Dimitris Kenourgios*
Department of Economics, University of Athens, 5 Stadiou Str., 10562 Athens,
Greece, Tel.: +30 210 3689449, e-mail: dkenourg@econ.uoa.gr
Aristeidis Samitas
Department of Business Administration, Business School, University of the Aegean, 8
Michalon Str., 82100 Chios, Greece, e-mail: a.samitas@ba.aegean.gr
Nikos Paltalidis
Faculty of Finance, Cass Business School, City University London, 105 Westbourne
Terrace, W2 6QT, Paddington, UK, e-mail: N.Paltalidis@city.ac.uk
Abstract
This paper proposes a new multivariate copula regime-switching model to capture
non-linear relationships in four emerging stock markets, namely Brazil, Russia, India,
China (BRIC) and two developed markets (U.S. and U.K.), during five recent
financial crises (the Asian crisis, the Russian crisis, the tech bust and the two episodes
in Brazil). We also investigate the presence of asymmetric responses in conditional
variances and correlations during these periods of negative shocks, applying the
asymmetric generalized dynamic conditional correlation (AG-DCC) approach. These
methods go beyond a simple analysis of correlation breakdowns, analyze the second
moment dynamics of stock market returns, and enable us to overcome problems
arising in the existing empirical research of financial contagion. Results provide
evidence that there is an asymmetric increase in dependence among stock markets
during all five crises. The jump in correlations and volatilities between stable and
crisis periods supports financial contagion when crises are spread to other markets.
Moreover, the differences in the magnitudes of the dependence changes show that the
industry-specific tech bust appears to have a larger impact than country-specific
crises. Furthermore, increases in tail dependence imply that the probability of markets
crashing together is higher during periods of financial turmoil. Our findings imply
that policy responses to a crisis are unlikely to prevent the spread among countries,
making fewer domestic risks internationally diversifiable when it is most desirable. A
further finding of this study is that the multivariate regime-switching copula model
captures higher level of dependence than the AG-DCC approach, possibly due to
econometric reasons.
*
Corresponding author: Dimitris Kenourgios, 5 Stadiou Street, 10562 Athens, Greece, e-mail:
dkenourg@econ.uoa.gr
The global extent of recent crises and the potential damaging consequences of being
correlations, beyond any fundamental link, has long been an issue of interest to
portfolio investment with important implications for portfolio allocation and asset
pricing. The past decade was marked by the Asian stock market crisis in 1997, the
Russian default in 1998, the Internet bubble collapse in 2000 and the Brazilian stock
market crash in 1997-1998 and 2002. According to Boyer et al. (2006), crisis is
spread from one country to the other generating the “contagion phenomenon”.
increase in correlation during the crash period). However, the existing literature
measuring correlations caused by volatility increases during the crisis. Second, there
due to the lack of availability of consistent and compatible financial data, especially in
the regressions, affecting at least in the second moments correlations and covariances
(Pesaran and Pick, 2006). Otherwise, a continued market correlation at high levels is
This study focuses on four major emerging stock markets, namely Brazil,
Russia, India and China (BRICs) and two developed stock markets, U.S. and U.K.
The relative importance of the BRICs as an engine of new demand growth and
spending power seems to shift more dramatically and quickly than expected. Higher
growth in these emerging market powerhouse economies could offset the impact of
greying population and slower growth in the advanced economies. The aim of this
paper is to investigate stock market contagion during five financial crises occurred the
last ten years: (i) the Asian stock market crisis in 1997, (ii) the Russian default in
1998, (iii) the collapse of the internet bubble of developed markets in 2000, (iv) the
Brazilian stock market crash in 1997–1998 and (v) the Brazilian crisis in 2002.
To avoid the limitations found in the existing literature, this paper proposes a
new multivariate copula function with Markov switching parameters. This approach
analyze the behaviour among stock markets when at least one of them is under
financial crisis. Following the methodology of Patton (2006a, b) and Bartram et al.
and the Gaussian copula for the joint distribution. Based on their work on time-
switching parameters. The dependence parameters in the copula function are modeled
2006). This process allows for series-specific news impact and smoothing parameters,
3
permits conditional asymmetries in correlation dynamics and accounts for
Furthermore, this model interprets asymmetries broader than just within the class of
GARCH models, since does not assume constant correlation coefficients over the
sample period. Ignoring the asymmetric frictions lead to overestimating the benefits
Similar to other studies in the literature (Forbes and Rigobon, 2002; Bekaert et
fundamentals, and the rapid increase in co-movements among markets during a crisis
episode. Given this definition, both methodologies applied in this paper allow us to
investors.
multivariate copula regime switching model for testing the contagion hypothesis; ii)
examining asymmetric contagion effects during the five extreme market downturns
applying the recently developed AG-DCC approach; iii) comparing the level of
dependence among stock markets obtained by using the above two methodologies; iv)
1
There is considerable disagreement regarding the definition of the fundamentals (for discussion on
this issue, see Bekaert et al. 2005). Our research uses multi-parameters that accommodate various
degrees of market co-movements.
