How Do Crises Spread?

Evidence from Investable and Non-investable Stock Indices∗

Brian H. Boyer, Tomomi Kumagai, and Kathy Yuan†

This version: January 2004

abstract

This paper provides empirical evidence that international investor asset holdings have

an effect on global transmission of stock market crisis. By separating stocks into two

categories, those eligible for purchase by foreigners (investable) and those that are

not (non-investable), we estimate and compare the degree to which investable and

non-investable stock index returns co-move with the crisis country index returns.

Our results indicate greater co-movement during high volatility periods, especially

for investable stock index returns. The more pronounced impact on investable stocks

than non-investable stocks provides evidence that crisis spreads through international

investors’ asset holdings rather than through changes in countries’ market fundamen-

tals.

∗Earlier versions of this paper circulated with the title “Are Investors Responsible for Stock

Market Contagion?ӠBrian H. Boyer (boyerb@umich.edu) and Kathy Yuan (kyuan@umich.edu) are at the University

of Michigan Business School, 701 Tappan Street, Ann Arbor, MI 48109-1234. Tomomi Kumagai

(tkumagai@wayne.edu) is at Department of Economics, Wayne State University, 2097 FAB, Detroit,

MI 48202. The authors wish to thank Xiangming Li at the International Monetary Fund for her

assistance, Mark Aguiar, Campbell Harvey, Jerry Hausman, Guido Kuersteiner, Matt Pritsker,

and Ruy Rubeiro for valuable comments. The authors also would like to thank participants at the

University of Michigan finance lunch, the University of Wisconsin-Madison finance seminar, the

Wayne State University economics seminar, the UCLA Brainstorming Conference on International

Capital Flows, and the AFA Winter 2003 meetings in Washington, DC, for helpful comments.

A series of emerging market stock market crises interspersed the decade of the 1990s:

the Mexican peso collapse in 1994, the East Asian crisis in 1997, and the Russian

default in 1998. One striking feature of these crises is how an initially country-

specific event seemed to rapidly transmit to markets around the globe. These events

have prompted a surge of theoretical and empirical interest in “contagion” and in

the determinants of a country’s vulnerability to crises that originate elsewhere.1

On the empirical side, Karolyi and Stulz (1996) and Connolly and Wang (1998a)

find that macroeconomic announcements and other public information do not affect

co-movements of Japanese and American stock markets. King, Sentana, and Wad-

hani (1994) find that observable economic variables explain only a small fraction of

international stock market co-movements. Forbes (2002) finds some evidence of trade

linkage: Countries that compete with exports from a crisis country and those that

export to the crisis country have significantly lower stock market returns. However,

these trade effects can only partially explain the sizably lower stock market returns

observed during crisis periods. In addition, the correlations between market returns

computed by Longin and Solnik (2001), Connolly and Wang (1998b), and Ang and

Chen (2002) are especially large in market downturns, suggesting asymmetry in the

contagion for market downturns and upturns.

Motivated by the lack of evidence that macroeconomic fundamentals serve as

determinants of contagion, researchers have sought for alternative explanations. In

particular, restrictions imposed on international investors’ asset holdings may lead

to limited arbitrage in the international asset market.2 Kodres and Pritsker (2002)

1Contagion, in this paper, is defined as a significant increase in cross-market link-

ages during periods of high volatility, similar to other work in this literature (for ex-

ample, Forbes and Rigobon (2002)). This is a more restrictive definition than conta-

gion defined as spillover effect from one country to another. For a detailed discussion of

the definition, go to ”Definition and Causes of Contagion” at the World Bank Website,

http://www1.worldbank.org/economicpolicy/managing%20volatility/contagion/index.html.2Our paper focuses on contagion among equity markets. Other contagion literature explores

contagion due to currency crises, as in Corsetti, Pesenti, and Roubini (1999), Goldfajn and Valdes

(1997), Chan-Lau and Chen (1998), and Kaminsky and Reinhart (2000). Another branch explores

contagion due to linkages among financial intermediaries, as in Allen and Gale (2000), Lagunoff

and Schreft (2001), and Van Rijckeghen and Weder (2000).

1

develop a theoretical model of financial contagion through cross-market portfolio re-

balancing. One implication of their model is that market co-movements should be

symmetrical in market upturns and downturns. Kyle and Xiong (2001) suggest that

contagion occurs through the wealth effects of investors. When investors suffer a large

loss in investment in the crisis country, they may have to liquidate their positions

in other countries, thus causing equity prices to depreciate elsewhere. Moreover,

Calvo (1999) and Yuan (2003) find that wealth effects persist even when only a

small fraction of investors are wealth-constrained, as long as they are relatively more

informed. They argue that uninformed rational investors are not able to distinguish

between selling based on liquidity shocks and selling based on fundamental shocks. In

the presence of margin-constrained informed investors, it is possible for contagion to

result from confused uninformed investors. Kyle and Xiong (2001), Calvo (1999), and

Yuan (2003) predict that crises spread to stock markets by their wealth-constrained

investors, and that correlations among markets are greater in market downturns.

Although theoretically convincing, there is little empirical evidence for the investor-

induced contagion hypothesis.

To test whether crises spread across countries by asset holding patterns or, in-

stead, by correlated country fundamentals, we utilize a unique attribute of emerging

stock markets. In emerging market countries, not all publicly listed stocks are eligible

for purchase by foreigners. By differentiating between those stocks that are readily

accessible to foreign investors (investable) and those that are accessible primarily to

local investors (non-investable), we form testable implications that allow us to distin-

guish between the investor-induced contagion hypothesis and the fundamental-based

contagion hypothesis. In order to illustrate these testable implications, we describe

a simple thought experiment.

Consider an economy with two emerging markets (A and B) that have indepen-

dent economic fundamentals. There are investable and non-investable stocks in each

country. Suppose there are three types of investors. First, there is a group of investors

who can only invest in investable stocks in countries A and B. For concreteness, one

might interpret these investors as institutional investors from developed countries.

Second, there is a group of investors from emerging market country A who only in-

2

vest in country A’s stocks (investable and non-investable). Similarly, there is a third

group of investors from emerging market country B who only invest in country B’s

stocks (investable and non-investable). One might think of these last two groups of

investors as investors from emerging market countries. We simply assume that in-

vestors from country A (B) face large transaction costs when investing in country B’s

(A’s) asset. These costs may be a reduced form of institutional constraints or infor-

mation asymmetries.3 For example, most governments in emerging market countries

impose capital outflow restrictions on their citizens. All investors face borrowing

constraints (i.e., they have to deposit collateral in margin accounts). When country

A is struck by a country-specific crisis, foreign investors suffer losses in country A’s

investable stock investment and have to withdraw country B’s investable investment

in order to meet margin calls. This selling of country B’s investable stocks is unre-

lated to country B’s economic fundamentals but, nevertheless, leads to a price drop

in country B’s investable stocks. If the price drop in country B’s investable stocks is

severe enough, it may affect country B investors’ wealth constraints and force them

to liquidate their holdings in non-investable stocks. Hence, the idiosyncratic shock

in country A, if severe enough, may spread to country B’s non-investable stocks.

We draw three testable implications from this thought experiment. First, if conta-

gion is investor-induced (either through portfolio re-balancing or wealth constraints),

stocks that are owned by international investors, i.e., investable stocks, should have

a higher than normal level of co-movement with the crisis country stock market

than non-investable stocks during the turmoil period. Alternatively, if contagion is

fundamental-based, the increase in co-movement between investable stocks with the

crisis country stock market during the turmoil period should not exceed the increase

in co-movement between non-investable stocks with the crisis country stock market.

Second, the investor-induced contagion hypothesis also predicts that within a coun-

try, movements of investable stock returns should lead movements of non-investable

stock returns through local investors’ wealth constraints during the turmoil periods.

3One simple story in the spirit of information asymmetry could be as follows. Suppose there is

a language barrier between the two countries, A and B. Then Country A investors will be reluctant

to hold an asset with information written in country B’s language for fear of holding a worthless

piece of paper. The reverse is true for investors in country B.

3

Finally, if crises spread through investors’ wealth constraints, correlations should be

asymmetrically higher in market downturns than in market upturns due to wealth

constraints.

To formally test these implications, we estimate and compare the degree to which

investable and non-investable stock returns co-move with the stock market returns

of the crisis country. If the investable stocks and non-investable stocks have no fun-

damental differences in cash flow, the differences in co-movement can be attributed

to differences in investor activity, with investable stocks more susceptible to outside

shocks through the need of portfolio re-balancing or wealth constraints. Although

seemingly straightforward to implement, calculating the equity market correlation

during the crisis period (in volatile times) is not a trivial statistical exercise. Some

misleading results have been reported in the past that ignore the relationship between

correlation and volatility. Stambaugh (1995), Boyer, Gibson, and Loretan (1999),

and Forbes and Rigobon (2002) note that calculating correlations conditional on high

(low) returns, or on high (low) volatility, induces a conditional bias in the correlation

estimate. To correct for the conditional bias, Forbes and Rigobon (2002) propose a

bias correction methodology based on the assumption of independent and identically

distributed (i.i.d) returns. However, Corsetti, Pericoli, and Sbracia (2002) show that

if the returns are not i.i.d. (for example, if variances increase during crisis periods)

the bias correction errs in the direction of not finding contagion. They modify the

Forbes-Rigobon (2002) test to account for changes in volatility. In this paper, we use

two different methodologies to calculate correlations. First, we estimate a regime-

switching model, in which return variances are allowed to change across regimes.

Second, we estimate correlations of tail observations using extreme value theory,

which is robust under any distributional assumption of returns.4

To determine the beginnings and ends of crisis and non-crisis regimes, we use

a dynamic model of stock returns that allows for endogenous structural breaks. In

particular, we estimate a regime-switching model where the unobserved state variable

4Earlier versions of this paper included another measure of correlation, following the bias cor-

rection suggested by Forbes and Rigobon (2002). The results are similar, and therefore omitted for

the clarity of exposition.

4

is assumed to follow a two-state Markov process. Regime-switching models have

been found to successfully exhibit important features of the correlation dynamics of

financial times series (Ang and Chen (2002), Gibson and Boyer (1997)). We find that

(1) an overwhelming portion of emerging markets in our sample have an increased

correlation in their investable returns with the crisis country during the volatile

regime; and (2) the correlation increase during the crisis time is more pronounced

with investable returns than with non-investable returns. These findings support the

investor-induced contagion hypothesis.

We also consider the time dimension in the transmission of crises between coun-

tries. As a preliminary measure, we test the lead and lag relationship between

investable and non-investable stock index returns for each country. If financial crisis

shocks spread through investors, the more accessible investable stocks should lead

the non-investable stocks during the crisis period. Specifically, we compare the sizes

of the cross-autocorrelations between investable and non-investable stock index re-

turns. Campbell, Lo, and MacKinlay (1997) use this methodology to determine if

large stocks lead small stocks. Our analysis shows that investable stock returns lead

non-investable stock returns during the crisis period. This suggests that crises are

absorbed into investable stocks first, thus providing additional support that interna-

tional investors are responsible for spreading the crises.5

After finding empirical evidence supporting the investor-induced contagion hy-

pothesis, we conduct further tests to determine how crises spread through investors.

We test whether market co-movements are symmetric in extreme market upturns

and downturns. If crises are transmitted due to investor constraints, such as investor

wealth constraints, market co-movements should be greater in market downturns

than upturns. Instead, if crises spread due to investor constraints such as the need

for portfolio re-balancing, co-movements should be symmetric. A test of symmetry

at extreme tails helps to distinguish between these two hypotheses. To do so, we

estimate exceedance correlation, implemented by Ledford and Tawn (1997) and Lon-

5This analysis of the timing of the transmission is preliminary. Further research into a more

structural time-dimension model is necessary to provide a better understanding of how crises spread

from one country to another.

5

gin and Solnik (2001). This methodology is based upon the non-degenerate limiting

distribution of the extreme tails, and is free of any assumptions on the underlying

joint distribution. In estimating exceedance correlation, we find that correlations in

negative tails are higher than correlations in positive tails, and that this effect is

more pronounced for investable returns than non-investable returns. This indicates

that some constraint, which causes investors to respond more to negative external

shocks than to positive external shocks, may be responsible for transmitting crises.

Investor wealth constraints could be one such constraint.

These tests of the investor-induced contagion hypothesis against the correlated

country fundamental hypothesis hinge on the crucial assumption that investable firms

share similar cash-flow fundamentals with non-investable firms, and differ only in in-

vestor ownership. Otherwise it is difficult to attribute the difference in co-movements

strictly to investor ownership. We establish the validity of this assumption prior to

our analysis of co-movements.

Our paper is closely related to the literature on limited arbitrage (Shleifer and

Vishny 1997). In contrast to the traditional asset pricing theory where the view

is that co-movement in prices reflects co-movement in fundamentals in an econ-

omy with rational investors, the limited arbitrage literature emphasizes that market

frictions or noise trader risks break the link between asset price movements and eco-

nomic fundamentals. For example, in a recent paper, Barberis, Shleifer, and Wurgler

(2003) find evidence for investor-induced co-movements by showing that some co-

movements of stock returns can be attributed to asset re-classification. Our paper

is complementary to this line of research since it provides cross-country evidence for

non-fundamental-based, investor-induced theory of co-movements.

Our paper also is related to the literature on measuring contagion across markets.

