Journal of Banking & Finance 35 (2011) 130141

Contents lists available at ScienceDirect

Journal of Banking & Finance

journal homepage: www.elsevier.com/locate/jbf

Global nancial crisis, extreme interdependences, and contagion effects: The role
of economic structure?
Riadh Aloui a, Mohamed Safouane Ben Assa a, Duc Khuong Nguyen b,*
a
LAREQUAD & FSEGT, University of Tunis El Manar, Boulevard du 7 novembre, B.P. 248, El Manar II, 2092 Tunis, Tunisia
b
ISC Paris School of Management, Department of Economics, Finance and Law, 22, Boulevard du Fort de Vaux, 75848 Paris Cedex 17, France

a r t i c l e i n f o a b s t r a c t

Article history: The paper examines the extent of the current global crisis and the contagion effects it induces by con-
Received 24 October 2009 ducting an empirical investigation of the extreme nancial interdependences of some selected emerging
Accepted 17 July 2010 markets with the US. Several copula functions that provide the necessary exibility to capture the
Available online 23 July 2010
dynamic patterns of fat tail as well as linear and nonlinear interdependences are used to model the
degree of cross-market linkages. Using daily return data from Brazil, Russia, India, China (BRIC) and
JEL classication: the US, our empirical results show strong evidence of time-varying dependence between each of the BRIC
F37
markets and the US markets, but the dependency is stronger for commodity-price dependent markets
G01
G17
than for nished-product export-oriented markets. We also observe high levels of dependence persis-
tence for all market pairs during both bullish and bearish markets.
Keywords: 2010 Elsevier B.V. All rights reserved.
Extreme comovements
Copula approach
BRIC emerging markets
Global nancial crisis

1. Introduction ature on measuring stock market comovements has been greatly

The modern portfolio theory, relying on the seminal work of
Markowitz (1952) and the underlying ideas of the Capital Asset
Pricing Model (CAPM), posits that investors can improve the per-
formance of their portfolios by allocating their investments into
different classes of nancial securities and industrial sectors that
would not move together in the event of a valuable new informa-
tion. Sub-perfectly correlated assets are thus particularly appropri-
ate for adding to a diversied portfolio. Subsequently, Solnik
(1974), among others, extends the domestic CAPM to an interna-
tional context and suggests that diversifying internationally
enables investors to reach higher efcient frontier than doing so
domestically.
Empirically, Grubel (1968) examines the potential benets of
international diversication and shows the superiority of portfolios
that are composed of both domestic assets and assets denominated
in foreign currencies from 11 developed markets. These ndings
are then conrmed by other earlier studies where the analysis of
market interdependence evolves both developed and developing
countries (Levy and Sarnat, 1970; Errunza, 1977). The recent liter-
* Corresponding author. Tel.: +33 140 539 999; fax: +33 140 539 898.
E-mail addresses: riadh.aloui@isg.rnu.tn (R. Aloui), safouane.benaissa@univ-
med.fr (M.S.B. Assa), dnguyen@groupeisc.com (D.K. Nguyen).
stimulated by the globalization of capital markets around the
world (Forbes and Rigobon, 2002; Gilmore et al., 2008; Abad
et al., 2010). Based on a wide variety of methodologies, the major-
ity of these works suggest that correlations of global stock returns
have increased in the recent periods as a result of increasing nan-
cial integration, leading to lower diversication benets especially
in the longer term. More importantly, the level of market correla-
tions varies over time.1
Modeling the comovement of stock market returns is, however,
a challenging task. The main argument is that the conventional
measure of market interdependence, known as the Pearson corre-
lation coefcient, might not be a good indicator. It represents only
the average of deviations from the mean without making any dis-
tinction between large and small returns, or between negative and
positive returns (Poon et al., 2004). Consequently, the asymmetric
correlation between nancial markets in bear and bull periods as
documented, for example, by Ang and Bekaert (2002), and Patton
(2004) cannot be explained.2 The Pearson correlation estimate is
further constructed on the basis of the hypothesis of a linear asso-
1
See Longin and Solnik (1995) and references therein.
2
By asymmetric correlations, we mean that negative returns are more correlated
than positive returns. This then suggests that nancial markets tend to be more
dependent in times of crisis. We also test this hypothesis within this paper.