4
shedding new light in the area of financial crises focusing our research on four major
emerging stock markets (BRICs) and two of the most dominant international stock
among stock markets during crises periods. Furthermore, the industry-specific tech
bust appears to have a larger impact than country-specific crises. Comparing the
results provided by the two methodologies, the multivariate regime switching copula
captures higher levels of dependence than AG-DCC model. Finally, results provide
support to the contagion phenomenon due to behavioral reasons rather than due to
methodology and the AG-DCC model, both applied to examine stock market
contagion. The data used for the empirical analysis is presented in Section 4. Section
2. Literature Review
Most financial decisions are based on the risk and return trade-off. A central issue in
asset allocation and risk management is whether financial markets become more
interdependent during financial crises. This issue has acquired great importance
among academics and practitioners in the last decade where five major crises were
obtained. Common to all these events was the fact that the turmoil originated in one
market extended to a wide range of markets and countries, in a way that was hard to
5
Generally, contagion refers to the spread of financial disturbances from one
increase in cross-market linkages after a shock to one country (or group of countries).
The study of financial contagion was conducted mostly around the notion of
“correlation breakdown”. Bertero and Mayer (1989) and King and Wadhwani (1990)
find evidence of an increase in stock returns’ correlation in 1987 crash. Calvo and
Reinhart (1996) report correlation shifts during the Mexican crisis. Baig and Goldfajn
(1999) find significant increases in correlation for several East Asian markets and
currencies during the East Asian crisis, supporting the contagion phenomenon.
correlation that do not take into account conditional heteroskedasticity. They argue
that the estimated correlation coefficient between the realized extreme values of two
random variables will likely suggest structural change, even if the true data generation
process has constant correlation. Forbes and Rigobon (2002) generalize the approach
of Boyer et al. (1999) and apply it to the study of three major crises (the 1987 crash,
the Mexican devaluation, and the East Asian crisis). They are unable to find evidence
and conclude that the phenomenon that has been labelled as “contagion” is non-
existent.
Evidence that equity returns have a dependence structure that is not consistent
with multivariate normality is presented by Lognin and Solnik (2001). They provide a
method based on extreme value theory and find that correlation generally increases in
periods of high-volatility of the U.S. market. Also, they find that correlation is
6
approach, Rachmand and Susmel (1998) and Ang and Bekaert (2002) estimate a
international conditional correlation between stock markets. They report that: (i)
regime; (ii) equity returns appear to be more correlated during downturns than during
upturns.
Boyer et al. (2006) use both regime switching model and extreme value theory
accessible and inaccessible stock indices. They identify that there is greater co-
movement during high volatility periods, suggesting that crises spread through the
However, Markov switching models have been limited to analyze the case of bivariate
dependence and pricing foreign exchange rate quanto options and multivariate
contingent claims)2. Patton (2006a) pioneers the study of time-varying copulas for
indices, following the methodology of Patton (2006a). Rodriquez (2007) uses a mixed
copula using returns from five East Asian stock indices during the Asian crisis and
from four Latin American stock indices during the Mexican crisis, and finds evidence
2
Cherubini et al. (2004) presents copulas applications in finance, while Bartram et al. (2007) provide
selective and recent international references.
7
A large literature investigate asymmetric effects in conditional covariances for
individual stocks, bonds, equity portfolios and stock market indices using different
approaches (e.g., Koutmos and Booth, 1995; Kroner and Ng, 1998; Christiansen,
and correlations. Chiang et al. (2007) apply a DCC model to nine Asian stock markets
from 1990 to 2003, confirming a contagion effect. Cappiello et al. (2006) extend the
original model of Engle (2002) along two dimensions: on the one hand, they allowed
for series- specific news impact and smoothing parameters and on the other hand,
group of countries (Europe, Australasia and North America), they provide evidence
Overall, the existing literature only partially explains the reaction of stock
stocks and the aggregate U.S. market, computed by Ang and Chen (2002), are
“asymmetric” (stronger during market downturns). As Bae et al. (2003) point out:
“The concerns (about contagion) are generally founded on the presumption that there
is something different about extremely bad events that leads to irrational outcomes,
excess volatility, and even panics. In the context of stock returns, this means that if
panic grips investors as stock returns fall and leads them to ignore economic
fundamentals, one would expect large negative returns to be contagious in a way that
8
3. Methodology: Conditional Copula and AG-DCC model
multivariate densities that are consistent with the univariate marginal densities.
