The Quarterly Review of Economics and Finance 50 (2010) 157166

Contents lists available at ScienceDirect

The Quarterly Review of Economics and Finance

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

Stock market linkages and nancial contagion: A cobreaking analysis

Niklas Ahlgren, Jan Antell
Hanken School of Economics, Department of Finance and Statistics, P.O. Box 479, Arkadiagatan 22, 00101 Helsingfors, Finland

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

Article history: Financial crises have shown that dramatic movements in one nancial market can have a powerful impact
Received 3 July 2009 on other markets. This paper proposes to use cobreaking to model comovements between stock markets
Received in revised form 13 October 2009 during crises and to test for contagion. We nd evidence of cobreaking between developed stock markets.
Accepted 11 December 2009
In emerging stock markets, the evidence of cobreaking is mainly due to the non-nancial event of the
Available online 23 December 2009
World Trade Center terrorist attacks in 2001. We nd evidence of short-term linkages during times of crisis
but not contagion. These short-term linkages have important implications for investors, risk managers
Keywords:
and regulators.
Cobreaking
Contagion 2009 The Board of Trustees of the University of Illinois. Published by Elsevier B.V. All rights reserved.
Financial crises

JEL classication:
G15
C32

1. Introduction
One of the salient features of globalisation and the rapid trans-
mission of information across markets is the spread of nancial
crises from one country to another. The experience of recent nan-
cial crises has shown that dramatic movements in one market can
have a powerful impact on other markets, even when the under-
lying economic fundamentals are different. Well-known examples
of such crises include the World Trade Center terrorist attacks in
September 2001 (referred to as 9/11), Russian bond market crash in
1998, Asian nancial crisis in 1997, Mexican devaluation in Decem-
ber 1994, and US and Hong Kong stock market crash in October
1987 (Black Monday).
The globalisation of nancial markets has important conse-
quences for the performance of international portfolios and risk
management. Investors are interested in international diversi-
cation of risks. However, if nancial markets are more closely
linked during times of crisis, then the opportunities for interna-
tional diversication are decreased when they would be needed the
most. For regulators of nancial markets, it is particularly impor-
tant to understand such linkages because of the perceived increase
in contagion among world nancial markets.
There is a long line of research on nancial market linkages.
Short-run linkages have been tested by correlation coefcients
Corresponding author. Tel.: +358 40 352 1384; fax: +358 9 4313 3393.
E-mail address: jan.antell@hanken. (J. Antell).
and vector autoregressive models for returns (see e.g. Karolyi &
Stulz, 2003). Long-run linkages have been modelled and tested
by cointegration relations between stock prices (see e.g. Ahlgren
& Antell, 2002). Volatility linkages have been studied by ARCH
and GARCH models, and their various extensions (see e.g. Baele,
2005). More recently, and as a consequence of the developments
described above, interest has shifted towards investigating short-
run linkages in times of crisis and the spread of nancial crises,
i.e. contagion (Claessens, Dornbusch, & Park, 2001; Dungey, Fry,
GonzalezHermosillo, & Martin, 2005; Forbes & Rigobon, 2001,
2002; Pericoli & Sbracia, 2003). Contagion is dened as the change
in the propagation mechanism of shocks between countries in peri-
ods of crisis.1 Forbes and Rigobon (2002) introduce the concept
shift contagion, by which is understood a signicant increase in
cross-market linkages following a shock to one country. Kaminsky,
Reinhart, and Vgh (2003) dene fast and furious contagion as
a signicant and immediate short-term transmission of shocks
between nancial markets. This paper uses their denition of con-
tagion.
1
The World Bank provides three denitions of contagion. According to the broad
denition, by contagion is understood the cross-country transmission of shocks or
the general cross-country spillover effects. This is a very general denition and does
not restrict contagion to take place only in times of crisis. The restrictive denition
states that contagion is the transmission or comovement of shocks in excess of
what can be explained by economic fundamentals. According to the very restrictive
denition, contagion is the change in the propagation mechanism of shocks between
countries in periods of crisis. The very restrictive denition is the one referred to
above.