Contagion in this literature is usually defined as correlation between markets in ex-

cess of what would be implied by economic fundamentals. Different approaches have

been adopted to model the linkage between economic fundamentals and asset mar-

ket co-movements. For example, Bae, Karolyi, and Stulz (2003) estimate the joint

occurrences of large absolute returns using multinomial logistic regression and define

contagion within regions as the fraction of exceedance events that is not explained

6

by fundamentals (exchange rates, interest rates, and market volatility). Instead of

studying correlations of returns, they focus on analyzing the joint occurrences of

extreme events. Another approach is to use a factor model. For example, Bekaert,

Harvey, and Ng (2003) apply a two-factor model with time-varying betas and mea-

sure contagion as correlation among model residuals. By defining the contagion as

excess correlation over and above what one would expect, they measure the extent

to which the crisis spreads due to reasons other than changes in economic fundamen-

tals. Tang (2001) uses a similar approach but restricts the factor model to a world

CAPM. In our analysis, by demonstrating that investable and non-investable stocks

share similar economic fundamentals, we can measure the excess correlation over and

above beyond what one would expect from economic fundamentals as the difference-

in-difference of co-movement with the crisis country for investable and non-investable

stock returns. Hence, we circumvent the need of specifying and estimating a factor

model.6

The remainder of the paper is organized as follows. Section I presents the hy-

potheses we consider and the methodologies we use to construct our tests. Section

II describes the data and the summary statistics of the index returns (separately for

investable and non-investable stock index returns). This section also compares the

industry and market microstructure characteristics of investable and non-investable

stocks. Section III discusses tests results and the hypotheses laid out in Section I.

6The theoretical literature on international asset pricing under mild segmentation (e.g., unequal

access to certain asset markets) predicts that restricted securities command a “super risk premium”

over unrestricted securities (Errunza and Losq, (1985) and (1989)), where the “super risk premium”

is related to three elements: (1) the risk aversion coefficients of two types of investors (those who

have restricted and those who have unrestricted access to the market), (2) the market capitalization

of all ineligible stocks (available only to those with unrestricted access to the market), and (3) the

return covariance of the ineligible stock with the portfolio of all ineligible stocks. This risk premium

does not vary with returns of eligible stocks. This theoretical finding lends support to the formation

of our testable hypotheses, since it predicts the “super risk premium” is the same over the volatile

and stable regimes. By examining the difference in pricing between investable (eligible) and non-

investable (ineligible) stocks (see Test 2A in Section II), we test two views of how crises spread:

fundamental-based versus investor-induced. We thank one of the participants at the AFA 2003

winter meetings in Washington, DC, for bringing this to our attention.

7

Section IV concludes.

I. Methodology

To gauge the cross-market transmission mechanism of financial crises, we em-

ploy two different methodologies: regime-switching model and exceedance correla-

tion. These methodologies are used to calculate the co-movements of investable and

non-investable stock index returns and to test our hypotheses on the transmission

mechanism of crises.

A. Test Hypotheses and Statistics

We first describe our hypotheses on the transmission mechanism of crises as a

series of tests and spell out the corresponding test statistics.

Our first test is a test for contagion. We estimate co-movement measures (ρ)

between the crisis country’s total market index returns with another country’s total

market index returns (indexed by T ), investable index returns (indexed by I), and

non-investable index returns (indexed by N) during the stable and crisis periods. We

compare correlation differences between stable and crisis periods. If contagion exists,

no matter what the transmission mechanism is (investor-induced or fundamental-

based), then the investable stock index returns should have a greater co-movement

with the crisis country market index returns during the turmoil period. Therefore,

the relevant test statistic for existence is the difference in correlations across peri-

ods between the crisis country’s total market index returns and another country’s

investable index returns (indexed by I).

Test 1 (Existence of stock market contagion) If stock market contagion exists, co-

movement is higher during the turmoil period for investable index returns (I).

T1 :

H0 : ρI, turmoil − ρI, stable ≤ 0

H1 : ρI, turmoil − ρI, stable > 0

For the implementation of this test and those below, we use one-sided tests where

the alternatives support our hypothesis. To summarize the results across countries,

8

we use the nonparametric sign test to compare the number of positive and negative

differences to determine which sign is more likely overall.

Our second set of tests is a test of how crises spread: through changes in country

fundamentals or through investor constraints (e.g., portfolio re-balancing or wealth

constraints). For this test, we compare how the crisis affects the index of investable

stocks to how the crisis affects the index of non-investable stocks. By the argu-

ment from the simple thought experiment described earlier, if crises spread through

correlated fundamentals, the effect on investable returns should be the same as the

effect on non-investable returns. In other words, investable co-movements should not

be significantly higher than non-investable co-movements. A finding of significantly

higher co-movements for investable returns during the turmoil period will reject the

fundamental-based hypothesis and indicate that contagion is investor-induced.

Test 2 A (Fundamentals vs. Investor-Induced Hypothesis) If stock market crises

spread due to investor constraints rather than fundamentals, the correlation in-

crease during turmoil periods is more pronounced for investable returns than for

non-investable returns.

T2− A :

H0 : (ρI, turmoil − ρI, stable)− (ρN, turmoil − ρN, stable) ≤ 0

H1 : (ρI, turmoil − ρI, stable)− (ρN, turmoil − ρN, stable) > 0

Furthermore, the investor-induced contagion hypothesis predicts that within a

country, movements of investable stock returns should lead movements of non-investable

stock returns through local investors’ constraints (either due to portfolio re-balancing

needs or wealth constraints) during market downturns. In other words, if exter-

nal shocks spread through investors, then investable stocks should respond first

and lead the non-investable stocks. In order to test the lead and lag relationship

between investable and non-investable returns for each country, we compare the

cross-autocorrelation between investable and non-investable returns as in Camp-

bell, Lo, and MacKinlay (1997) in their comparison between large and small size

stocks. Higher correlation between non-investable returns and lagged investable re-

turns (than correlation between investable returns and lagged non-investable returns)

suggests that investable returns lead non-investable returns.

9

Test 2 B (Fundamentals vs. Investor-Induced: Lead-Lag) If stock market crises

spread due to investor constraints rather than fundamentals, the effect first appears

in the investable stock index, followed by the non-investable stock index.

T2 − B :

H0 : ρIt−1,Nt − ρIt,Nt−1 ≤ 0

H1 : ρIt−1,Nt − ρIt,Nt−1 > 0where It denotes the investable stock

index return at time t, and Nt denotes the non-investable stock index return at time

t.

Our third test is a further test of how crises spread. It tests whether market

co-movements are symmetric in extreme market upturns and downturns. If crises

are transmitted due to investors’ portfolio re-balancing needs, market co-movements

should be symmetric in extreme market upturns and downturns. On the other hand,

if crises are transmitted due to investor wealth constraints, market co-movements

should be greater in extreme market downturns than market upturns. A test of

whether market co-movements are symmetric or not at extreme tails allows us to

distinguish these two hypotheses. More specifically, we estimate and compare co-

movements at two extreme tails: jointly positive, denoted by (+), and jointly nega-

tive, denoted by (−). If the cross-market linkage is through investor stock ownership,

the relevant test statistic is the co-movement difference for investable index returns.

Despite the similarity of this set to Test 1, there is one important difference: Instead

of comparing co-movement between the turmoil period and the stable period, we

compare correlations at two extreme tails: extreme market upturns (+) and down-

turns (−).

Test 3 (Portfolio Re-Balancing vs. Wealth Constraints) If stock market crises are

spread by constraints such as investor wealth constraints rather than the need for

portfolio re-balancing, co-movement is higher during extreme market downturns than

upturns, especially for investable stock index returns.

T3 :

H0 : ρI, (−) − ρI, (+) ≤ 0

H1 : ρI, (−) − ρI, (+) > 0

The rest of this section describes in detail the methodologies we employ to esti-

mate the co-movement (ρ).

10

B. A Dynamic Model of Correlation: Regime-Switching Model

Our first methodology of estimating the correlation between asset returns is to

specify a dynamic model of stock returns that captures the potential differences in

the co-movements of asset returns. We estimate a parametric model of the joint

data-generating process that allows for endogenous structural breaks, but also nests

a simple model with fixed moments. We then determine if the joint distribution

changes between regimes by testing the differences in estimated parameters.

The model is the regime-switching model where the unobserved state variable

is assumed to follow a Markov process. Regime-switching models have been found

to successfully exhibit important features of the correlation dynamics of financial

time series (Ang and Chen (2002), Gibson and Boyer (1997)). We model the data

generating process of four time-series: the total returns of the crisis source coun-

try, investable and non-investable returns of another country, and the world index

returns.

The unobserved state variable in our model, st, is allowed to take on one of two

values, st ∈ {1, 2}. Conditional on being in state s, returns are assumed normally

distributed,

(rt|st = s) ∼ N (αs,Σs) (1)

where rt is a 4× 1 vector of index returns realized at time t and t ranges from 0 to

T , αs is a vector of expected returns, and Σs is a 4 × 4 covariance matrix. Let ψt

incorporate all information at time t. The unobserved state variable st defines the

regime at time t and is assumed to follow a two-state Markov process:

P (st = b|ψt−1) = P (st = b|st−1 = a) = pab (2)

where a, b = 1, 2 thus resulting in a 2 × 2 transition matrix. For any two indices

out of four (crisis, world, investable, and non-investable), let σji,s be the covariance

between index i and index j given st = s (an element in Σs):

σij,s = Cov(rit, rjt|st = s) (3)

where rit is the return for index i. In addition, to clear up ambiguity in the definition

of states, let st = 1 be the regime in which estimated volatility for the crisis index

11

is higher. The correlation estimate ρij,s between rit and rjt given state st = s is as

follows:

ρij,s =σij,s√

σii,sσjj,s

. (4)

Once we have estimates of Σs for each state, the test for contagion is a test for a

significant increase in the correlation in the crisis period.

The parameters of the regime-switching model are estimated using a two-step

estimation process, using weekly data.7 In the first step, we define the two regimes

by estimating a regime-switching model using two index returns: the crisis country

index return and the world index return. The crisis country index helps associate the

regimes with a particular financial crisis; while the world index allows for a clearer

distinction between the regimes by helping to define the high volatility regime as

a regime of high global volatility rather than country-specific volatility. Maximum

likelihood estimates of the model parameters, including the elements of the transition

matrix, are obtained using the approach discussed in Hamilton (1989), and robust

standard errors are generated using the method of White (1980). This involves

estimating a total of 13 parameters for each country: two vectors of means (each

with two elements), two covariance matrices (each with three unique elements), the

two diagonal elements of the transition matrix, and the initial state probability.

Once we have estimated the parameters that define the two regimes, we estimate

the probability of being in the high volatility (correlation) regime for each period

in the sample. This is done by estimating smoothed inferences using an algorithm

developed by Kim (1994) and discussed in Hamilton (1989). This algorithm provides

an estimate of P (st = a|ψT ) under the model assumptions described above. That

is, we are able to estimate the probability of being in state a given all information

of the entire sample. Hamilton (1989) estimates smoothed inferences using data on

GNP growth to determine time periods of economic recession. Hence, the regime

switching model estimated in this paper not only provides us with a basis to test the

existence of two separate regimes, but also allows us to objectively distinguish the

7We have revised the estimation method to maintain the same endogenously specified regimes

across all countries.

12

time periods that are more likely representative of a crisis (or stable) regime.8

In the second step, we estimate parameters of a conditional regime-switching

model of the four index returns: the crisis country index return, the investable and

non-investable index returns of another country, and the world index return, condi-

tioned on the estimated regime probabilities. The resulting standard errors account

for second-step estimation variance as well as the first-step estimation variance, using

the conventional delta method.

C. Exceedance Correlation

An alternative estimate of correlation between returns is obtained from the ex-

treme tails of the joint distribution. We estimate the extreme tails of the joint

distribution using extreme value distribution methodology as in Longin and Solnik

(2001). This method is attractive in that we do not need to specify a joint distri-

bution function and do not need to define a crisis period. Instead, we rely strictly

on the tail distribution, which is a non-degenerate limit distribution that converges

to a generalized Pareto distribution, regardless of the underlying joint distribution.

This enables us to estimate the correlation without enforcing the underlying joint

distribution of returns to be normal. The result is also robust for non-i.i.d. returns.

In addition, this methodology allows us to distinguish the differences in correlation

during market upturns and downturns within periods of high volatility.

We estimate this limiting distribution to characterize the exceedance correlation

between returns of the crisis country and another country for the two tail ends of the

joint distribution function: positive tail (for jointly positive returns or bull market)

and negative tail (for jointly negative returns or bear market). If the joint distri-

bution is indeed a bivariate normal, the exceedance correlation is asymptotically

independent, which means that for high values of threshold (away from the mean),

the exceedance correlation will tend to zero, symmetrically in both tails. In their

study of correlations between equity markets, Longin and Solnik (2001) find decreas-

8In the earlier versions of this paper, we estimated a separate four-variable regime-switching

model for each country, with the crisis country. We noted that the average regime probabilities

obtained from that approach are very similar to the regime probabilities of this approach.

13

ing exceedance correlation in the bull market (positive tail) and increasing or flat

exceedance correlation in the bear market (negative tail), thus providing evidence

for asymmetry. Here, in addition to estimating the exceedance correlations of inter-

national equity markets at both tails, we estimate separate correlation measures for

the investable and non-investable portions of the markets, to see if the asymmetry

also appears at the disaggregate level.

C.1. Tail Distribution: Generalized Pareto Distribution

For a univariate random variable (X), the tail probability beyond the threshold

level (θ) is obtained from the generalized Pareto distribution:

Pr (X > x|X > θ) =

(1 +

ξ(x− θ)

σ

)−1/ξ

+

(5)

for x > θ,9 where ξ and σ are shape parameters, also called the tail index and

dispersion parameters, respectively. The tail index characterizes the tail distribution:

ξ > 0 corresponds to fat-tailed distributions, ξ < 0 to thin-tailed distributions, and

ξ = 0 to no-tail or finite-distributions.10 Therefore, the univariate tail distribution

(F ∗(x)), conditional on the observation being in the tail, is fully characterized by

two parameters: the tail index (ξ) and the dispersion parameter (σ):

F ∗(x) = 1−(

1 +ξ(x− θ)

σ

)−1/ξ

+

(6)

For the bivariate (or more generally, multivariate) case, the resulting multivariate

distribution of extreme returns is a function of the univariate generalized Pareto

distributions. As stated before, this distribution does not depend on the underlying

true distribution. Let (X1, X2, . . . Xq)′ be a q-dimensional vector of random variables,

9The plus-sign notation is defined as s+ = max(s, 0).10Longin and Solnik (2001) cite some known tail index values for various distributions. ξ = 0

for normal and log-normal distributions, ξ > 0 for student-t distribution, and ξ < 0 for uniform

distribution (Embrechts, Kluppelberg, and Mikosch (1997)). Also, various processes based on the

normal distribution - autocorrelated processes, discrete mixtures of normal distributions and mixed

diffusion jump process – have thin tails, so that ξ = 0 (Leadbetter, Lindgren, and Rootzen, (1983)).