Hence, they allow separation of the marginal distributions from the dependence
structure that is entirely represented by the copula function. This separation enables
marginal distributions that avoid the common assumption of normality for either
Copulas have certain properties that are very useful in the study of
random variables. Second and most important in the study of financial contagion,
asymptotic tail dependence is also a property of the copula. A switching copula can
capture increases in tail dependence, reflecting for example, that the probability of
dependence between two equity index returns, conditional upon the historical
employ the copula function with the return and volatility from one country against the
other. We estimate the univariate distributions and then the joining distributions. The
9
been developed and applied to financial market data by Patton (2006a, b) and Bartram
et al. (2007).
Let Xt and Yt be random variables that represent two returns for period t and
Gt(yt|Φt-1) respectively, with Φt-1 denoting all previous returns, i.e. {xt-1, yt-1, i>0}.
Moreover, let Mt and Nt be random variables that represent two volatilities for period
t and their c.d.f.s are Ft*(mt|Λt-1) and Gt*(nt|Λt-1) respectively, with Λt-1, denoting all
previous volatilities, i.e. {mt-i, nt-I,,i>0}. Define two further random variables by Ut =
interval from zero to one. Then the conditional copula density function, denoted by
Also, the conditional bivariate density functions of Xt , Yt, and Mt, Nt are given by the
product of their copula density and their two marginal conditional densities,
−
Introducing regime switches in variance, ut is modeled as u = g st * u t and
−
u t follows a GJR-GARCH-MA- t specification process (Glosten et al., 1993). The
level of the variance can occasionally change, depending on the values of gst, where
gst is a scaling parameter that changes in time as a function of a latent variable st. This
Chain:
10
p11 p12….pk1
P= p12 p22…pk2 ,
… … …
p1k p2k…pkk
where, pij= p(st=i|st-1=j) and st is regarded as the regime that the process is in at date t.
Hence, variable st can be in any K states at time t and q is the number of lags in the
state 2, and so on. Hamilton (1989) describes how to estimate the parameters in
through the maximization of a likelihood function and also how to do inference about
the state in which the process has been at date t. Inferences based on information up to
time t are called “filtered probabilities”, while inferences based on information from
in the markets considered in this paper follows a (2, 1) process. Although the selection
of the number of states and lags has been based on practical reasons of avoiding
specification tests on the copulas suggest that the GJR-GARCH-MA-t −switching (2,
In the first stage, the parameters of the marginal distributions parameters are
θ χ
≡ arg max ∑ log
t =1
f (x |Φ θ
t t t −1, x
)
(3)
11
^ n
θ y
≡ arg max ∑ log g ( y
t =1
t t
| Φ t −1,θ y )
t =1
θ = ⎡⎢θ x ;θ y ;θ c ⎤⎥
^ ^ ^ ^
(5)
⎣ ⎦
^
matrix of θ has to be obtained from numerical derivatives.
The conditional densities of equity index returns are leptokurtic and have variances
that are asymmetric functions of previous returns. Therefore, we obtain our marginal
distributions.
We let Ri,t and hi,t denote the return from equity index i and its conditional
variance for period t. Also, Li,t and λi,t denote the actual volatility from equity index i
and its conditional variance for period t. The ARCH model for the returns from index
i is defined by:
12
εi,t|Φt-1 ~ tvi (0,λi,t),
with si,t-1 = 1 when εi,t-1 is negative and otherwise si,t-1 = 0. In the first stage of
parameters estimation, all of the parameters, including the degrees of freedom vi, are
estimated separately for each equity index by maximizing the log-likelihood for each
Patton (2006a) proposes that current dependence between markets is explained by the
for the two assets. Following Patton (2006a) and Bartram et al. (2007), we suppose
that pt depends on the previous dependence pt-1 to capture persistence, and historical
absolute differences, |ut-i – vt-i|, i>0, to capture variation in the dependence process.