0378-4266/$ - see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jbankn.2010.07.021
 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141
ciation between the nancial return series under consideration
whereas their linkages may well take nonlinear causality forms
(Beine et al., 2008). Other complications refer directly to stylized
facts related to the distributional characteristics of stock market
returns, in particular the departure from Gaussian distribution
and tail dependence (or extreme comovement). Solutions for han-
dling these problems include either the use of multivariate GARCH
models with leptokurtic distributions which allow for both asym-
metry and fat tails (S erban et al., 2007 ) or the use of multivariate
extreme value theory and copula functions (Longin and Solnik,
2001; Pais and Stork, in press). Notice that the rst modeling ap-
proach allows for capturing deviations from conditions of normal-
ity, while the last two approaches deal essentially with the
extreme dependence structure of large (negative or positive) stock
market returns, all in multivariate frameworks.
Since the investigation of dependence structure is crucial for
risk management and portfolio diversication, this paper also fo-
cuses on the issue of interactions between nancial markets. We
are particularly interested in modeling the co-exceedances of stock
market returns below or above a certain threshold.3 Our main
objective is thus to look at the margins of stock market return distri-
butions and test for both the degree and type of their dependence at
extreme levels conditionally on the possibility of extreme nancial
events (e.g., nancial turbulence, and crisis). Although we do not
explicitly search for the determinants of cross-market nancial
dependence, we think that differences in the economic structure
would be an important candidate for possible explanations and build
our intuition on the basis of some prevailing economic indicators.
For doing so, we combine the so-called conditional multivariate cop-
ula functions with extreme value theory as well as generalized auto-
regressive conditional heteroscedasticity process (hereafter extreme
value copula-based GARCH or EVC-GARCH models).4 In this nested
setting, the GARCH models with possibly skewed and fat tailed re-
turn innovations are applied to lter the stock market returns and
to draw their marginal distributions, while the multivariate depen-
dence structure between markets is modeled by parametric family
of extreme value copulas which are perfectly suitable for non-nor-
mal distributions and nonlinear dependencies. The model thus cap-
tures not only the tail dependence, but also the asymmetric tail
dependence (i.e., the strength of market dependence may be differ-
ent for extreme negative returns and for extreme positive returns).
With regard to the methodological choice, our work is broadly sim-
ilar to that of Jondeau and Rockinger (2006) who study the dynamics
of dependency between four major stock markets,5 but it is more
general in terms of GARCH specications and copula functions. In
addition, we demonstrate that portfolio managers will have an inter-
est in employing EVC-GARCH models to estimate the value at risk in
their internationally diversied portfolios during widespread market
panics.
This present study also contributes to the related literature in
that we provide a general framework for addressing the extent of
extreme interdependences and contagion effects between emerg-
ing and US markets, and among emerging markets themselves in
the context of the 20072009 global nancial crisis. This is impor-
3 6
The exceedance can be dened as the occurrence of an extreme return
observation, i.e., a return value that is below (extreme negative return) or above
(extreme positive return) a prespecied threshold of a nancial market at a certain
point in time (Teiletche and Xu, 2008). We then refer to the joint occurrence of
exceedances in two particular markets at the same point in time as co-exceedance,
which typically provides a measure of extreme comovement in nancial markets. See
also Christiansen and Ranaldo (2009) and Beine et al. (2010).
4
Copulas are functions that describe the dependencies between variables, and
enable modeling their joint distribution when only marginal distributions are known.
The main applications of copulas in nance can be found in Cherubini et al. (2004).
5
Jondeau and Rockinger (2006) focus on the US, the UK, German and French stock
markets represented respectively by the S & P500, FTSE, DAX and CAC40 indices.
131
tant since knowing only the degree of time-varying comovement is
actually not sufcient to make international investment decisions
because stock market returns might exhibit common extreme vari-
ations. A number of past studies have reported the existence of sig-
nicant linkages both between emerging and developed markets,
and among emerging markets (e.g., Gallo and Otranto, 2005 for
Asian emerging markets; Fujii, 2005 for Latin American emerging
markets), but little is known about their extreme comovements.6
For instance, the work of De Melo Mendes (2005) investigates
the asymmetric extreme dependence in daily log-returns for seven
most important emerging markets using extreme value copula
functions and shows some evidence of asymmetry in the joint
co-exceedances for the majority of 21 pairs of markets considered.
The cross-market tail dependence is also found to be stronger dur-
ing bear market. Caillault and Guegan (2005) apply the Student
and Archimedean copulas (Gumbel and Clayton) to daily data of
three Asian emerging markets over the period from July 1987 to
December 2002. They document that dependence structure is sym-
metric for ThailandMalaysia pair, while it is asymmetric for Thai-
landIndonesia, and MalaysiaIndonesia pairs. More recently, Hu
(2008) examines the tail dependence between the Chinese stock
market and the seven major developed markets by making use of
the Normal and Generalized JoeClayton copulas.7 The author re-
ports that time-varying dependence models are not always better
than constant dependence models and that the upper tail depen-
dence may be much higher than the lower tail dependence in some
short periods. Note that our study differs from De Melo Mendes
(2005) and Caillault and Guegan (2005) by allowing the marginal
distributions of stock market returns to follow appropriate GARCH
dynamics as well as the GARCH-in-Mean (GARCH-M) effects to
control for the risk-return trade-off. Compared to Hu (2008), our
GARCH specications are more exible since negative and positive
shocks can affect the conditional variance differently. Further, the
fact that we focus on the most important markets in the emerging
universe (Brazil, Russia, India and China) with their differing eco-
nomic systems allows us to shed light on the impact of economic
structure on the extreme nancial dependencies. Indeed, among
our BRIC markets, Brazil and Russia can be viewed as commod-
ity-price dependent countries, whereas India and China are n-
ished-product export-oriented countries. The comparison of
comovement levels among these markets is quite interesting be-
cause both commodity and nished-products prices have experi-
enced lengthy swings during recent times.
Using daily returns on stock market indices over the period
from March 22, 2004 to March 20, 2009, we mainly nd that
the GARCH-M specication which allows for asymmetric effects
from negative and positive shocks is the most appropriate for
the data, and that stock market volatility is highly persistent
over time. With regard to copula modeling, the Gumbel extreme
value copula appears to t at best the tail dependence of the
markets studied. More importantly, our results provide strong
evidence of extremely negative and positive co-exceedances for
all market pairs, but extreme comovement with the US is higher
for commodity-price dependent markets than export-price
In a recent study, Rodriguez (2007) uses a copula approach with regime-
switching parameters to model the dependence of daily returns from ve Asian
emerging markets, and four Latin American emerging markets during the 19971998
Asian crisis and the 19941995 Mexican crisis. The author nds evidence of changing
dependence during times of crisis. However, the methodology adopted basically
consists of tting the copulas-based regime-switching ARCH models to the whole
distribution of market returns, which is not the focus of our paper albeit the objective
is somewhat similar.
7
Hu (2008) considers the following developed markets: France, Germany, Hong
Kong, Japan, the United Kingdom, and the United States. Data are daily stock market
indices covering the period from January 1991 to December 2007. A comprehensive
description of the generalized JoeClayton copula can be found in Patton (2006).
132 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141
dependent markets. Within the universe of BRIC markets, the re-
sults indicate that they are more dependent in the bull markets
than in the bear markets.
The remainder of this paper is organized as follows: Section 2
presents the theoretical background of the copula functions and
shows how they can be applied to study the extreme comove-
ments between the BRIC markets and the US, especially over
the 20072009 period of the global nancial crisis. In Section
3, the empirical results are reported and interpreted with
reference to the economic structure of the emerging markets
considered. We provide summary of our conclusions in
Section 4.
2. Copula functions and their applications
Copulas are functions that link multivariate distributions to
their univariate marginal functions. A good introduction to copula
models and their fundamental properties can be found in Joe
(1997) and Nelsen (1999). Formally, we refer to the following
denition:
Denition 1. A d-dimensional copula is a multivariate distribution
function C with standard uniform marginal distributions.
Theorem 1 (Sklars theorem). Let X1, . . . , Xd be random variables
with marginal distribution F1, . . . , Fd and joint distribution H, then
there exists a copula C: [0, 1]d ? [0, 1] such that:
Hx1 ; . . . ; xd CF 1 x1 ; . . . ; F d xd :
Conversely if C is a copula and F1, . . . , Fd are distribution functions,
then the function H dened above is a joint distribution with margins
F1, . . . , Fd.
Therefore copulas functions provide a way to create distribu-
tions that model correlated multivariate data. Using a copula,
one can construct a multivariate distribution by specifying mar-
ginal univariate distributions, and then choose a copula to detect
a correlation structure between the variables. Bivariate distribu-
tions, as well as distributions in higher dimensions are possible.
If we are particularly concerned with extreme values, the con-
cept of tail dependence can be very helpful in measuring the
dependence in the tails of the distribution. The coefcient of tail
dependence is, in this case, a measure of the tendency of markets
to crash or boom together.
Let X and Y be random variables with marginal distribution
functions F and G. Then the coefcient of lower tail dependence
kL is
h i
kL lim Pr Y 6 G1 tjX 6 F 1 t
t!0
which quanties the probability of observing a lower Y assuming
that X is lower itself. In the same way, the coefcient of upper tail
dependence kU can be dened as
h i
kU lim

