the unrestricted (restricted) maximum likelihood estimator of the parameters and L is the
log-likelihood function. It is proved that this statistics has a c2 ðmÞ probability distribution, where m is
the rank of the linear system, which defines the restrictions. See Hamilton (1994) for details on the LR
test.

C Tests on the standard errors of the parameters

We used the tests based on the quasi-maximum likelihood estimation proposed by White (1982),
1
which uses the matrix T 1 ð^I 2D^I OP^I 2D Þ as an approximation of the variance–covariance of maximum
likelihood estimates.
In this approximation the second-derivative estimate of the information matrix is given by (A2.2):

2 
^I ¼ T 1 v L ðqÞ (A2.2)
2D 0  ^
vq v q q¼q
and ^I OP is the outer-product estimate of the same information matrix.
X
T
! 0 !
^I ¼ T 1
OP grad Lt ð^qÞ 5 grad Lt ð^qÞ (A2.3)
t¼1
PT
where Lt represents the tth term of the log-likelihood function LðqÞ ¼ t ¼ 1 Lt ðqÞ, and where the
! ^
gradient line vector grad Lt ðqÞ is the vector of scores of the tth observation.

C Specification tests

The specification tests of regime-switching time series models allow us to test for omitted auto-
correlations, omitted ARCH and Markovian effects on errors. These are described in detail in Hamilton
(1996), that presents specification tests for Markov-switching models based on special properties of
the vector of scores (defined above). The main property is that the vector series of scores have zero
time-correlation when evaluated at the true parameters, namely:
 ! 0 ! 
E grad Lt ðqÞ 5 grad Lt1 ðqÞ ¼ 0 for t ¼ 2; 3; .; T (A2.4)

As proposed by White (1987), if Ct(q) is any k-dimensional line vector whose coordinates are
elements of the corresponding outer-product matrix of scores (inside the expectation above) and if the
model is correctly specified, then:
" #
X
T h X i0
T 1=2 Ct ð^qÞ B^ T 1=2 Ct ð^qÞ zc2 ðkÞ (A2.5)
t¼1

^ is the (2,2) sub-block of the inverse of the following partitioned matrix:


where B

2  3
PT  ! 0  !  PT  ! 0
6 grad Lt b
q 5 grad Lt bq grad Lt b
q 5Ct bq 7
b 1 6 t ¼ 1 t ¼1 7
A ¼ T 6 T   0 ! 
7 (A2.6)
4 P P b0   5
Ct bq 5grad Lt bq Ct q b
5Ct q
t ¼1

In our analysis we use six specifications for the vector Ct(q): two for AR(1)’s (one for each regime)
which correspond to the derivatives of the log-likelihood with respect to bS00 and bC0, two for
ARCH(1)’s (one for each regime) which correspond to the derivatives with respect to sS and sC, one for
1058 W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059

Markovian effects which corresponds to the derivative with respect to bq0 and one for all (five) previous
specifications, considering all the derivatives above.

Appendix 3
The results of the PERRON97 procedure or RATS 7.0 (Estima, 2007) are presented in Table B1.

Table B1
The results of the PERRON97 procedure or RATS 7.0 (Estima, 2007).

Break date TB ¼ 2005:05 statistic t(alpha ¼ ¼ 1) 3.79603

Critical values at 1% 5% 10% 50% 90% 95% 99%


For 100 obs. 6.21 5.55 5.25 4.22 3.35 3.13 2.63
Infinite sample 5.57 5.08 4.82 3.98 3.25 3.06 2.72

Number of lag retained: 12.