GARCH process have ξ < 0.5. (De Haan, Resnick, Rootzen, and DeVries, (1989)).

14

and (θ1, θ2, . . . θq)′ denote the vector of threshold level. Following Ledford and Tawn

(1997) and Tawn (1988), the joint tail distribution is as follows:

F ∗(x1, x2 . . . xq) = exp

[−D

( −1

logFm1 (x1)

,−1

logFm2 (x2)

, . . .−1

logFmq (xq)

)](7)

where D() is some dependence function and Fmj () is the marginal distribution of the

jth variable:

Fm(xj) = (1− λj) + λj

(1 +

ξj(xj − θj)

σj

)−1/ξj

+

(8)

with λj as the tail probability. The marginal distribution combines two parts: prob-

ability (1 − λj) that the observation is not in the tail and probability λj that the

observation is in the tail and represented by the Pareto distribution of the tail. In the

multivariate case, although the limiting marginal distributions are known, the depen-

dence function is not known. Therefore, the estimation of the joint tail distribution

requires choosing a functional form for the dependence function.

C.2. Dependence Function and Estimation

The models of the dependence function can be either nonparametric or paramet-

ric. In the class of parametric functions, Tawn (1988) describes two distinct models:

mixed and logistic. These mixed and logistic models have equal-weight (symmet-

ric) and weighted (asymmetric) versions where equal weight means simply that the

weight given to each of the variables are all the same.11

The most widely used dependence function is the equal-weighted (symmetric)

logistic function:

D(z1, z2 . . . zq) = (z−1/α1 + z

−1/α2 + · · ·+ z−1/α

q )α (9)

where the dependence parameter is α, 0 < α ≤ 1. With only one parameter to

estimate, this dependence function is simple and tractable. Furthermore, with this

11The use of the word “symmetry” here refers to the weights given to each variable (in this case,

asset returns) in the model. This “symmetry” is not the same as it appears elsewhere in this paper,

where it is used to describe the difference in the tail distribution between market upturns and

downturns.

15

dependence function for q = 2, the measure of exceedance correlation is 1 − α2,

which is also very simple to compute. We proceed with this dependence function to

estimate the correlation between markets. With the symmetric logistic dependence

function, the bivariate tail distribution has seven parameters: two tail probabilities

(λ), two tail indices (ξ), and two dispersion parameters (σ) and parameters from the

dependence function, which in this case is one (α).12

We estimate parameters of the bivariate extreme value distribution by maximum

likelihood, assuming independent observations. The likelihood function for each ob-

servation is one of the four functions, determined by whether the pair of observations

each satisfy the threshold condition, x > θU for upper tail exceedance (and x < θL

for lower tail exceedance). The details on the construction of the likelihood function

can be found in Prescott and Walden (1980), Ledford and Tawn (1997), and Lon-

gin and Solnik (2001) and it is straightforward to implement. We use asymptotic

properties of the maximum likelihood estimator to obtain our parameter covariance

matrix, and perform inference based on asymptotic normality.

II. Data

This section offers a discussion of the data. Our analysis requires two types

of stock index returns for each country: returns for investable stocks and for non-

investable stocks. In developed financial markets, such as the U.S., practically all

stocks are accessible to both foreign and local investors, so that the entire market is

investable. However, in many emerging markets, some stocks are only accessible to

local investors. Section A explains the construction of the index returns for these two

categories. Section B compares the characteristics of the firms that are designated

investable and non-investable. Section C describes the overall sample correlation

between investable and non-investable stock index returns for each country.

12A future extension to this study is to consider nonparametric dependence functions to see if

the choice of the symmetric logistic dependence functions is appropriate for the data.

16

A. Investable and Non-Investable Stock Index Returns

We use two sources of financial market data for this paper: International Finance

Corporation’s (IFC) Emerging Markets Data Base (EMDB) and Datastream. Table

1 reports the size of the markets for the 31 emerging market countries and the 16

developed countries, measured in total market capitalization and total value traded

(in millions of U.S. dollars at year-end).

In the EMDB, IFC provides two stock market index series for each emerging

market: IFC Global (IFCG), representing the total market,13 and IFC Investable

(IFCI), consisting of firms that are designated to be investable. Both IFC index

series are dividend-inclusive and U.S. dollar-denominated. The IFC selects stocks

for inclusion in its IFCG index by reviewing the trading activities and targets market

coverage of 60 percent to 75 percent in market capitalization.14 The stock returns

included in the IFCI index consist of a subset of firms that are in IFCG, and selection

for this subset is a three-step process. First, the IFC determines which securities may

be legally held by foreigners.15 Next, the IFC applies two further screening criteria

for practicality of investment. The first criterion screens for a minimum investable

market capitalization of $50 million or more over the 12 months prior to a stock’s

addition to an IFCI index. The second criterion screens firms for liquidity. A stock

must trade at least $20 million over the prior year for inclusion in an IFCI index, and

also must have traded on at least half the local exchange’s trading days. Thus, the

13The totals reported for the emerging market countries are the totals in IFC Global index and

are not totals of all stocks.14The IFC trading criteria are as follows: Any share selected must be among the most actively

traded shares in terms of value traded; must have traded frequently during the annual review

period (i.e., one large block trade might skew the value traded statistics); and must have reasonable

prospects for a continued trading presence in the stock exchange without imminent danger of being

suspended or de-listed. Stocks are selected in the order of the trading criteria until the market

capitalization coverage target of 60 percent to 75 percent of total market capitalization is met.15The first legal test of a stock’s investability is to determine whether the market is open to

foreign institutions. IFC determines the extent to which foreign institutions can (1) buy and sell

shares on local exchanges, and (2) repatriate capital, capital gains, and dividend income without

undue constraint. The second legal test is to determine whether there are any corporate bylaw,

corporate charter, or industry limitations on foreign ownership of the stock.

17

IFCI index is designed to measure the composite stock market index of what foreign

investors might receive from investing in emerging market securities that are legally

and practically available to them (IFC 1999).

Since IFC does not provide an explicit series for non-investable stocks (IFCN),

we construct the index return using the global (IFCG) and the investable (IFCI)

series of index return (R) and market capitalization (M) as follows:

RIFCN =MIFCG ×RIFCG −MIFCI ×RIFCI

(MIFCG −MIFCI)(10)

The stocks in the global index that are not in the investable stock index are desig-

nated to be non-investable. We henceforth refer to the two stock index returns as

investable return and non-investable return.16

For the developed countries, we treat all stocks as investable. We use the U.S.

dollar-denominated country index returns from Datastream (Datastream Total Mar-

ket Index) as the investable index for these countries. The returns are calculated as

the first difference of the natural logarithm of the prices. Friday prices are used for

the weekly returns.17 We also use the Datastream World Market Index in weekly

frequency as a measure for the overall world market in our estimation of the regime

switching model.

We have 665 observations of weekly (Friday-to-Friday returns) data for the period

January 1989 to September 2001, for both EMDB and Datastream data, with the

summary statistics in Table 2.

The percent of market capitalization attributed to non-investable firms varies

greatly across countries. The average, weighted by total market capitalization, is 45%

in 1994 and declines to 20% in 1999 as shown in Table 3. From 1994 to 1999, some of

the countries with consistently high non-investable percentages are in Asia: China,

16For precision, the non-investable index returns are calculated for observations for which the

non-investable portion of the total market capitalization is greater than 0.01%. When the fraction

of non-investable is extremely small, the precisions of the global and investable returns are not

sufficient for accuracy. This criterion was sufficient in deleting most of the very large returns for

weekly frequency data.17If Friday observations are not available due to closings or insufficient data, we use Thursday

observations.

18

Sri Lanka, India, Taiwan, and Thailand. Outside Asia, other countries with high non-

investable percentages are Czech Republic, Jordan, and Zimbabwe. The emerging

market countries with low non-investable percentages are Argentina, Mexico, Peru,

Malaysia, Greece, South Africa, and Turkey.

We also use the stock level data from EMDB, in monthly frequency, for 1994

to 2001. IFC assigns a value ranging from 0 to 1 to each stock, representing how

accessible a firm is to foreign investors (“degree of open factor”). When this degree

of open factor is 1, the stock of the firm is completely investable (accessible to

foreigners); at 0, the stock of the firm is completely non-investable (inaccessible) to

foreigners. At the end of 1994, we have data on 1505 stocks across all emerging

market countries, with 480 completely investable, 534 completely non-investable and

491 in between. At the end of 1997, there are 2005 stocks, with 535 completely

investable, 581 completely non-investable and 889 in between.

Note that the degree of investor accessibility may not be a good proxy for the

degree of actual foreign ownership - a stock that is designated as investable may or

may not be owned by foreign investors. If the measure of investability overstates for-

eign ownership, there will be a bias against finding any systematic difference between

investable and non-investable stocks. In other words, our estimated differences in

co-movements are conservative measures of the more pronounced actual differences

that would be found if exact investor ownership were known.

B. Characteristics of Investable and Non-Investable Firms

In order to compare the impact of crises on the emerging market’s investable index

and non-investable index, it is necessary to understand the differences in the sets of

firms that comprise the indices. Our test of the investor-induced contagion hypoth-

esis against the correlated country fundamental hypothesis requires the assumption

that investable firms share similar cash-flow fundamentals with non-investable firms.

That is, investable and non-investable firms differ primarily in investor holdings. To

explore structural differences that may affect our results, we examine the character-

istics of investable and non-investable firms. In these comparisons, investable firms

are those with a degree of open factor equal to one, while non-investable firms are

19

those with a degree of open factor equal to zero.18

One concern in our comparison is that the difference in co-movement may be due

to the fact that non-investable firms are in different industries from investable firms.

In particular, tradable sectors are more vulnerable to exogenous shocks originated

elsewhere in the world. Table 4 compares the distribution of investable firms with

non-investable firms among industries during the sample period (1994-2001). It

shows that investable firms feature more in the Finance, Insurance, and Real Estate

sector (a predominantly non-tradable sector), while the non-investable firms feature

more in the Manufacturing sectors (a predominantly tradable sector). If investable

firms are more concentrated in the tradable sectors (such as in the manufacturing

sector), the higher correlation can be caused by external shocks to tradable firms’

balance sheets (since these firms have a higher β with global portfolio). This is

disputed by the data because non-investable firms are more concentrated in the

manufacturing sector. Overall, there is no significant difference in the distribution of

investable firms with non-investable firms among these industries. Bae, Chan, and

Ng (2002) also find no pronounced variation in the investability across industries.

Therefore, any differences we find in correlations cannot be attributed to sector

differences in investable and non-investable firms.

To determine if there are any other differences between the two sets of firms,

investable vs. non-investable, we compare two market microstructure characteristics

of the stocks: size (market capitalization as a fraction of total market capitalization)

and liquidity (ratio of value-traded to market capitalization). Table 5 reports the

one-sided test of the means to see if the average size or liquidity is significantly higher

for completely investable firms than completely non-investable firms. We conduct

this test using January data for 1994 (the year of Mexican crisis), 1997 (the year of

Asian crisis), and 1998 (the year of Russian crisis) respectively. In 1994, out of 16

countries for which comparison was possible, we find that average liquidity for in-

vestable firms is significantly greater in four countries (Mexico, Venezuela, Pakistan,

and South Africa) and average size is significantly greater in seven countries (Colom-

18For the comparisons in this section, we do not include firms with the degree open factor between

zero and one.

20

bia, Peru, Pakistan, Sri Lanka, Czech Republic, Greece, and Hungary). In 1997, out

of 16 countries, average liquidity for investable firms is significantly greater in four

countries (Mexico, Venezuela, Pakistan, and Czech Republic) and average size is sig-

nificantly greater in six countries (Colombia, Peru, Venezuela, Pakistan, Sri Lanka,

and Hungary). In 1998, out of 19 countries, average liquidity for investable firms is

significantly greater in three countries (Colombia, Mexico, and Czech Republic) and

average size is significantly greater in seven countries (Mexico, Venezuela, Pakistan,

Sri Lanka, Hungary, Slovakia, and Egypt).19 Our findings of average liquidity across

two types of firms are similar to those of Bae, Chan, and Ng (2002): Average liquid-

ity of investable stocks is not significantly greater than that of non-investable stocks;

but they found that the average size of investable stocks is significantly higher than

that of non-investable stocks.20 Large exogenous shocks generally have a smaller

impact on large stocks. From this perspective, investable stocks are less susceptible

to external shocks, and, hence, we may expect their co-movements to change less

during crises. Therefore, any finding of increased co-movements for investable firms

during crises cannot be due to size. Furthermore, since cash flows of investable firms

are similar to cash flows of non-investable firms, a finding of higher cross-country

co-movements for investable stocks during crises would reject the correlated country

fundamental hypothesis.

19The monthly EMDB stock-level data used in this analysis contain more stocks than are con-

tained in the Global index. At the index level, the difference in liquidity and size between investable

and non-investable may be even less pronounced, because liquidity and size are in the criteria for

selecting stocks into the indexes (which includes both investable and non-investable stocks). Given

that our hypothesis tests are conducted at the index level, the less pronounced differences in market

microstructure characteristics lead to a lesser concern.20Chari and Henry (2003) found the opposite results (they found higher liquidity for investable

but no difference in size between investable and non-investable stocks) using data before stock

market liberalization dates. Bae, Chan, and Ng (2002) conduct the analysis on liquidity and size

using the whole EMDB data sample (1994-2001), and we conduct the analysis using January of

crisis years: Mexico (1994), Asia (1997), and Russia (1998).