1062-9769/$ see front matter 2009 The Board of Trustees of the University of Illinois. Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.qref.2009.12.004
158 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166
Many researchers have tested for contagion based on cor-
relation coefcients. If correlations between returns increase
signicantly after a crisis, it is taken as evidence of contagion.
However, as shown by Forbes and Rigobon (2002), correlation coef-
cients are conditional on market volatility, and during a crisis
when stock market volatility increases, estimates of cross-market
correlations will be biased upwards. Increased correlation coef-
cients may be induced by heteroskedasticity (in this case higher
volatility during a crisis compared to stable periods), which biases
tests for contagion. Forbes and Rigobon develop a heteroskedas-
ticity correction, and nd no contagion during the 1997 Asian
crisis, 1994 Mexican devaluation, and 1987 US and Hong Kong
stock market crash. Dungey et al. (2005) propose implementing the
change in correlation test of Forbes and Rigobon within a regres-
sion framework as a Chow test for a structural break of a regression
coefcient. The regression framework is extended by Dungey et
al. to multivariate testing for contagion (see also below). Rigobon
(2003) discusses econometric issues in tests of shifts in transmis-
sion mechanisms. For 36 stock markets he nds no changes in
the transmission mechanisms during the 1998 Russian crisis, 1997
Asian crisis and 1994 Mexican crisis. Corsetti, Pericoli, and Sbracia
(2005) point out that the heteroskedasticity correction is based on
unrealistic assumptions about the idiosyncratic variance, which
bias the test towards nding no contagion. Using a factor model
of returns, they nd contagion more often than one would nd
applying the heteroskedasticity correction. Bekaert, Harvey, and
Ng (2005) employ a two-factor asset pricing model, and nd no
excess correlations above what one would expect from economic
fundamentals caused by the Mexican crisis, but some evidence of
excess correlations caused by the Asian crisis.
Dungey, Fry, GonzalezHermosillo, and Martin (2007) model
contagion by a factor model incorporating contagion as an addi-
tional factor linking global equity markets during periods of
nancial crisis. They nd evidence of contagion during the 1998
Russian bond market crash. Dungey and Martin (2007) model
spillovers and contagion between currency and equity markets
during the 19971998 Asian crisis. They nd that both spillover and
contagion effects are statistically signicant. An interesting result
is that the main source of contagion in most equity markets is from
the US, i.e. nancial crises are propagated through the US market.
Favero and Giavazzi (2002) and Pesaran and Pick (2007) develop
tests of contagion in vector autoregressive (VAR) models. The test
of Favero and Giavazzi is carried out by checking the signicance of
country-specic crisis indicators in equations for other countries.
Pesaran and Pick (2007) base their test on contagion indices con-
structed from indicator variables, and take the endogeneity of the
crises into account by using instrumental variable estimation. The
tests of Favero and Giavazzi, and Pesaran and Pick are applied in
Dungey et al. (2005) to test contagion during the 19971998 Asian
crisis. They nd some evidence of contagion during the Asian crisis.
Bae, Karolyi, and Stulz (2003) propose to model contagion
between nancial markets by the joint occurrences of extreme
returns, or coexceedances. They nd that contagion from Latin
America to other markets is more important than contagion from
Asia.
Dungey et al. (2005) is a survey of tests of contagion. They show
that alternative models and tests may lead to different conclu-
sions concerning the existence of contagion. In particular, they nd
that the test of Forbes and Rigobon (2002) based on correlation
coefcients is conservative and generally fails to nd evidence of
contagion. The multivariate version of the Forbes and Rigobon test
(Dungey et al., 2005) points to stronger evidence of contagion link-
ages. In contrast, the test of Favero and Giavazzi nds contagion in
almost all cases where it is applied. The test of Pesaran and Pick
gives similar results as the test of Dungey et al. (2007). Finally, the
test of Bae et al. (2003) tends to give conicting results with the
other tests.
In this paper we propose a new method to model comovements
between nancial markets during crises and to test for contagion.
Our approach is based on the statistical concept of cobreaking.
Cobreaking is a special case of common features introduced by
Engle and Kozicki (1993). A more well-known example of com-
mon features is cointegration, which is used by e.g. Ahlgren and
Antell (2002) to test long-run linkages between international stock
markets. Cobreaking is dened as the cancellation of breaks across
linear combinations of variables. More specically, if two or more
stationary time series have a break in their conditional mean but
some linear combination does not have the break, then the series
are said to be cobreaking. Tests for the number of such linear com-
binations in vector autoregressive models, or the cobreaking rank,
are developed by Hendry and Massmann (2007), and Krolzig and
Toro (2002).
Cobreaking is particularly revealing about large comovements
in nancial markets during a crisis because it provides information
on the instantaneous spread of a crisis from one market to another
(cf. the denition of fast and furious contagion in Kaminsky et al.,
2003). Tests based on correlation coefcients (Corsetti et al., 2005;
Forbes & Rigobon, 2002) may be uninformative about a specic
event like a stock market crash, since correlation coefcients are
based on averages over a longer period of time. The test of contagion
in multi-factor models for returns represents a multivariate exten-
sion of the correlation change test of Forbes and Rigobon (Dungey
et al., 2007, p. 156), so a similar comment applies to it. Related
to our test for cobreaking are the tests of contagion proposed by
Favero and Giavazzi (2002), and Pesaran and Pick (2007). The cru-
cial difference between these tests and the test for cobreaking is
that whereas the former check the signicance of country-specic
indicators or contagion indices in equations for other countries or
markets, the latter is a test of nancial crises being related across
markets and over time, so that they are cancelled in a linear com-
bination of returns. More specically, by cobreaking we model and
test the co-occurrence of large negative stock returns during nan-
cial crises.
The approach of Bae et al. (2003) based on coexceedances is
closely related to our approach, but has the drawback that their
denition of an extreme return is somewhat arbitrarily chosen to be
a particular quantile of the marginal return distributions. Another
disadvantage is that the distribution of extreme returns depends on
the underlying distribution of returns, and critical values have to
be simulated. In our approach there is no need to restrict the breaks
to be of a particular size for all markets in the sample. Moreover,
the distribution of the test statistic for cobreaking rank does not
depend on the distributional assumptions of the returns.
We apply our methodology to data on stock prices from devel-
oped, as well as emerging markets. We nd cobreaking between
stock returns in developed markets, but not in emerging markets.
Crises in emerging markets do not spread to other emerging mar-
kets or to developed markets. There are comovements between
nancial markets during crises but not contagion.
The reminder of the paper is organised as follows. Section 2
reviews modelling and testing for cobreaking. Section 3 describes
the data and Section 4 reports the empirical results. Section 5 con-
cludes.
2. Cobreaking
In this section we describe the test for cobreaking rank (Hendry
& Massmann, 2007; Krolzig & Toro, 2002) and report the results
from a small Monte Carlo simulation to examine the size and power
of the test.
 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166
2.1. Test for cobreaking
Let Xt be a p 1 vector of time series of logarithms of stock
prices, assumed to be integrated of order one. The model for the
individual series Xit , i = 1, . . ., p, allows for q-time changes in the
level of the series:
q
Xit = it + ij djt + it , it IID(0, i2 ), t = 1, . . . , T,
j=1
where it is the level of the series. The error terms it may be
mutually correlated but are assumed to be IID across t. The for-
mulation (Eq. (1)) allows the series to be quite general processes.
Later we will model the joint dynamics of the return series by a vec-
tor autoregressive (VAR) model. The dates  1 , . . .,  q are the break
dates of nancial crises, modelled by indicator variables djt , which
take the value 1 if t =  j and 0 otherwise, j = 1, . . ., q. Perron (1989)
refers to the model with a one-time change (q = 1) in the level of
the series as the crash model. Here we allow for q-time changes in
the level of the series.
We want to test if the series have breaks in common, i.e. whether
they are cobreaking. The test for cobreaking is based on the p-
dimensional VAR model in error correction form:

k1
Xt = Xt1 + i Xti + + 1 D1t + 2 D2t + t ,
i=1
t NID(0, ),
where all the parameters are unrestricted. We write the VAR model
in error correction form because Xt is assumed to be integrated
of order one and possibly cointegrated. If stock prices are coin-
tegrated, then the matrix  has reduced rank (see e.g. Ahlgren
& Antell, 2002). If stock prices are not cointegrated, then  = 0,
and the model reduces to a VAR model in the differences Xt , or
returns. The q 1 vector D1t contains the indicator variables for
nancial crises and the vector D2t other indicator variables that we
condition on.
It is worth noting that our model is slightly different from
Hendry and Massmann (2007), and Krolzig and Toro (2002). The
former are mainly concerned with breaks in deterministic trends,
whereas the latter analyse permanent breaks in the unconditional
mean of stationary series. We allow the series to contain a unit root
and to possibly be cointegrated. From Eq. (1) it is seen that if the
breaks are of the crash type and prices contain a unit root, then the
crashes are cumulated into breaks in the levels of the series.
Cobreaking requires that the p q matrix 1 has reduced
rank, rank(1 ) = s, s < min(p,q) (Hendry & Massmann, 2007). Since
cobreaking is equivalent with the matrix 1 being of reduced rank,
1 can be decomposed as 1 =  , where and are p s and q s
matrices of rank s, respectively. We can nd the p (q s) orthogo-
nal complement of , , such that  = 0. Hendry and Massmann
(2007) call the number q s the cobreaking rank. The q s columns
of the matrix contain the cobreaking vectors, and the cobreaking
relations are the linear combinations  Xt that do not contain the
breaks in the series. To see this, multiply model (2) by  , which
annihilates 1 D1t .
Some instances that induce cobreaking without the reduced
rank condition on 1 becoming a restriction are identied in
Hendry and Massmann (2007). Here it sufces to note the following
two special cases. First, for cobreaking not to be trivial, no subvector
of Xt must be free from breaks, because otherwise we can choose
to be a matrix of zeroes apart from ones in the positions of series
with no breaks. Second, if there are fewer breaks in the sample
than there are series in the system, i.e. if 1 has fewer columns
than rows, so that q < p, then at least p q cobreaking vectors can
always be found.
159
The null hypothesis of cobreaking H0 : rank(1 ) s against H1 :
rank(1 ) = s can be tested by a likelihood ratio test (Hendry &
Massmann, 2007; Krolzig & Toro, 2002). The likelihood ratio statis-
tic for testing the reduced rank hypothesis is given by

