ICAPM VECH Asian Market

Looking for risk premium and contagion in
Asia-Pacific foreign exchange markets

$
Chu-Sheng Tai
*
Department of Economics and Finance, College of Business Administration,
Texas A&M University-Kingsville, MSC 186, 700 University Boulevard, Kingsville,
TX 78363-8203, USA
Abstract
This article tests pure contagion effects among four Asian foreign exchange markets,
namely, Japan, Hong Kong, Singapore, and Taiwan during the 1997 Asian crisis. A conditional
version of international capital asset pricing model (ICAPM) in the absence of purchasing
power parity (PPP) is used to control for economic fundamentals or systematic risks. The
empirical results show strong contagion effects in both conditional means and volatilities of
those markets after systematic risks have been accounted for. Specifically, the contagion-in-
mean effects are mainly driven by the past return shocks in Hong Kong, Singapore, and
Taiwan. As for contagion in volatility, the lead/lag relationships appear to be multidirectional
among Japan, Singapore, and Taiwan, but between Hong Kong and Singapore, and between
Hong Kong and Taiwan, they are unidirectional, with Hong Kong playing the dominant role in
generating negative volatility shocks. In addition, the conditional ICAPM with asymmetric
multivariate general autoregressive conditional heteroscedastic in mean (MGARCH(1,1)-M)
structure is able to explain/predict on average 17.28% of the return variations in those markets.
Therefore, this study provide a further evidence that the time-varying risk premium is a very
strong candidate in explaining the predictable excess return puzzle [Lewis, K. K. (1994).
Puzzles in international financial markets. NBER Working Paper No. 4951] since the risk
1057-5219/$ - see front matter D 2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.irfa.2004.02.020
$
An earlier version of the article was presented at the 2002 National Taiwan University International
Conference on Finance in Taipei, Taiwan, ROC, the 2002 Financial Management Association International
Annual Meeting in San Antonio, TX, and the 2003 Midwest Finance Association Annual Meeting in St. Louis,
MO, and was awarded the Outstanding Paper in International Finance in 2003 by the Midwest Finance
Association. I would like to thank the conference participants for their comments and suggestions. Any remaining
errors are my own.
* Tel.: +1-361-593-2355; fax: +1-361-593-3912.
E-mail address: ctai@tamuk.edu (C.-S. Tai).
International Review of Financial Analysis
13 (2004) 381409
premia founded in this article are not only statistically significant but also economically
significant.
D 2004 Elsevier Inc. All rights reserved.
JEL classification: C32; F31; G12
Keywords: Contagion; Spillover; Time-varying risk premium; Multivariate GARCH-M
1. Introduction
Due to a series of financial crises in 1990s, the study of the transmission of financial
shocks/crisis across markets/countries has become one of the most intensive research
topics in international financial literature in recent years. Previous articles on this topic
have failed to take into account an important distinction between the two concepts of
interdependence and contagion. Masson (1998) argues that there are three main channels
that financial markets turbulence can spread from one country to another. They are
monsoonal effects, spillovers, and pure contagion effects. Monsoonal effects, or
contagions from common causes, tend to occur when affected countries have similar
economic fundamentals or face common external shocks. The second type of financial
market interlinkages arises from spillover effects, which may be due to trade linkages or
financial interdependence. The first two channels of financial crises can be categorized as
fundamental-driven crises since the affected countries share some macroeconomic
fundamentals, which implies that the transmission of financial crises is due to the
interdependence among those countries and not necessarily due to contagion. The third
transmission channel is the pure contagion effect. Contagion here refers to the cases where
crisis in one country triggers a crisis elsewhere for reasons unexplained by macroeconomic
fundamentals. For instance, a crisis in one country may lead creditors and investors to pull
out from other countries over which they have a poor understanding resulting from
information asymmetries.
The goal of this article is to test the pure contagion effects among four Asian foreign
exchange markets, namely, Japan, Hong Kong, Singapore, and Taiwan during the 1997
Asian crisis. Specifically, in this article, I define contagion as significant spillovers of
country-specific idiosyncratic shocks during the crisis after economic fundamentals or
systematic risks have been accounted for. In testing for contagion, its existence depends
on the economic fundamentals used. Unfortunately, there is disagreement on the
definitions of the fundamentals. To control for the economic fundamentals, most
empirical studies tend to choose those fundamentals arbitrarily, such as by using
macroeconomic variables, dummies for important events, and time trends. The problem
with these control variables is that contagion is not well defined without reference to a
theory. To overcome this problem, I rely on an international capital asset pricing model
(ICAPM) in the absence of purchasing power parity (PPP), which provides a theoretical
basis in selecting economic fundamentals. The economic fundamentals under ICAPM
are the world market and foreign exchange risks, so the evidence of contagion is based
on testing whether idiosyncratic risksthe part that cannot be explained by the world
market and foreign exchange risksare significant in describing the dynamics of
C.-S. Tai / International Review of Financial Analysis 13 (2004) 381409 382
conditional mean and volatility in Asian foreign exchange markets during the crisis.
The ICAPM used in this article also provides another avenue to test the existence of
risk premium in foreign exchange markets since previous empirical studies using
consumption-based asset pricing model to test the existence of risk premium in ex-
plaining the predictable excess return puzzle (Lewis, 1994) have not been very
successful.
1
In addition to the contribution in overcoming the drawback of arbitrarily choosing
economic fundamentals in testing contagion effects in previous studies, another
contribution of this article is methodology used to test those effects. In particular, I
utilize an asymmetric multivariate general autoregressive conditional heteroscedastic in
mean (MGARCH-M) approach to model the conditional mean and asymmetric
volatility spillovers during the crisis, in addition to capturing the time dependencies
in the second moments of asset returns, a stylized property found in most financial
time series, which has been ignored by most empirical studies on contagion.
2
Furthermore, the ICAPM in the absence of PPP with MGARCH-M parameterization
adopted in this article also overcomes the drawbacks in previous studies in testing risk
premium hypothesis in explaining the predictable excess returns puzzle, and thus
provides a new inside on the test of risk premium hypothesis. For example, Mark
(1988) uses a single-beta CAPM to price forward foreign exchange contracts and
specifies the betas as ARCH-like process. He estimates the model jointly for four
currencies using a generalized method of moments (GMM) procedure. His results
show significant time variation for the betas, and tests of the overidentifying
restrictions are not rejected. However, as pointed out by Mark, the GMM estimator
is robust, but, in general, is not asymptotically efficient. Consequently, instead of
using GMM estimation, McCurdy and Morgan (1991) also apply the single-beta
CAPM with a bivariate GARCH parameterization to price deviations from UIP for
five major currencies. They estimate their model currency by currency, while Mark
2
According to Dornbusch, Park, and Claessens (1999), Forbes and Rigobon (1999), and Kaminsky and
Reinhart (in press), previous empirical studies on contagion can be categorized by methodology into four groups:
(1) the testing of significant increases in correlation (Baig & Goldfajn, 1999; Calvo & Reinhart, 1996; Forbes &
Rigobon, 1998, 1999; Park & Song, 1999); (2) the testing of significance in innovation correlation (Baig &
Goldfajn, 1999); (3) the testing of significant volatility spillover (Edwards, 1998; Edwards & Susmel, 1999); (4)
crisis prediction regression (Bae, Karolyi, & Stulz, 2000; Eichengreen, Ross, & Wyplosz, 1996; Kaminsky &
Reinhart, in press; Sachs, Tornell, & Velasco, 1996; Van Rijckeghem & Weder, 1999). None of the contagion
studies mentioned above explicitly takes the time dependencies in the second moment into account. A recent
paper by Bekaert, Harvey, and Ng (2002) applies three-stage univariate GARCH model to study contagion in
equity markets by testing whether there is evidence of significant increase in cross market residual correlation
during the crisis. Although they model conditional second moments, they cannot answer whether return shocks
originated from one market will significantly affect the other markets during the crisis.
1
For example, Backus, Gregory, and Telmer (1993), Cumby (1988), Kaminsky and Peruga (1990), and Mark
(1985) use an intertemporal asset pricing model to test the existence of a time-varying risk premium in foreign
exchange markets. In this model, the risk premium is due to consumption risk measured by the covariance between
returns and the marginal utility of money. The results from these studies are disappointing because the observable
ingredients in the risk premium models do not vary sufficiently to explain the high degree of variability in asset
returns without implausibly large estimates of the coefficient of relative risk aversion (see Engelm 1996).
(1988) estimates his model jointly across currencies, so the efficiency might be
sacrificed in McCurdy and Morgans study because cross-asset correlations and
parameter restrictions are ignored. To maintain the efficiency, Tai (2001) applies
MGARCH-M to test risk premium hypothesis, but similar to Mark (1988) and
McCurdy and Morgan (1991), he assumes that PPP holds, and thus ignores foreign
exchange risk. Therefore, under the fully parameterized multivariate model adopted in
this article, not only is the maximum efficiency gain retained in controlling the
systematic risks when testing the contagion effects, but also some interesting statistics
are recovered, which are mostly ignored in previous studies.
3
In addition, the test of
ICAPM in the absence of PPP provides a new insight on the sources of time-varying
risk premium in foreign exchange markets.
The empirical results show strong contagion effects in both the conditional means
and volatilities of foreign exchange returns after systematic risks have been accounted
for. Specifically, the contagion-in-mean effects are mainly driven by the past return
shocks in Hong Kong, Singapore, and Taiwan. As for contagion in volatility, the lead/
lag relationships appear to be multidirectional among Japan, Singapore, and Taiwan,
but between Hong Kong and Singapore, and between Hong Kong and Taiwan, they
are unidirectional, with Hong Kong playing the dominant role in generating negative
volatility shocks.
The remainder of the article is organized as follows. Section 2 presents the theoretical
asset pricing model used to control for systematic risks when testing pure contagion
effects. Section 3 describes the econometric methodology employed to estimate the model
and several test hypotheses are presented in Section 4. Section 5 describes the data and
empirical results are reported in Section 6. Some conclusions are offered in Section 7.
2. The theoretical motivation
We know that the first-order condition of any consumer investors portfolio optimi-
zation problem can be written as:
EM
t
R
i;t
j X
t1
1; bi 1 . . . N 1
where M
t
is known as a stochastic discount factor or an intertemporal marginal rate of
substitution; R
i,t
is the gross return of asset i at time t and X
t 1
is market information
known at time t 1. Without specifying the form of M
t
, Eq. (1) has little empirical content
since it is easy to find some random variable M
t
for which the equation holds. Thus, it is
the specific form of M
t
implied by an asset pricing model that gives Eq. (1) further
3
Previous studies of finding time-varying risk premia in foreign exchange markets using GMM, which does
not require researchers to model conditional second moments of asset returns, include Dumas and Solnik (1995)
and Tai (1999). Other studies using multivariate GARCH approach include Baillie and Bollerslev (1990), De
Santis and Gerard (1998), Giovannini and Jorion (1989), and Tai (2001). Among the four studies, only De Santis
and Gerard (1998) and Tai (2001) are able to detect significant time-varying risk premia, but they do not exam
how a financial crisis would affect the dynamics of the risk premia.
empirical content (e.g., Ferson, 1995; Tai, 2000). Suppose M
t
and R
i,t
have the following
factor representations:
M
t
a
K
k1
b
k
F
k;t
u
t
2
r
i;t
a
i