The intuition for the use of |ut-1 – vt-1| is the smaller (larger) the difference between the
realized cumulative probabilities, the higher (lower) is the dependence. So, equation
(8) describes an AR (2) model when extra assumptions are made, namely that a linear
function of the previous absolute difference, |ut-1 – vt-1|, provides a white noise
Our approach is different from Patton (2006a), since we use all historical
information about the absolute differences, rather than arbitrarily truncating the
use a constraint in the estimation procedure to keep the dependence process within
13
zero and plus one. The use of a logistic transformation function would unhelpfully
restrict the volatility of the dependence term when it is near its limiting values.
Volatilities and correlations measured from historical data may miss changes in risk.
which is assumed to be conditionally normal with mean zero and covariance matrix
Ht:
rt|Ћt-1~N(0,Ht) (9)
where, Ћt-1 is the time t-1 information set. All DCC class models use the fact that Ht
Ht = DtPtDt (10)
univariate GARCH models with √hit on the ith diagonal and Pt is the time-varying
correlation matrix.
stationary and assumes normally distributed errors (irrespective of the true error
distribution) can be used to model the variances (Engle and Sheppard, 2001). In this
1986), and the standardised residuals, εi,t = ri,t / √hi,t, are used to estimate the
14
correlation parameters4. The evolution of the correlation in the standard DCC model
[ ] [ ]
−
where, P = E[ε t ε 't ] and α and b are scalars such that α+b<1. Qt* = qiit* = qiit is a
diagonal matrix with the square root of the ith diagonal element of Qt on its ith
Pt = Qt*−1Qt Qt*−1 is a correlation matrix with ones on the diagonal and every other
element≤1 in absolute value. The model described by equations (11) and (12),
however, does not allow for asset-specific news and smoothing parameters or
asymmetries.
− _ − −
Qt = ( P − A' P A − B ' P B − G ' N G ) + A' ε t −1ε ' t −1 A + G ' n t −1n ' t −1 G + B ' Qt −1 B (13)
indicator function which takes on value 1 if the argument is true and 0 otherwise,
[ ]
−
while “ o ” indicates the Hadamard product) and N = E nt nt' . Equation (13) is the AG-
DCC model. In order the Qt to be positive definite for all possible realizations, the
− − − −
intercept, P − A ' P A − B ' P B − G ' N G must be positive semi-definite and the initial
covariance matrix Q0 positive definite. The scalar A-DCC model has the following
form:
4
The first stage of the DCC model building consists of fitting univariate GARCH specifications to each
of the return series and selecting the best one according to the Bayesian information criterion. For a
description of the three-stage estimation procedure, see Cappiello et al. (2006).
15
− _ − −
Qt = ( P − a 2 P − b 2 P − g 2 N ) + a 2ε t −1ε ' t −1 + g 2 nt −1n' t −1 + b 2Qt −1 (14)
_ −1 / 2 _ _ −1 / 2
α2+b2+δg2<1, where δ is the maximum eigenvalue [ P NP ] . If the matrices A,
B and G are assumed to be diagonal, the AG-DCC model reduces to the following
form:
− −
Qt = Po (ii ' − aa ' − bb ' ) − N o gg ' + aa ' o ε t −1ε 't −1 + gg ' o nt −1 n' t −1 +bb ' o Qt −1 (15)
where, i is a vector of ones and a, b and g are vectors containing the diagonal
4. Data
The data employed consist of stock indices for four developing countries, namely
Brazil, Russia, India and China and two developed (U.S. and U.K.) The stock indices
are the following: Bovespa index (Brazil), RTS index (Russia), BSE Sensex index
(India), Shanghai index (China), S&P500 (U.S.) and FTSE-100 (U.K.). We construct
weekly log returns, in U.S. dollar terms, dividend un-adjusted, and cross–country
differences in volatilities. This paper uses all historical information generated from
equity index returns, rather than arbitrarily truncating the historical information. This
point makes our approach different from Patton (2006a, b). Non-synchronous trading
can cause a bias in the estimation of correlations. Therefore, weekly data provide the
The sample period is from January, 2, 1995 till October, 31, 2006 and
excludes holidays. We split our data as follows: (i) Asian crisis: 1997; (ii) Brazilian
16
crisis: 1997-1998; (iii) Russian crisis: 1998; (iv) Internet collapse in developed
markets: 2000; (v) Brazilian crisis: 2002. Then, we compare the difference in returns,
actual and asymmetric volatilities and correlations between stable and crisis periods.