min(p,q)
Q (s) = T i ),
log(1  (3)
(1) i=s+1
where  1 
min(p,q) are the squared partial canonical corre-
lations between Xt and D1t , i.e. corrected for a constant, Xt1 ,
. . ., Xtk+1 and D2t . The use of canonical correlation analysis is
similar to the approach of Johansen (1996) for determining the
cointegration rank.
The LR statistic is approximately distributed as 2 with
(p s)(q s) degrees of freedom under the null hypothesis. We
remark that there is no asymptotic distribution involved. The
matrix D = (D11 , D12 , . . . , D1T ) contains only a nite number of
ones and the rest of the elements are zeroes, so D D/T converges to
a zero matrix as T tends to innity, and the asymptotic distribution
would be degenerate (see e.g. Davidson, 2000, Section 7.2.1). As
indicated above, it is not necessary to assume a distribution for the
returns. Since the errors are assumed to be normal, the Q statistic
has an approximate 2 distribution. The test is therefore sensitive
to the assumption of normally distributed errors.
The sequence of tests begins by testing the most restricted
(2) hypothesis s = 0 and continues by testing s = 1, and so on until the
rst non-rejected null hypothesis.
The model with rank s < q is estimated by reduced rank regres-
sion. The estimate is found by reduced rank regression, and then
the estimate conditional on by regression (see Krolzig & Toro,
2002, for details). The orthogonal complement is found by a
2
singular value decomposition (SVD) of .
It is helpful to comment here on the relation between the tests
of contagion suggested by Favero and Giavazzi (2002), and Pesaran
and Pick (2007) on the one hand, and our test for cobreaking as a test
of contagion on the other hand. The tests of Favero and Giavazzi, and
Pesaran and Pick are tests of the signicance of indicator variables
or contagion indices in equations for other markets than the market
which is the origin of the crisis in question. Our test for cobreaking
is a test of reduced rank of the coefcient matrix of the indicator
variables, conditional on the indicator variables being signicant
in the equations for the other markets. Finding cobreaking implies
that nancial markets are subject to the same shocks or crises, so
that they are cancelled in a linear combination of returns.
2.2. Simulation study
The version of the test for cobreaking rank used here is new, as
pointed out above. We therefore present the results from a small
Monte Carlo experiment to examine the size and power of the test.
Data are generated from the data-generation process (DGP):
Xt = 1 Xt1 + 1 D1t + t , (4)
with p = 4 and q = 2. This specication of the DGP has q < p, and there-
fore corresponds to one of the two special cases mentioned in the
previous section. It is used for simplicity. The test for cobreaking
rank is valid even when q < p, which is all that matters here. The
errors are generated as NID(0,I4 ). We found that the distribu-
tion of the Q statistic does not depend on 1 , and in the DGP we
can take 1 = 0. The break points are dened as  1 T and  2 T, and
2
The orthogonal complement is computed as = UDV, where U is p p and the
last q s columns of U are (see e.g. Hendry, 1995, 631632).
160 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166

Table 1
Simulated size and power of the test for cobreaking rank.

d 2 3 4 6 2 3 4 6

T DGP has rank s = 1 DGP has rank s = 2

Power of test of s = 0 Power of test of s = 0

100 0.990 1.000 1.000 1.000 0.909 0.999 1.000 1.000

200 0.991 1.000 1.000 1.000 0.912 1.000 1.000 1.000
400 0.992 1.000 1.000 1.000 0.913 1.000 1.000 1.000

d 2 3 4 6 2 3 4 6

T Size of test of s = 1 Power of test of s = 1

100 0.059 0.064 0.065 0.066 0.166 0.393 0.648 0.955

200 0.051 0.055 0.056 0.057 0.156 0.382 0.642 0.956
400 0.047 0.052 0.052 0.053 0.151 0.377 0.640 0.956

Notes: The DGP is given by Eqs. (4) and (5) with p = 4 and q = 2. The number of replications is 1,000,000.

are modelled by the indicator variables D1t = (d1t , d2t ) , where d1t = 1
for t =  1 T and 0 otherwise, and d2t = 1 for t =  1 T and 0 otherwise.
In the simulations,  1 = 0.3 and  2 = 0.7, so that the rst break point
occurs in the rst half of the series and the second break point in
the second half of the series. Following Krolzig and Toro (2002), we
consider the following two specications for the matrix 1 :
1 1 0.5
0.5 1 to be manageable, four markets are included in the VAR system. In
= d
= d
0.5
11 and 12 . (5)
1 1
0.5 1 wide. The rst data set contains four large, leading markets in
The matrix 11 has reduced rank s = 1 and the matrix 12 has
full rank s = q = 2. In the DGP (5) we can choose = = 1. The values
for d are d = 2, 3, 4, 6. Then a value of d = 2 corresponds to the breaks
being equal to two standard deviations of the errors, since the errors
have unit variances. The series lengths are T = 100, 200, 400.
Table 1 reports the results from the Monte Carlo experiments.
The number of Monte Carlo replications is 1,000,000. We begin by
testing s = 0 against s > 0. We nd that the test rejects, as it should,
since the rank of 1 is at least s = 1. The test of s = 1 against s = 2 with
11 gives the size of the test and with 12 the power of the test.
We nd that the test is slightly oversized when T = 100, but has the
correct size when T = 200 and T = 400. Turning to power, when d = 2
the power is below 20 per cent, when d = 3 the power is about 40
per cent, when d = 4 the power is above 60 per cent and when d = 6
the power approaches 100 per cent.
It is worth noting that the null hypothesis s = 1 implies cobreak-
ing. Consequently, if the test has low power it may falsely accept
cobreaking. The power does not increase with the series length.
The breaks are modelled by indicator variables D1t , which only con-
tain a nite number of ones and the rest of the elements are zeros.
The information about the breaks is not increasing with the series
length. The powers of the test are therefore constant for different
values of T in Table 1.
3. Data
We use data on international stock prices. For the VAR analysis
choosing the markets, the objective is to include markets world-
industrialised countries, namely Germany, Japan, the UK and the
US. For all markets, the Morgan Stanley Capital International (MSCI)
total return indices denominated in USD are used. The sample
period for the returns is January 1980 to August 2006 and the
number of month-end observations is T = 320. The second data set
contains the emerging markets Hong Kong, Korea, Mexico, and in
addition the US. There are one Latin American and two Asian mar-
kets among the emerging markets. The US is included to capture
the effects of major markets on emerging markets and emerging
markets on developed markets. For most of the emerging markets,
the MSCI indices are not available before 1988. The sample period
for the returns is therefore January 1988 to August 2006 and the
number of month-end observations is T = 224.
Table 2 provides descriptive statistics for the stock mar-
ket returns computed as logarithmic differences. The annualised
returns in the developed markets are between 8.9 per cent for
Japan and 12.6 per cent for the UK, and the volatilities are between
15.0 per cent for the US and 22.8 per cent for Japan. The monthly
minimum returns range between 20 and 30 per cent, and the
maximum returns are around 20 per cent. For the UK and the US,
the minima coincide with the stock market crash in October 1987.