K
k1
b
ik
F
k;t
e
i;t
bi 1 . . . N 3
where r
i,t
= R
i,t
R
0,t
is the raw returns of asset i in excess of the risk-free rate, R
0,t
, at
time t, and E[u
t
F
k,t
jX
t1
] = E[u
t
jX
t 1
] = E[e
i,t
F
k,t
jX
t 1
] = E[e
i,t
jX
t 1
] = 0 bi,k; F
k,t
are
common risk factors that capture systematic risk affecting all assets r
i,t
including M
t
; b
ik
are the associated time-invariant factor loadings that measure the sensitivities of the asset
to the common risk factors, while u
t
is an innovation, and e
i,t
are idiosyncratic terms that
reflect unsystematic risk. The risk-free rate, R
0,t 1
, must also satisfy Eq. (1).
EM
t
R
0;t1
j X
t1
1 4
Subtracting Eq. (4) from Eq. (1), we obtain
EM
t
r
i;t
j X
t1
0 bi 1 . . . N 5
Apply the definition of covariance to Eq. (5), obtaining:
Er
i;t
j X
t1

Covr
i;t
; M
t
j X
t1
EM
t
j X
t1
bi 1 . . . N 6
Substitute Eq. (2) into Eq. (6):
Er
i;t
j X
t1

k
b
k
EM
t
j X
t1
Covr
i;t
; F
k;t
j X
t1

k
k
k;t1
Covr
i;t
; F
k;t
j X
t1
7
where k
k,t 1
is the time-varying price of factor risk. Eq. (7) is a general conditional
multifactor asset pricing model derived from the intertemporal consumptioninvestment
optimization problem.
In empirical tests, the SDF is projected onto five factors: the world market portfolio
and four currency returns.
4
The selection of those five factors is theoretically justified
based on either the intertemporal CAPM of Merton (1973) or an international version of
CAPM in the absence of PPP developed by Adler and Dumas (1983). To extend
domestic CAPM into an international setting, previous researchers assume that either
4
In this article, I consider four Asian foreign exchange rates: Japanese yen (JP), Hong Kong dollar (HK),
Singapore dollar (SG), and New Taiwan dollar (TA), so there are four currency risks plus one world market
risk (WD).
investors have logarithmic utility or PPP holds. However, many empirical studies have
documented that the violation of PPP is a norm although PPP at best tends to hold in the
long run. In the absence of PPP resulting from either different consumption tastes or
violation of the law of one price, investors from different countries face different prices
when holding the same asset. In this situation, international asset pricing model will
contain risk premia, which are related to the covariances of asset returns with exchange
rates, besides the traditional market risk premium.
5
Therefore, a conditional multifactor
asset pricing model containing world market and foreign exchange risks (Eq. (7)) seems
to be reasonable and will be used to control for systematic risks in testing contagion
(e.g., De Santis & Gerard, 1998; Dumas & Solnik, 1995; Ferson & Harvey, 1994; Tai,
1999, among others). I can now rewrite the conditional multifactor asset pricing model
in Eq. (7) as
r
i;t
k
w;t1
Covr
i;t
; r
w;t
j X
t1

c
k
c;t1
Covr
i;t
; r
c;t
j X
t1
e
i;t
bi 1 . . . N
8
where w denotes world market risk and c is the currency risk.
6
3. Econometric methodology
The conditional ICAPM in Eq. (7) has to hold for every asset. However, the model
does not impose any restrictions on the dynamics of the conditional second moments.
Several MGARCH models have been proposed to model the conditional second
moments, such as the diagonal VECH model of Bollerslev, Engle, and Wooldridge
(1988), the constant correlation (CCORR) model of Bollerslev (1990), the factor
ARCH (FARCH) model of Engle, Ng, and Rothschild (1990), and the BEKK model
of Engle and Kroner (1995). Among these four popular MGARCH models, the BEKK
model is better suited for the purpose of this article because it not only guarantees that
the covariance matrices in the system are positive definite, but also allows the
conditional variances and covariances of different markets to influence each other,
which is very important for testing contagion in this article. Although it is easy to
understand, the VECH model might not yield a positive definite covariance matrix.
FARCH assumes that the covariance matrix is driven by the conditional variance
process of one portfolio (the market portfolio), and this assumption does not hold in
this article since the conditional covariance matrix is assumed to be driven not only
by market portfolio but also by foreign exchange returns. As for the CCORR model,
it restricts the correlation between two assets returns to be constant over time, which
is unlikely to hold as suggested by Longin and Solnik (1995) and Karolyi and Stulz
6
In this article, the factors are market portfolio and short-term currency deposits, which are traded assets, so I
can replace F
k,t
with r
k,t
.
5
See Adler and Dumas (1983), Solnik (1974), and Stulz (1981, 1984).
(1996). As a result, a BEKK structure with asymmetric volatility effects is selected
over the other MGARCH specifications to model the conditional second moments of
asset returns and to test contagion effects among Asian foreign exchange markets.
7
Specifically, the dynamic process for the conditional variancecovariance matrix of
asset returns is specified as:
H
t
CVCAV H
t1
A BV e
t1
e
t1
V B DV h
t1
h
t1
V D GV w
t1
w
t1
V
G KV n
t1
n
t1
V K LV l
t1
l
t1
V L MV m
t1
m
t1
V M PV
1
t1
1
t1
V P QV s
t1
s
t1
V Q SV t
t1
t
t1
V S VV f
t1
f
t1
V V 9
where H
t
is 5 5 time-varying variancecovariance matrix of asset returns, C is
restricted to be a 5 5 upper triangular matrix, and A, B, D, G, K, L, M, P, Q, S,
and V are diagonal matrices whose general form, X, is given by:
X
x
JP; j
0 0 0 0
0 x
HK; j
0 0 0
0 0 x
SG; j
0 0
0 0 0 x
TA; j
0
0 0 0 0 x
WD; j
_
_
_
_
10
The 5 1 vector, g
t 1
, captures the asymmetric impact that the vector of past
negative shocks has on the conditional covariance matrix in a manner similar to that
of Glosten, Jagannathan, and Runkle (1993), and is defined as:
g
t1

mine
JP;t1
; 0
mine
HK;t1
; 0
mine
SG;t1
; 0
mine
TA;t1
; 0
mine
WD;t1;
0
_
_
_
_
11
7
The asymmetric volatility effects in variances and covariances have been documented in recent articles by,
among others, Kroner and Ng (1998) and Bekaert and Wu (2000).
The effects of past shocks of other markets on a markets conditional variance or
conditional covariances (volatility spillovers) are captured by the vectors w
t 1
, n
t 1
,
l
t 1
, and m
t 1
, which are as follows:
w
t1

e
HK;t1
e
SG;t1
e
TA;t1
e
WD;t1
e
JP;t1
_
_
_
_
; n
t1

e
SG;t1
e
TA;t1
e
WD;t1
e
JP;t1
e
HK;t1
_
_
_
_
; l
t1
e
TA;t1
e
WD;t1
e
JP;t1
e
HK;t1
e
SG;t1
_
_
_
_
; m
t1

e
WD;t1
e
JP;t1
e
HK;t1
e
SG;t1
e
TA;t1
_
_
_
_
12
Several articles in the literature show that volatility spillovers between markets are
asymmetric in the sense that negative innovations in a market increase volatilities in other
markets more than do positive innovations in that market. Consequently, it will be interesting
to see whether such asymmetric volatility spillovers do occur during the crisis. The vectors
1
t 1
, s
t 1
, t
t 1
, and f
t 1
, capture this asymmetry and are defined as:
1
t1