Pr Y > G1 tjY > F 1 t :
t!1
There is a symmetric tail dependence between two assets when
the lower tail dependence coefcient equals the upper one, other-
wise it is asymmetric. The tail dependence coefcient provides a
way for ordering copulas. One would say that copula C1 is more
concordant than copula C2 if kU of C1 is greater than kU of C2.
In order to measure the time-varying degrees of interdepen-
dence among markets, we employ an empirical method based on
the combination of copulas and extreme value theory. At the esti-
mation level, we will proceed as follows:
(i) We rst test the presence of ARCH effects in raw returns
using the ARCH LM test. Various GARCH specications that
allow for the leverage effect are estimated and compared
using the usual information criteria such as AIC, BIC and Log-
lik statistics. We choose the GARCH-M model as it gives the
best t. This model extends the basic GARCH model by
allowing the conditional mean to depend directly on the
conditional variance. The conditional variance specication
considered allows for a leverage effect, i.e. it may respond
differently to previous negative and positive innovations.
Instead of assuming normal distributions for the errors, we
use the Student-t distribution to capture the fat tails usually
present in the models residuals. The GARCH-M model may
be expressed as:
yt c kr2t t ;
1
r2t x ajet1 j  cet1 2 br2t1 ;
where c is the mean of yt and t is the error term which fol-
lows a Student-t distribution with m degrees of freedom. A
positive GARCH-in-Mean term k implies that higher risk is
positively associated with higher return. The conditional var-
iance equation depends upon both the lagged conditional
standard deviations and the lagged absolute innovations.
Here the GARCH model works like a lter in order to remove
any serial dependency from the returns.
(ii) We consider the innovations computed in step 1 and we t
the generalized Pareto distribution (GPD) to the excess
losses over a high threshold. We note that in the extreme
value theory (EVT) the tail of any statistical distribution
can be modeled by the GPD. The use of EVT is of great impor-
tance for emerging markets since they are signicantly inu-
enced by extreme returns (Harvey, 1995). The main
difference between emerging and developed markets resides
in the tail of the empirical distribution produced by extreme
events. More specically, stock returns from emerging mar-
kets have signicantly fatter tails than stock returns from
developed markets.
(iii) The uniform variates are obtained by plugging the GPD
parameter estimates into the GPD distribution function
and the following selected copula models belonging to the
extreme value copula family (the Gumbel, Galambos, and
Husler-Reiss copulas) are tted.8
The Gumbel Copulaof Gumbel (1960) is probably the best-known
extreme value copula. It is an asymmetric copula with higher prob-
ability concentrated in the right tail. By contrast, the Gumbel cop-
ula retains a strong relationship even for the higher values of the
density function in the upper right corner. It is given by
Cu; v expf ln ud  ln v d 1=d g; d P 1:
The parameter d controls the dependence between the vari-
ables. When d = 1 there is independence and when d ? +1, there
is perfect dependence. The coefcient of upper tail dependence
for this copula is
kU 2  21=d :
The Galambos copulaintroduced by Galambos (1975) is
Cu; v uv expf ln ud  ln v d 1=d g; 0 6 d < 1:
The HuslerReiss Copula introduced by Husler and Reiss (1987)
has the following form:
8
See Joe (1997) for complete references for Gumbel (1960), Galambos (1975), and
Husler and Reiss (1987).
 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141 133

       Z
1 1 e
u 1 1 ve fC n u; v  C hn u; v g2 dC n u; v
Cu; v exp  u
e / d ln v
e / d ln ; Sn n
d 2 ve d 2 e
u
0 6 d 6 1; and
Z 1
where / is a CDF of a standard Gaussian distribution, u e  lnu
Tn n fK n w  K hn wg2 dK n w:
and ve  lnv . Fig. 1 shows the contour plots of the selected cop- 0
ula models. These plots are very informative about the dependence
The rst statistics Sn measures how close the tted copula C hn is
properties of the copulas. For this reason, one often uses contour
from the empirical copula Cn, while the second statistics Tn mea-
plots to visualize differences between various copulas and possibly
sures the distance between an empirical distribution Kn and a para-
to assist in choosing appropriate copula functions.
metric estimation K hn of K. The null hypothesis that the copula C
In order to t copulas to our data, we use the method proposed
belongs to a class C0 is rejected for high values of the computed test
by Joe and Xu (1996) called inference functions for margins (IFM).
statistics. The p-values associated with the tests are computed
This method rst determines the estimate of the margin parame-
using a parametric bootstrap procedure and the validity of such
ters by making an estimate of the univariate marginal distributions
an approach is established in Genest and Rmillard (2008).
and then the parameters of the copula. The IFM method has the
advantage of solving the maximization problem for cases of high
dimensional distributions. 3. Empirical results
Two goodness-of-t tests are used to compare copula models.
These tests, qualied by Genest et al. (2009) as blanket tests, 3.1. Data and stochastic properties
are based on empirical copula and on Kendalls process.
Specically, the statistics considered use both the CramrVon We empirically investigate the interaction between various
Mises distances as stock market indices. Specically, the data consist of ve indices

Gumbel Copula with normal margins Gumbel Copula with t4 margins

0.001

0.002
2 0.02

10 0.003
0.004
0.04

0.005
0.006 0.08 0.06
0.007
0.008
1 0.10
0.009 0.12
0.010 0.14
0.011
0.16
5 0.012
0.013
0.18

0
y

0
-1

-5 -2

-5 0 5 10 -2 -1 0 1 2
x x

Galambos Copula with normal margins Galambos Copula with t4 margins

3
0.002
0.02
0.004
10 0.006 2 0.04
0.008 0.06
0.010 0.08
0.012 0.10
0.12
0.014 1 0.14
0.16
0.016 0.18
5 0.20
0.22
0.018
0.24
y

0.018
0.018
0
0
-1

-5 -2

-5 0 5 10 -2 -1 0 1 2 3
x x

Husler-Reiss Copula with normal margins Husler-Reiss Copula with t4 margins

0.001

0.002 2 0.02
0.003
10 0.004 0.04
0.005 0.06
0.006
0.007
0.008 1 0.12
0.10
0.08
0.009
0.010
0.011 0.16 0.14
0.012
5 0.013
0.18

0
y

0.014

0
-1

-5 -2

-5 0 5 10 -2 -1 0 1 2
x x

Fig. 1. Contour plots of copula models with normal margins and Students t margins with 4 df.
134 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141

Table 1
Descriptive statistics for daily stock market returns.

Brazil Russia India China US

Min 0.183 0.255 0.120 0.128 0.095
Max 0.166 0.239 0.088 0.140 0.110
Mean 6.783e004 2.161e004 2.627e004 3.607e004 2.59e004
Standard deviation 2.68e002 2.813e002 2.043e002 2.168e002 1.407e002
Skewness 0.431 0.513 0.640 0.049 0.348
Ex. kurtosis 7.836 17.28 4.573 6.404 13.198
Q (12) 26.346* 78.254* 55.870* 18.820*** 63.609*
Q2(12) 1584.60* 688.78* 709.44* 1080.83* 1650.27*
JB 3348.016* 16163.85* 1214.66* 2209.21* 9412.38*
ARCH(12) 505.211* 266.090* 238.288* 354.732* 454.186*

Notes: The table displays summary statistics for daily returns for the ve countries. The sample period is from March 22, 2004 to March 20, 2009. Q(12) and Q2(12) are the
JarqueBera statistics for serial correlation in returns and squared returns for order 12. ARCH is the Lagrange multiplier test for autoregressive conditional heteroskedasticity.
*
The rejection of the null hypotheses of no autocorrelation, normality and homoscedasticity at the 1% levels of signicance respectively for statistical tests.
**
The rejection of the null hypotheses of no autocorrelation, normality and homoscedasticity at the 5% levels of signicance respectively for statistical tests.
***
The rejection of the null hypotheses of no autocorrelation, normality and homoscedasticity at the 10% levels of signicance respectively for statistical tests.