Table 2
Descriptive statistics for the stock market returns in US dollars (in per cent).

Mean Standard deviation Minimum Maximum Skewness Excess kurtosis JB

Developed markets. Sample period January 1980 to August 2006 and T = 320
Germany 10.59 22.36 27.90 21.26 0.602** 2.024** 73.942**
Japan 8.93 22.78 21.22 21.76 0.065 0.358 1.933
UK 12.59 18.26 24.25 16.38 0.415** 2.033** 64.308**
US 12.25 15.04 23.86 12.47 0.841** 3.403** 192.170**

Emerging markets. Sample period January 1988 to August 2006 and T = 224
Hong Kong 12.74 26.30 34.06 28.69 0.200 2.477** 58.757**
Korea 7.63 38.10 37.22 53.41 0.294 2.969** 85.502**
Mexico 22.24 33.42 40.24 25.91 0.972** 2.979** 118.000**
US 11.28 13.90 14.97 10.82 0.568** 0.996** 21.308**

Notes: The means and standard deviations are annualised, while the minimum and maximum are monthly. The standard error for the coefcient of skewness is 6/T and

for kurtosis 24/T . JB is the JarqueBera statistic for normality.
**
Statistically signicant at the 1 per cent level.
 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166 161

Table 3
Misspecication diagnostics for the VAR systems.

LM(1) LM(12) ARCH(1) LM JB

Developed markets. VAR(5), sample period January 1980 to August 2006 and T = 320
Germany 0.336 2.272** 0.989 1.348 4.424
Japan 0.002 1.325 3.066 1.441 1.127
UK 0.451 0.627 2.075 1.938** 0.155
US 0.877 0.955 0.042 0.972 1.053
System 1.706* 1.107 1.292** 14.288

Emerging markets. VAR(5), sample period January 1988 to August 2006 and T = 224
Hong Kong 0.425 1.375 1.098 1.379 5.106
Korea 1.571 1.486 0.213 2.048** 2.328
Mexico 0.317 1.081 0.212 1.925** 6.730*
US 0.107 0.570 0.747 1.388 3.800
System 1.941* 1.585** 1.289** 35.507**

Notes: LM(1) and LM(12) are the Lagrange multiplier tests for residual autocorrelation of order 1 and up to order 12, respectively. ARCH(1) is the test for autoregressive
conditional heteroskedastic residuals of order 1. LM is the White test for heteroskedasticity. JB is the JarqueBera test for normality. All statistics are F, except JB which is 2 .
*
Statistically signicant at the 5 per cent level.
**
Statistically signicant at the 1 per cent level.

With the exception of Japan, the returns are skewed to the left
and leptokurtic. The JarqueBera statistics show that the returns
are not normal, with the exception of Japan. In the emerging mar-
kets the annualised returns are between 7.6 per cent for Korea and
22.2 per cent for Mexico. The volatilities are much higher than
in the developed markets, ranging from 26.3 per cent for Hong
Kong to 38.1 per cent for Korea. The monthly minimum returns
are 40 per cent for Mexico and the maximum returns 53 per
cent for Korea. The minimum for Hong Kong coincides with the
Asian crisis in October 1997, and the minima for Mexico and the
US with the Russian bond market crash in August 1998. With the
exception of Korea, the returns are skewed to the left. All series are
leptokurtic. The JarqueBera statistics show that the returns are not
normal.
4. Empirical results
We report the empirical results of the tests for cobreaking rank
for the developed and emerging markets. We also perform some
robustness checks with respect to the dates of the crises, the rank of
 in the model (2), the inuence of a single crisis on the cobreaking
results, and the data frequency.
4.1. Estimated VAR systems
We choose a VAR(5) specication for both the developed and
emerging markets. The test for cobreaking rank requires that we
dene a set of indicator variables for nancial crises. For the devel-
oped markets, we include in D1t four indicator variables. They
take the value one for the crisis date and zero otherwise. D8710
is the indicator variable for the October 1987 stock market crash.
The Asian crisis in October 1997 is modelled by D9710 and the
Russian bond market crash in August 1998 by D9808. The latter
one coincides with the default of the Long-Term Capital Manage-
ment hedge fund. The indicator variable D0109 is for the terrorist
attacks in September 2001. For the emerging markets the indica-
tor variable D8710 is exchanged for the indicator variable D9412
for the Mexican Peso crisis in December 1994. The other crisis
indicators are the same. To improve the misspecication diag-
nostics of the models, we include in D2t the non-crisis indicator
D0209 for the developed markets and D9801 for the emerging
markets.
The crisis dates correspond to major nancial crises and are
exogenously determined based on the nancial calendar and previ-
ous research.3 The standardised residuals reported in Table 4 and
the absence of any other large standardised residuals in the VAR
models conrm the crisis dates and the absence of any neglected
nancial crises. The crisis dates can be determined endogenously,
if desired, by dening indicator variables for observations with
large standardised residuals from the estimated VAR model with-
out indicator variables (see e.g. Favero & Giavazzi, 2002). However,
as pointed out by Pesaran and Pick (2007), assuming that the cri-
sis dates are endogenously determined leads to an endogeneity
problem, which requires generalised instrumental variable esti-
mation (GIVE). Here we shall focus on the relatively simple case
where the crisis dates are exogenous, but our approach and argu-
ments can be readily extended to the case of endogenous crisis
dates by adopting the GIVE procedure with lagged dependent vari-
ables as instruments proposed in Pesaran and Pick. The reduced
rank regression technique used in the paper would then have to be
modied, something which would require another paper to work
out the statistical details.4
Table 3 reports misspecication diagnostics for the estimated
VAR systems. The individual equations are acceptable, with the
exception of autocorrelation up to order 12 for Germany and
heteroskedasticity for the UK in the system for the developed mar-
kets, and heteroskedasticity for Korea and Mexico in the system
for the emerging markets. The vector tests indicate that some
autocorrelation remains in the systems. The vector LM tests for het-
eroskedasticity are signicant for both the developed and emerging
markets. More importantly, the residuals of the estimated VAR
model for the developed markets are approximately normally dis-
tributed. For the emerging markets, the JB statistic is signicant for
Mexico at the ve per cent level and for the system at the one per
cent level.
To gauge the magnitude of the breaks, Table 4 displays the stan-
dardised residuals from the estimated VAR models without the
indicator variables for the nancial crises. We see that most of the
breaks are between 2 and 4 standard deviations, so the test for
cobreaking may have low power to reject some of the reduced rank
hypotheses. The t-statistics for the crisis indicators in the VAR sys-
tem for the developed markets are all negative and signicant (not
reported). For the emerging markets, most of the t-statistics are
negative and signicant. The only exception is D9412 for the Mex-
3
See e.g. Tables 3, 6 and 8 of Forbes and Rigobon (2002), and Table 1 of Kaminsky
et al. (2003).
4
We are grateful to an anonymous referee for pointing out this interesting exten-
sion.
162 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166

Table 4
The standardised residuals from the VAR systems.