One difficulty in testing for contagion during these periods is that there is no a
single event to act as a definite catalyst behind the turmoil periods. For instance, the
Thai market declined sharply in June 1997, the Indonesian market fell in August
1997, and the Hong-Kong market crashed in mid-October 1997. Moreover, the
collapse of the Long Term Capital Management (LTCM) played an important role in
the Russian and the Brazilian crisis in 1998. Finally, the bust of the tech bubble in
2000 seems that began in March. However, a review of American and British
newspapers and periodicals during this period shows that there was a belief that the
market will recover soon to its all time high. It is obvious that contagion possibly
occurred during other periods of time. Therefore, this study takes into account a
period before and after the actual crash, using all historical information contained in
5. Empirical Results
(2001) to evaluate the goodness-of-fit of our marginal densities, specified by the GJR-
GARCH-MA-t model given by (6) and (7). The residual series pass the goodness-of-
5
Following Forbes and Rigobon (2002), a sensitivity analysis shows that period definition does not
affect the central results.
17
Table 1 reports the summary statistics for the data which exhibit the standard
properties of financial returns. All markets exhibit a positive mean return for the
whole sample period. As for equity returns, all are negatively skewed and leptokurtic
Table 2 shows estimates of the copula dependence model equity index returns.
The time-varying dependence model is estimated for each country index. Across all
equity indices, parameter β1 is always larger than 0.5, and even as high as 0.92,
implying high dependence persistence. The other autoregressive parameter, β2, is also
larger than 0.5, but smaller than β1. Comparing the parameters during normal and
crises periods, β1 and β2 appear to increase significantly during turmoil periods. This
negative and highly significant, indicating that the latest absolute difference of returns
to adjust their portfolios accordingly, since markets share higher correlation during
market crashes. Portfolio diversification across markets might be less useful than
crises periods. However, the differences in the magnitudes of the dependence changes
countries are much larger than those between BRICs and the developed U.S. and
U.K., suggesting that the emerging countries are more prone to financial contagion.
However, in the case of China, the jump in dependence is higher with the developed
markets. This may be due to the convergence of market consensus and index returns.
It seems that, given the increasing uncertainty in the markets, investors in the Chinese
18
market preferred to follow information and financial movements from developed
the valuation in the other markets that is driven by movements of the S&P500
increases significantly. The dependence between the U.S. and U.K. is the higher
shared during all crises examined (0.92). The bust of the tech bubble in 2000 makes
dependence among all markets to be constantly higher than 0.8. This implies that
stock market co-movements increased dramatically in 2000. One reason for this is the
leading role of the U.S. market, since is considered as a “global factor” influencing all
countries. Analyzing the crisis in the U.K. market over the same period, we also
U.K. and BRICs is lower when compared to the effect of the U.S. market on the
On the contrary, the Asian crisis in 1997 does not spread with the same
magnitude. Correlations between the Chinese and all other stock markets range from
0.53 (U.K.) to 0.643 (Brazil). However, the co-movement with its neighbored Indian
market is higher than 0.78, implying that the crisis spreads with a larger magnitude
due to regional factors. The Brazilian crisis in 1997-1998 also spreads with a large
magnitude, ranging from 0.783 (Russia) to 0.813 (U.S.). During the Brazilian crisis in
2002, stock market correlations are also high, but to a lesser extent compared to the
crisis in 1997-1998. During the Russian crisis (1998), co-movements between Russia
and other countries increase significantly. Spread of contagion is higher with the U.S.
(0.728) and Brazil (0.723). However, the correlation between Russia and Brazil may
be misleading since the Brazilian stock market was in crisis during 1998.
19
Table 3 shows results for the multivariate copula with switching regime
parameters during all five crises periods. At least two regimes exist (s11 and s22)
during each of the five crises periods. The maximum value of the restricted likelihood
(LF) shows that a crisis in one market affects the movement to the other markets.
Also, all variances (v) are significantly different from one. The shifts in variances
during crisis periods affect the value of a market portfolio, since betas become time-
dependence imply that the probability of markets crashing together is higher during
periods of financial turmoil. Figure 1 shows the regimes listed in Table 3. The higher
regime among stock markets observed during the 1997-1998 Brazilian crisis
(smoothed prob. 0.868), implying that this turmoil affected strongly the movements of
other markets.
The findings for dependence changes between stable and crises periods
suggest that all crises drive higher market dependence. While different economies
macro fundamentals. Equally, the policy responses to a crisis depend on the perceived
nature of transmission of shocks across the financial markets. Hence, it seems that
do not help to avoid a crisis. Contagion effects result when pessimism for stocks in
fundamentals’ development.
20
increase significantly during all crises, supporting the contagion hypothesis.