21

C. Sample Correlations: Investable vs. Non-Investable

Our methodology tests whether the co-movements of investable returns and cri-

sis country index returns are significantly higher than the co-movements of non-

investable returns and crisis country index returns. For this methodology to be

meaningful, investable and non-investable returns should be different and not highly

correlated. In Table 2, we show that investable returns are not identical to non-

investable returns. The sample correlations between investable and non-investable

returns for each country also are reported in the right-most column. The average

sample (contemporaneous) correlation between investable and non-investable returns

is 0.647.

In the weekly data, all the sample correlations are positive. The countries with

especially high positive correlations (over 0.9) are India, Korea, Malaysia, Taiwan,

Thailand, and Jordan. We are cautious in interpreting the results for these high

correlation countries. Other moderately high correlations are in Brazil and Philip-

pines. The low correlation countries are China, Slovakia, and Turkey. We conduct

all analyses at the weekly frequency to avoid market microstructure complications

that appear in daily level data.21

III. Contagion in the 1997 Asian Crisis

We consider the 1997 Asian crisis as our empirical analysis of transmission mech-

anisms of crises. As stated in Forbes and Rigobon (2002), there were multiple events

behind this Asian financial crisis, beginning with the Thailand stock market crash in

June, the Indonesian market crash in August, and then the Hong Kong market crash

in mid-October. As a benchmark case, we first adopt Forbes and Rigobon’s definition

of the Asian crisis, using the Hong Kong equity market as the source of contagion.

We use the measures of co-movement described in the previous section: from the

21In addition to positive contemporaneous correlation with weekly data, we also find but do

not report that contemporaneous correlations are positive negative with daily data. This pattern

is recognized as a market microstructure effect in the literature (Campbell, Lo, and MacKinlay

(1997)). Moreover, another market microstructure issue is that daily market closing prices are

non-synchronous across the globe.

22

regime-switching model estimation and from exceedance correlation estimation. For

comparison, we then consider Thailand as the source of crisis.

A. Test 1: Existence of Stock Market Contagion

To test for the existence of stock market contagion, we examine whether there are

any increases in correlation from the stable regime to the crisis regime. We measure

the increase in correlation between all three index returns (investable index returns,

total market index returns, and non-investable index returns) and the crisis country

returns. If the cross-market linkage is through investor ownership, the correlation

change between investable index returns with the crisis country index returns is the

most relevant.

We analyze a regime-switching model where volatile and stable periods are es-

timated rather than exogenously specified. Note that high volatility regimes may

not be limited to the periods surrounding the Asian crisis. We report test results

in Table 6A with Hong Kong as the crisis source country. For each country in the

table we separately estimate the parameters of the regime switching model outlined

in Section I.B. Since we are mainly interested in the change in volatilities and corre-

lations across regimes and disaggregate index returns, we report only differences in

these estimated moments and do not report levels.22 Changes in volatility have been

annualized by multiplying by√

52.

The crisis regime is defined as the state with higher volatility in the crisis country

index returns. In Table 6A, among the 30 emerging market countries, the differences

in volatility (column Volatility) are significantly positive in 26 countries for investable

returns and in 24 countries for non-investable returns, at five percent significance

level. This positive increase in volatility with investable returns and non-investable

returns is strong evidence for the existence of a crisis regime during the sample

estimation period.23

22Full set of parameter results, as well as our original results from estimating a regime switching

model for each country paired with the crisis country, are available from the authors upon request.23In fact, there is a marked increase in volatilities of all four indices: crisis country, world market,

investable and non-investable in this regime. The results are similar to our results from estimating

23

We test for the existence of contagion by determining if the differences in cor-

relation are positive for investable index returns, as compared with non-investable

index returns. For investable returns, 24 countries show statistically significant in-

creases in correlation with Hong Kong stock market index returns (column Corr C),

and four additional countries show statistically significant increases in correlation

with the world index returns (column Corr W), at five percent significance level.

For non-investable returns, slightly fewer number of 16 countries show a statistically

significant increase in correlation with Hong Kong stock market index returns, and

24 countries show statistically significant increase in correlation with the world index

returns. The sign test rejects the null that there are no increases in correlation with

the crisis country for investable returns but not for non-investable returns.

The similarity in the increased correlation with Hong Kong stock market index

returns and with the world index returns seems to suggest that the increase in cor-

relation with Hong Kong index returns is due to the increased correlation with the

world index returns. Alternatively, we consider Thailand as the possible source of

the contagion, with the direction of contagion being from Thailand to Hong Kong,

and not vice versa.

Figure 1 is a graph of time-series of the stock index returns for some Asian

countries from January 1997 to December 1997. It shows that the volatile period

started much earlier than October 1997. According to Nouriel Roubini’s account of

the Asian Crisis and its global contagion,

July 2 - The Bank of Thailand announces a managed float of the Baht

and calls on the International Monetary Fund for “technical assistance.”

The announcement effectively devalues the Baht by about 15-20 percent.

It ends at a record low of 28.80 to the dollar. This is a trigger for the

Asian crisis.24

Radelet and Sachs (1998) also document that the Thailand crisis started with this

devaluation: “... the Thai Baht devaluation triggered the capital outflows from the

rest of East Asia.” Given the consensus in the literature that the triggering event

a separate regime switching model for each country paired with the crisis country.24http://www.stern.nyu.edu/nroubini./asia/AsiaChronology1.html

24

started on July 2, we modify the regime switching model by redefining the crisis

country to be Thailand.25

Results from estimating the regime-switching model using Thailand as the crisis

country are reported in Table 6B. Again we find that volatilities of investable and

non-investable returns are significantly higher in the crisis regime, the state in which

estimated index return volatility of the crisis country, Thailand, is higher at five per-

cent significance level. The pattern of increased correlations with the Thailand stock

market index returns and the world index returns during the turmoil regime is also

similar to what we find when we use Hong Kong as the crisis source country. The

statistically significant increase in correlation with Thailand stock market returns

during the volatile regime is more pronounced in investable index returns (18 coun-

tries) than in non-investable index returns (12 countries). For investable returns, a

total of 15 countries show a statistically significant increase in correlation with the

world index returns (column Corr W). For non-investable returns, 10 countries show

statistically significant increase in correlation with the world index (column Corr

W).

The above results indicate that, in general, periods of high volatility are associ-

ated with periods of high correlation. Moreover, in the case of emerging markets,

high correlation is especially characteristic of the investable indices. In Figure 2, we

plot the probability of being in the high volatility (correlation) regime across time.

These are estimates of smoothed inferences (Kim 1994) using Hong Kong as the crisis

country in one and using Thailand as the crisis country in another for a sub-sample

period from January 1996. In Hong Kong (figure 2A), the probability of being in

the high volatility (correlation) regime jumps from near zero to near unity during

1997. The probability increases from sub-ten percent in May and June of 1997 to

100 percent in November 1997, in three jumps: first in July, then September, and

then finally in November. The downward spikes during 1997 measure at near zero

percent in August and 15 percent in October. From November, 1997, the probability

of being in the crisis regime is between 70 and 100 percent until February 1998. In

25Baur (2003) also shows evidence of Thailand shock transmitting volatility to other countries in

between July and September, prior to the Hong Kong shock in October.

25

Thailand (Figure 2B), the probability of being in the high volatility (correlation)

regime jumps dramatically from near zero to near unity slightly earlier in 1997, in

May. Then, it stays near 100 percent level for a longer duration, only dipping down

to 80 percent in April 1999. In comparing figures 2A and 2B, we note that the crisis

regime begins much earlier for Thailand, and the regime probabilities graph is much

smoother (hence the two regimes are more clearly defined) for Thailand. In contrast

to these results, Forbes and Rigobon (2002) narrowly define their crisis period from

October 17, to November 16, 1997. Hence, our results indicate the Asian crises pe-

riod could have started much earlier and lasted much longer than one month. This

also points out the difficulties in correctly specifying the regimes with exogenously

defined periods.

In summary, when we disaggregate the total market return index into investable

and non-investable stock return index, we find that (1) there are more emerging

market countries that have significant increases in correlation with the crisis country

market returns during the turmoil period (or the volatile regime) and this pattern

is more pronounced for investable returns than for non-investable returns; and (2)

the sign test indicates that correlations generally increase for investable returns, but

not for non-investable returns. These results provide supporting evidence for the

existence of stock market contagion during the 1997 Asian crisis period.

B. Test 2: Fundamental or Investor-Induced?

In this section, we report results from our second set of tests, which are tests of

how contagion spreads: either through correlated country fundamentals or through

investor constraints such as portfolio re-balancing and wealth constraints. We exam-

ine whether the increase in co-movement is more pronounced for investable stocks

than for non-investable stocks during the turmoil period (T2-A), and whether in-

vestable index returns lead non-investable index returns during the turmoil period

(T2-B). We report results based on correlations estimated from the regime-switching

model in Table 7 and summarize the results on the lead-lag comparison between in-

vestable and non-investable stocks during the crisis period in Table 8.

With Hong Kong as the crisis country, out of 30 emerging markets, 17 coun-

26

tries show positive difference-in-difference statistics for correlation with the Hong

Kong stock market index returns, although not all are significantly positive. With

Thailand as the crisis country, 24 countries show positive (but not all significant)

difference-in-difference statistics for correlation. The sign test rejects the null hy-

pothesis of equal probability of the difference-in-difference statistics being negative

or positive, favoring the positive, for Thailand, but not for Hong Kong. The in-

creases in correlation with Hong Kong stock market index returns affect investable

and non-investable similarly, while the increases in correlation with Thailand stock

market index returns are higher for investable. This finding corresponds with the

results shown in Figure 2, where the high volatility regime in Hong Kong is found to

start much later in September-November 1997 compared with May 1997 in Thailand.

Our finding here may reflect that crisis is originated in Thailand and spread mostly

through investable stocks, and by the time crisis reaches Hong Kong, the crisis has

affected both investable and non-investable stocks global-wide.

We report the results on lead-lag comparisons as specified in Test 2-B in Table

8. If investable stocks are a larger channel to transmit external shocks, investable

stocks should lead non-investable stocks during the crisis period. In Table 8, the

crisis period is defined by the regime-switching model using Thailand as the source

country.26 Table 8 shows that, indeed, the correlation between lagged investable

returns and non-investable returns (ρ(It−1,Nt), column 2) is higher than correlation

between lagged non-investable returns and investable returns (ρ(It,Nt−1), column 4)

for most countries: 21 out of 31 countries for one-week lag and 22 countries for two-

week lag. The differences (columns 6 and 7) are small, although positive, relative to

the contemporaneous correlation (column 1). This suggests that investable stocks

lead non-investable stocks, but the effect is small.

To summarize, we find strong statistical evidence that during the turmoil period,

increase in co-movement is more pronounced with investable than non-investable

returns; and investable returns lead non-investable returns during the crisis period.

These findings indicate that investable returns act as a larger channel for crisis

transmission and provide supporting evidence for the investor-induced contagion

26We obtain similar results by defining the crisis period to be from July 2 to November 16, 1997.

27

hypothesis.

C. Test 3: Asymmetric Exceedance Correlation

In this section, we conduct a further test to determine whether crises spread

through investors, by examining market co-movements in market downturns (ex-

treme tails of negative returns) and in market upturns (extreme tails of positive re-

turns). Instead of comparing correlations between the turmoil period and the stable

period as in previous sections, we compare exceedance correlations between negative

and positive tails, which are extreme events. Although this set of comparisons is

similar to those in the previous sections, the interpretations are slightly different.

If crises are transmitted by investor restrictions such as wealth constraints, market

co-movements should be asymmetric for upturns and downturns; in contrast, if crises

spread through the need for portfolio re-balancing, market co-movements should be

symmetric.

We use the extreme value estimation approach to estimate exceedance correlation.

The advantage of this method is that it requires neither distributional assumption

nor exogenously defined turmoil period. We follow Ang and Chen (2002) in set-

ting the threshold values to be multiples of sample deviations away from the sample

mean.27 We estimate exceedance correlations both for the positive tail and the neg-

ative tail at weekly frequency using the whole sample. We then check for asymmetry

in correlations in the positive and the negative tail for total market, investable and

non-investable returns.

C.1. Exceedance Correlation with Hong Kong

First, we estimate the positive and negative tails of the joint distribution of

the total market index returns with Hong Kong as the crisis country. Table 9A

contains correlation estimates at 1.5 deviations away from the mean, the difference

27A tradeoff exits in selecting the threshold values: The sample distribution converges toward

the true extreme tail distribution as threshold increases, but the precisions of the estimates suffer

from smaller number of observation satisfying the tail criteria.

28

between negative and positive tail correlations, and the corresponding t-statistics.28

At the total market index level, the comparison of the exceedance correlation across

tails, positive vs. negative, yields evidence of asymmetry. Overwhelmingly, most of

the differences show higher correlations in market downturns. Out of 45 countries

(including both developed and emerging markets), 39 have higher correlations during

market downturns, and the sign test clearly rejects the null hypothesis of equal

probability of positive and negative differences. The test favors the alternative of

increased correlations in the market downturns. In general, the correlations in the

positive tail are smaller than the correlations in the negative tail. This indicates

that it is a global phenomenon that stock index returns co-move more with the

crisis country stock index returns during market downturns. The countries with

statistically higher correlations with Hong Kong during the market downturns, at

five percent significance level, are Germany, Switzerland, Argentina, Chile, Mexico,

Malaysia, Pakistan, Philippines, and Greece.