1987(10) 1997(10) 1998(8) 2001(9)

Developed markets. VAR(5), sample period January 1980 to August 2006 and T = 320
Germany 3.164 1.185 3.237 2.677
Japan 0.997 1.622 1.939 1.274
UK 4.552 0.934 1.691 1.470
US 5.848 0.966 3.515 2.150

1994(12) 1997(10) 1998(8) 2001(9)

Emerging markets. VAR(5), sample period January 1988 to August 2006 and T = 224
Hong Kong 0.554 4.288 1.626 2.639
Korea 0.275 3.423 1.303 1.813
Mexico 4.205 2.408 3.965 1.830
US 0.344 0.797 3.922 2.221

Notes: 1987(10) is the date of the October 1987 stock market crash, 1994(12) is the Mexican crisis, 1997(10) is the Asian crisis, 1998(8) is the Russian crisis and 2001(9) is
the 9/11 terrorist attacks.

Table 5
Tests for cobreaking rank.

Developed markets Emerging markets

s s+1
 Q (s)  p-Value s s+1
 Q (s)  p-Value

0 0.208 86.590** 16 0.000 0 0.263 119.834** 16 0.000

1 0.039 14.315 9 0.112 1 0.143 53.100** 9 0.000
2 0.004 1.731 4 0.785 2 0.084 19.182** 4 0.001
3 0.002 0.572 1 0.449 3 0.000 0.026 1 0.872

Notes: The VAR systems are estimated with q = 4 crisis indicators. The cobreaking rank is q s,  s+1 are the eigenvalues, Q(s) is the likelihood ratio statistic for cobreaking
rank and  are the degrees of freedom. The indicator variables for the developed markets are D8710 for the October 1987 stock market crash, D9710 for the Asian crisis,
D9808 for the Russian crisis and D0109 for the 9/11 terrorist attacks. The sample period is January 1980 to August 2006 and T = 320. The indicator variables for the emerging
markets are D9412 for the Mexican crisis, D9710 for the Asian crisis, D9808 for the Russian crisis and D0109 for the 9/11 terrorist attacks. The sample period is January 1988
to August 2006 and T = 224.
**
Statistically signicant at the 1 per cent level.

ican crisis. While all other crises affect all markets, the Mexican Table 6
Orthogonal complements or cobreaking vectors.
crisis is mainly conned to Mexico itself.
We note that the conditions to avoid trivial cobreaking and Developed markets Emerging markets
induced cobreaking are satised, since all markets are affected by
qs=2 1 2 qs=1 1
at least one crisis, and there are exactly as many crises or breaks in
the series as there are markets or series in the VAR system, i.e. we Germany 0.437 0.842 Hong Kong 1.000
Japan 1.000 0.412 Korea 0.918
have q = p.
UK 0.833 0.519 Mexico 0.021
US 0.347 1.000 USA 0.009
4.2. Tests for cobreaking
Notes: The VAR systems are estimated with q = 4 crisis indicators. The cobreaking
rank is q s. For the developed markets the sample period is January 1980 to August
If there exists cobreaking between the returns Xt , the matrix 2006 and T = 320. For the emerging markets the sample period is January 1988 to
1 in the model (2) has reduced rank. Note that if rank(1 ) = 0, August 2006 and T = 224.
none of the crisis indicators is signicant, and if rank(1 ) = 4,
there is no cobreaking. Table 5 presents the results of the tests
Then q s = 1 and there is only one cobreaking relation. The emerg-
for cobreaking. For the developed markets, there are two eigen-
ing markets therefore exhibit less cobreaking than the developed
values which are different from zero and two eigenvalues close to
markets.
zero. We would therefore expect that s = 2. The test for cobreaking
The orthogonal complements , or the cobreaking vectors, are
rank rejects s = 0 but not s = 1, entailing that the cobreaking rank,
provided in Table 6. The cobreaking vectors are normalised on the
or the number of cobreaking relations, is q s = 3, i.e. the maximal
largest element. The unrestricted vectors are not uniquely identi-
number of cobreaking relations. The p-value of the hypothesis s = 1
ed. Further, it is not known how to test hypotheses about .6 For
is 0.112. In the simulations in Section 2.2 we found that the test
may have low power. This suggests that it may underestimate s the developed markets, the orthogonal complement consists of
and, consequently, overestimate the cobreaking rank q s. Inspec- two vectors in a four-dimensional system of stock returns such that

tion of the estimated cobreaking relations with q s = 3 reveals X t do not contain the breaks in the series. All the estimated

that the rst cobreaking relation contains the breaks in the series, coefcients are of the same magnitude and different from zero.

whereas with q s = 2 we get cobreaking relations that do not con- Fig. 1 plots the cobreaking relations. Instead of plotting X t ,
tain the breaks.5 We have therefore decided to choose s = 2 and 
we plot X t to better see the cobreaking relations. Note that

q s = 2, i.e. two cobreaking relations. The results for the emerging 
markets are different. The test rejects s = 0, 1, 2 and accepts s = 3. X t contains a unit root in the stock prices Xt , but no breaks.

5 6
Note that the third relation is the relation corresponding to the smallest eigen- Deriving a test for restrictions on the orthogonal complement is a difcult sta-
value, i.e. the relation containing the most evidence of cobreaking, and so on. tistical problem, which it would be useful to solve.
 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166 163

 
Fig. 1. The cobreaking relations for the developed and emerging markets. Notes: The upper row shows the cobreaking relations 1 Xt and 2 Xt for the developed markets.
 
The lower row shows the cobreaking relation 1 Xt and the hypothesised cobreaking relation 1 Xt for the emerging markets, where 1 = (1, 1, 0, 0) .