Furthermore, stock markets share the higher correlated level during the bust of the
The magnitude of the tech bust spread is higher between the U.S. and U.K.
markets (0.896). Furthermore, the effect of U.S. stock returns on BRIC stock markets
0.872 (Brazil). Correlations between the U.K. and BRICs are also significantly high
but to a lesser extent compared to the U.S. effect. The Asian turmoil appears to be
spread with a large magnitude between India and China (0.896), though not
surprisingly. On the contrary, low correlations appear between China and Russia
(0.523), India and Russia (0.539) during the Asian crisis, as well as between Russia
and India (0.527) and Russia and China (0.536) during the Russian turmoil. These
stock markets seem to follow weaker co-movements during the above two episodes.
Finally, the Brazilian crisis in 2002 spreads with a larger magnitude than the crisis
during 1997-1998. This result comes in contrast to copula findings regarding those
second moments. Results indicate that covariances are higher during negative shocks
than stable periods, supporting the contagion phenomenon6. During the more volatile
negative shocks. Overall, comparing the levels of dependence during normal and
crises periods, this paper finds, consistent with the existing literature and the findings
6
Conditional correlations between equity index returns, not presented here, are also highly increased
during crises periods. Results are available upon request.
21
of Cappiello et al. (2006), that during crises, stock market dependence is heightened
Table 6 reports results for testing whether we should adopt the symmetric or
the asymmetric DCC model. Four different parameterizations are estimated for the
dynamics of correlation. The first model is a standard scalar DCC model with no
(Equation 13, where matrices A and B are diagonal and matrix G is set equal to zero).
model (Equation 14) and the full diagonal version of a scalar A-DCC (Equation 15).
All parameters are significantly different from zero. The g2i term in the asymmetric
model is always higher than zero, implying the presence of asymmetric movements.
Furthermore, the asymmetric b2i term is always higher than the symmetric b2i,
providing further evidence to support the use of the asymmetric model in the study.
Therefore, results show that correlations among stock markets are much greater
markets obtained by the two methodologies. Although both models provide results
supporting the contagion hypothesis, they report different level of dependence. One
interpretation of this result is that the models may fail under specific circumstances to
capture fully asymmetric volatility and its potential effects on covariances and
multivariate regime switching copula, we find that the latter captures higher levels of
dependence. This may be due to its ability to connect the multivariate distribution to
22
its marginals in such a way that it captures the entire dependence structure in the
increases. As a result, the multivariate copula manages to separate out the temporal
dependence from the marginal behavior of the time series capturing higher levels of
6. Conclusions
The financial crises of the late 1990s prompted extensive empirical studies on
contagion. However, the existing literature suffers from certain limitations caused
during the crisis. This paper applies two methodologies to investigate stock market
contagion, using returns for BRIC countries as well as for the U.S. and U.K., during
five recent financial crises (the Asian, Russian, Tech bust and the two episodes in
pairs of sample countries, account for asymmetries in the second moments of stock
market returns and enable us to overcome the problems arising in the existing
markets during crises periods. We find that conditional correlations among stock
other markets with a big order of magnitude, while increases in tail dependence imply
that the probability of markets crashing together is higher during periods of financial
turmoil. Furthermore, the changes in dependence between BRICs are larger than those
23
between BRICs and the developed U.S. and U.K., suggesting that emerging markets
dynamics for all countries during crises periods, applying the AG-DCC model.
The AG-DCC results provide further evidence for higher joint dependence during
stock market crashes. When bad news hit stock markets, conditional equity
phenomenon. This implies that policy responses to a crisis are unlikely to prevent the
behavioural reasons. Moreover, the jump in correlations is higher during the tech
bubble crash implying that the industry–specific turmoil has a larger impact than the
methodologies, this paper finds that the multivariate copula regime switching model
captures higher level of dependence than the AG-DCC approach. This may be due to
the fact that the multivariate copula captures the entire dependence structure in the
multivariate distribution, managing to separate out the temporal dependence from the
This study shows that a crisis episode spreads with a large magnitude even
though some countries belong in different regional blocks. This finding has important
returns and volatilities exhibit higher correlations among stock markets (i.e. contagion
24
risks. As a result, an investment strategy focused solely on international
A possible explanation for the contagion effect during recent crises is that
investors retrenched from many emerging markets after the series of crises in the late
banks reduced their volume of short-term loans to emerging markets, while portfolio
investors also reduced their exposure to emerging markets due to the shift in their
attitudes generated from the loss of confidence. This created a “domino effect”,
Acknowledgement
The authors would like to thank an anonymous referee and the participants of the
European Financial Management Association Annual Conference, 27th – 30th June
2007, Vienna, Austria, for their valuable comments in an earlier version of this paper.