representing four emerging economies, namely Brazil, Russia, In-
dia, and China, together with the US market index. All data are
the MSCI total return indices expressed in US dollars on a daily ba-
sis from March 22, 2004 to March 20, 2009. The returns are calcu-
lated by taking the log difference of the stock prices on two
consecutive trading days, yielding a total of 1304 observations.
To assess the distributional characteristics and stochastic prop-
erties of the return data, we must rst examine some descriptive
statistics reported in Table 1. The reported statistics show that all
the data series are negatively skewed and exhibit excess kurtosis,
which indicates evidence that the returns are not normally distrib-
uted. The JarqueBera statistics are highly signicant for all return
series and just conrm that an assumption of normality is unreal-
istic. Moreover, the Ljung-Box statistics (lags 12) suggest the exis-
tence of serial correlations in all the data series. Both the Ljung-Box
statistics for the squared returns and the ARCH LM test are highly
signicant, which indicates the presence of ARCH effects in all the
series.
Fig. 2 illustrates the variation of stock returns in ve markets.
From the graph, we can see that the stock prices were fairly stable
during the period from March 2004 to the third quarter of 2008.
After this date all return series displayed more instability due in
particular to the global nancial crisis.
Table 2 reports the unconditional correlations for all return ser-
ies. As expected, there is a positive correlation between the US and
BRIC markets. The highest correlation is between the US and Brazil
(0.63) and the lowest one is between US and China (0.20). The
same is true for emerging markets, although the ChinaIndia and
the RussiaBrazil markets are more correlated than other BRIC
markets with correlations of 0.57 and 0.53, respectively.

Brazil Russia
0.15
0.05
-0.15

-0.25

Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2
2004 2005 2006 2007 2008 2009 2004 2005 2006 2007 2008 2009

India China
0.10
0.04

-0.10
-0.12

Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2
2004 2005 2006 2007 2008 2009 2004 2005 2006 2007 2008 2009

US
0.08
-0.08

Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2 Q4 Q2
2004 2005 2006 2007 2008 2009

Fig. 2. Time paths of daily returns for Brazil, Russia, India, China and the US.
 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141 135

Table 2 the extreme value theory is less likely to be violated by the ltered
Unconditional correlations. series. Fitting the GPD to the ltered returns requires specication
Brazil Russia India China US of the lower and upper thresholds. Following De Melo Mendes
Brazil 1.000 0.537 0.363 0.424 0.639 (2005), we set the threshold values such that 10% of the residuals
Russia 1.000 0.391 0.449 0.306 are reserved for the lower left and the upper right quadrants (i.e.,
India 1.000 0.570 0.254 the empirical 10% and 90% quantiles in each margin). In what fol-
China 1.000 0.206 lows, we refer the tail dependence coefcient which describes how
US 1.000
large (or small) returns of a stock market occur with large (or
Notes: This table gives the unconditional correlation between daily returns of BRIC small) returns of another to as bull and bear markets (Nelsen,
and US markets. 1999; Longin and Solnik, 2001; Cherubini et al., 2004). To evaluate
the GPD t in the tails of the distribution, in Fig. 4 we show the qq-
plots of the upper and lower tail exceedances against the quantiles
Table 3 obtained from the GPD t. The approximate linearity of these plots
Estimates of the GARCH-M parameters for all returns series.
indicates that the GPD model seems to be a good choice.
Brazil Russia India China US Next, we consider the following market pairs: BrazilUS, Rus-
c 0.11e2 0.411e3 0.34e2 2.626e3 0.231e3 siaUS, IndiaUS and ChinaUS, and we estimate the copula model
(9.845e4) (1.285e4)** (9.083e4)* (3.585e3)** (3.187e4) parameters. Selection of the best copula t is based on the two
k 0.182 0.344 3.988 2.173 1.345 goodness-t tests discussed above. In Fig. 5, we show the scatter-
(7.925e1) (1.413e1)** (2.045)*** (3.359)*** (2.260)
plot of the markets pairs studied. The joint behavior of the returns
x 0.17e4 0.145e4 0.10e4 4.025e6 0.147e5
(4.507e6)* (4.087e6)* (2.495e6)* (1.452e6)* (3.830e7)* witnesses some extreme comovements in the lower left and upper
a 0.070 0.140 0.130 8.923e2 0.020 right quadrant.
(2.876e2)** (3.195e2)* (3.212e2)* (1.882e2)* (9.592e1) For each pair, the estimated parameters of the best copula mod-
b 0.868 0.837 0.817 0.897 0.928
el and the values of the lower and upper tail dependence coef-
(2.377e2)* (2.493e2)* (2.616e2)* (1.638e2)* (1.531e2)*
c 0.540 0.298 0.437 0.232 0.992
cients are reported in Table 4. The results are ordered according
(2.60e1)** (7.679e2)* (1.305e1)* (8.817e2)* (4.693e+1) to the value of their tail dependence coefcient. All the pairs con-
sidered are mutually dependent during bear and bull markets.
Notes: The table summarizes the GARCH-M estimation results. The values between
brackets represent the standard error of the parameters.
Most of the symmetric t are based on the Gumbel copula which
*
Signicance at the 1% levels. gives the best t in most cases. We rst note that the pair Bra-
**
Signicance at the 5% levels. zilUS is the strongest tail dependent pair for both positive and
***
Signicance at the 10% levels. negative co-exceedances. The second and the third position are
occupied by the RussiaUS and IndiaUS pairs; the ChinaUS mar-
3.2. Estimation results kets show the smallest degree of tail dependence. For ChinaUS,
we note that dependence during bull markets is stronger than
In the rst step we t by using the quasi-maximum likelihood dependence in bear markets.
estimation (QMLE) method a GARCH model for the return data. In Table 4 we also report the results for the joint losses and joint
Akaike Information Criterion (AIC), Bayesian Information Criterion gains for the following pairs: BrazilRussia, BrazilIndia, Brazil
(BIC) and the log-likelihood function are used to compare various China, RussiaIndia, RussiaChina and IndiaChina. We plot a scat-
specications of the GARCH models. The GARCH-M volatility mod- terplot of the considered markets pairs in Fig. 6. We observe that
el is indeed selected according to these statistics. The results of the the rst three positions are occupied by IndiaChina, BrazilRussia
GARCH-M tting are reported in Table 3.9 For all return series, we and RussiaChina. For BrazilRussia, BrazilChina, and RussiaChi-
note that the parameter b is close to 0.9 with an extremely signi- na the dependence in the left lower tail is smaller than the depen-
cant t-statistics. This result indicates that volatility is highly persis- dence in the right tail, i.e. these markets tend to comove more
tent, i.e. large changes in the conditional variance are followed by closely during bull markets than during bear markets.
other large changes and small changes are followed by other small
changes. Use of the asymmetric GARCH-M model seems to be justi-
3.3. Estimating the value at risk
ed. The estimated c coefcients for the leverage term are signicant
at the 5% level except for the US returns, which indicates the exis-
Financial institutions are exposed to risk from movements in
tence of an asymmetric response of volatility to shocks. The negativ-
the prices of many instruments and across many markets. To
ity and signicance of the GARCH-in-Mean terms estimated for
examine and measure market risk, the most commonly used tech-
Russia, India and China imply that high volatility leads to lower
nique is the value at risk (VaR), dened as the maximum loss in a
expected returns. The persistence of this market situation will cause
portfolio value of given condence level over a given time period.
the so-called ight-to-quality phenomenon, i.e. investment com-
During currency crises and stock market crashes, traditional VaR
munity will move away from stock markets to bond markets.
methods fail to provide a good evaluation of the risk because they
Fig. 3 shows the ACF of the standardized residuals obtained
assume a multivariate normal distribution of the risk factors. Here
from the GARCH-M t. When compared to the ACFs of the raw re-
we propose the use of copula to quantify the risk of three equally-
turns, it can be easily seen that the standardized residuals are
weighted portfolios. The market pairs considered are BrazilUS, In-
approximately i.i.d, thus far more suitable for GPD tail estimation.
diaChina, and BrazilRussia. These pairs showed the strongest
In the second step we extract the ltered residuals from each
extreme dependence during both bear and bull markets. We now
returns series with an asymmetric GARCH-M model, and then we
consider a portfolio composed of two assets; the one period log-re-
construct the marginal of each series using the empirical CDF for
turn for this portfolio is given by
the interior and the GPD estimates for the upper and lower tails.
The advantage of this method is that the i.i.d assumption behind R logk1 eX k2 eY ;