For the emerging markets, the results are more interesting. Only
two coefcients, namely those for Hong Kong and Korea, are rea-
sonably large, of the same magnitude and with opposite signs, and
the remaining two coefcients are close to zero. Suppose that in
fact = (1, 1,0,0) . Then we have a cobreaking relation involving
only the difference between the returns in Hong Kong and Korea,
which are both Southeast Asian markets. Cobreaking for the emerg-
ing markets is restricted to these two markets, and there is no
cobreaking involving the other emerging markets, or the emerg-
ing markets and the US. The US market is therefore insulated from
cobreaking between the emerging markets. To see whether the
restrictions are compatible with the data, we plot the cobreaking

relation X t together with the hypothesised cobreaking relation
 the Asian crisis should be dated to 1997(10). The Mexican crisis
X t . Fig. 1 shows that they are almost indistinguishable, which is
strong evidence in favour of the hypothesis.
 dardised residuals in the system for the emerging markets. We
Table 7 reports the cobreaking relations X t , together with
the returns Xt for the break dates. For the developed markets, we
 
see that the estimated cobreaking relations X t and X t do
1 2
not contain the breaks in the return series. For the emerging mar-

kets, the relation X t does not contain the breaks in the return
1 results involves the rank of  in the model (2). If stock prices
series, indicating cobreaking between Hong Kong and Korean stock
returns.
4.3. Robustness checks
We perform four sets of robustness tests. In each of the four
robustness checks the main results of our base analysis do not
change. The rst robustness check is concerned with alternative
dates for some of the nancial crises. The Asian crisis has been dated
by some authors to June 1997, coinciding with the devaluation
of the Thai Bath. For both 1997(6) and 1997(7), the standard-
ised residuals for the developed markets are positive and smaller
than one (with the exception of 1997(7) for Japan and the US,
which are 1.154 and 1.289, respectively). For the emerging mar-
kets, the standardised residuals are positive and smaller than one
(with the exception of 1997(7) for Korea which is 0.611). Clearly,
can be dated to 1995(1). Moreover, there are some large stan-
therefore computed the test for cobreaking with an indicator for
1995(1) instead of 1994(12), and found that the results are similar
(q s = 1).
The second robustness check pertaining to our cobreaking
are cointegrated, then  is different from zero and we should
account for cointegration before testing the cobreaking rank.

Table 7
The returns and cobreaking relations  Xt for the break dates.

Developed markets Emerging markets

1987(10) 1997(10) 1998(8) 2001(9) 1994(12) 1997(10) 1998(8) 2001(9)

Germany 19.21 7.22 17.92 17.73 Hong Kong 3.13 34.06 6.92 16.77
Japan 7.58 9.77 12.09 9.74 Korea 6.43 36.77 17.10 17.79
UK 24.25 3.87 7.52 6.81 Mexico 39.78 20.67 40.24 16.95
US 23.85 2.75 14.97 7.95 US 1.48 2.75 14.97 7.95
 
X t
1 8.27 7.00 5.17 4.49 X t
1 2.67 0.09 6.99 0.11

2 X t 2.12 1.91 0.80 5.73

Notes: The VAR systems are estimated with q = 4 crisis indicators. 1987(10) is the date of the October 1987 stock market crash, 1994(12) is the Mexican crisis, 1997(10) is the
Asian crisis, 1998(8) is the Russian crisis and 2001(9) is the 9/11 terrorist attacks. For the developed markets the sample period is January 1980 to August 2006 and T = 320.
For the emerging markets the sample period is January 1988 to August 2006 and T = 224.
164 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166

Table 8
Tests for cointegration rank.

Developed markets Emerging markets

r r+1
 Q (r) p-Value r r+1
 Q (r) p-Value

0 0.104 55.60** 0.007 0 0.042 21.97 0.971

1 0.046 21.00 0.368 1 0.023 12.12 0.910
2 0.014 6.02 0.696 2 0.019 7.42 0.536
3 0.005 1.56 0.212 3 0.015 3.23 0.072

Notes: The cointegration rank is r,  r+1 are the eigenvalues and Q (r) is the likelihood ratio statistic for cointegration rank. For the developed markets the sample period is
January 1980 to August 2006 and T = 320. For the emerging markets the sample period is January 1988 to August 2006 and T = 224.
**
Statistically signicant at the 1 per cent level.

Table 9
Tests for cobreaking rank with cointegration rank r = 1.

Developed markets Emerging markets

s s+1
 Q (s)  p-Value s s+1
 Q (s)  p-Value

0 0.210 88.235** 16 0.000 0 0.261 117.772** 16 0.000

1 0.039 14.164 9 0.117 1 0.146 51.563** 9 0.000
2 0.004 1.495 4 0.828 2 0.074 16.976** 4 0.002
3 0.001 0.322 1 0.570 3 0.000 0.053 1 0.818
s+1 are the eigenvalues, Q (s) is the likelihood ratio statistic for cobreaking rank
Notes: The VAR systems are estimated with q = 4 crisis indicators. The cobreaking rank is q s, 
and  are the degrees of freedom. The indicator variables are D8710 for the October 1987 stock market crash, D9412 for the Mexican crisis, D9710 for the Asian crisis, D9808
for the Russian crisis and D0109 for the 9/11 terrorist attacks. For the developed markets the sample period is January 1980 to August 2006 and T = 320. For the emerging
markets the sample period is January 1988 to August 2006 and T = 224.
**
Statistically signicant at the 1 per cent level.

Table 10
Tests for cobreaking rank with one indicator variable removed.

s  Developed markets Emerging markets

D8710 D9710 D9808 D0109 D9412 D9710 D9808 D0109

** ** ** ** ** ** **
0 12 36.800 (<0.001) 83.480 (<0.001) 63.302 (<0.001) 70.844 (<0.001) 82.227 (<0.001) 80.989 (<0.001) 82.726 (<0.001) 105.607** (<0.001)
1 6 2.035 (0.916) 12.025 (0.061) 8.264 (0.219) 10.587 (0.102) 32.433** (<0.001) 24.062** (0.001) 29.566** (<0.001) 47.734** (<0.001)
2 2 0.597 (0.742) 0.660 (0.719) 0.929 (0.628) 1.203 (0.548) 1.176 (0.555) 4.899 (0.086) 3.648 (0.161) 14.601** (0.001)

Notes: The VAR systems are estimated with q = 3 crisis indicators. The cobreaking rank is q s and  are the degrees of freedom of the test for cobreaking rank. The indicator
variables for the developed markets are D8710 for the October 1987 stock market crash, D9710 for the Asian crisis, D9808 for the Russian crisis and D0109 for the 9/11
terrorist attacks. The sample period is January 1980 to August 2006 and T = 320. The indicator variables for the emerging markets are D9412 for the Mexican crisis, D9710
for the Asian crisis, D9808 for the Russian crisis and D0109 for the 9/11 terrorist attacks. The sample period is January 1988 to August 2006 and T = 224.
**
Statistically signicant at the 1 per cent level.