References

Ball, L., 2001. Another look at long-run money demand. Journal of Monetary Economics 47, 31–44.
Blanchard, O., 1979. Speculative bubbles, crashes and rational expectations. Economics Letters 3, 387–389.
Blanchard, O., Fisher, S., 1989. Lectures on Macroeconomics. MIT Press, Cambridge, Massachusetts.
Blanchard, O., Watson, M., 1982. Bubbles, rational expectations and financial markets. In: Wachtel, P. (Ed.), Crises in the
Economic and Financial Structure. D.C. Heath and Company, Lexington, MA, pp. 295–316.
Brooks, C., Katsaris, A., 2005. A three-regime model of speculative behaviour: modelling the evolution of bubbles in the S&P 500
composite index. Economic Journal 115 (505), 767–797.
Charemza, W., Deadman, D., 1995. Speculative bubbles with stochastic explosive roots: the failure of unit root testing. Journal of
Empirical Finance 2, 153–163.
Diba, B.T., Grossman, H.I., 1988a. Explosive rational bubbles in stock prices? American Economic Review 78, 520–530.
Diba, B.T., Grossman, H.I., 1988b. The theory of rational bubbles in stock prices. Economic Journal 98, 746–754.
Doldado, J., Jenkinson, T., Sosvilla-Rivero, S., 1990. Cointegration and unit roots. Journal of Economic Surveys 4, 249–273.
Enders, W., 1995. Applied Econometric Time Series. Wiley.
Engsted, T., 1993. Testing for rational inflationary bubbles: the case of Argentina, Brazil and Israel. Applied Economics 25, 667–
674.
Estima, 2007. RATS 7.0 Reference Manual. Estima Corp, Evanston, Illinois.
Evans, G.W., 1986. A test for speculative bubbles in the sterling-dollar exchange rate: 1981–84. American Economic Review 76
(4), 621–636.
Evans, G.W., 1991. Pitfalls in testing for explosive bubbles in asset prices. The American Economic Review 81 (4), 922–930.
Evans, M.D.D., 1996. Peso problems: their theoretical and empirical implications. In: Maddala, G.S., Rao, C.R. (Eds.), Handbook of
Statistics. Statistical Methods in Finance, vol. 14. Elsevier Science, Amsterdam, pp. 613–646.
Funke, M., Hall, S., Sola, M., 1994. Rational bubbles during Poland’s hyperinflation: implications and empirical evidence.
European Economic Review 38 (6), 1257–1276.
Goldfeld, S., Quant, R., 1973. A Markov model for switching regressions. Journal of Econometrics 1, 3–16.
Hamilton, J., 1994. Time Series Analysis. Princeton University Press, Princeton, NJ.
Hamilton, J., 1996. Specification and testing in Markov-switching time-series models. Journal of Econometrics 70, 127–157.
Hodrick, R., Prescott, E., 1980. Postwar U.S. business cycles: an empirical investigation. Carnegie Mellon University discussion
paper no. 451.
Hodrick, R., Prescott, E., 1997. Postwar U.S. business cycles: an empirical investigation. Journal of Money, Credit and Banking 29
(1), 1–16.
Kiefer, N.M., 1978. Discrete parameter variation: efficient estimation of a switching regression model. Econometrica 46, 427–
434.
Meese, R., 1986. Testing for bubbles in exchange markets: a case of sparkling rates? Journal of Political Economy 94 (2), 345–
373.
Obstfeld, M., Rogoff, K., 1983. Speculative hyperinflations in maximizing models: can we rule them out? Journal of Political
Economy 91 (4), 675–687.
Perron, P., 1997. Further evidence on breaking trend functions in macroeconomic variables. Journal of Econometrics 80 (2), 355–
385.
Roche, M., 2001. Fads versus fundamentals in farmland prices: comment. American Journal of Agricultural Economics 83 (4),
1074–1077.
Stock, J.H., Watson, M.W., 1993. A simple estimator of cointegrating vectors in higher order integrated systems. Econometrica 61,
783–820.
Tirole, J., 1982. On the possibility of speculation under rational expectations. Econometrica 50 (5), 1163–1180.
Tirole, J., 1985. Asset bubbles and overlapping generations. Econometrica 53 (6), 1499–1528.
Tourinho, O., 1995. A demanda por moeda em processos de inflação elevada. Pesquisa e Planejamento Econômico 25 (1), 7–68,
English translation available as: Demand for money in high inflation. Brazilian Economic Studies (New Series), IPEA, Rio de
Janeiro, RJ, Chapter 7, pp. 187–246. 1998.
W.L. Maldonado et al. / Journal of International Money and Finance 31 (2012) 1033–1059 1059

Van Norden, S., 1996. Regime switching as a test for exchange rate bubbles. Journal of Applied Econometrics 11, 219–251.
Van Norden, S., Schaller, H., 1993. The predictability of stock market regime: evidence from the Toronto stock exchange. The
Review of Economics and Statistics 75 (3), 505–510.
West, K., 1987. A specification test for speculative bubbles. Quarterly Journal of Economics CII, 553–580.
White, H., 1982. Maximum likelihood estimation of misspecified models. Econometrica 50, 1–25.
White, H., 1987. Specification testing in dynamics models. In: Bewley, T.F. (Ed.), Advances in Econometrics. Annals of the Fifth
World Congress of the Econometric Society, vol. II. Cambridge University Press, Cambridge.

You might also like