9
The estimation results for other GARCH specications are not reported here in
order to save spaces, but they are made available upon request addressed directly to
the corresponding author.
where X and Y denote the continuously compounded log-returns,
and k1 and k2 are the fractions of the portfolio invested in the two
assets.
136 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141
Brazil
0.8
ACF
0.4
0.0
0 5 10 15 20
Lag
India
0.8
ACF
0.4
0.0
0 5 10 15 20
Lag
US
0.8
ACF
0.4 0.0
0 5
Lag
Fig. 3. ACF of the standardized residuals.
In order to compute risk measures such as value at risk and
expected shortfall, we have to use Monte Carlo simulations be-
cause analytical methods exist only for a multivariate normal dis-
tribution (i.e., a Gaussian copula). When copula functions are used,
it is relatively easy to construct and simulate random scenarios
from the joint distribution of X and Y based on any choice of mar-
ginals and any type of dependence structure.
Our strategy consists of rst simulating dependent uniform
variates from the estimated copula model and transforming them
into standardized residuals by inverting the semi-parametric mar-
ginal CDF of each index. We then consider the simulated standard-
ized residuals and calculate the returns by reintroducing the
GARCH volatility and the conditional mean term observed in the
original return series. Finally, given the simulated return series,
for each pair (xi, yi) we compute the value of the global portfolio
R. The VaR for a given level q is simply the 100qth percentile of
the loss distribution, expressed analytically as
VaRq F 1
R q:
Consequently, the expected shortfall, dened as the expected
loss size given that VaRq is exceeded, is given by
ES ER j R > VaRq :
In order to assess the accuracy of the VaR estimates, a backtest
for the 95%, 99%, and 99.5% VaR estimates was applied. First, at
time t0 we estimate the whole model (GARCH + GPD + Copula)
using data only up to this time. Then by simulating innovations
from the copula we obtain an estimate of the portfolio distribution
and estimate the VaR using model (1). This procedure can be re-
peated until the last observation and we compare the estimated
VaR with the actual next-day value change in the portfolio. The
whole process is repeated only once in every 50 observations ow-
ing to the computational cost of this procedure and because we did
not expect to see large modications in the estimated model when
only a fraction of the observations is modied. However, at each
Russia
0.8
ACF
0.4
0.0
0 5 10 15 20
Lag
China
0.8
ACF
0.4
0.0
0 5 10 15 20
Lag
10 15 20
new observation the VaR estimates are modied because of
changes in the GARCH volatility and the conditional mean. If
selected models are well suited for calculating VaR, the numbers
of exceedances from these models should be then close to the
expected numbers. Notice that the expected number of excee-
dances at (1  a) condence level over a backtesting period of N
days is equal to a  N.
We started by estimating the model using a window of 750
observations. Then we simulate 5000 values of the standardized
residuals, estimate the VaR and count the number of the losses that
exceeds these estimated VaR values. At observations t = 800, 850,
. . . , 1300, we re-estimate the model and repeat the whole
procedure.
We also estimate the VaR using two other approaches: the var-
iancecovariance (also known as analytical) and the historical sim-
ulation methods. The rst approach estimates the VaR assuming
that the joint distribution of the portfolio returns is normal. The
VaR can be then computed as follows:
VaR l  za  r;
where l and r are the mean and the standard deviation of the port-
folio returns, and za denotes the (1  a) quantile of the standard
normal distribution for our chosen condence level. The main
advantage of this method is its appealing simplicity. However, it
suffers from several drawbacks. Among these, there is the fact that
it gives a poor description of extreme tail events because it assumes
that the risk factors are normally distributed. Also, the parametric
method inadequately measures the risk of nonlinear instruments,
such as options or mortgages.
The second approach considered is non-parametric, which
means that it does not require any distributional assumptions for
the probability distribution. The historical simulation estimates
the VaR by means of ordered LossProt observations. More gener-
ally, assume that we have N sorted simulated LossProt observa-
tions, then the VaR at the desired condence level (1  a)
 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141 137

Brazil Upper Tail Fit Russia Upper Tail Fit

0 1 2 3 4 5
Excess over threshold
Excess over threshold
0.0 0.5 1.0 1.5 2.0 2.5

0.0 0.5 1.0 1.5 2.0 2.5 0 1 2 3 4 5

GPD Quantiles GPD Quantiles

Brazil Lower Tail Fit Russia Lower Tail Fit

Excess over threshold
0 1 2 3 4 5

Excess over threshold

0 1 2 3 4 5 6
0 1 2 3 4 0 2 4 6 8
GPD Quantiles GPD Quantiles

India Upper Tail Fit China Upper Tail Fit

Excess over threshold

4
4

Excess over threshold

3

3
2

2
1

1
0

0.0 0.5 1.0 1.5 2.0 2.5 3.0 0 1 2 3 4

GPD Quantiles GPD Quantiles

India Lower Tail Fit China Lower Tail Fit

Excess over threshold

Excess over threshold

4
4

3
3

2
1 2

1
0

0 1 2 3 4 5 0 1 2 3 4 5
GPD Quantiles GPD Quantiles

US Upper Tail Fit

Excess over threshold
0 1 2 3

0 1 2 3 4
GPD Quantiles

US Lower Tail Fit

Excess over threshold
0 2 4 6

0 1 2 3 4 5 6
GPD Quantiles

Fig. 4. Estimated tails from GPD models t to lower and upper tail exceedances.

corresponds to the a  N th order statistics of the sample. Like the will naturally be omitted from the window and the estimated
parametric method, historical simulation may be very easy to VaR may change abruptly from one day to another.
implement, but it suffers from a major drawback related to its In case of the variancecovariance and historical simulation
dependence on a particular historical moving window. So when methods, the model parameters were updated for every observa-
running this method immediately after a special crisis, this event tion. The results for the backtesting are reported in Table 5. We
138 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141

0.10

0.10
0.05

0.05
MSCI US

MSCI US
0.0

0.0
-0.10

-0.10
-0.1 0.0 0.1 -0.2 -0.1 0.0 0.1 0.2
MSCI Brazil MSCI Russia
0.10

0.10
0.05

0.05
MSCI US

MSCI US
0.0

0.0
-0.10

-0.10
-0.10 -0.05 0.0 0.05 -0.10 -0.05 0.0 0.05 0.10 0.15
MSCI India MSCI China

Fig. 5. The scatterplot of four pairs of markets: BrazilUS, RussiaUS, IndiaUS and ChinaUS.