Table 8 presents the results from tests for cointegration rank
(Johansen, 1996). For the developed markets, the cointegration
rank r = 0 is rejected, and the cointegration rank is found to be
r = 1. This outcome is in line with the results in Ahlgren and
Antell (2002, 2008). For the emerging markets, there is no coin-
tegration, since r = 0 cannot be rejected. In Table 9 we report the
results of the tests for cobreaking rank with the cointegration rank
r = 1. The results show that the cobreaking tests are not sensi-
tive to the choice of the cointegration rank. This is not surprising,
given that the test for cobreaking rank is approximately indepen-
dent of the remaining parameters , 1 , . . ., k1 , and 2 in
the model (2).
The third robustness check deals with the robustness of the
cobreaking results with respect to the indicator variables by
redening one indicator in turn as a non-crisis indicator, so that
q = 3. We are interested in checking how a single crisis impacts
on the cobreaking results. For the developed markets, in all tests
with q = 3 in Table 10, s = 1 is found, which is consistent with the
outcome for q = 4. However, for the emerging markets, s = 3 in the
tests with the indicator variable D0109 removed, and then there
is no cobreaking. The removal of this important crisis makes the
cobreaking results for the emerging markets insignicant. Thus,
what drive the cobreaking results for the emerging markets are
the 9/11 terrorist attacks.
In the fourth robustness check we change the data frequency
from monthly to weekly returns.7 For weekly data we choose
both a VAR(5) and VAR(20) model. It turns out that the lag length
in the VAR model is unimportant for the conclusions, so we
only report the results for the VAR(5) model. Instead of den-
ing a crisis date as with monthly data, we now dene a crisis
period. The indicator variables take the value one for one or two
consecutive observations, depending on the crisis. For example,
the Mexican crisis triggered by the devaluation of the Mexican
Peso on 20 December 1994 is modelled by an indicator vari-
able taking the value one for 21 and 28 December 1994. The
specications of the crisis indicators are provided in Table 11.
In general, the denitions of the crisis periods coincide with
the denitions of the turmoil periods in Forbes and Rigobon
(2002).
Table 11 presents the results of the tests for cobreaking. For the
developed markets, the test rejects s = 0 and s = 1, and accepts s = 2,
which corroborates the choice of s = 2 in our base analysis with
7
The returns are calculated from Wednesday to Wednesday. Different from
monthly data, we use the MSCI price indices instead of the total return indices,
the latter not being available on a weekly frequency for the whole sample period.
As a sensitivity test, we performed the cobreaking analysis with monthly data using
the price indices, and found that the results remain unchanged.
 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166 165

Table 11
Tests for cobreaking rank with weekly data.

Developed markets Emerging markets

s s+1
 Q (s)  p-Value s s+1
 Q (s)  p-Value

0 0.0832 141.422** 16 0.000 0 0.0629 116.952** 16 0.000

1 0.0141 21.052* 9 0.001 1 0.0534 54.047** 9 0.000
2 0.0009 1.375 4 0.849 2 0.0009 0.904 4 0.924
3 0.0001 0.104 1 0.747 3 0.0000 0.048 1 0.826

Notes: The data are weekly observations from Wednesday to Wednesday. The VAR systems are estimated with q = 4 crisis indicators. The cobreaking rank is q s,  s+1 are the
eigenvalues, Q (s) is the likelihood ratio statistic for cobreaking rank and  are the degrees of freedom. The indicator variables for the developed markets are D8710, D9710,
D9808 and D0109 and for the emerging markets D9412, D9710, D9808 and D0109. D8710 for the October 1987 stock market crash takes the value one for 21 and 28 October
1987, and zero otherwise. D9710 for the Asian crisis in October 1997 takes the value one for 22 and 29 October, and zero otherwise. D9808 for the Russian crisis takes the
value one 19 and 26 August 1998. D9412 for the Mexican crisis takes the value one for 21 and 28 December 1994, and zero otherwise. D0109 for the terrorist attacks on 11
September 2001 takes the value one for 12 September and zero otherwise. For the developed markets the sample period is January 1980 to August 2006 and T = 1391. For
the emerging markets the sample period is January 1988 to August 2006 and T = 974.
*
Statistically signicant at the 5 per cent level.
**
Statistically signicant at the 1 per cent level.

monthly data. The estimated cobreaking vectors are similar to the 5. Discussion and conclusions
ones obtained using monthly data and are therefore not reported.
For the emerging markets, the test rejects s = 0 and s = 1, and accepts This paper suggests cobreaking as a new method for modelling
s = 2. The orthogonal complement , or the cobreaking vectors, and testing comovements between nancial markets in times of
with q s = 2 are = (1.000, 1.135, 0.121, 0.071) and =
1 2
crisis. We nd cobreaking between developed stock markets. There
is some evidence of cobreaking between emerging markets. How-
(0.098, 0.146, 0.029, 1.000) .8 The rst cobreaking vector 1 is
ever, cobreaking is conned to the Southeast Asian markets Hong
very similar to the one estimated from monthly data, but now we
Kong and Korea. Moreover, the results are driven by the 9/11 terror-
have in addition the cobreaking vector 2 . The second cobreaking
ist attacks. Thus, there is cobreaking between large, industrialised
relation closely corresponds to US returns. In weekly data the neg-
markets, while nancial crises in emerging markets do not lead
ative returns for the US in times of crisis tend to be much smaller
to cobreaking between emerging markets, or between emerging
than for the emerging markets. The test for cobreaking therefore
markets and developed markets.
selects US returns as a trivial cobreaking vector. Suppose as before
Financial contagion is dened as a signicant immediate short-
that the rst cobreaking vector is 1 = (1, 1, 0, 0) and that the
term transmission of shocks between nancial markets in times
second vector is 2 = (0, 0, 0, 1) . Then we have as in the base
of crisis. In addition, the origin of the shock in question should be
analysis with monthly data a cobreaking relation involving only
a nancial event such as a severe liquidity crisis, default on debt
the difference between the returns in Hong Kong and Korea, and
obligations or large devaluation. For example, in their list of nan-
in addition a trivial cobreaking vector consisting of US returns.9
cial crises with immediate international repercussions, Kaminsky
The conclusion that the US market is insulated from cobreaking
et al. (2003) exclude the 9/11 terrorist attacks and the October 1987
between the emerging markets continues to hold. The only differ-
stock market crash. Of course, the 9/11 terrorist attacks are not
ence is that the US market is now modelled as a trivial cobreaking
a nancial crisis, but an important non-nancial event affecting
relation.
nancial markets worldwide. Since contagion is related to nan-
We also experimented with daily returns, calculated as rolling-
cial crises, cobreaking around an event like a terrorist attack is
average two-day returns as in Forbes and Rigobon (2002), but found
not evidence of contagion, but comovements in general. The same
that it is not possible to model the crises by indicator variables
argument applies to the October 1987 stock market crash.
because in daily data returns can be either positive, close to zero
For the developed markets, none of the nancial crises origi-
or negative during a crisis period. Moreover, cobreaking therefore
nated in these markets, and some of the crises are non-nancial.
seems unlikely in daily data.
We interpret the evidence of cobreaking as short-run linkages in
In view of our discussion of tests of contagion in Section 1, it
times of crisis, not contagion. The evidence of cobreaking is weak
is clear that tests based on correlation coefcients require den-
for the emerging markets and the results are driven by the 9/11
ing stable and crisis periods in daily data. Cobreaking analysis
terrorist attacks. We nd no evidence of contagion during the 1998
is based on modelling a crisis by an indicator variable taking
Russian bond market crash, 1997 Asian crisis and 1994 Mexican
the value one for the crisis and zero otherwise. It is developed
crisis. Different from Bae et al. (2003), we do not nd contagion
for low frequency data like monthly or weekly data. For exam-
from Latin America to other regions. Among the developed mar-
ple, in the original contribution, Hendry and Massmann (2007)
kets, at least the US is largely insulated from contagion from both
use annual macroeconomic data from a period of over 100 years,
Latin America and Asia. The empirical result that the evidence of
which roughly corresponds to between 10 and 20 years of monthly
cobreaking is weak for the emerging markets is in line with the
nancial data.
nding in Dungey et al. (2007) for the Asian crisis that nancial
crises are propagated through the US market.
The results on cobreaking between developed markets should
be relevant for international investors, risk managers and regu-
lators. Since stock markets are cobreaking, they are more closely
8
linked during crises. Investors should develop diversication
To be consistent with the results for monthly data in Table 6, the rst cobreaking
vector is normalised on the rst element. strategies, and risk managers and regulators should develop risk
9
To see whether the restrictions are compatible with the data, we plotted the measures that take cobreaking into account. The nancial crisis that
 