crisismine
HK;t1
; 0
crisismine
SG;t1
; 0
crisismine
TA;t1
; 0
crisismine
WD;t1
; 0
crisismine
JP;t1;
0
_
_
_
_
; s
t1

crisismine
SG;t1
; 0
crisismine
TA;t1
; 0
crisismine
WD;t1
; 0
crisismine
JP;t1
; 0
crisismine
HK;t1;
0
_
_
_
_
;
t
t1

crisismine
TA;t1
; 0
crisismine
WD;t1
; 0
crisismine
JP;t1
; 0
crisismine
HK;t1
; 0
crisismine
SG;t1;
0
_
_
_
_
; f
t1

crisismine
WD;t1
; 0
crisismine
JP;t1
; 0
crisismine
HK;t1
; 0
crisismine
SG;t1
; 0
crisismine
TA;t1;
0
_
_
_
_
13
where crisis is a dummy variable, which is equal to 1 during the crisis and 0 otherwise.
8
The
difference between the first set of innovation vectors (w
t 1
, n
t 1
, l
t 1
, m
t 1
) and the
8
I assume that Asian crisis began in the first week of July 1997 (July 4, 1997) and ended in the last week of
December 1998 (December 25, 1998).
second set of innovation vectors (1
t 1
, s
t 1
, t
t 1
, f
t 1
) is that the first set captures overall
volatility spillovers during the entire sample period, while the second set captures the
asymmetric volatility spillovers during the crisis period. By including vectors 1
t 1
, s
t 1
,
t
t 1
, and f
t 1
, I can then test the incremental influences of volatility shocks on the foreign
exchange markets, which is a true test of contagion in volatility. In this model, for example,
the conditional variance of excess Japanese yen returns, h
JP,t
, depends on its past conditional
variance, h
JP,t 1
, through the parameter, a
JP
, its own past shocks, e
JP,t 1
, through the
parameter, b
JP
, and past shocks of the other markets through the parameters, g
JP
, k
JP
, l
JP
, and
m
JP
. This conditional variance also depends on its own past negative shocks through the
parameter, d
JP
, and on past negative shocks of the other markets through the parameters, p
JP
,
q
JP
, s
JP
, and v
JP
during the crisis. Here, these parameters measure the incremental amounts
by which bad news in one market at time t 1 affect the conditional variance of excess
Japanese yen returns at time t.
The parameterization of the conditional covariance matrix can therefore be viewed as
an extension of the diagonal BEKK representation of Engle and Kroner (1995) that allows
for past shocks from other markets to influence conditional variances and covariances, for
asymmetries in the impacts of these shocks.
9
This representation of the conditional
covariance matrix differs from the most general BEKK form in that conditional variances
are not permitted to depend on cross-products of lagged shocks, lagged conditional
variances of other markets, and lagged conditional covariances with other markets.
Similarly, conditional covariances are not influenced by lagged squared shocks and lagged
conditional variances in other markets. The parameterization presented here facilitates
testing of the null hypothesis of no volatility spillover effects against the alternative that
conditional variances depend on other markets only through their past squared shocks.
Even with this diagonal BEKK parameterization, it still requires the estimation of 70
parameters in the conditional covariance matrix.
Under the assumption of conditional normality, the log-likelihood to be maximized can
be written as
lnLq
TN
2
ln2p
1
2
T
t1
lnAH
t
qA
1
2
T
t1
e
t
qVH
t
q
1
e
t
q 14
where q is the vector of unknown parameters in the model. Since the normality assumption
is often violated in financial time series, I use quasi-maximum likelihood estimation
(QML) proposed by Bollerslev and Wooldridge (1992), which allows inference in the
presence of departures from conditional normality. Under standard regularity conditions,
the QML estimator is consistent and asymptotically normal and statistical inferences can
be carried out by computing robust Wald statistics. The QML estimates can be obtained by
maximizing Eq. (14) and calculating a robust estimate of the covariance of the parameter
estimates using the matrix of second derivatives and the average of the period-by-period
9
Ebrahim (2000) also uses this diagonal BEKK model to test volatility spillover effects between foreign
exchange and money markets, but in this article, I not only test the usual volatility spillover effects within foreign
exchange markets and between foreign exchange and world equity markets, but also test contagion in asymmetric
volatility spillover effects within and between those markets during a crisis.
outer products of the gradient. Optimization is performed using the Broyden, Fletcher,
Goldfarb, and Shanno algorithm.
4. Hypothesis testing
4.1. Testing time-varying risk premium
Many empirical studies have shown that the prices of risks are time varying (e.g., De
Santis & Gerard, 1997, 1998; Dumas & Solnik, 1995; Harvey, 1991; Tai, 1999, 2001,
among others). This time-varying price of risk is economically appealing in the sense that
investors use all available information to form their expectations about future economic
performance, and when the information changes over time, they will adjust their expect-
ations and thus their expected risk premia when holding different risky assets. Therefore,
to test time-varying risk premium hypothesis, I allow not only the conditional second
moments (covariance risks) to change over time, but also the prices of covariance risks to
be time varying (Eq. (8)).
The dynamics of prices of risks are chosen according to the theoretical international
asset pricing model developed by Adler and Dumas (1983). In their model, the price of
world market risk is a weighted average of the coefficients of risk aversion of all national
investors. Since the weights measure the relative wealth of each country and if all investors
are risk averse, the world price of market risk should be positive. Thus, similar to Bekaert
and Harvey (1995) and De Santis and Gerard (1997, 1998), an exponential function is
used to model the dynamic of k
w,t 1
and for the dynamics of k
c,t 1
, a linear specification
is adopted because the model does not restrict the price of currency risk to be positive.
10
k
w;t1
expu
w
VZ
t1
15
k
c;t1
u
c
VZ
t1
16
where Z
t 1
is a vector of information variables observed at the end of time t 1 and us
are time-invariant vectors of weights. Thus, the price of each source of currency risk is
assumed to be a linear function of the information variables in Z
t 1
, and the price of
world market risk is assumed to be an exponential function of information variables in
Z
t 1
. Given the dynamics of prices of risks, I can then test the time-varying risk premium
hypothesis by testing whether the information variables in Z
t 1
are significant in addition
to significant GARCH parameters.
10
As pointed out by De Santis and Gerad (1997), the conditional ICAPM is only a partial equilibrium model,
and the theory does not help identify the state variables that affect the prices of market and currency risks, so
inevitably any parameterization of the dynamics of k
w,t 1
and k
c,t 1
can be criticized for being ad hoc.
4.2. Testing contagion in mean and volatility
To test whether a countrys past idiosyncratic shocks have significant impact on the other
countries condition returns (contagion in mean) during the Asian crisis, I incorporate past
country-specific innovations into Eq. (8). Specifically, Eq. (8) can be modified as
r
i;t
k
m;t1
Covr
i;t
; r
m;t
j X
t1

c
k
c;t1
Covr
i;t
; r
c;t
j X
t1

i; j
/
ij
e
j;t1
crisis
_
i; j
x
ij
e
j;t1
_
e
i;t
; bi; j 17
where crisis is a dummy variable, which is equal to 1 during the crisis and 0 otherwise. In
testing the contagion-in-mean effects, I allow the past country-specific innovations to affect
currency returns in the entire sample period, and then test whether there are any incremental
influences of past innovations on currency returns during the crisis period. Thus, the
contagion-in-mean hypothesis can be examined by testing whether the parameters, x
ij
(i p
j), are individually or jointly significant after the systematic risks have been accounted for.
To test contagion-in-volatility hypothesis, one can test whether the elements in matrices
P, Q, S, and Vare individually or jointly significant. For example, a test of null hypothesis
that p
JP, j
is 0 (H
0
: p
JP, j
= 0) means that there is no contagion-in-asymmetric volatility
shocks from country j to Japan. Similarly, a test of null hypothesis of H
0
: p
i,HK
= q
i,SG
=
s
i,TA
= v
i,WD
= 0; bi = JP implies that the conditional volatility of Japanese yen returns is
not affected by the other countries negative idiosyncratic shocks. Finally, one can test the
source of negative idiosyncratic shocks. For example, to test whether the negative shocks
originate from Japan, one can test the null hypothesis of H
0
: v
HK,j
= s
SG,j
= q
TA,j
= p
WD,j
= 0;
bj = JP.
5. Data and summary statistics
Weekly observations on four Asian currency spot prices and 1-week interest rates (JP,
HK, SG, and TA) are used to construct a time-series of weekly deviations from UIP.
11
The
Datastream world total market return index (WD) is used to proxy world market risk.
12
Seven-day eurodollar interest rate is used as conditionally risk-free rate to compute the
excess equity returns on WD and the four excess Asian currency returns (or the ex post
deviations from UIP). In particular, the excess world market return is computed as r
i;t

ln
p
t
p
t1
_ _
1
52
ln1 i
US$
t1
, where p
t
is the Datastream world total return index (dividend
11
Ideally, it would be interesting to use currency and interest rate data from other Asia-Pacific countries,
such as Indonesia, the Philippines, South Korea, Thailand, Malaysia, etc., but the short-term interest rate data for
those countries are not available during the sample period considered in this study.
12
I prefer to use the world total market return index complied by Datastream, which captures more than 75%
of the market, as opposed to the widely used Morgan Stanley Capital International index, which captures only
approximately 60% of the total market.
included) at time t and i
t 1
US$
is the annualized 7-day eurodollar interest rate known at
timet 1. The excess currency return is computed as r
i;t