Table 4 Formally, let rt denotes the prot or loss on a portfolio over a

Copula parameters and tail dependence coefcients for dependent positive and xed time interval and VaRtjt1(a) denotes the ex ante VaR antici-
negative co-exceedances.
pated conditionally to an information set Xt1 and for an a% cover-
Pairs CopulaEstimates (standard errors) Upper tail age rate. Then, the hit variable It(a) associated with the ex post
Panel A: Copula parameters and upper tail dependence estimates observation of an a% VaR at time t is
BrazilUS Gumbel 1.66 (0.045) 0.482

IndiaChina Gumbel 1.420 (0.035) 0.370 1 if r t < VaRtjt1 a;
BrazilRussia Gumbel 1.370 (0.032) 0.341 It a
RussiaChina Gumbel 1.252 (0.028) 0.260 0 else:
RussiaIndia Gumbel 1.224 (0.027) 0.239
BrazilChina Gumbel 1.206 (0.027) 0.223
We rst evaluate the performance of our model by using Kupiec
BrazilIndia Gumbel 1.193 (0.026) 0.212 (1995)s proportion of failures (POF) test which focuses only on the
RussiaUS Gumbel 1.18 (0.025) 0.205 property of unconditional coverage. Under the null hypothesis that
IndiaUS Gumbel 1.12 (0.023) 0.152 the model is correct, the number of failures or exceptions X ob-
ChinaUS Gumbel 1.11 (0.029) 0.136
served in a sample of size T must follow a binomial distribution
CopulaEstimates (standard errors) Lower tail with parameter (T, a), and the Kupiecs Likelihood Ratio test statis-
Panel B: Copula parameters and lower tail dependence estimates tics conforms to the Chi-square distribution with 1 degree of free-
BrazilUS Gumbel 1.66 (0.046) 0.482 dom v2(1).
IndiaChina Gumbel 1.420 (0.035) 0.370
We also apply the dynamic quantile (DQ) test introduced by
BrazilRussia Galambos 0.638 (0.034) 0.337
RussiaChina Galambos 0.508 (0.031) 0.256 Engle and Manganelli (2004) to check whether the VaR model
RussiaIndia Gumbel 1.224 (0.026) 0.239 exhibit both correct unconditional coverage and serial indepen-
BrazilChina Galambos 0.457(0.031) 0.219 dence. More precisely, let Hit(a) be the de-meanded process on a
BrazilIndia Gumbel 1.193 (0.026) 0.212 associated with It(a) and consider a linear regression linking the
RussiaUS Gumbel 1.18 (0.026) 0.205
variable Hitt to its past, to the current VaR and any other variables,
IndiaUS Gumbel 1.129 (0.023) 0.152
ChinaUS Galambos 0.342 (0.029) 0.132 the null hypothesis can be tested by a Wald test for joint signi-
cance. In this study, we retain ve lags and the current VaR to carry
Notes: This table presents the copula parameters estimates and the tail dependence
out the test and to compute the DQ test statistics.
coefcients for dependent positive and negative co-exceedances. Standard errors
are given in parenthesis. Pairs are ranked according to the strength of their tail Table 5 reports the results from backtesting procedure as well
dependence coefcient. as the corresponding p-values of the POF statistics to test the null
hypothesis of correct unconditional coverage and the DQCC statis-
tics to test the null hypothesis of conditional efciency, using the
can see that copula models outperform both the alternative mod- 99% VaR condence level and a sample size of 250 observations.
els, and provide more accurate estimate of the VaR at the 95%, According to Kupiec backtest, the EVC-GARCH model performs
99% and 99.5% condence intervals. well for all considered portfolio. The reported p-values of the DQ
Testing the accuracy of VaR models is a critical issue in the test are satisfactory for the BrazilUS and BrazilRussia pairs and
acceptance of internal models for market risk management. How- coherent with the results obtained from Kupiecs test. The only
ever, the validation of risk models empirically is still a challenge. exception is for the IndiaChina portfolio, since the p-value of
Lopez and Saidenberg (2001) report some of the difculties in eval- the DQ test is inferior to the conventional signicance level.
uating the quality of the VaR estimate produced by these models. Overall, the ndings lead to not reject the accuracy of our VaR
Christoffersen (1998) points out that an accurate VaR model must model. Effectively, the number of exceedances from the copula-
produce a hit sequence with the unconditional coverage and inde- based GARCH model (Panel A) is closer to the expected number
pendence properties. than for alternative models, while both statistic tests (Panel B)
 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141 139

0.2

MSCI India
MSCI Russia

0.0
0.0
-0.2

-0.10
-0.1 0.0 0.1 -0.1 0.0 0.1
MSCI Brazil MSCI Brazil
MSCI China

MSCI India
0.10 0.0

0.0
-0.10

-0.10
-0.1 0.0 0.1 -0.2 -0.1 0.0 0.1 0.2
MSCI Brazil MSCI Russia
MSCI China
0.10 0.0

0.10 0.0
MSCI China
-0.10

-0.10
-0.2 -0.1 0.0 0.1 0.2 -0.10 -0.05 0.0 0.05
MSCI Russia MSCI India

Fig. 6. The scatterplot of ve pairs of markets: BrazilRussia, BrazilIndia, BrazilChina, RussiaIndia, RussiaChina, and IndiaChina.