At the disaggregate level (i.e., measuring co-movements of Hong Kong with other

country’s investable returns and non-investable returns), once again, most of the dif-

ferences show higher correlations in market downturns. Out of 31 countries for

investable index returns of emerging markets, 20 have higher correlations during

market downturns turns, and the sign test clearly rejects the null hypothesis of

equal probability of positive and negative differences. The countries with statis-

tically higher correlations with Hong Kong during the market downturns, at five

percent significance level, are Argentina, Brazil, Chile, Mexico, and Malaysia for in-

vestable returns, and just Chile for non-investable returns. In general, the correlation

increases in the market downturns are somewhat smaller in non-investable returns,

especially for the non-Asian countries. In fact, the Asian countries’ correlations are

very similar across investable returns and non-investable returns, expect for China.

These results on asymmetric exceedance correlation are consistent with hypothe-

28We also estimate exceedance correlations at 1 and 2 deviations away from the mean. The

results are similar and therefore are not reported. We do report, however, the shape of asymmetry

whether exceedance correlations are increasing (denoted as +) as the thresholds are moved away

from the mean. An increase in the exceedance correlation away from the mean suggests that the

bivariate distribution is not normal.

29

ses that predict asymmetry in co-movements. One such hypothesis is that crises

spread through investor wealth constraints. On the other hand, these results are

not consistent with hypotheses such as portfolio re-balancing needs which predict

symmetric co-movements in market upturns and downturns.29

C.2. Exceedance Correlation with Thailand

We also estimate the exceedance correlation using Thailand as the crisis country.

Once again, we find some evidences of asymmetry between the positive and negative

tails, as reported in Table 9B at the total index returns level. Moreover, when we

disaggregate country’s total index returns into investable and non-investable returns,

the pattern is clearer. In general, the correlations in the negative tails are higher than

the correlations in the positive tails, more so for investable than for non-investable.

Similar to the findings in the Hong Kong case, across the board, the correlation

of crisis country returns with investable returns is significantly higher than that

with non-investable returns. In the case of China, the increased correlation in the

market downturns appears only in investable index, not for non-investable index nor

even total index. Also, the pattern of correlations for negative and positive tails

are remarkably similar for both investable and non-investable index returns in most

Asian countries, thus indicating the severity of regional contagion.

Hence, we again find contrasting contagion patterns when we define Thailand as

the crisis source country versus when we define Hong Kong as the source country.

Using Thailand as the source country, we find that Asian countries have higher

29Moreover, the shape of the positive and negative tails implied by the exceedance estimates

could further our understanding of the distributional properties of underlying stochastic processes

and assist us in building a structural empirical model of correlations. For example, Longin and

Solnik (2001) document a particular shape of asymmetry in their sample – namely, asymptotic

independence in the positive tail and flat or increasing dependence in the negative tail, and conclude

that the data process cannot be described by a bivariate normal distribution. We find this shape

for total index returns of Switzerland, Mexico, Indonesia, and Russia. However, we do find that

for most countries, the correlation estimates decrease further out into the tails, with some more

substantially than the others. The shape implied by these exceedance correlations is a tent-shape,

as could be described by a bivariate normal distribution that has slightly different properties in its

positive and negative tails.

30

exceedance correlations with the crisis country, and larger increases in exceedance

correlations during market downturns. However, using Hong Kong as the source

country, we find larger overall correlations. This is consistent with the idea that the

Asian crisis spread regionally from Thailand primarily to Asian countries.30

In summary, the results on asymmetric exceedance correlation are robust to alter-

native crisis country specifications and hence provide further support for the investor-

induced contagion hypotheses that predict asymmetric co-movements in market up-

turns and downturns.

IV. Conclusion

We set out to empirically test the hypothesis that crises spread through investors’

asset holdings. We estimate and compare the degree to which investable and non-

investable stock index returns co-move with the crisis country returns in the case of

the 1997 Asian crisis. Our analysis using Thailand as the crisis country reveals that

the increase in correlation is more prevalent in investable stock index returns rather

than non-investable stock index returns, especially in Asian countries, indicating that

the crisis was transmitted through investable stocks. Our analysis using Hong Kong

as the crisis country finds that the increase in correlation is also more pronounced in

investable stocks than non-investable stocks and more so among non-Asian countries.

Since the Thailand stock market crash precedes the Hong Kong stock market crisis,

we interpret the finding as the crisis was first transmitted regionally to Asian coun-

tries and then globally to non-Asian countries. We also find that investable stocks

are not fundamentally different from non-investable stocks in terms of industry dis-

30Technically, the primary difference in these results from those with Hong Kong as the source

country is in the shape of the exceedance correlation. We observe more cases of increased exceedance

correlation, implying that the asymptotic independence may no longer hold and bivariate normal

distribution may not describe the data generating process. Whereas the predominant shape with

Hong Kong as the crisis country is tent-shape, this is not so with Thailand. With the total index

return, increasing exceedance correlation is found in the positive tails for three countries - France,

Spain, and Korea – and in the negative tails for most other countries. Although the correlations

are larger with Hong Kong as the source country, evidence of increasing dependence in the negative

tail is much more prevalent using Thailand as the crisis country.

31

tributions, so the cash flow fundamentals cannot explain the increase in correlations.

We also compare cross-autocorrelations between investable and non-investable re-

turns for each country during the turmoil period, and evidence that investable stocks

lead non-investable stocks, further suggesting that crises spread first to investable

stocks through investors’ international holdings. Finally, we test whether market co-

movements are symmetric in extreme market upturns and downturns. Our finding

of high levels of asymmetry in the co-movements between investable and the crisis

country is consistent with the behavioral theory of investors who are constrained by

their wealth.

REFERENCES

Allen, Franklin, and Douglas Gale, 2000, Financial contagion, Journal of Political

Economy 108, 1–33.

Ang, Andrew, and Joseph Chen, 2002, Asymmetric correlations of equity portfolios,

Journal of Financial Economics 63, 443–494.

Bae, Keehong, Kalok Chan, and Angela Ng, 2002, Investibility and return volatility

in emerging equity markets, Working Paper, Hong Kong University of Science and

Techology.

Bae, Kee Hong, Andrew Karolyi, and Rene Stulz, 2003, A new approach to measuring

financial market contagion, Review of Financial Studies.

Barberis, Nicholas, Andrei Shleifer, and Jeffrey Wurgler, 2003, Comovement, Journal

of Financial Economics Forthcoming.

Baur, Dirk, 2003, Testing for contagion: Mean and volatility contagion, Journal of

Multi-national Financial Management 13, 405–422.

Bekaert, Geert, Campbell R. Harvey, and Angela Ng, 2003, Market integration and

contagion, Journal of Business forthcoming.

Boyer, Brian H., Michael. S. Gibson, and Mico Loretan, 1999, Pitfalls in tests for

changes in correlations, Working Paper.

32

Calvo, Guilermo A., 1999, Contagion in emerging markets: When Wall Street is a

carrier, AEA (1999) Conference Working Paper.

Campbell, John Y., Andrew W. Lo, and A. Craig Mackinlay, 1997, The Econometrics

of Financial Markets (Princeton University Press: New Jersey).

Chan-Lau, Jorge A., and Chen Zhaohui, 1998, Financial crisis and credit crunch as

a result of ineffecient financial intermediation – with reference to Asian financial

crisis, IMF Working Paper.

Chari, Anusha, and Peter Blair Henry, 2003, Risk sharing and asset prices: Evidence

from a natural experiment, Journal of Finance Forthcoming.

Connoly, Robert, and Albert Wang, 1998a, International equity market co-

movements: Economic fundamentals or contagion?, Mimeo.

, 1998b, On stock market return co-movements: Macroeconomic news, dis-

persion of beliefs, and contagion, Mimeo.

Corsetti, Giancarlo, Marcello Pericoli, and Massimo Sbracia, 2002, Some contagion,

some interdependence: More pitfalls in tests of financial contagion, Working Paper.

Corsetti, Giancarlo, Paolo Pesenti, and Nouriel Roubini, 1999, Paper tigers? A

model of the Asian crisis, European Economic Review 43, 1211–1236.

DeHaan, L., I.S. Resnick, H. Rootzen, and C.G. de Vries, 1989, Extremal behavior

of solutions to a stochastic difference equations with applications to arch process,

Stochastic Processes and Their Applications 32, 213–224.

Embrechts, P., C. Kluppelberg, and T. Mikosch, 1997, Modeling Extremal Events:

For Insurance and Finance (Springer: Berlin, Heidelberg).

Errunza, Vihang, and Etienne Losq, 1985, International asset pricing under mild

segmentation: Theory and test, Journal of Finance 40, 105–124.

, 1989, Capital flow controls, international asset pricing, and investor’s wel-

fare: A multi-country framework, Journal of Finance 44, 1025–1037.

33

Forbes, Kristin J., 2002, Are trade linkages important determinants of country vul-

nerability to crises, in Sebastian Edwards, and Jeffrey Frankel, ed.: Preventing

Currency Crises in Emerging Markets. pp. 77–124 (University of Chicago Press:

Chicago).

, and Roberto Rigobon, 2002, No contagion, only interdependence: Measur-

ing stock market co-movements, Journal of Finance 56, 2223–2261.

Gibson, M. S., and B. H. Boyer, 1997, Evaluating forecasts of correlation using option

pricing, Working Paper.

Goldfajn, Ilan, and Rodrigo O. Valdes, 1997, Capital flows and the twin crises: the

role of liquidity, IMF Working Paper.

Hamilton, James D., 1989, A new approach to the economic analysis of nonstationary

time series and the business cycle, Econometrica 57, 357–384.

IFC, 1999, The IFC indexes: Methodology, definitions, and practices, User Guide.

Kaminsky, Gaciela L., and Carmen M. Reinhart, 2000, On crises, contagion, and

confusion, Journal of International Economics 51, 145–168.

Karolyi, Andrew, and Rene Stulz, 1996, Why do markets move together? An investi-

gation of U.S.-Japan stock return co-movements, Journal of Finance 51, 951–986.

Kim, Chang-Jin, 1994, Dynamic linear models with markov-switching, Journal of

Econometrics 60, 1–22.

King, Mervyn, Enrique Sentana, and Sushil Wadhwani, 1994, Volatility and links

between national stock markets, Econometrica 62, 901–933.

Kodres, Laura E., and Matthew Pritsker, 2002, A rational expectations model of

financial contagion, Journal of Finance 57, 769–799.

Kyle, Albert S., and Wei Xiong, 2001, Contagion as a wealth effect, Journal of

Financial Economics 62, 247–292.

34

Lagunoff, R.D., and S.L. Schreft, 2001, A model of financial fragility, Journal of

Economic Theory 99, 220–264.

Leadbetter, M.R., G. Lindgren, and H. Rootzen, 1983, Extremes and Related Prop-

erties of Random Sequences and Processes (Springer Verlag: New York).

Ledford, M. R., and J.A. Tawn, 1997, Statistics for near independence in multivariate

extreme values, Biometrika pp. 169–187.

Longin, Franois, and Bruno Solnik, 2001, Extreme correlation of international equity

markets, Journal of Finance 56, 649–676.

Prescott, P., and A. T. Walden, 1980, Maximum likelihood estimation of the param-

eters of the generalized extreme-value distribution, Biometrika 67, 723–724.

Radelet, S., and J. Sachs, 1998, The onset of asian crisis, Working Paper.

Rijckeghem, C. Van, and B. Weder, 2000, Spillovers through banking centers: A

panel data analysis, Working Paper.

Shleifer, Andrew, and Robert W. Vishny, 1997, The limit of arbitrage, Journal of

Finance 52, 35–55.

Stambaugh, Robert F., 1995, Unpublished discussion of Karolyi and Stulz (1996),

National Bureau of Economic Research Universities Research Conference on Risk

Management.

Tang, Jon Wongswan, 2001, Contagion: An empirical test, Working Paper, Duke

University.

Tawn, Jonathan, 1988, Bivariate extreme value theory; models and estimation,

Biometrika 75, 397–415.

White, Halbert, 1980, A heteroskedasticity-consistent covariance matrix estimator

and a direct test for heteroskedasticity, Econometrica 48, 817–838.

35

Yuan, Kathy, 2003, Asymmetric price movements and borrowing constraints: A ra-

tional expectations equilibrium model of crisis, contagion, and confusion, Journal

of Finance Forthcoming.

36

This table reports the total market capitalization and annual value-traded of each stock market in the sample.All statistics are reported as millions of U.S. dollars as of year-end 1996. For emerging market countries,statistics are reported based on IFCG index in emerging market database (EMDB) from the InternationalFinance Corporation (2000). For developed countries, data source is Datastream.

Region Country

Emerging MarketsLatin Argentina 26,564 339 America Brazil 94,879 7,422

Chile 35,780 390 Colombia 9,287 119 Mexico 72,197 2,311 Peru 7,605 166 Venezuela 7,153 120

Asia-Pacific China 38,754 19,962 South Asia India 50,802 1,735

Indonesia 49,638 2,808 Korea 73,108 3,245 Malaysia 169,941 3,835 Pakistan 4,630 335 Philippines 48,633 1,163 Sri Lanka 1,126 11 Taiwan, China 151,413 22,375 Thailand 57,750 1,482

Emerging Czech Rep. 9,558 24 Europe Greece 10,210 647

Hungary 3,569 165 Poland 5,730 191 Portugal 16,437 853 Russia 19,273 154 Slovakia 1,311 27

Others Egypt 1,884 322 Israel 12,609 - Jordan 2,890 44 Morocco 4,551 45 South Africa 86,122 1,362 Turkey 16,276 1,744 Zimbabwe 2,386 33

Developed MarketsAsia-Pacific Australia 311,988 145,482

Hong Kong 449,381 166,419Japan 3,088,850 1,251,998Singapore 150,215 42,739

Europe Belgium 119,831 26,120France 591,123 277,100Germany 670,997 768,745Italy 258,160 102,351Netherlands 378,721 339,500Spain 242,779 249,128Sweden 247,217 136,898Switzerland 402,104 392,783United Kingdom 1,740,246 578,471

North America Canada 486,268 265,360United States 8,484,433 7,121,487

Market Capitalization and Value Traded, 1996

Total Market Capitalization Total Value Traded

Table 1

This table contains summary statistics for weekly data from EMDB and Datastream. Full 665 observations are from January 1989 to September 2001.For the emerging markets, the overall and investable returns are as reported in the EMDB database from IFC. Non-investable returns are calculated.The last column is the contemporaneous correlation between investable (IR) and non-investable (NR) weekly returns.