1
52
ln1 i
t1
* ln
s
t
s
t1
_ _
1
52
ln
1 i
US$
t1
, where s
t
is the spot rate at time t expressed as domestic price (the U.S. dollar)
of one unit of Asian currency and i
t 1
* is the annualized short-term Asian currency
interest rate known at time t 1.
I select a set of conditioning variables that have been widely used in the international asset
pricing literature (e.g., Bekaert & Harvey, 1995; Bekaert & Hodrick, 1992; De Santis &
Gerard, 1997, 1998; Ferson &Harvey, 1993; Harvey, 1991; Tai, 1999, 2000, among others).
They are excess dividend yield measured by the dividend yield on WDin excess of the 7-day
eurodollar interest rate (DIV), the change in the U.S. term premium, measured by the yield
difference between 10-year Treasury constant maturity rate and 7-day eurodollar rate
(DUSTP), the U.S. default premium, measured by the yield difference between Moodys
Table 1
Panel A: Summary statistics of foreign exchange returns and world equity return
Returns JP HK SG TA WD
Mean (%) 0.020 0.007 0.017 0.021 0.068
S.D. (%) 1.683 0.093 0.773 0.723 1.900
Minimum (%) 6.174 0.603 4.851 4.319 13.755
Maximum (%) 14.500 0.457 8.292 5.875 7.608
Skewness 1.138** 0.971** 1.239** 0.242** 0.713**
Excess kurtosis 7.972** 9.892** 23.411** 13.180** 5.375**
BJ 2122.542** 3138.013** 17,111.99** 5371.034** 954.907**
LB(20) 25.299 54.833** 124.861** 40.245** 20.448
LB
2
(20) 52.276** 195.194** 286.789** 56.706** 38.934**
Panel B: Unconditional correlation of conditioning variables
DIV DUSTP USDP WD
DIV 1
DUSTP .122 1
USDP .159 .038 1
WD .052 .017 .034 1
The statistics are based on weekly data from January 16, 1987 to March 23, 2001 (741 observations). The interest
rates are 1-week Euroyen deposit rate (JP), 1-week Hong Kong deposit rate (HK), 1-week Singapore deposit rate
(SG), and 10-day Taiwan money market rate (TA). WD is the Datastream world total market return index in
excess of the 7-day Eurodollar interest rate.
The BeraJarque (BJ) tests normality based on both skewness and excess kurtosis and is distributed v
2
with
two degrees of freedom.
LB(20) and LB
2
(20) denote the LjungBox test statistics for up to the 20th-order autocorrelation of the raw and
squared returns, respectively.
The conditioning variables are the excess dividend yield, measured by the dividend yield on Datastream world
total market return index in excess of the 7-day Eurodollar deposit rate (DIV), the change in the U.S. term
premium, measured by the first difference of the yield difference between 10-year Treasury constant maturity rate
and 7-day Eurodollar rate (DUSTP), the U.S. default premium, measured by the yield difference between
Moodys Baa- and Aaa-rated U.S. corporate bonds (USDP), and the lagged excess return on Datastream world
total market return index (WD).
*Statistical significance at the 5% level.
**Statistical significance at the 1% level.
Baa- and Aaa-rated U.S. corporate bonds (USDP), the lagged excess return on Datastream
world total market return index (WD), and a constant (CONSTANT).
13
The weekly data ranges from January 2, 1987, to March 23, 2001, which is a 743-data-
point series. However, I work with rates of return and use the first difference of
conditioning variables, and finally all the conditioning variables are used with a 1-week
lag, relative to the excess return series, which leaves 741 observations expanding from
January 16, 1987, to March 23, 2001. All the data are extracted from Datastream.
Table 1 presents summary statistics of the continuously compounded excess world
equity returns and currency returns. As can be seen from panel A, WD not only has the
highest weekly mean returns, 0.068%, but also the highest standard deviation, 1.900%.
Comparing the performance of four excess currency returns, TA is the best one with a
mean of 0.021% per week and a standard deviation of 0.723%, and JP, on the other hand,
performs worst with a return of 0.020% per week and a standard deviation of 1.683%.
Table 1 also reports skewness, excess kurtosis, BeraJarque, and LjungBox statistics.
In all cases, skewness and excess kurtosis are significantly higher than they should be
under normality, and hence BeraJarque test rejects normality for all excess return series.
The LjungBox test statistics for raw returns (LB(20)) are significant at the 1% level in
three excess currency returns, implying strong linear dependencies among those returns.
For squared returns, LB
2
(20) is significant at the 1% level for all the return series,
indicating strong nonlinear dependencies in both currency and equity returns. This is
consistent with the volatility clustering observed in most stock and foreign exchange
markets, suggesting that the use of a conditional heteroscedasticity model is advisable.
The unconditional correlation coefficients for the conditioning variables are reported in
panel B of Table 1. The correlation coefficients are small, and all are below .5, indicating
that the selected conditioning variables contain sufficiently orthogonal information.
6. Empirical evidence
The QML estimation of the conditional ICAPM (Eq. (17)) is reported in Table 2. The
hypothesis tests regarding the prices of risks and the predictability of conditioning
variables are presented in Table 3. The hypothesis tests concerning the contagion in mean
and volatility are shown in Tables 4 and 5, respectively. Finally, summary statistics about
the sources of risk premia and diagnostic test statistics for the standardized residuals are
reported in Table 6.
6.1. The evidence of time-varying risk premia
First, considering the test results for the existence of time-varying risk premium. The
results are very encouraging. For example, the joint hypothesis of zero prices of market and
currency risks is strong rejected by Wald statistic (Wald =2300.169) with a P value of 0.
13
The excess dividend yield (DIV) is highly correlated with the U.S. term premium (USTP), so similar to
De Santis and Gerard (1997, 1998), I use first difference of the U.S. term premium (DUSTP) as one of the
conditioning variables.
Table 2
Quasi-maximum likelihood estimation of the conditional ICAPM
a
Conditional mean process
World prices of market and currency risks
CONSTANT DIV DUSTP USDP WD
u
w
3.932
(3.399)
1.807
(3.173)
4.137
(3.415)
60.930
(10.862)**
24.863
(4.326)**
u
JP
1.625
(4.403)
11.726
(5.055)*
176.896
(42.429)**
66.220
(67.145)
150.697
(71.683)*
u
HK
412.099
(113.249)**
326.048
(108.925)**
1123.212
(530.127)*
2450.446
(1224.145)*
2140.210
(957.788)*
u
SG
35.537
(11.663)**
87.034
(13.466)**
401.491
(56.286)**
378.680
(165.597)*
431.662
(86.084)**
u
TA
30.215
(5.741)**
9.306
(5.105)
304.737
(41.050)**
300.213
(46.236)**
333.379
(43.048)**
Mean spillovers
j =JP j =HK j =SG j =TA WD
/
JP, j
0.013 (0.031) 0.598 (0.499) 0.092 (0.073) 0.142 (0.077)
/
HK, j
0.001 (0.002) 0.232 (0.043)** 0.003 (0.005) 0.010 (0.006)
/
SG, j
0.008 (0.014) 0.347 (0.240) 0.018 (0.034) 0.049 (0.035)
/
TA, j
0.025 (0.011)* 0.415 (0.209)* 0.224 (0.041)** 0.076 (0.034)*
Contagion in mean
x
JP, j
0.222 (0.082)** 3.604 (1.908) 0.573 (0.106)** 0.475 (0.134)**
x
HK, j
0.005 (0.004) 0.160 (0.064)* 0.017 (0.006)** 0.037 (0.006)**
x
SG, j
0.001 (0.040) 3.170 (1.073)** 0.115 (0.072) 0.207 (0.080)**
x
TA, j
0.042 (0.034) 0.100 (0.826) 0.172 (0.063)** 0.490 (0.101)**
Conditional variance process
JP HK SG TA WD
a
j
0.924 (0.018)** 0.911 (0.059)** 0.797 (0.028)** 0.055 (0.041) 0.750 (0.067)**
b
j
0.183 (0.034)** 0.347 (0.089)** 0.213 (0.058)** 1.345 (0.162)** 0.234 (0.035)**
d
j
4.958 (1.974)* 2.626 (6.150) 3.224 (3.728) 2.302 (5.790) 12.641 (5.066)*
Volatility spillovers
JP HK SG TA WD
j =JP 0.001 (0.001) 0.054 (0.018)** 0.006 (0.032) 0.178 (0.103)
j =HK 0.141 (0.403) 0.499 (0.343) 2.126 (0.836)* 1.468 (1.189)
j =SG 0.002 (0.049) 0.002 (0.003) 0.255 (0.053)** 0.024 (0.096)
j =TA 0.080 (0.062) 0.009 (0.008) 0.161 (0.048)* 0.043 (0.054)
j =WD 0.091 (0.032)** 0.005 (0.003)* 0.046 (0.011)** 0.040 (0.055)
Contagion in asymmetric volatility
JP HK SG TA WD
j =JP 0.221 (0.193) 26.782 (3.904)** 32.096 (4.925)** 17.484 (7.656)*
j =HK 1116.507
(1512.059)
7505.472
(1667.680)**
7813.758
(2745.339)**
1607.466
(1669.696)
The joint hypothesis of constant prices of market and currency risks is also significantly
rejected (Wald = 1001.038). Next, the joint hypothesis of constant prices of currency risks is
strongly rejected by Wald statistic (Wald = 332.470), and the joint hypothesis of constant
price of market risk is also rejected (Wald = 151.039). These test results imply that both
market and currency risks are not only priced but also time varying. Finally, the hypothesis
of constant price of currency risk for each currency is tested individually, and Wald test
statistic rejects the null at the 1% level in every case, implying that all four currencies are
sources of the time-varying currency risk premium. These results are consistent with the
findings of De Santis and Gerard (1998) and Dumas and Solnik (1995).
14
The conditioning
variables selected in this article are all very useful in predicting the dynamics of the risk
prices as can be seen from the hypothesis tests (1013) reported in Table 3. That is, the null
hypothesis of zero predictability of conditioning variable is strongly rejected by Wald
statistic at the 1% level in all cases. Although statistically significant time-varying risk
premia are found, an interesting question to ask is to what extent these predictable risk