Table 5
More precisely, a combination of large human capital, competitive
VaR backtesting results.
work force, access to natural resources, and a sustainable revitali-
a = 0.05 a = 0.01 a = 0.005 zation of internal demand has substantially increased the role of
Panel A the BRIC economies in the global economy. Today, they collectively
Copula account for nearly 30% of global output, and only China and India
BrazilUS 0.08844 (49) 0.02707 (15) 0.01083 (6)
contribute about 1.16% and 0.41% to the global GDP growth accord-
IndiaChina 0.08664 (48) 0.02527 (14) 0.01805 (10)
BrazilRussia 0.07581 (42) 0.01805(10) 0.01083 (6)
ing to the IMFs World Economic Outlook Report of April 2008.10
Growth projections for these economies in the coming years are
Historical simulation
BrazilUS 0.12635 (70) 0.05234 (29) 0.02707 (15)
also substantially above the average growth of the world and
IndiaChina 0.13176 (73) 0.03790 (21) 0.01805 (10) developed economies, leading economists and experts to expect
BrazilRussia 0.11010 (61) 0.03068 (17) 0.01624 (9) that, based on their potential of internal demand expansion and
Variance covariance spending power, they could provide a cushion against slower
BrazilUS 0.13718 (76) 0.07942 (44) 0.06137 (34) growth in the global economy.
IndiaChina 0.13176 (73) 0.07220 (40) 0.05956 (33) On the other hand, the rationales for equity and foreign direct
BrazilRussia 0.10469 (58) 0.05234 (29) 0.04332 (24)
investments rely particularly on the specicities of the BRIC mar-
Kupiec test DQ test kets, which can be considered as traditional emerging markets,
compared for example to those of the most advanced emerging
Panel B
BrazilUS 0.102 0.1304 markets (e.g., Mexico, South Korea, and Taiwan). The BRIC markets
IndiaChina 0.3701 1.17e9 have experienced spectacular increase in size as measured by ra-
BrazilRussia 0.9919 0.9976 tios of stock market capitalization to GDP over the recent years,
Notes: This table reports the VaR backtesting results obtained from EVC-GARCH, are less correlated with developed markets, and display higher idi-
historical simulation and variance covariance method (Panel A) as well as the p- osyncratic risk due to the low level of their market sensitivities to
values for the Kupiec (1995)s test and the DQ test of Engle and Manganelli (2004) global factors. Additionally, it is worth noting that growing internal
using ve lags and the current VaR as explicative variables. A model which is said to markets are helping the BRIC countries to substantially reduce
be best suited for calculating VaR refers to the one with the number of exceedances
their strong dependence on external markets such as the European
(given in brackets) closest to the expected number of exceedances. The latter are
37.5, 7.5, and 3.75 for 95%, 99% and 99.5% condence levels, respectively. Union, and the US, and thus their vulnerability to external shocks.
As a result, this creates incentives for investors to consider dedi-
cated strategies of asset allocations to the BRIC markets.
are, in almost all cases, unable to reject the null hypothesis we Although the BRIC markets have many features in common as
examine at the conventional signicance levels. discussed previously, the empirical evidence we report here indi-
cates that they do not behave similarly in regard to their nancial
linkages with the US. If we refer to Table 4, we observe that ex-
3.4. The effects of economic structure on nancial comovements treme nancial dependency on the US during the 20072009 glo-
bal nancial crisis is much stronger for Brazil and Russia than for
Among emerging markets, the BRIC group of markets represents
a different and dynamic set of investment opportunities. On the
one hand, their economic rationales are straightforwardly linked 10
The contribution of the US, Euro area, Japan and other developed countries
to their size and their contributions to global economic growth. reached only 1.53%, while other developing countries accounted for 1.76%.
140 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141

Table 6 4. Concluding remarks

Trade and nancial characteristics of BRIC markets.