Start Standard Start Standard Start StandardEmerging Markets Obs Date Mean Deviations Obs Date Mean Deviations Obs Date Mean Deviations Obs CorrelationArgentina 665 Jan-89 0.0060 0.0800 665 Jan-89 0.0063 0.0820 665 Jan-89 0.0079 0.1085 665 0.464Brazil 665 Jan-89 0.0045 0.0710 665 Jan-89 0.0052 0.0778 665 Jan-89 0.0047 0.0694 665 0.890Chile 665 Jan-89 0.0034 0.0314 665 Jan-89 0.0036 0.0330 665 Jan-89 0.0029 0.0326 665 0.646Colombia 665 Jan-89 0.0028 0.0376 556 Feb-91 0.0031 0.0444 556 Feb-91 0.0015 0.0400 556 0.623Mexico 665 Jan-89 0.0036 0.0429 665 Jan-89 0.0042 0.0450 665 Jan-89 0.0033 0.0441 665 0.621Peru 456 Jan-93 0.0020 0.0362 456 Jan-93 0.0017 0.0376 456 Jan-93 0.0035 0.0427 456 0.510Venezuela 665 Jan-89 0.0037 0.0576 612 Jan-89 0.0056 0.0692 612 Jan-89 0.0038 0.0639 612 0.542

China 456 Jan-93 0.0024 0.0622 455 Jan-93 -0.0005 0.0544 455 Jan-93 0.0029 0.0707 455 0.158India 665 Jan-89 0.0011 0.0411 464 Jan-89 -0.0003 0.0398 464 Jan-89 -0.0002 0.0383 464 0.989Indonesia 574 Oct-90 -0.0002 0.0739 574 Oct-90 -0.0001 0.0746 574 Oct-90 -0.0004 0.0857 574 0.807Korea 665 Jan-89 0.0000 0.0553 508 Jan-89 0.0006 0.0608 508 Jan-89 0.0005 0.0595 508 0.920Malaysia 665 Jan-89 0.0013 0.0466 665 Jan-89 0.0014 0.0471 665 Jan-89 0.0018 0.0469 665 0.910Pakistan 548 Apr-91 0.0006 0.0420 548 Apr-91 0.0011 0.0474 548 Apr-91 0.0008 0.0365 548 0.751Philippines 665 Jan-89 0.0003 0.0440 665 Jan-89 0.0002 0.0465 665 Jan-89 0.0006 0.0452 665 0.842Sri Lanka 456 Jan-93 -0.0011 0.0326 456 Jan-93 -0.0011 0.0403 456 Jan-93 -0.0010 0.0310 456 0.785Taiwan 333 May-95 -0.0011 0.0433 333 May-95 -0.0010 0.0436 333 May-95 -0.0012 0.0434 333 0.992Thailand 665 Jan-89 0.0006 0.0560 665 Jan-89 0.0005 0.0549 665 Jan-89 0.0007 0.0575 665 0.984

Czech Republic 404 Jan-94 -0.0020 0.0361 404 Jan-94 -0.0014 0.0438 404 Jan-94 -0.0030 0.0371 404 0.561Greece 665 Jan-89 0.0035 0.0478 665 Jan-89 0.0039 0.0502 584 Jan-89 0.0148 0.1828 584 0.226Hungary 456 Jan-93 0.0019 0.0466 455 Jan-93 0.0028 0.0486 455 Jan-93 0.0019 0.0481 455 0.405Poland 456 Jan-93 0.0054 0.0663 455 Jan-93 0.0053 0.0664 401 Jan-93 0.0081 0.0760 401 0.775Portugal* 534 Jan-89 0.0025 0.0292 534 Jan-89 0.0027 0.0297 534 Jan-89 0.0027 0.0326 534 0.804Russia 300 Jan-96 0.0073 0.0913 242 Feb-97 0.0033 0.1016 242 Feb-97 0.0046 0.1022 242 0.628Slovakia 300 Jan-96 -0.0015 0.0365 242 Feb-97 -0.0025 0.0426 242 Feb-97 -0.0020 0.0422 242 0.216

Egypt 300 Jan-96 -0.0007 0.0337 242 Feb-97 -0.0038 0.0369 242 Feb-97 -0.0045 0.0226 242 0.523Israel 247 Jan-97 0.0022 0.0357 247 Jan-97 0.0022 0.0357 247 Jan-97 0.0088 0.0691 247 0.570Jordan 665 Jan-89 0.0014 0.0216 665 Jan-89 0.0018 0.0236 665 Jan-89 0.0013 0.0211 665 0.930Morocco 300 Jan-96 0.0020 0.0199 242 Feb-97 0.0010 0.0213 242 Feb-97 -0.0006 0.0242 242 0.634South Africa 456 Jan-93 0.0021 0.0381 455 Jan-93 0.0020 0.0382 229 Jan-93 0.0039 0.0607 229 0.598Turkey 665 Jan-89 0.0053 0.0853 634 Jan-89 0.0043 0.0864 406 Jan-89 0.1833 1.5389 408 0.062Zimbabwe 430 Jul-93 0.0037 0.0655 430 Jul-93 0.0054 0.0590 430 Jul-93 0.0034 0.0692 430 0.703

Emerging Market Average 0.0020 0.0486 0.0019 0.0510 0.0082 0.1046 0.6474

Developed MarketsAustralia 665 Jan-89 0.0009 0.0234Hong Kong 665 Jan-89 0.0027 0.0378Japan 665 Jan-89 -0.0006 0.0331Singapore 665 Jan-89 0.0010 0.0309

Belgium 665 Jan-89 0.0011 0.0225France 665 Jan-89 0.0015 0.0269Germany 665 Jan-89 0.0019 0.0297Italy 665 Jan-89 0.0008 0.0323Netherlands 665 Jan-89 0.0020 0.0218Spain 665 Jan-89 0.0012 0.0273Sweden 665 Jan-89 0.0017 0.0347Switzerland 665 Jan-89 0.0024 0.0237United Kingdom 665 Jan-89 0.0015 0.0231

Canada 665 Jan-89 0.0015 0.0225United States 665 Jan-89 0.0024 0.0216

Developed Market Average 0.0015 0.0274

World Index 665 Jan-89 0.0008 0.0192Note: (*) Portugal data ends in Mar-99.

Table 2Summary Statistics

IR and NROverall Return Investable Return Non-Investable Return

This table reports percentages of year-end market capitalization that are investable and non-investable according to the IFC index, for the emerging market countries. For the developedcountries, the investable are 100%. The market capitalizations are year-end values in millions of U.S. dollars.

Region Country 1994 1996 1997 1998 1999 1994 1996 1997 1998 1999 1994 1996 1997 1998 1999Latin Argentina 98% 99% 100% 100% 100% 2% 1% 0.4% 0.1% 0.5% 18,751 26,564 35,142 24,894 23,319 America Brazil 63% 82% 87% 92% 94% 37% 18% 13% 8% 6% 111,906 94,879 102,965 65,871 107,803

Chile 22% 99% 99% 95% 95% 78% 0.8% 1.1% 5% 5% 45,058 35,780 44,498 31,837 42,639 Colombia 78% 76% 80% 78% 72% 22% 24% 20% 22% 28% 11,437 9,287 11,452 6,337 5,371 Mexico 89% 93% 95% 92% 99% 11% 7% 5% 8% 1% 83,279 72,197 108,941 67,174 118,418 Peru 81% 93% 94% 88% 97% 19% 7% 6% 12% 3% 5,271 7,605 9,657 6,151 7,583 Venezuela 78% 89% 91% 98% 98% 22% 11% 9% 2% 2% 3,352 7,153 9,138 4,689 4,237

Asia-Pacific China 10% 12% 10% 25% 37% 90% 88% 90% 75% 63% 19,307 38,754 49,981 52,452 78,532 India 22% 26% 29% 26% 28% 78% 74% 71% 74% 72% 65,358 50,802 50,856 49,797 92,526 Indonesia 50% 62% 98% 98% 96% 50% 38% 2% 2% 4% 22,199 49,638 13,553 10,682 23,260 Korea 11% 22% 54% 91% 94% 89% 78% 46% 9% 6% 125,065 73,108 25,157 61,203 181,989 Malaysia 83% 92% 94% 94% 96% 17% 8% 6% 6% 4% 125,883 169,941 45,911 44,849 68,937 Pakistan 53% 80% 89% 84% 81% 47% 20% 11% 16% 19% 7,668 4,630 5,979 2,400 3,115 Philippines 54% 46% 48% 48% 47% 46% 54% 52% 52% 53% 30,222 48,633 18,650 21,969 25,953 Sri Lanka 30% 32% 34% 33% 53% 70% 68% 66% 67% 47% 1,710 1,126 1,293 1,036 29,232 Taiwan, China 8% 33% 40% 35% 55% 92% 67% 60% 65% 45% 160,212 151,413 153,176 138,369 230,955 Thailand 28% 34% 47% 54% 53% 72% 66% 53% 46% 47% 79,660 57,750 10,921 16,735 29,232

Emerging Czech Rep. 35% 35% 38% 71% 80% 65% 65% 62% 29% 20% 8,735 9,558 4,847 4,204 4,205 Europe Greece 93% 99% 100% 100% 100% 7% 1.3% 0.2% 0.2% 0.0% 8,021 10,210 16,255 40,891 90,076

Hungary 63% 90% 100% 99% 99% 37% 10% 0.5% 0.6% 0.5% 747 3,569 11,175 10,933 13,076 Poland 100% 100% 100% 100% 100% 0.0% 0.2% 0.0% 0.1% 0.1% 1,478 5,730 6,234 8,134 11,770 Portugal 64% 77% 90% 91% - 36% 23% 10% 9% - 11,185 16,437 24,745 34,574 -Russia - 0% 63% 67% 77% - 100% 37% 33% 23% - 19,273 54,744 7,558 25,252 Slovakia - 0% 86% 85% 91% - 100% 14% 15% 9% - 1,311 1,320 600 433

Others Egypt - 0% 79% 82% 93% - 100% 21% 18% 7% - 1,884 8,141 5,848 11,433 Israel - 94% 99% 99% 100% - 6% 0.6% 0.6% 0.3% - 12,609 18,812 16,373 26,527 Jordan 35% 26% 34% 34% 51% 65% 74% 66% 66% 49% 2,787 2,890 3,261 4,277 4,109 Morocco - 0% 86% 81% 95% - 100% 14% 19% 5% - 4,551 7,562 9,674 8,141 South Africa 96% 100% 99% 100% 100% 4% 0.0% 0.5% 0.0% 0.0% 137,866 86,122 90,297 68,672 110,672 Turkey 100% 100% 100% 100% 99% 0.0% 0.0% 0.0% 0.0% 1.3% 15,106 16,276 33,732 16,947 58,708 Zimbabwe 12% 26% 40% 30% 32% 88% 74% 60% 70% 68% 1,341 2,386 1,123 653 1,613

Total 55% 61% 73% 77% 80% 45% 39% 27% 23% 20% 946,506 965,348 816,590 713,342 1,217,913

Investable Non-Investable Market Capitalization

Table 3Percentage of Investable and Non-Investable by Country

This table reports percentage distributions of investable and non-investable firms in each industry (defined bythe SIC code in parentheses) for all firms in the EMDB sample across countries. Industry groupings aresimilar to those found in Bae, Chan, Ng (2001).

Industry (SIC codes) Investable % Non-Investable % Difference

Agriculture, Forestry and Fishing 2.20 3.46 -1.26(1,2,7,8,9)

Mining 3.46 4.78 -1.32(10,12,13,14)

Construction 4.09 3.98 0.11(15,16,17)

Manufacturing 42.59 45.28 -2.69(20-36)

Trasnportation and Public Utilities 15.93 15.81 0.12(38-49)

Wholesale Trade 0.69 1.06 -0.37(50, 52)

Retail Trade 2.64 3.32 -0.68(53,54,55,58,59)

Finance, Insurance and Real Estate 17.93 13.42 4.51(51,60,61,62,63,65)

Services 5.16 5.44 -0.28(67,70,73,75,78,79,89)

Others 5.35 3.45 1.9(99)

TOTAL (Median) 4.63 4.38 -0.33

Table 4Percentage of Investable and Non-Investable by Industry

This table compares investable firms and non-investable firms for each country, by comparing the averages of two measures of firm characteristics: (1) Liquidity: turnover rate of

stock holdings as represented by the ratio of value-traded to total market capitalization, and (2) size: ratio of the firm market capitalization to total country market capitalization.

A firm is defined to be investable if the investable fraction reported in the EMDB database is 1 and non-investable if the fraction is 0. All observations with interim values are deleted

from this analysis. The test of comparison is a one-sided t-test to see if liquidy or size of investable firms is significantly larger than that of the non-investable firms, at 95% significance

level. All values are taken at the beginning of the year.