premia are economically significant. In answering this question, I computed pseudo-R
2
statistic calculated as the ratio between the explained sum of squares and the total sum of
squares. As can be seen from Table 6, the reported pseudo-R
2
s for four excess currency
returns range from 9.957% for HK to 24.361% for TA, with an average of 17.284%, which
is significant higher than those reported in similar studies.
15
These relatively high pseudo-
Conditional variance process
Contagion in asymmetric volatility
JP HK SG TA WD
j = SG 15.325 (7.376)* 0.264 (0.275) 8.547 (6.641) 17.142 (5.645)**
j = TA 4.601 (1.747)** 0.125 (0.150) 0.397 (1.774) 9.181 (4.435)*
j = WD 10.450 (2.118)** 0.328 (0.205) 0.885 (1.552) 0.145 (1.067)
Log-likelihood function: 17,376.527
The reported parameter estimates for both the volatility spillover and contagion in asymmetric volatility
coefficients can be interpreted as follows. For example, if x
ij
represents the volatility spillover coefficient from
market j to market i, then the volatility spillover coefficient estimate from HK to JP is .141, which corresponds
to g
JP,HK
in matrix G in the variancecovariance matrix in Eq. (9). Similarly, the volatility spillover coefficient
estimate from SG to JP is .002, which corresponds to k
JP,SG
in matrix K in the variancecovariance matrix in Eq.
(9), and so on. The reported parameter estimates for the contagion in asymmetric volatility coefficients have the
same interpretation as those for volatility spillover coefficients.
a
Robust standard errors are given in parentheses.
*
Statistical significance at the 5% level.
**
Statistical significance at the 1% level.
Table 2 (continued)
14
Both De Santis and Gerard (1998) and Dumas and Solnik (1995) use excess returns on 1-month European
currency deposit rates to proxy for currency risks; however, in this article, 1-week Asian currency deposit rates are
employed to test the existence of time-varying currency risk premia.
15
Using bivariate GARCH model to price deviations from UIP for five major currencies, McCurdy and
Morgan (1991) find that their model on average can only explain 2.22% of the return variations in their sample.
Two possible explanations for their relatively low pseudo-R
2
s are that they assume PPP and thus ignore currency
risk as a potential source of risk premium, and they estimate the bivariate model for each currency instead of
estimating all currencies jointly with the world equity portfolio.
R
2
s indicate that the model perform very well in explaining the deviations from UIP. Based
on these empirical results, one can safely conclude that the predictable excess return puzzle
is due to the existence of time-varying risk premia and the sources of the risk premia come
not only from market risk, but also from currency risk.
6.2. Evidence of mean spillover and contagion in mean
After controlling the systematic currency and world market risks, I can then test the
contagion-in-mean effects. However, before that I need to the overall mean spillover
effects in the entire sample period, so any incremental mean spillover effects can be tested
during the crisis period. It can be seen from Table 4 that the joint hypothesis of no mean
Table 3
Hypothesis tests concerning prices of risks and predictability of conditioning variables
Null hypothesis Wald df P value
1. Are the prices of market and currency risks equal to zero?
H
0
: u
w
= u
c
=0; bc 2300.169 25 .000
Z
t 1
={CONSTANT,DIV,DUSTP,USDP,WD}
2. Are the prices of market and currency risks constant?
H
0
: u
w
= u
c
=0; bc 1001.038 20 .000
Z
t 1
={DIV,DUSTP,USDP,WD}
3. Are the prices of currency risks equal to zero?
H
0
: u
c
= 0; bc 1478.836 20 .000
Z
t 1
={CONSTANT,DIV,DUSTP,USDP,WD}
4. Are the prices of currency risks constant?
H
0
: u
c
= 0; bc 332.470 16 .000
Z
t 1
5. Is the price of market risk constant?
H
0
: u
w
= 0 151.039 4 .000
Z
t 1
6. Is the price of the Japanese yen risk constant?
H
0
: u
JP
=0 51.563 4 .000
Z
t 1
= DIV,DUSTP,USDP,WD
7. Is the price of the Hong Kong dollar risk constant?
H
0
: u
HK
=0 26.712 4 .000
Z
t 1
8. Is the price of the Singapore dollar risk constant?
H
0
: u
SG
= 0 91.051 4 .000
Z
t 1
9. Is the price of the New Taiwan dollar risk constant?
H
0
: u
TA
=0 76.311 4 .000
Z
t 1
10. Is there no predictability from excess dividend yield?
H
0
: u
m,k
= u
c,k
= 0; bc; k = DIV 101.186 5 .000
11. Is there no predictability from the change in term premium?
H
0
: u
m,k
= u
c,k
= 0; bc; k = DUSTP 110.935 5 .000
12. Is there no predictability from the U.S. default premium?
H
0
: u
m,k
= u
c,k
= 0; bc; k = USDP 106.507 5 .000
13. Is there no predictability from the world market portfolio?
H
0
: u
m,k
= u
c,k
= 0; bc; k = WD 169.883 5 .000
spillover (14) is rejected at the 1% level only for TA. To find out the sources of mean
spillover for TA, one can check statistical significance of individual mean spillover
coefficient, /, reported in Table 2. Table 2 indicates that the sources of mean spillover for
TA come from JP, HK, and SG, and SG appears to be the major market in generating
return shocks for the other three markets based on the hypothesis test (11) reported in
Table 4.
Now, I can test the contagion-in-mean effects. Table 4 shows that these effects are
statistically significant at the 1% level for three markets: JP, HK, and SG. For example, the
joint hypothesis of no contagion-in-return shocks for JP (H
0
:x
JP, j
=0; bj = HK,SG,TA)
during the crisis is strongly rejected by Wald statistic (Wald = 40.589) at the 1% level. The
same rejection also applies to HK and SG. To find out the sources of contagion-in-return
shocks for JP, one can examine the individual significance of contagion-in-mean
coefficient, x
JP, j
, reported in Table 2 based on robust standard errors. Basically, the
Table 4
Hypothesis tests concerning contagion in mean
1. Is there no mean spillover for JP?
H
0
: /
JP, j
=0; bj =HK,SG,TA 5.608 3 .132
2. Is there no mean spillover for HK?
H
0
: /
HK, j
= 0; bj = JP,SG,TA 3.106 3 .375
3. Is there no mean spillover for SG?
H
0
: /
SG, j
=0; bj =JP,HK,TA 3.744 3 .290
4. Is there no mean spillover for TA?
H
0
: /
TA, j
=0; bj = JP,HK,SG 47.796 3 .000
5. Is there no contagion in return shocks for JP?
H
0
: x
JP, j
= 0; bj = HK,SG,TA 40.589 3 .000
6. Is there no contagion in return shocks for HK?
H
0
: x
HK,
j =0; bj =JP,SG,TA 36.898 3 .000
7. Is there no contagion in return shocks for SG?
H
0
: x
SG, j
= 0; bj = JP,HK,TA 17.306 3 .000
8. Is there no contagion in return shocks for TA?
H
0
: x
TA, j
= 0; bj = JP,HK,SG 7.724 3 .052
9. Is there no mean spillover from JP?
H
0
: /
i,JP
=0; bi =HK,SG,TA 5.551 3 .135
10. Is there no mean spillover from HK?
H
0
: /
i,HK
=0; bi =JP,SG,TA 6.093 3 .107
11. Is there no mean spillover from SG?
H
0
: /
i,SG
= 0; bi = JP,HK,TA 31.310 3 .000
12. Is there no mean spillover from TA?
H
0
: /
i,TA
=0; bi =JP,HK,SG 5.217 3 .156
13. Is there no contagion in return shocks from JP?
H
0
: x
i,JP
= 0; bi = HK,SG,TA 3.477 3 .323
14. Is there no contagion in return shocks from HK?
H
0
: x
i,HK
= 0; bi = JP,SG,TA 12.682 3 .005
15. Is there no contagion in return shocks from SG?
H
0
: x
i,SG
= 0; bi = JP,HK,TA 42.669 3 .000
16. Is there no contagion in return shocks from TA?
H
0
: x
i,TA
= 0; bi = JP,HK,SG 51.321 3 .000
current returns in JP are affected by past return shocks in SG (x
JP,SG
= 0.573) and TA
(x
JP,TA
= 0.475), in addition to its own past return shocks (x
JP,JP
= 0.222). Similarly, the
current return shocks in HK are due to the past return shocks in SG (x
HK,SG
= 0.017)
and TA (x
HK,TA
=0.037), in addition to its own past return shocks (x
HK,HK
= 0.160).
The current return shocks in SG are due to the past return shocks in HK (x
SG,HK
= 3.170)
and TA (x
SG,TA
= 0.207), and finally, the current return shocks in TA are due to the past
return shocks in SG (x
TA,SG
= 0.172). By examining the significance of those individual
contagion-in-mean coefficients, one can conclude that basically all the contagion-in-mean
effects originate from three markets: HK, SG, and TA and none from JP, and this
conclusion has been confirmed by the hypothesis tests (1316) reported in Table 4.
Table 5
Hypothesis tests concerning contagion in volatility
1. Is there no volatility spillover for JP?
H
0
: g
i,HK
= k
i,SG
=l
i,TA
= m
i,WD
= 0; bi = JP 10.008 4 .040
2. Is there no volatility spillover for HK?
H
0
: m
i,JP
=g
i,SG
= k
i,TA
= l
i,WD
= 0; bi = HK 4.794 4 .309
3. Is there no volatility spillover for SG?
H
0
: l
i,JP
= m
i,HK
= g
i,TA
= k
i,WD
= 0; bi = SG 43.629 4 .000
4. Is there no volatility spillover for TA?
H
0
: k
i,JP
= l
i,HK
= m
i,SG
= g
i,WD
= 0; b= TA 30.587 4 .000
5. Is there no contagion in asymmetric volatility shocks for JP?
H
0
: p
i,HK
= q
i,SG
= s
i,TA
= v
i,WD
= 0; bi = JP 32.508 4 .000
6. Is there no contagion in asymmetric volatility shocks for HK?
H
0
: v
i,JP
= p
i,SG
=q
i,TA
= s
i,WD
= 0; bi = HK 5.632 4 .228
7. Is there no contagion in asymmetric volatility shocks for SG?
H
0
: s
i,JP
= v
i,HK
= p
i,TA
= q
i,WD
=0; bi =SG 83.254 4 .000
8. Is there no contagion in asymmetric volatility shocks for TA?
H
0
: q
i,JP
= s
i,HK
= v
i,SG
= p
i,WD
= 0; bi = TA 59.588 4 .000
9. Is there no volatility spillover from JP?
H
0
: m
HK, j
= l
SG, j
= k
TA, j
= g
WD, j
= 0; bj = JP 11.539 4 .021
10. Is there no volatility spillover from HK?
H
0
: g
JP, j
= m
SG, j
=l
TA, j
= k