All remaining errors are the sole responsibility of the authors. Dimitris Kenourgios
would like to acknowledge financial support through the Operational Programme for
Education and Initial Vocational Training (O.P. "Education"), in the framework of the
project “Reformation of undergraduate programme of the Department of Economics /
University of Athens”.
References
Ang, A. and Bekaert, G., ‘International asset allocation with regime shifts’, Review of
25
Bae, K-H, Karolyi, G. A. and Stulz, R. M., ‘A new approach to measuring financial
market contagion’, Review of Financial Studies, Vol. 16, pp. 2003, pp. 717–
64.
Baig, T. and Goldfajn, I., ‘Financial market contagion in the Asian crisis’, IMF Staff
Bartram, S. M., Taylor, S. J. and Wang, Y-H, ‘The Euro and European financial
market dependence’, Journal of Banking & Finance, Vol. 31, 2007, pp. 1461-
81.
Bekaert, G., Harvey, C. R. and Ng, A., ‘Market integration and contagion’, Journal of
Journal of Business and Economic Statistics, Vol. 19, 2002, pp. 465-74.
stock markets around the crash of October 1987’, C.E.P.R. Discussion Papers,
No. 307.
Bollerslev, T., ‘Modelling the coherence in short run nominal exchange rates: a
Boyer, B. H., Kumagai, T. and Yuan K., ‘How do crises spread? Evidence from
accessible and inaccessible stock indices’, Journal of Finance, Vol. 66, 2006,
pp. 957-1003.
Calvo, S. and Reinhart, C. M., ‘Capital flows to Latin America: is there evidence of
contagion effects?’, in: Calvo, G., Goldstein, M., and Hochreiter, E. (eds.),
26
Private Capital Flows to Emerging Markets, Institute for International
Cherubini, U., Luciano, E. and Vecchiato, W., Copula methods in finance, John
Chiang, T. C., Jeon, B. N. and Li, H., ‘Dynamic correlation analysis of financial
479–507.
GARCH models’, Journal of Business and Economic Statistics, Vol. 20, 2002,
pp. 339–50.
Forbes, K. J. and Rigobon, R., ‘No contagion, only interdependence: Measuring stock
Glosten, R. T., Jagannathan, R. and Runkle, D., ‘On the relation between the expected
value and the volatility of the nominal excess return on stocks’, Journal of
27
Hamilton, J. D., ‘A new approach to the economic analysis of nonstationary time
series and the business cycle’, Econometrica, Vol. 57, 1989, pp. 357–84.
stock markets’, Journal of International Money and Finance, Vol. 14, 1995
pp. 747–62.
Patton, A. J., ‘Estimation of multivariate models for time series of possibly different
Pesaran, M. H., and Pick A., ‘Econometric issues in the analysis of contagion’,
Journal of Economic Dynamics & Control, Vol. 31, 2007, pp. 1245-77.
Ramchand, L. and Susmel, R., ‘Volatility and cross correlation across major stock
28
Table 1
Descriptive statistics
This table reports summary statistics for the 6 equity index returns (at weekly frequency, dividend un-
adjusted) during the period 1995-2006. All of the indices are denominated in USD.