Region Country Inv. Non-Inv. Liquidity Size Inv. Non-Inv. Liquidity Size Inv. Non-Inv. Liquidity Size

Latin Argentina 24 6 N N 25 3 N N 26 1 - -America Brazil 44 26 N N 25 12 N N 44 10 Y N

Chile 0 15 N N 35 3 N N 30 11 N YColombia 10 14 N N 4 12 N Y 4 10 N YMexico 63 13 Y N 5 6 Y Y 51 13 Y NPeru 11 24 N Y 16 13 N Y 13 14 N YVenezuela 12 5 N N 10 7 N Y 13 5 N Y

Asia-Pacific China 18 99 N Y 43 152 N N 47 156 N NIndia 0 43 - - 0 61 - - 0 56 - -Indonesia 0 9 - - 45 1 - - 31 17 N NKorea 0 6 - - 57 13 N N 52 27 N YMalaysia 99 0 - - 88 0 - - 33 18 N YPakistan 15 56 Y Y 10 31 N Y 0 36 - -Phillipines 16 22 Y N 3 10 N N 1 21 - -Sri Lanka 4 27 N Y 4 45 N Y 0 46 - -Taiwan 0 2 - - 5 0 - - 0 0 - -Thailand 0 6 - - 0 9 - - 0 21 - -

Emerging Czech Rep 5 55 N Y 3 35 Y N 5 24 N NEurope Greece 25 11 N Y 43 2 N N 43 2 N N

Hungary 5 8 N Y 8 2 N Y 9 2 N YPoland 12 0 - - 13 0 - - 12 1 - -Portugal 19 7 N N 10 0 - - 10 0 - -Russia -- -- -- -- 11 7 N N 5 19 Y NSlovakia -- -- -- -- 3 15 N Y 2 14 N Y

Others Egypt -- -- -- -- 11 26 N Y 5 32 N YIsrael -- -- -- -- 10 3 N N 13 3 N NJordan 0 29 - - 1 38 - - 2 35 N YMorocco -- -- -- -- 5 6 N N 5 8 Y NSouth Africa 58 4 N N 17 1 - - 32 0 - -Turkey 40 0 - - 30 0 - - 31 0 - -Zimbabwe 0 19 - - 0 12 - - 0 9 - -

1994 1997Observations

Table 5

Observations Observations1998

Firm-Level Comparison of Investable vs. Non-Investable

Significantly Greater? Significantly Greater? Significantly Greater?

This table reports differences in estimated moments across two regimes. The T-test statistics reported are of tests of equivalence in volatility (Vol), correlation with crisis country (Corr C), and correlation with world index (Corr W) between the turmoil regime and stable regime separately for investable and non-investable stock index returns. The rejections of the null, against the one-sided alternative that the turmoil regime volatility or correlation is greater, at the 10%, 5% and 1% significance level rejections are denoted by (*) , (**) and (***), respectively. Mean LL is the mean of likelihood function and N is the number of observations used in estimation.

Emerging Markets: N Diff T-Stat Diff T-Stat Diff T-Stat Diff Vol. Diff T-Stat Diff T-StatArgentina 665 0.106 29.662 *** 0.136 3.738 *** 0.094 4.963 *** 0.036 0.832 0.006 0.311 -0.003 -0.139Brazil 665 0.107 15.547 *** 0.165 4.026 *** 0.290 12.003 *** 0.058 7.153 *** 0.171 4.469 *** 0.234 10.032 ***

Chile 665 0.059 0.987 0.145 3.135 *** 0.202 2.029 ** 0.068 2.069 ** 0.044 0.590 0.004 0.130Colombia 556 -0.001 -2.003 -0.013 -0.540 0.185 4.334 *** -0.033 -11.842 0.033 3.181 *** 0.109 5.861 ***

Mexico 665 0.105 5.097 *** 0.202 5.095 *** 0.336 6.472 *** 0.060 3.003 *** -0.049 -2.804 -0.004 -0.158Peru 456 0.016 13.621 *** 0.214 13.492 *** 0.236 10.103 *** 0.020 6.067 *** 0.045 3.671 *** 0.060 2.169 **

Venezuela 612 0.046 15.680 *** 0.132 18.507 *** 0.013 2.320 ** 0.035 11.110 *** 0.007 2.001 ** -0.045 -7.214

China 455 0.200 155.476 *** 0.141 5.455 *** 0.146 10.621 *** -0.252 -34.109 -0.139 -34.357 -0.063 -30.084India 464 0.064 8.470 *** 0.174 2.837 *** 0.159 1.929 ** 0.058 7.810 *** 0.185 2.973 *** 0.167 1.952 **

Indonesia 574 0.434 77.343 *** 0.206 9.243 *** 0.142 5.400 *** 0.432 95.083 *** 0.166 11.201 *** 0.127 9.007 ***

Korea 508 0.274 78.762 *** 0.142 8.227 *** 0.153 12.246 *** 0.275 60.618 *** 0.129 4.949 *** 0.159 8.055 ***

Malaysia 665 0.278 15.363 *** 0.122 4.694 *** 0.101 3.230 *** 0.278 17.029 *** 0.149 4.610 *** 0.106 2.615 ***

Pakistan 548 0.090 162.317 *** -0.051 -7.631 -0.051 -3.411 0.085 79.912 *** -0.048 -3.691 -0.053 -4.963Philippines 665 0.141 7.103 *** 0.243 9.521 *** 0.238 22.103 *** 0.171 8.259 *** 0.293 15.855 *** 0.278 11.665 ***

Sri Lanka 456 0.072 42.710 *** 0.047 1.783 ** 0.130 2.687 *** 0.044 115.459 *** 0.070 3.060 *** 0.202 4.949 ***

Taiwan 333 0.088 18.727 *** 0.175 8.623 *** 0.179 7.972 *** 0.086 15.725 *** 0.190 6.806 *** 0.183 5.651 ***

Thailand 665 0.256 25.315 *** 0.218 7.498 *** 0.250 9.161 *** 0.286 18.673 *** 0.223 8.107 *** 0.255 10.354 ***

Czech Republic 584 0.128 314.084 *** 0.129 5.193 *** 0.328 21.670 *** 0.083 48.330 *** 0.029 11.181 *** 0.211 44.804 ***

Greece 404 0.122 6.303 *** 0.096 1.382 * 0.253 3.541 *** 0.122 0.466 0.159 1.310 * 0.255 1.728 **

Hungary 455 0.152 20.829 *** 0.175 3.068 *** 0.289 5.127 *** 0.156 29.175 *** 0.053 0.998 0.125 1.731 **

Poland 401 -0.018 -5.099 0.135 8.630 *** 0.210 19.885 *** 0.056 9.746 *** -0.049 -11.112 0.122 6.988 ***

Portugal 534 0.100 87.767 *** 0.239 2.110 ** 0.297 2.588 *** 0.114 56.223 *** 0.188 1.418 * 0.272 2.231 **

Russia 242 0.297 32.652 *** 0.136 8.487 *** 0.199 5.038 *** 0.273 34.725 *** -0.058 -2.386 0.088 1.783 **

Slovakia 242 -0.018 -4.375 0.024 5.773 *** 0.138 9.838 *** -0.031 -16.444 -0.128 -3.720 0.070 1.747 **

Egypt 242 0.037 4.030 *** -0.053 -0.819 0.038 0.576 0.009 10.680 *** -0.027 -1.781 0.071 14.285 ***

Israel 247 0.097 23.147 *** 0.032 1.317 * 0.160 4.754 *** 0.099 14.254 *** 0.053 11.607 *** 0.141 15.198 ***

Jordan 665 0.013 1.866 ** 0.202 12.436 *** 0.258 11.611 *** 0.015 3.084 *** 0.214 13.714 *** 0.313 8.781 ***

Morocco 242 -0.042 -22.013 0.166 13.241 *** 0.109 6.821 *** -0.060 -22.121 0.030 2.011 ** 0.109 15.450 ***

South Africa 229 0.156 16.442 *** 0.328 4.695 *** 0.332 4.538 *** 0.278 18.925 *** 0.104 1.541 * 0.348 4.104 ***

Turkey 408 0.235 48.285 *** 0.013 0.292 0.126 3.481 *** 2.868 18.491 *** -0.035 -8.340 0.035 8.910 ***

Zimbabwe 430 0.163 105.418 *** -0.065 -2.832 -0.042 -1.568 0.003 0.668 -0.079 -39.542 -0.046 -27.291Countries with Higher Correlation in Turmoil: 27 27 29 27 24 25Sign Test (P-value) 0.000 0.000 0.000 0.000 0.000 0.000

Corr WCorr CVolatility Volatility Corr CCorr W

Table 6A

Non-Investable Stock Index ReturnsInvestable Stock Index Returns

Test 1: Existence of Contagion (Regime-Switching Model)Crisis Country: Hong Kong

This table reports differences in estimated moments across two regimes. The T-test statistics reported are of tests of equivalence in volatility (Vol), correlation with crisis country (Corr C), and correlation with world index (Corr W) between the turmoil regime and stable regime separately for investable and non-investable stock index returns. The rejections of the null, against the one-sided alternative that the turmoil regime volatility or correlation is greater, at the 10%, 5% and 1% significance level rejections are denoted by (*) , (**) and (***), respectively. Mean LL is the mean of likelihood function and N is the number of observations used in estimation.

Emerging Markets: N Diff T-Stat Diff T-Stat Diff T-Stat Diff T-Stat Diff T-Stat Diff T-StatArgentina 665 -0.290 -14.739 0.283 2.503 *** 0.293 3.019 *** 0.122 1.336 * 0.087 1.181 0.111 0.797Brazil 665 -0.120 -3.700 0.356 1.961 ** 0.305 6.811 *** -0.049 -1.455 0.277 1.646 * 0.203 2.727 ***

Chile 665 0.027 0.243 0.241 1.366 * 0.388 2.432 *** 0.064 1.124 -0.026 -0.109 0.030 0.978Colombia 556 -0.008 -1.587 0.108 0.923 0.199 0.950 -0.108 -4.055 0.018 0.320 0.120 1.139Mexico 665 0.062 1.180 0.239 2.781 *** 0.400 2.793 *** 0.040 0.957 0.113 17.903 *** 0.123 2.555 ***

Peru 456 -0.082 -6.987 0.142 7.132 *** 0.212 1.977 ** -0.045 -1.459 -0.020 -0.158 -0.083 -0.798Venezuela 612 -0.083 -2.206 0.287 3.855 *** 0.000 0.016 0.126 2.760 *** 0.053 2.893 *** -0.071 -1.838

China 455 0.169 23.327 *** 0.084 8.370 *** 0.190 3.521 *** -0.405 -6.667 -0.134 -17.923 -0.085 -5.880India 464 0.034 0.501 0.214 1.158 0.202 0.866 0.027 0.405 0.212 1.154 0.202 0.839Indonesia 574 0.530 17.116 *** 0.298 4.877 *** 0.218 1.523 * 0.681 12.413 *** 0.148 4.098 *** 0.106 1.421 *

Korea 508 0.360 20.240 *** 0.389 4.292 *** 0.216 14.501 *** 0.346 62.440 *** 0.385 3.873 *** 0.209 3.068 ***

Malaysia 665 0.300 7.469 *** 0.147 2.080 ** 0.082 0.826 0.302 7.344 *** 0.202 2.303 ** 0.089 0.679Pakistan 548 0.083 13.188 *** 0.018 0.146 0.027 0.333 0.091 27.697 *** 0.045 0.233 0.011 0.231Philippines 665 0.137 3.119 *** 0.348 10.858 *** 0.323 11.245 *** 0.178 3.970 *** 0.470 21.125 *** 0.325 3.572 ***

Sri Lanka 456 0.060 4.624 *** 0.164 7.244 *** 0.115 0.573 -0.030 -30.907 0.104 6.024 *** 0.193 1.062Taiwan 333 0.089 4.056 *** 0.247 1.787 ** 0.180 2.574 *** 0.088 2.819 *** 0.234 1.567 * 0.168 1.703 **

Czech Republic 584 0.022 3.051 *** 0.248 2.198 ** 0.287 4.502 *** 0.052 1.826 ** 0.132 1.658 ** 0.014 0.928Greece 404 0.120 2.127 ** 0.351 1.781 ** 0.353 2.297 ** -0.163 -0.115 0.112 0.716 0.256 1.268Hungary 455 0.103 1.749 ** 0.223 2.252 ** 0.332 2.056 ** 0.098 2.677 *** -0.039 -0.244 0.021 0.077Poland 401 -0.195 -5.276 0.199 1.375 * 0.300 5.411 *** -0.072 -0.987 0.017 0.465 0.128 1.894 **

Portugal 534 0.115 31.221 *** 0.341 1.419 * 0.353 1.510 * 0.127 69.064 *** 0.368 1.201 0.321 1.243Russia 242 0.377 3.190 *** 0.348 9.810 *** 0.305 3.611 *** 0.253 2.305 ** 0.294 21.517 *** 0.155 0.825Slovakia 242 0.037 0.938 -0.105 -3.493 -0.145 -2.906 -0.032 -1.274 -0.169 -5.395 0.182 1.021

Egypt 242 0.022 0.516 -0.044 -2.792 0.047 0.275 -0.044 -4.845 -0.275 -17.231 0.073 4.704 ***

Israel 247 0.082 9.106 *** -0.020 -0.144 0.182 1.288 * 0.040 0.510 0.081 4.539 *** 0.035 0.900Jordan 665 -0.012 -0.836 0.113 2.272 ** 0.127 1.422 * -0.008 -0.463 0.119 2.731 *** 0.175 1.260Morocco 242 -0.047 -2.927 -0.096 -1.199 0.073 0.841 -0.028 -1.337 -0.153 -8.194 0.267 7.245 ***

South Africa 229 0.170 3.380 *** 0.475 8.120 *** 0.338 2.134 ** 0.435 5.799 *** 0.108 1.063 0.344 1.898 **

Turkey 408 0.235 7.650 *** 0.303 1.398 * 0.223 1.472 * -11.458 -7.509 0.061 1.465 * 0.139 7.499 ***

Zimbabwe 430 0.153 17.979 *** 0.021 0.158 -0.017 -0.132 -0.058 -1.015 0.097 1.680 ** 0.000 -0.009Countries with Higher Correlation in Turmoil 22 23 28 17 24 26Sign Test (P-value) 0.003 0.001 0.000 0.181 0.000 0.000

Table 6BTest 1: Existence of Contagion (Regime-Switching Model)

Crisis Country: Thailand

Corr C Corr WNon-Investable Stock Index ReturnsInvestable Stock Index Returns

Volatility Corr C Corr W Volatility

This table reports the difference between turmoil and stable regime correlations in the investable returns with the crisis countryindex returns and difference-in-difference (difference between investable and non-investable in the difference between turmoil andstable) corresponding to test 2. The rejections of the null, against the one-sided alternative that turmoil period correlation is greater,at the 10%, 5% and 1% significance levels are denoted by (*), (**), and (***), respectively.