WD, j
= 0; bj = HK 8.808 4 .066
11. Is there no volatility spillover from SG?
H
0
: k
JP, j
= g
HK, j
=m
TA, j
= l
WD, j
= 0; bj = SG 23.962 4 .000
12. Is there no volatility spillover from TA?
H
0
: l
JP, j
= k
HK, j
=g
SG, j
= m
WD, j
=0; bj =TA 14.385 4 .006
13. Is there no volatility spillover from WD?
H
0
: m
JP, j
=l
HK, j
= k
SG, j
= g
TA, j
= 0; bj = WD 30.392 4 .000
14. Is there no contagion in asymmetric volatility shocks from JP?
H
0
: v
HK, j
= s
SG, j
=q
TA, j
= p
WD, j
= 0; bj = JP 71.069 4 .000
15. Is there no contagion in asymmetric volatility shocks from HK?
H
0
: p
JP, j
= v
SG, j
= s
TA, j
= q
WD, j
= 0; bj = HK 28.590 4 .000
16. Is there no contagion in asymmetric volatility shocks from SG?
H
0
: q
JP, j
= p
HK, j
= v
TA, j
= s
WD, j
= 0; bj =SG 20.485 4 .000
17. Is there no contagion in asymmetric volatility shocks from TA?
H
0
: s
JP, j
= q
HK, j
= p
SG, j
= v
WD, j
= 0; bj = TA 10.754 4 .029
18. Is there no contagion in asymmetric volatility shocks from WD?
H
0
: v
JP, j
= s
HK, j
=q
SG, j
=p
TA, j
= 0; bj = WD 27.150 4 .000
6.3. Evidence of volatility spillover and contagion in volatility
Turning to volatility spillovers and contagion effects on the conditional variance of
excess currency returns, it can be seen from Table 5 that the hypothesis of no volatility
spillover (14) is rejected in all cases except HK. In particular, by examining the robust
standard errors of volatility spillover parameters reported in Table 2, it can be seen that the
Table 6
Summary statistics and residual diagnostics
JP HK SG TA WD
Full sample period
Predicted total risk premium (%) 0.046 0.006 0.029 0.037 0.079
S.D. 0.389 0.018 0.347 0.357 0.388
Market risk premium (%) 0.022 0.000 0.005 0.003 0.106
S.D. 0.057 0.002 0.020 0.036 0.335
Currency risk premium (%) 0.068 0.006 0.034 0.034 0.027
S.D. 0.384 0.018 0.347 0.355 0.195
Conditional volatility (%) 1.579 0.081 0.611 0.791 1.941
S.D. 0.414 0.037 0.419 0.805 1.406
Price of risk 2.101 62.013 1.681 3.137 2.195
S.D. 8.968 124.582 30.894 16.581 4.348
Crisis period
Predicted total risk premium (%) 0.199 0.005 0.214 0.012 0.019
S.D. 0.747 0.011 1.007 0.882 0.448
Market risk premium (%) 0.024 0.000 0.013 0.012 0.127
S.D. 0.067 0.002 0.045 0.080 0.398
Currency risk premium (%) 0.222 0.006 0.226 0.000 0.108
S.D. 0.761 0.011 1.008 0.895 0.327
Conditional volatility (%) 2.345 0.104 1.611 1.721 2.230
S.D. 0.742 0.029 0.704 1.397 0.659
Price of risk 2.972 63.730 9.037 1.084 2.052
S.D. 7.296 56.320 19.391 11.348 5.172
Residual diagnostics
BJ 61.243** 5425.881** 46.233** 1679.036** 395.827**
LB(20) 21.674 25.941 27.070 24.297 15.560
LB
2
(20) 17.300 68.825** 7.408 15.496 10.009
Pseudo-R
2
(%) 12.087 9.957 22.732 24.361 4.371
Engle and Ng (1993) asymmetric tests
Sign bias test 0.904 1.551 0.258 0.101 0.049
Negative size bias test 0.297 0.032 0.207 0.599 0.214
Positive size bias test 0.321 5.532** 0.338 0.946 1.749
The BeraJarque (BJ) tests normality based on both skewness and excess kurtosis and is distributed v
2
with
two degrees of freedom.
LB(20) and LB
2
(20) are the LjungBox test statistics of order 20 for serial correlation in the standardized
residuals and standardized residuals squared.
Pseudo-R
2
is computed as the ratio between the explained sum of squares and total sum of squares.
*Statistical significance at the 5% level.
**Statistical significance at the 1% level.
conditional variance of SG depends positively on lagged shocks in JP (l
SG,JP
= 0.054) and
negatively on lagged shocks in TA ( g
SG,TA
=0.161) and WD (k
SG,WD
= 0.046). For
TA, its conditional variance is negatively affected by lagged shocks in HK (l
TA,HK
=
2.126) and positively affected by SG (m
TA,SG
= 0.255). Although the hypothesis of
no volatility spillover for JP (H
0
: g
JP,HK
=k
JP,SG
= l
JP,TA
= 0) is rejected at the 5% level
( P value = .04), none of the volatility spillover parameters is statistically significant. The
same result also applies to HK. However, both their conditional variances are positively
affected by the lagged world market shocks (m
JP,WD
= 0.091; l
HK,WD
=0.005), implying
that the world equity market is the major source in generating the volatility shocks for JP
and HK.
It would be interesting to examine next where the dynamics of conditional variances of
these Asian currency returns behave differently during the crisis. In particular, I test
whether countries negative idiosyncratic shocks become contagious during the crisis after
controlling the overall volatility spillovers in the entire sample period. That is, I test
contagion-in-volatility hypothesis. The results are again reported in Tables 2 and 5.
As shown in Table 5, the joint hypothesis of no contagion-in-asymmetric volatility
shocks during the crisis is strongly rejected by Wald statistic at the 1% level in all cases
except HK. To investigate the possible sources of asymmetric volatility shocks, one can
test the individual significance of contagion-in-asymmetric volatility parameters reported
in Table 2 based on robust standard errors. For example, JP is affected positively by
negative shocks generated from SG ( q
JP,SG
= 15.325) and TA (s
JP,TA
= 4.601), in addition
to WD (v
JP,WD
= 10.450). SG is affected positively by the negative shocks from JP
(s
SG,JP
= 26.782), but negatively by the negative shocks from HK (v
SG,HK
= 7505.472).
Finally, TA is affected both positively by the negative shocks from JP ( q
TA,JP
= 32.096)
and HK (s
TA,HK
= 7813.75). Overall, the contagion effects in conditional second moments
are very significant in Asian foreign exchange markets, and lead/lag relationships appear
to be multidirectional between JP and SG, and between JP and TA since the negative
volatility shocks generated by JP are later transmitted to SG (TA), and vice versa.
However, this multidirectional lead/lag relationship does not apply to HK, SG, and TA
because negative shocks from HK are transmitted to SG and TA, but the negative shocks
generated from these two markets have no effect on HK. If one checks what really
happened during the 19971998 Asian crisis, it is not surprising to find this unidirec-
tional lead/lag relationship especially between HK and TA. Hong Kong dollar came first
under speculative attack in August 1997, and it was attacked again on October 17, 1997,
which forced the New Taiwan dollar to devalue. In June 1998, Hong Kong dollar suffered
the third speculative attack, and again in August 1998. At this time, government in
Taiwan enacted a 2-year, $5.7 billion economic stimulus package due to the slowing
growth.
Finally, in addition to the volatility spillover and contagion-in-volatility parameter
estimates shown in Table 2, the table also reports the estimates for GARCH and ARCH
parameters (a
j
, b
j
) and own asymmetric volatility shock parameters (d
j
). The GARCH
parameters are all significant at the 1% level, implying that all the conditional variance
processes are highly persistence. The asymmetric volatility shock parameters, d
j
, are
significant positive for JP and WD, implying that negative return shocks induce higher
conditional volatility than positive return shocks for these two markets.
6.4. Residual diagnostics
To access the fit of the conditional ICAPM with MGARCH(1,1)-M specification,
Table 6 reports the LjungBox statistics for 20th-order serial correlation in the level
(LB(20)) and squared standardized residuals (LB
2
(20)) as well as the asymmetry test
developed by Engle and Ng (1993). Under the multivariate framework, the standardized
residuals at time t is computed as Z
t
=H
t
1/2
e
t
, where H
t
1/2
is the inverse of the
Cholesky factor of the estimated variancecovariance matrix. The LjungBox statistics
show no serious linear and nonlinear dependencies for the standardized residuals of
excess currency returns, with one exception in which LB
2
(20) is significant at the 1%
level for HK. Thus, based on the LjungBox statistics, the volatility process is correctly
specified. However, as suggested by Engle and Ng, the LjungBox test may not have
much power in detecting misspecifications related to the asymmetric effects. For this
purpose, the set of diagnostics proposed by Engle and Ng are used.
16
These tests are
based on the news impact curve implied by a particular ARCH-type model used. The
premise is that if the volatility process is correctly specified, then the squared
standardized residuals should not be predictable based on observed variables. The
results reported in Table 6 show no evidence of misspecification. As for BJ test
statistics, they are all significant, indicating departures from normality, which justifies
the use of robust standard errors computed from using the QML method of Bollerslev
and Wooldridge (1992). Overall, the MGARCH(1,1)-M specification fits the data very
well.
6.5. Asian crisis and time-varying risk premia
One advantage of modeling the conditional second moments via MGARCH approach
is that it enables one to recover some interesting statistics, such as conditional volatility,
and, more importantly, the size of different risk premia. These interesting statistics will not
be available if one leaves the condition second moments unspecified, such as the pricing
kernel approach employed by Dumas (1993), Dumas and Solnik (1995), and Tai (1999).
17
Table 6 reports those statistics. For example, the predicted weekly total risk premium is
measured by
TRP
i;t
k
m;t1
h
im;t