cobreaking relations X t and X t , and compared them with the hypothesised
1 2
originated with the US subprime mortgage market and in 2008 and
 
cobreaking relations 1 X t and 2 X t (not reported), and found that they are almost 2009 spread to other markets makes such a conclusion imperative.
indistinguishable, which is strong evidence in favour of the hypothesis.
166 N. Ahlgren, J. Antell / The Quarterly Review of Economics and Finance 50 (2010) 157166
Acknowledgements
The authors thank Kate Phylaktis, Pentti Saikkonen, the edi-
tor, and an anonymous referee for their helpful comments. The
paper was presented in the Annual Meeting of the Finnish Sta-
tistical Society in Tammerfors 2007, the Annual Meeting of the
European Finance Association in Ljubljana 2007, the Arne Ryde
Workshop in Financial Economics in Lund 2008, and in seminars
in Helsingfors. Niklas Ahlgren acknowledges nancial support in
the form of a research grant from Brderna Lars och Ernst Krogius
forskningsfond, managed by Svenska litteratursllskapet i Finland.
References
Ahlgren, N., & Antell, J. (2002). Testing for cointegration between international stock
prices. Applied Financial Economics, 12, 851861.
Ahlgren, N., & Antell, J. (2008). Bootstrap tests of cointegration rank with nancial
time series. Computational Statistics & Data Analysis, 52, 47544767.
Bae, K. H., Karolyi, G. A., & Stulz, R. M. (2003). A new approach to measure nancial
contagion. The Review of Financial Studies, 16, 717763.
Baele, L. (2005). Volatility spillover effects in European equity markets. Journal of
Financial and Quantitative Analysis, 40, 373401.
Bekaert, G., Harvey, C. R., & Ng, A. (2005). Market integration and contagion. Journal
of Business, 78, 3969.
Claessens, S., Dornbusch, R., & Park, Y. C. (2001). Contagion: How it spreads and
how it can be stopped. In S. Claessens, & K. J. Forbes (Eds.), International nancial
contagion. Norwell, MA: Kluwer Academic Publishers.
Corsetti, G., Pericoli, M., & Sbracia, M. (2005). Some contagion, some interdepen-
dence: More pitfalls in tests of nancial contagion. Journal of International Money
and Finance, 24, 11771199.
Davidson, J. (2000). Econometric theory. Oxford: Basil Blackwell.
Dungey, M., Fry, R., GonzalezHermosillo, B., & Martin, V. L. (2005). Empirical mod-
elling of contagion: A review of methodologies. Quantitative Finance, 5, 924.
Dungey, M., Fry, R., GonzalezHermosillo, B., & Martin, V. L. (2007). Contagion in
global equity markets in 1998: The effects of the Russian and LTCM crises. North
American Journal of Economics and Finance, 18, 155174.
Dungey, M., & Martin, V. L. (2007). Unravelling nancial market linkages during
crises. Journal of Applied Econometrics, 22, 89119.
Engle, R. F., & Kozicki, S. (1993). Testing for common features. Journal of Business &
Economic Statistics, 11, 369380.
Favero, C. A., & Giavazzi, F. (2002). Is the international propagation of nancial
shocks non-linear? Evidence from the ERM. Journal of International Economics,
57, 231246.
Forbes, K. J., & Rigobon, R. (2001). Measuring contagion: Conceptual and empirical
issues. In S. Claessens, & K. J. Forbes (Eds.), International nancial contagion (pp.
4366). Norwell, MA: Kluwer Academic Publishers.
Forbes, K. J., & Rigobon, R. (2002). No contagion, only interdependence: Measuring
stock market comovements. Journal of Finance, 57, 22232261.
Hendry, D. F. (1995). Dynamic econometrics. Oxford: Oxford University Press.
Hendry, D. F., & Massmann, M. (2007). Co-breaking: Recent advances and a synopsis
of the literature. Journal of Business & Economic Statistics, 25, 3351.
Johansen, S. (1996). Likelihood-based inference in cointegrated vector autoregressive
models. Oxford: Oxford University Press.
Kaminsky, G. L., Reinhart, K. M., & Vgh, C. A. (2003). The unholy trinity of nancial
contagion. Journal of Economic Perspectives, 17, 5174.
Karolyi, G. A., & Stulz, R. M. (2003). Are nancial assets priced locally or globally? In
G. Constantinides, M. Harris, & R. M. Stulz (Eds.), Handbook of the economics of
nance. Amsterdam: North-Holland.
Krolzig, H. M., & Toro, J. (2002). Testing for super-exogeneity in the presence
of common deterministic shifts. Annales dEconomie et de Statistique, 6768,
4171.
Pericoli, M., & Sbracia, M. (2003). A primer on nancial contagion. Journal of Economic
Surveys, 17, 571608.
Perron, P. (1989). The great crash, the oil price shock, and the unit root hypothesis.
Econometrica, 57, 13611401.
Pesaran, M. H., & Pick, A. (2007). Econometric issues in the analysis of contagion.
Journal of Economic Dynamics & Control, 31, 12451277.
Rigobon, R. (2003). On the measurement of the international propagation of
shocks: Is the transmission stable? Journal of International Economics, 61,
261283.