Trade proles Brazil Russia India Chine World Studies of the transmission of return and volatility shocks
Exports of 80470 276857 52847 80741 3782058 from one market to another as well as studies of the cross-market
commodities correlations are essential in nance, because they have many
% of total exports 51.8 79.5 36.0 6.6 28.4 implications for international asset pricing and portfolio alloca-
Exports of 74903 71186 93859 1134733 9520232
tion. Indeed, a higher degree of comovement (or correlation) be-
manufactures
% of total exports 48.2 20.5 64.0 93.4 71.6 tween markets would reduce the diversication benets and
Total exports 155373 348043 146706 1215474 13302290 imply that at least a partially integrated asset pricing model is
% of World total 1.2 2.6 1.1 9.1 100.0 appropriate for modeling the risk-return prole of the assets is-
exports sued by the considered countries. With the advent of the current
Total imports 112538 220740 198236 953489 13752278
global nancial crisis in the aftermath of the US housing market
% of World total 0.8 1.6 1.4 6.9 100.0
imports failures, not only academic researchers but also investors and pol-
Trade to GDP ratio (%) 25.8 54.1 44.9 71.3 icymakers have shifted their attention to the extreme dependence
Relative stock market 102.8 116.5 154.6 184.1 structure of nancial markets. This is explained by their shared
size (%)
concerns regarding the harmful consequences of contagion
Turnover ratio (%) 56.2 63.9 95.9 197.5
effects.
Notes: This table presents the trade proles and some nancial characteristics of This paper employs a multivariate copula approach to examine
BRIC markets in 2007. The total exports of commodities products are equal to the the extreme comovement for a sample composed of four emerging
sum of the exports of agricultural, and fuel and mining products. The trade to GDP
ratios represent the average of trade to GDP ratios over the period 20052007.
markets and the US markets during the 20042009 period. The use
Relative stock market size refers to the market capitalization of listed companies in of this method is advantageous in that it satisfactorily captures the
percentage of GDP. Turnover ratio corresponds to the value of shares traded in tail dependencies between the markets studied, when univariate
percentage of market capitalization. Data are expressed in millions of US dollar and distributions are complicated and cannot be easily extended into
obtained from World Trade Organization statistics database and World Develop-
a multivariate analysis (Jondeau and Rockinger, 2006). The copula
ment Indicators database.
functions also provide an interesting alternative to the traditional
assumption of jointly normal distribution series, which appears
to be unrealistic given the stochastic properties of the return data.
China and India; meanwhile the latter have established important We rst provide evidence of the superiority of a Student-t
trade links with the world economy. Table 6 shows that both China GARCH-in-Mean specication which allows for leverage effects in
and India have a high degree of economic openness with trade to explaining the time-variations of daily returns on stock market
GDP ratios of 71.3% and 44.9%. In particular, the shares of Chinas indices. When calibrating several well-known copulas based on
exports and imports in the world total trade activities are 9.1% the marginal distributions of the ltered returns from the selected
and 6.9%, respectively. A careful inspection of the trade proles GARCH model, we nd evidence of extreme comovement for all
of the markets studied reveals that Brazil and Russia are more market pairs both in the left (bearish markets) and right tails (bull-
dependent on the revenues from exports of commodities products ish markets). Further, the results suggest that dependency on the
(51.8% and 79.5% of the respective economies total exports), US is higher and more persistent for Brazil and Russia countries
whereas the economic performance of China and India depends which are highly dependent on commodity prices than for China
greatly on exports of manufactured products (93.4% and 64.0% of and India whose economic growth is largely inuenced by n-
the respective economys total exports).11 In other words, coun- ished-products export-price levels. Finally, the extreme depen-
tries with higher sensitivity to commodity-price changes tend to dence between emerging market pairs is found to be generally
comove closely with the US in both bull and bear markets. Our re- smaller in bearish markets than in bullish markets, which might
sults contrast, to some extent, the ndings of Teiletche and Xu indicate a low probability of simultaneous crashes. As a practical
(2008) according to which trade linkage variables such as import exercise to check the usefulness of the copula models developed
demand and trade competition appear as the most important in the paper, we estimate the value at risk for three equally-
determinants of extreme dependence between nancial markets. weighted portfolios for three couples of countries exhibiting tigh-
Taken together, we think that the heterogeneity of the BRICs ter extreme comovements over the study period. The results
economic structure and especially the trade proles could, to this indicate that the extreme value copula-based VaR model outper-
extent, be a prevailing explanatory factor for the cross-market ex- forms the analytical approach and historical simulation method.
treme interdependences. It would be, for future research, interest- Undoubtedly, copula models t at best nancial data during wide-
ing to quantify the impact of different types of economic structure spread market panics and frictions, where the approximations of
on market comovement by running cross-sectional studies. Of the usual probability distributions are likely to be strongly biased.
course, an equal attention should be given to nancial characteris- Given the increasing interest in detecting potential gains from
tics of sample markets since they also constitute important chan- international portfolio diversication in a globalized context, further
nels for nancial contagion. Extreme nancial dependence is for investigation of stock market relationships is needed. Future exten-
example expected to be stronger for countries with deeper and sions of this work could focus on an explicit explanation of extreme
more liquid nancial markets. However, one can remark that nancial interdependence, using country-specic fundamentals.
nancial variables might be less relevant in view of the two market
indicators we show in the lower part of Table 6. For instance, stock
markets in both China and India are less dependent on the US References
shocks, while they have higher ratios of market capitalization
and liquidity to GDP. Abad, P., Chuli, H., Gmez-Puig, M., 2010. EMU and European government bond
market integration. Journal of Banking and Finance 34, 28512860.
Ang, A., Bekaert, G., 2002. International asset allocation with time-varying
correlations. Review of Financial Studies 15, 11371187.
Beine, M., Capelle-Blancard, G., Raymond, H., 2008. International nonlinear causality
11
The total commodities exports include the exports of agricultural, and fuels and between stock markets. European Journal of Finance 14, 663686.
mining products. The interested readers can refer to the WTOs statistics database for Beine, M., Cosma, A., Vermeulen, R., 2010. The dark side of global integration:
a complete list of the highest-value exported products for BRIC countries. Increasing tail dependence. Journal of Banking and Finance 34, 184192.
 R. Aloui et al. / Journal of Banking & Finance 35 (2011) 130141
Caillault, C., Guegan, D., 2005. Empirical estimation of tail dependence using
copulas: Application to Asian markets. Quantitative Finance 5, 489501.
Cherubini, U., Luciano, E., Vecchiato, W., 2004. Copula Methods in Finance. John
Wiley & Sons, New York.
Christiansen, C., Ranaldo, A., 2009. Extreme coexceedances in new EU member
states stock markets. Journal of Banking and Finance 33, 10481057.
Christoffersen, P., 1998. Evaluating interval forecasts. International Economic
Review 39, 841862.
De Melo Mendes, B.V., 2005. Asymmetric extreme interdependence in emerging
equity markets. Applied Stochastic Models in Business and Industry 21, 483
498.
Engle, R.F., Manganelli, S., 2004. CAViaR: Conditional autoregressive value at risk by
regression quantiles. Journal of Business and Economic Statistics 22, 367381.
Errunza, V., 1977. Gains from portfolio diversication into less developed countries
securities. Journal of International Business Studies 52, 8399.
Forbes, K., Rigobon, R., 2002. No contagion, only interdependence: Measuring stock
market comovements. Journal of Finance 57, 22232261.
Fujii, E., 2005. Intra and inter-regional causal linkages of emerging stock markets:
Evidence from Asia and Latin America in and out of crises. Journal of
International Financial Markets, Institutions and Money 15, 315342.
Gallo, G.M., Otranto, E., 2005. Volatility spillovers, interdependence and
comovements: A Markov switching approach. Computational Statistics and
Data Analysis 52, 30113026.
Genest, C., Rmillard, B., 2008. Validity of the parametric bootstrap for goodness-of-
t testing in semiparametric models. Annales de lInstitut Henri Poincar:
Probabilits et Statistiques 44, 10961127.
Genest, C., Rmillard, B., Beaudoin, D., 2009. Goodness-of-t tests for copulas: A
review and a power study. Insurance: Mathematics and Economics 44, 199213.
Gilmore, C.G., Lucey, B., McManus, G.M., 2008. The dynamics of central European
equity market comovements. Quarterly Review of Economics and Finance 48,
605622.
Grubel, H., 1968. Internationaly diversied portfolios: Welfare gains and capital
ows. American Economic Review 58, 12991314.
Harvey, C.R., 1995. Predictable risk and returns in emerging markets. Review of
Financial Studies 8, 773816.
Hu, J., 2008. Dependence structures in Chinese and US nancial markets: A time-
varying conditional copula approach. Working Paper, Southern Methodist
University.
141
Joe, H., 1997. Multivariate Models and Dependence Concepts. Chapman & Hall/CRC,
London.
Joe, H., Xu, J.J., 1996. The estimation method of inference functions for margins for
multivariate models. Technical Report No. 166, Department of Statistics,
University of British Columbia.
Jondeau, E., Rockinger, M., 2006. The Copula-GARCH model of conditional
dependencies: An international stock market application. Journal of
International Money and Finance 25, 827853.
Kupiec, P., 1995. Techniques for verifying the accuracy of risk measurement models.
Journal of Derivatives 2, 7384.
Levy, H., Sarnat, M., 1970. International diversication of investment portfolios.
American Economic Review 60, 668675.
Longin, F., Solnik, B., 1995. Is the correlation in international equity returns
constant: 19601990? Journal of International Money and Finance 14, 326.
Longin, F., Solnik, B., 2001. Extreme correlation of international equity markets.
Journal of Finance 56, 649676.
Lopez, J.A., Saidenberg, M.R., 2001. Evaluating credit risk models. Journal of Banking
and Finance 25, 16351664.
Markowitz, H., 1952. Portfolio selection. Journal of Finance 7, 7791.
Nelsen, R.B., 1999. An Introduction to Copulas. Springer, New York.
Pais, A., Stork, P.A., in press. Contagion risk in the Australian banking and property
sectors. Journal of Banking and Finance. doi:10.1016/j.bankn.2010.05.012.
Patton, A., 2004. On the out-of-sample importance of skewness and asymmetric
dependence for asset allocation. Journal of Financial Econometrics 2, 130168.
Patton, A.J., 2006. Modelling asymmetric exchange rate dependence. International
Economics Review 47, 527556.
Poon, S.H., Rockinger, M., Tawn, J., 2004. Extreme-value dependence in nancial
markets: Diagnostics, models and nancial implications. Review of Financial
Studies 17, 581610.
Rodriguez, J., 2007. Measuring nancial contagion: A copula approach. Journal of
Empirical Finance 14, 401423.
S
erban, M., Brockwell, A., Lehoczky, J., Srivastava, S., 2007. Modelling the dynamic
dependence structure in multivariate nancial time series. Journal of Time
Series Analysis 28, 763782.
Solnik, B., 1974. Why do not diversify internationally rather than domestically.
Financial Analyst Journal 30, 4854.
Teiletche, J., Xu, B., 2008. The determinants of extreme dependence in nancial
markets. Working Paper, Bordeaux IV University.