c
k
c;t1
h
ic;t
; i JP; HK; SG; TA 18
16
Engle and Ng (1993) asymmetric tests include the sign bias, the negative size bias, and the positive size
bias tests. The sign bias test examines the impact of positive and negative innovations on volatility not predicted
by the model. The squared standardized residuals are regressed against a constant and a dummy S
t
that takes
the value of unity if e
t 1
is negative and 0 otherwise. The test is based on the t statistic for S
t
. The negative
(positive) size bias test examines how well the model captures the impact of large and small negative (positive)
innovations, and it is based on the regression of the squared standardized residuals against a constant and
S
t
e
t 1
((1 S
t
)e
t 1
). The computed t statistic for S
t
e
t 1
((1 S
t
)e
t 1
) is used in this test.
17
See the comments provided by Campbell Harvey in Dumas (1993).
and ranges from 0.046% for JP to 0.079% for WD. The TRP
i,t
can be decomposed into
two components: currency risk premium (CRP
i,t
) and market risk premium (MRP
i,t
). The
currency risk premium is measured by
CRP
i;t

c
k
c;t1
h
ic;t
; i JP; HK; SG; TA 19
and the market risk premium is measured by
MRP
i;t
k
m;t1
h
im;t
; i JP; HK; SG; TA: 20
The predicted average total risk premium and its variability are basically dominated by
the currency risk premia. For instance, the currency risk premium (and its standard
deviation) is 0.068% (0.384%) for JP, 0.006% (0.018) for HK, 0.034% (0.347%)
for SG, and 0.034% (0.355%) for TA. However, the predicted market risk premium and its
standard deviation are relatively small as compared to those of predicted currency risk
premium. This empirical evidence points out that an international asset pricing model
under PPP would not deliver economically significant risk premia in foreign exchange
markets.
In light of the result by Flood and Rose (1996) that the forward premium puzzle or
predictable excess return puzzle is less severe for countries with fixed exchange rates, it
will be interesting to examine whether the reported predicted risk premium confirm their
result.
18
If the exchange rate regime for a country is completely flexible (independent
float), we would expect a more severe predictable excess return puzzle in terms of grater
magnitudes in both the size and variability of the predicted risk premium (ex post
deviations from UIP). On the other hand, if the exchange rate regime is completely fixed
(hard peg), we should expect smaller magnitudes of the size and variability of the
predicted risk premium. However, if the exchange rate regime is between above two polar
regimes, such as managed float, we would expect that the magnitudes of the size and
variability of the predicted risk premium fall into the ranges of those for the two polar
regimes. In other words, the absolute size and the variability of predicted risk premium for
each country in general should positively related the degree of the flexibility of its
exchange rate regime. The predicted risk premium for each currency reported in Table 6
confirms our conjecture, and is consistent with the result by Flood and Rose. For example,
the exchange rate regimes for the countries for the period studied in this article range from
independent float (Japan), managed float (Taiwan and Singapore) to hard peg (Hong
Kong), and the size (variability) of their respective predicted total risk premium ranges
from 0.046% (0.389%), to 0.037% (0.357%), to 0.029% (0.347%), to 0.006%
(0.018%).
18
If covered interest parity holds, the predictable excess return puzzle will be the same as forward premium
puzzle since the deviations from UIP can be expressed as the difference between expected future spot rates and
current forward rates (i.e., forward bias or forward forecast error).
F
i
g
.
1
.
A
c
t
u
a
l
a
n
d
p
r
e
d
i
c
t
e
d
r
i
s
k
p
r
e
m
i
a
.
F
i
g
.
2
.
C
u
r
r
e
n
c
y
a
n
d
m
a
r
k
e
t
r
i
s
k
p
r
e
m
i
a
.
F
i
g
.
3
.
C
o
n
d
i
t
i
o
n
a
l
v
o
l
a
t
i
l
i
t
y
.
To see if the dynamics of predicted excess currency returns behave differently during
the crisis, the same statistics are calculated for the crisis period. As can be seen from the
table, the averages of the estimated total risk premia and their standard deviations change
dramatically, especially for JP, SG, and TA. For these three markets, the average of the
predicted risk premium decreases from 0.046% to 0.199% for JP, from 0.029% to
0.214% for SG, and from 0.037% to 0.012% for TA, and theses decreases are mainly
due to the drops in their respective currency risk premium components. Since the
estimated model is fully parameterized, the impact of the dynamics of the prices of
currency risk can be separated from the estimated covariances in determining the estimated
currency risk premia. Because the estimated covariances are on average positive during the
entire sample and the average price of each currency risk is negative except TA, the
estimated currency risk premia are negative, implying that investors are willing to give up
some of the total risk premium when the hedging value of the assets in the portfolio
becomes predominant. By comparing the prices of currency risk in two different sample
periods, it is not surprised to see that the currency risk premium decreases since the prices
of currency risk are lower during the crisis period. Another observation worth mentioning
is that the predicted risk premium for excess currency returns is basically dominated by the
currency risk premium; however, the predicted risk premium for the world equity market is
dominated by the market risk premium, suggesting the importance of incorporating
currency risk in pricing the deviations from UIP.
A useful complement to Table 6 is to display the time-series plots of those interesting
statistics. Fig. 1 contains the plots of actual risk premia (deviations from UIP) and
predicted risk premia. It can be seen that the dynamics of the predicted risk premia follow
very closely to those of actual risk premia, especially during the period of Asian crisis.
These close resemblances have been confirmed by the relatively high pseudo-R
2
statistics
reported in Table 6. Fig. 2 displays the plots of time-varying currency risk and market risk
premia. Clearly, the currency risk premia are not only more volatile but also higher than
the market risk premia. Finally, Fig. 3 contains the plots of conditional volatility. The
conditional volatility for each currency shows significant time variation, especially during
the crisis period.
7. Summary and concluding remarks
This article tests whether there are pure contagion effects in both conditional means
and volatilities of Asian foreign exchange markets during the 1997 Asian crisis. Previous
studies on contagion have failed to take into account the important distinction between
the two concepts of interdependence and contagion. Specifically, in this article, I define
contagion as significant spillovers of country-specific idiosyncratic shocks during the
crisis after economic fundamentals or systematic risks have been accounted for. To
control for the economic fundamentals, I rely on an ICAPM in the absence of PPP, which
provides a theoretical basis in selecting economic fundamentals. The economic funda-
mentals under ICAPM are the world market and foreign exchange risks, so the evidence
of contagion is based on testing whether idiosyncratic risksthe part that cannot be
explained by the world market and foreign exchange risksare significant in describing
the dynamics of conditional mean and volatility in Asian foreign exchange markets
during the crisis.
The empirical results indicate strong contagion effects in both the conditional means
and volatilities of those markets after systematic risks have been accounted for.
Specifically, the contagion-in-mean effects are mainly driven by the past return shocks
in Hong Kong, Singapore, and Taiwan. As for contagion in volatility, the lead/lag
relationships appear to be multidirectional among Japan, Singapore, and Taiwan, but
between Hong Kong and Singapore, and between Hong Kong and Taiwan, they are
unidirectional, with Hong Kong playing the dominant role in generating negative volatility
shocks. In addition, the conditional ICAPM with asymmetric MGARCH(1,1)-M structure
is able to explain/predict on average 17.28% of the return variations in those markets.
Therefore, this study provides a further evidence that the time-varying risk premium is a