29
Table 2
Estimates of the multivariate regime-switching copula dependence model for stock
market indices
Country ω β1 β2 γ LLF(c)
Crisis
period
Brazil 1997-1998 Russia 0.0468** 0.783** (0.639)* 0.702** (0.612)* -0.0371** 385.19
India 0.0392** 0.794** (0.608)* 0.715** (0.572)* -0.0316** 347.11
China 0.0386** 0.785** (0.539)* 0.710** (0.518)* -0.0349** 321.17
U.S. 0.0472 0.813** (0.697)** 0.802** (0.625)** -0.0303** 390.06
U.K. 0.0419 0.791** (0.630)** 0.784** (0.592)** -0.0455** 418.77
Russia 1998 Brazil 0.0374** 0.723** (0.639)* 0.680** (0.612)* -0.0347** 385.99
India 0.0348** 0.631** (0.544)* 0.607** (0.493)* -0.0388** 401.42
China 0.0365** 0.627** (0.537)* 0.601** (0.491)* -0.0479** 314.60
U.S. 0.0421 0.728** (0.642)** 0.716** (0.583)** -0.0597** 396.12
U.K. 0.0480 0.683** (0.611)** 0.628** (0.587)** -0.0521** 402.71
India 1997 Brazil 0.0368** 0.623** (0.608)* 0.597** (0.572)* -0.0738** 631.43
Russia 0.0372** 0.629** (0.544)* 0.615** (0.493)* -0.0583** 431.00
China 0.0829** 0.782** (0.679)* 0.751** (0.615)* -0.0694** 579.42
U.S. 0.0548 0.587** (0.511)** 0.560** (0.476)** -0.0428** 468.93
U.K. 0.0620 0.564** (0.503)** 0.540** (0.428)** -0.0676** 350.04
China 1997 Brazil 0.0576** 0.643** (0.539)* 0.619** (0.518)* -0.0686** 578.93
Russia 0.0527** 0.628** (0.537)* 0.620** (0.491)* -0.0517** 476.31
U.S. 0.0479 0.527** (0.409)** 0.516** (0.370)** -0.0438** 495.47
U.K. 0.0468 0.530** (0.428)** 0.518** (0.388)** -0.0691** 416.30
U.S. 2000 Brazil 0.0942** 0.874* (0.697)* 0.826* (0.625)* -0.0807** 682.22
Russia 0.0752 0.839* (0.642)** 0.810* (0.583)* -0.0641** 529.03
India 0.0740 0.816* (0.511)* 0.793* (0.476)* -0.0431** 368.89
China 0.0781 0.827** (0.409)** 0.798** (0.370)** -0.0588** 402.36
U.K. 0.104** 0.920* (0.587)* 0.845* (0.542)* -0.0836** 814.74
U.K. 2000 Brazil 0.089** 0.724* (0.630)* 0.718* (0.592)* -0.0732** 527.90
Russia 0.064 0.674* (0.611)** 0.639* (0.587)* -0.0613** 498.86
India 0.052 0.619* (0.503)** 0.600* (0.428)** -0.0539** 471.55
China 0.047 0.584* (0.428)** 0.566 * (0.388)** -0.0526** 470.03
Brazil 2002 Russia 0.0376 0.721** (0.639)* 0.689** (0.612)* -0.0582** 361.47
India 0.0437 0.708** (0.608)* 0.682** (0.572)* -0.0544** 421.73
China 0.0482 0.693** (0.539)* 0.658** (0.518)* -0.0371** 366.19
U.S. 0.0475 0.724* (0.697)** 0.700* (0.625)** -0.0466** 385.42
U.K. 0.0439 0.711* (0.630)** 0.695* (0.592)** -0.0427** 437.88
This table reports estimates of the multivariate copula dependence model during crises periods. The correlation
parameter pt is given by Equation 8. Estimates of parameters β1 and β2 during stable periods are reported in
parentheses.
LLF(c) is the maximum of the copula component of the log-likelihood function.
* and ** indicate significance at the 1% and 5% level, respectively.
30
Table 3
Brazilian crisis Russian default Asian crisis Tech bust Brazilian crisis
(1997-98) (1998) (1997) (2000) (2002)
This table reports results using the multivariate regime-switching copula for each stock market against
all others when at least one of them is under financial crisis. LF is the restricted likelihood function, v
is the variance, while s11 and s22 are two regimes. P-value shows the statistical significance of the
results.
31
Table 4
32
Table 5
33
Table 6
This table reports parameter estimates for the symmetric and asymmetric DCC GARCH models. Four different
parameterizations are estimated: a) a standard scalar DCC model with no asymmetric terms (Equation 11); b) a
symmetric diagonal version (Equation 13, where matrices A and B are diagonal and matrix G is set equal to zero;
c) an asymmetric scalar DCC model (Equation 14); d) the full diagonal version of a scalar A-DCC (Equation 15).
* indicates significance at the 5% level.
34
Figure 1
1
0,9 Brazilian crises
0,8
0,7 Russian crisis
0,6
0,5 Asian crisis - India
0,4
0,3 Asian crisis - China
0,2
0,1
Developed markets
0 crisis - U.S.
Jan. 1995
Jan. 1996
Jan. 1997
Jan. 1998
Jan. 1999
Jan. 2000
Jan. 2001
Jan. 2002
Jan. 2003
Jan. 2004
Jan. 2005
Jan. 2006
Developed markets
crisis - U.K.
35
Figure 2
1
0,8
Dependence
AG-DCC
0,6
0,4 Copula Switching
regime parameters
0,2
0
95
96
97
98
99
00
01
02
03
04
05
06
19
19
19
19
19
20
20
20
20
20
20
20
Date
36