very strong candidate in explaining the predictable excess return puzzle since the risk
premia founded in this article are not only statistically significant but also economically
significant.
References
Adler, M., & Dumas, B. (1983). International portfolio choice and corporate finance: A synthesis. Journal of
Finance, 38, 925984.
Backus, D. K., Gregory, A. W., & Telmer, C. I. (1993). Accounting for forward rates in markets for foreign
currency. Journal of Finance, 48, 18871909.
Bae, K. -H., Karolyi, G. A., & Stulz, R. (2000). A new approach to measuring financial contagion. NBER
Working Paper No. 7913.
Baig, T., & Goldfajn, I. (1999). Financial market contagion in the Asian Crisis. IMF Staff Papers, 46, 167195.
Baillie, R. T., & Bollerslev, T. (1990). A multivariate generalized ARCH approach to modeling risk premia in
forward foreign exchange rate markets. Journal of International Money and Finance, 9, 309124.
Bekaert, G., & Harvey, C. R. (1995). Time-varying world market integration. Journal of Finance, 50, 403444.
Bekaert, G., Harvey, C. R., & Ng, A. (2002). Market integration and contagion. Working paper.
Bekaert, G., & Hodrick, R. J. (1992). Characterizing predictable components in excess returns on equity and
foreign exchange markets. Journal of Finance, 47, 467508.
Bekaert, G., & Wu, G. (2000). Asymmetric volatility and risk in equity markets. Review of Financial Studies, 13,
142.
Bollerslev, T. (1990). Modelling the coherence in short run nominal exchange rates: A multivariate generalized
ARCH Model. Review of Economics and Statistics, 72, 498505.
Bollerslev, T., Engle, R. F., & Wooldridge, J. M. (1988). A capital asset pricing model with time-varying
covariance. Journal of Political Economy, 96, 116131.
Bollerslev, T., & Wooldridge, J. (1992). Quasi-maximum likelihood estimation and inference in dynamic models
with time varying covariances. Econometric Review, 11, 143172.
Calvo, S., & Reinhart, C. R. (1996). Capital flows to Latin America: Is there evidence of contagion
effects? In G. A. Calvo, M. Goldstein, & E. Hochreittter (Eds.), Private capital flows to emerging markets
( pp. 151171). Washington, DC: Institution for International Economics.
Cumby, R. E. (1988). Is it risk? Explaining deviations from uncovered interest rate parity. Journal of Monetary
Economics, 22, 279299.
De Santis, G., & Gerard, B. (1997). International asset pricing and portfolio diversification with time-varying
risk. Journal of Finance, 52, 18811912.
De Santis, G., & Gerard, B. (1998). How big is the premium for currency risk? Journal of Financial
Dornbusch, R., Park, Y. C., & Claessens, S. (1999). Contagion: How it spreads and how it can be stopped.
Mimeo, MIT.
Dumas, B. (1993). A test of the international CAPM using business cycles indicators as instrumental
variables. In J. Frankel (Ed.), The internationalization of equity markets ( pp. 2358). Chicago, IL:
University of Chicago Press.
Dumas, B., & Solnik, B. (1995). The world price of exchange rate risk. Journal of Finance, 50, 445479.
Ebrahim S. (2000). Volatility transmission between foreign exchange and money markets. Bank of Canada
Edwards, S. (1998). Interest rate volatility, capital controls and contagion. NBER Working Paper No. 6756.
Edwards, S., & Susmel, R. (1999). Interest rate volatility in emerging markets: Evidence from the 1990s. Mimeo,
UCLA and University of Houston.
Eichengreen, B., Rose, A. K., & Wyplosz, C. (1996). Contagious currency crises. NBER Working Paper
No. 5681.
Engel, C. (1996). The forward discount anomaly and the risk premium: A survey of recent evidence. Journal of
Empirical Finance, 3, 123192.
Engle, R. F., & Kroner, K. (1995). Multivariate simultaneous GARCH. Econometric Theory, 11, 122150.
Engle, R. F., & Ng, V. K. (1993). Measuring and testing the impact of news on volatility. Journal of Finance, 48,
17491778.
Engle, R. F., Ng, V. K., & Rothschild, M. (1990). Asset pricing with a FACTOR-ARCH covariances structure:
Empirical estimates for treasury bills. Journal of Econometrics, 45, 213237.
Ferson, W. E. (1995). Theory and testing of asset pricing models. In R. A. Jarrow, V. Maksimovic, & W. T.
Ziemba (Eds.), Finance, handbooks in operation research and management science, vol. 9 ( pp. 145200).
Amsterdam: North-Holland.
Ferson, W. E., & Harvey, C. R. (1993). The risk and predictability of international equity returns. Review of
Financial Studies, 6, 527567.
Ferson, W. E., & Harvey, C. R. (1994). Sources of risk and expected returns in global equity markets. Journal of
Banking and Finance, 18, 775803.
Flood, R. P., & Rose, A. K. (1996). Fixes: Of the forward discount puzzle. Review of Economics and
Statistics, 78, 748752.
Forbes, K., & Rigobon, R. (1998). No contagion, only interdependence: Measuring stock market co-movements.
NBER Working Paper No. 7267.
Forbes, K., & Rigobon, R. (1999). Measuring contagion: Conceptual and empirical issues. Mimeo, MIT.
Giovannini, A., & Jorion, P. (1989). The time variation of risk and return in the foreign exchange and stock
markets. Journal of Finance, 44, 307325.
Glosten, L. R., Jagannathan, R., & Runkle, D. (1993). On the relation between expected value and the volatility
of the nominal excess return on stocks. Journal of Finance, 48, 17791801.
Harvey, C. R. (1991). The world price of covariance risk. Journal of Finance, 46, 117157.
Kaminsky, G. L., & Peruga, R. (1990). Can time-varying risk premium explain excess returns in the forward
market for foreign exchange? Journal of International Economics, 28, 4770.
Kaminsky, G. L., & Reinhart, C. M. (2000). On crises, contagion, and confusion. Journal of International
Karolyi, G. A., & Stulz, R. (1996). Why do markets move together? An investigation of U.S. Japan stock return
comovements. Journal of Finance, 51, 951986.
Kroner, K. F., & Ng, V. K. (1998). Modeling asymmetric comovements of asset returns. Review of Financial
Studies, 11, 817844.
Lewis, K. K. (1994). Puzzles in international financial markets. NBER Working Paper No. 4951.
Longin, F., & Solnik, B. (1995). Is the correlation in international equity returns constant: 19701990? Journal
of International Money and Finance, 14, 326.
Mark, N. C. (1985). On time-varying risk premia in the foreign exchange market: An econometric analysis.
Journal of Monetary Economics, 16, 318.
Mark, N. C. (1988). Time-varying betas and risk premia in the pricing of forward foreign exchange contracts.
Journal of Financial Economics, 22, 318.
Masson, P. (1998). Contagion: Monsoonal effects, spillovers, and jumps between multiple equilibria. IMF
Working Paper WP/98/142.
McCurdy, T. H., & Morgan, I. G. (1991). Tests for a systematic risk component in deviations from uncovered
interest rate parity. Review of Economic Studies, 58, 587602.
Merton, R. C. (1973). An intertemporal capital asset pricing model. Econometrica, 41, 867888.
Park, Y. C., & Song, C. -Y. (1999). Financial contagion in the East Asian crisisWith special reference to the
Republic of Korea. Mimeo, Korea University.
Sachs, J., Tornell, A., & Velasco, A. (1996). Financial crises in emerging markets: The lessons from 1995. NBER
Solnik, B. H. (1974). The international pricing of risk: An empirical investigation of the world capital market
structure. Journal of Finance, 29, 365378.
Stulz, R. M. (1981). A model of international asset pricing. Journal of Financial Economics, 9, 383406.
Stulz, R. M. (1984). Pricing capital assets in an international setting: An introduction. Journal of International
Business Studies, 15, 5574.
Tai, C. S. (1999). Time-varying risk premia in foreign exchange and equity markets: Evidence from Asia-Pacific
countries. Journal of Multinational Financial Management, 9, 291316.
Tai, C. S. (2000). Time-varying market, interest rate, and exchange rate risk premia in the U.S. commercial bank
stock returns. Journal of Multinational Financial Management, 10, 397420.
Tai, C. S. (2001). A multivariate GARCH in mean approach to testing uncovered interest parity: Evidence from
Asia-Pacific foreign exchange markets. Quarterly Review of Economics and Finance, 41, 441460.
Van Rijckeghem, C. V., & Weder, B. (1999). Financial contagion: Spillovers through banking centers. Mimeo,
IMF.

ICAPM VECH Asian Market

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ICAPM VECH Asian Market

Uploaded by

Copyright:

Available Formats

Looking for risk premium and contagion in

Asia-Pacific foreign exchange markets

You might also like