Developed Markets: Investable Diff-in-Diff T-Stat Investable Diff-in-Diff T-StatAustralia 0.115 ** 0.302 ***

Japan 0.133 *** 0.231 ***

Hong Kong -- 0.371 ***

Singapore 0.240 *** 0.341 ***

Belgium 0.157 * 0.170 *

France 0.183 *** 0.299 ***

Germany 0.118 * 0.242 ***

Italy 0.287 *** 0.215 ***

Netherlands 0.122 * 0.245 ***

Spain 0.134 * 0.316 ***

Sweden 0.103 * 0.246 ***

Switzerland 0.094 0.185 *

UK 0.153 ** 0.259 ***

Canada 0.029 0.216 *

US 0.164 *** 0.271 ***

Emerging Markets:Argentina 0.136 *** 0.130 4.332 *** 0.283 *** 0.197 3.096 ***

Brazil 0.165 *** -0.006 -0.513 0.356 ** 0.079 3.778 ***

Chile 0.145 *** 0.101 3.299 *** 0.241 * 0.267 3.263 ***

Colombia -0.013 -0.046 -3.396 0.108 0.091 1.456Mexico 0.202 *** 0.251 5.791 *** 0.239 *** 0.126 1.440Peru 0.214 *** 0.169 29.265 *** 0.142 *** 0.162 1.126Venezuela 0.132 *** 0.125 23.206 *** 0.287 *** 0.234 4.165 ***

China 0.141 *** 0.280 9.614 *** 0.084 *** 0.218 63.281 ***

India 0.174 *** -0.011 -11.003 * 0.214 0.001 2.004 **

Indonesia 0.206 *** 0.041 5.038 ** 0.298 *** 0.150 5.992 ***

Korea 0.142 *** 0.013 1.299 0.389 *** 0.004 0.476Malaysia 0.122 *** -0.027 -3.724 0.147 ** -0.056 -3.196Pakistan -0.051 -0.003 -0.171 0.018 -0.027 -0.378Philippines 0.243 *** -0.049 -4.341 0.348 *** -0.121 -2.513Sri Lanka 0.047 ** -0.023 -5.745 0.164 *** 0.060 11.166 ***

Taiwan 0.175 *** -0.015 -1.761 0.247 ** 0.013 1.179Thailand 0.218 *** -0.005 -1.092

Czech Republic 0.129 *** 0.100 4.217 *** 0.248 ** 0.117 0.607Greece 0.096 ** -0.063 -0.827 0.351 ** 0.240 5.001 ***

Hungary 0.175 *** 0.122 17.131 *** 0.223 ** 0.263 4.285 ***

Poland 0.135 *** 0.184 10.291 *** 0.199 * 0.182 1.672Portugal 0.239 *** 0.050 2.434 *** 0.341 * -0.027 -0.410Russia 0.136 *** 0.194 17.775 *** 0.348 *** 0.054 2.420 ***

Slovakia 0.024 *** 0.151 4.968 *** -0.105 0.064 16.319 ***

Egypt -0.053 -0.025 -0.518 -0.044 0.230 54.141 ***

Israel 0.032 -0.021 -1.007 -0.020 -0.101 -0.822Jordan 0.202 *** -0.011 -1.037 0.113 ** -0.006 -0.071Morocco 0.166 *** 0.135 46.582 *** -0.096 0.057 0.574South Africa 0.328 *** 0.224 43.546 *** 0.475 *** 0.367 2.300 **

Turkey 0.013 0.048 1.118 0.303 * 0.242 1.382Zimbabwe -0.065 0.014 0.650 0.021 -0.076 -0.993

Countries with Higher Corr. in Turmoil 38 17 41 24Sign Test (P-value) 0.000 0.237 0.000 0.000

ThailandHong Kong

Table 7Test 2: Fundamentals vs. Investor-Induced Contagion (Regime-Switching Model)

Crisis Country: Hong Kong and Thailand

This table exhibits the contemporaneous correlations and lag correlations between investable and non-investablereturns for weekly data, during the crisis period, as defined by the regime-switching model using Thailand as thecrisis period. Contemporaneous correlations between investable and non-investable are reported in (1), differences between the correlations of non-investable with lagged investable (2, 3), and of investable with laggednon-investable (4, 5) are reported in (6, 7).

Contemporaneous

Correlation 1 week 2 weeks 1 week 2 weeks 1 week 2 weeks

Emerging Markets: (1) (2) (3) (4) (5) (6) (7)Argentina 0.206 0.150 0.039 -0.091 -0.002 0.111 -0.089Brazil 0.868 0.153 0.078 0.193 0.182 0.076 0.012Chile 0.302 0.107 0.144 0.120 -0.023 -0.037 0.143Colombia 0.531 0.116 0.187 0.204 0.095 -0.071 0.110Mexico 0.479 0.248 -0.001 0.192 0.079 0.249 0.114Peru 0.423 0.033 0.149 0.015 -0.098 -0.116 0.113Venezuela 0.467 0.100 0.063 0.072 0.115 0.037 -0.043

China 0.127 0.113 0.010 -0.071 0.125 0.103 -0.196India 0.991 0.100 0.087 0.010 0.006 0.013 0.004Indonesia 0.791 -0.041 -0.154 0.164 0.084 0.112 0.079Korea 0.909 -0.105 -0.120 0.123 0.130 0.015 -0.007Malaysia 0.952 -0.061 -0.030 0.086 0.080 -0.031 0.006Pakistan 0.794 0.301 0.184 0.061 0.042 0.118 0.019Philippines 0.904 0.119 0.008 0.184 0.256 0.110 -0.072Sri Lanka 0.799 0.214 0.188 0.105 0.015 0.026 0.091Taiwan 0.990 -0.007 0.001 0.088 0.081 -0.008 0.007Thailand 0.985 0.014 -0.023 0.218 0.193 0.037 0.025

Czech Rep. 0.454 0.076 0.046 0.018 0.081 0.030 -0.063Greece 0.313 0.027 -0.049 -0.097 0.009 0.075 -0.106Hungary 0.347 0.043 0.022 0.372 0.041 0.022 0.331Poland 0.494 -0.079 -0.096 0.134 0.014 0.017 0.120Portugal 0.875 0.112 -0.033 -0.067 0.036 0.145 -0.103Russia 0.644 0.120 0.152 0.271 0.111 -0.032 0.160Slovakia 0.199 0.129 0.105 0.045 0.092 0.024 -0.047

Egypt 0.528 -0.083 0.134 0.081 -0.130 -0.217 0.210Israel 0.598 -0.082 -0.064 0.131 0.125 -0.018 0.007Jordan 0.922 0.056 0.030 0.026 -0.048 0.026 0.074Morocco 0.628 0.209 0.139 0.054 -0.014 0.070 0.068South Africa 0.602 0.407 0.287 0.264 0.097 0.119 0.168Turkey 0.506 -0.114 -0.078 0.020 -0.044 -0.036 0.064Zimbabwe 0.947 0.070 0.074 0.104 0.081 -0.004 0.023

Average 0.631 0.079 0.048 0.098 0.058 0.031 0.039Number of Positive Differences 21 22Sign Test (P-value) 0.0041 0.0012

Table 8Lead and Lag Relationship

Lag Investable Lag Non-Investable Difference

Crisis Period

This table reports the correlation estimates and the difference between negative and positive tail, at 1.5 standard deviations away from the mean. The rejection of the null, againstthe one-sided alternative that the negative tail correlation is greater, at the 10%, 5% and 1% significance level are denoted by (*), (**), (***). A plus sign (+) denotes tails thatincrease in correlation as the thresholds are moved away from the mean.

Positive Negative Difference T-stat Positive Negative Difference T-stat Positive Negative Difference T-statDeveloped Markets

Australia 0.214 0.342 0.128 1.273Japan 0.036 0.118 0.082 1.033Singapore 0.393 0.485 0.092 0.904

Belgium 0.146 0.243 0.097 1.041France 0.228 0.398 0.170 1.390 *Germany 0.126 0.276 0.150 1.674 **Italy 0.106 0.162 0.056 0.646Netherlands 0.234 0.305 0.071 0.723Spain 0.316 0.353 0.037 0.293Sweden 0.248 0.222 -0.026 -0.248Switzerland 0.154 0.346 + 0.192 1.983 **UK 0.163 0.241 0.078 0.830

Canada 0.184 0.246 0.062 0.651US 0.201 0.291 0.090 0.945

Emerging Markets:Argentina 0.000 0.173 0.173 1.813 ** 0.000 0.172 0.172 1.775 ** 0.000 0.000 0.000 -Brazil 0.000 0.116 0.116 1.443 * 0.008 0.164 0.156 1.729 ** 0.001 0.019 0.018 0.242Chile 0.043 0.248 0.205 1.745 ** 0.048 0.316 + 0.268 2.370 *** 0.045 0.252 + 0.207 1.775 **Colombia 0.000 0.015 0.015 0.219 0.002 0.000 -0.002 -0.113 0.002 0.002 0.000 -0.021Mexico 0.084 0.402 + 0.319 2.839 *** 0.075 0.398 + 0.323 2.965 ** 0.001 0.090 0.090 1.153Peru 0.120 0.246 0.126 0.855 0.107 0.200 0.093 0.658 0.024 0.034 0.009 0.098Venezuela 0.015 0.079 0.064 0.801 0.009 0.037 0.028 0.326 0.003 0.037 0.034 0.519

China 0.000 0.000 0.000 0.020 0.512 0.489 -0.023 -0.148 0.001 0.000 -0.001 -0.051India 0.121 0.150 0.029 0.249 0.148 + 0.246 + 0.098 0.679 0.154 + 0.249 + 0.095 0.646Indonesia 0.209 0.385 + 0.176 1.206 0.166 0.318 0.152 1.083 0.159 0.313 0.154 1.083Korea 0.178 0.295 0.117 0.884 0.188 0.354 + 0.166 1.094 0.173 0.308 + 0.135 0.882Malaysia 0.149 0.402 0.253 1.900 ** 0.148 0.403 0.255 1.906 ** 0.143 0.311 0.168 1.267 *Pakistan 0.002 0.045 0.043 3.624 *** 0.011 0.043 0.032 0.345 0.004 0.036 0.033 0.420Philippines 0.167 0.381 0.214 1.685 ** 0.239 0.363 0.124 0.935 0.163 0.366 0.203 1.602 *Sri Lanka 0.050 0.037 -0.013 -0.101 0.000 0.053 0.053 0.530 0.057 0.001 -0.056 -0.719Taiwan 0.342 0.358 0.016 0.086 0.351 0.350 -0.001 -0.005 0.335 0.357 0.022 0.118Thailand 0.297 0.450 0.153 1.179 0.326 0.461 + 0.135 1.060 0.309 0.414 0.105 0.788

Czech Republic 0.208 0.031 -0.177 -1.186 0.138 0.067 -0.071 -0.499 0.106 0.004 -0.102 -Greece 0.104 0.122 0.018 0.247 ** 0.102 + 0.063 -0.039 -0.373 0.084 0.243 + 0.159 -Hungary 0.328 0.241 -0.087 -0.557 0.225 0.157 -0.068 -0.458 0.071 0.177 0.106 0.644Poland 0.199 0.005 -0.194 -1.744 0.198 0.005 -0.193 -1.752 0.004 0.004 0.000 0.010Portugal 0.251 0.260 0.009 0.062 0.253 0.279 + 0.026 0.180 0.183 0.173 + -0.010 -0.075Russia 0.144 0.178 + 0.034 0.188 0.200 0.001 -0.200 -1.330 0.067 0.067 0.001 0.002Slovakia 0.001 0.043 0.042 0.380 0.003 0.005 0.002 0.051 0.000 0.000 0.000 -0.055

Egypt 0.000 0.060 0.060 0.660 0.003 0.066 0.062 0.616 0.003 0.000 -0.003 -0.168Israel 0.203 0.230 0.027 0.186 0.252 0.230 -0.022 -0.152 0.000 0.070 0.069 0.652Jordan 0.048 0.053 0.006 0.058 0.058 0.051 -0.006 -0.065 0.039 0.061 + 0.022 0.222Morocco 0.001 0.000 -0.001 -0.075 0.028 0.000 -0.027 -0.233 0.002 0.000 -0.002 -0.147South Africa 0.286 0.362 0.076 0.486 0.284 0.370 0.086 0.547 0.196 0.000 + -0.196 -1.431Turkey 0.004 0.125 0.121 1.320 * 0.004 0.131 0.127 1.390 * -- 0.971 -- --Zimbabwe 0.001 0.143 0.142 1.354 * 0.003 0.124 0.121 1.255 0.001 0.119 0.118 1.130

Total Number of Countries Estimated: 45 31 30Countries with Positive Difference: 39 20 21Sign Test (P-value) 0.000 0.035 0.008

Total Index ReturnNegative-PositiveCorrelation Correlation Correlation

Table 9ATest 3: Portfolio Balancing vs. Wealth Constraint (Exceedance Correlation)

Crisis Country: Hong Kong

Investable Index Return Non-Investable Index ReturnNegative-PositiveNegative-Positive

This table reports the correlation estimates and the difference between negative and positive tail, at 1.5 standard deviations away from the mean. The rejection of the null, againstthe one-sided alternative that the negative tail correlation is greater, at the 10%, 5% and 1% significance level are denoted by (*), (**), (***). A plus sign (+) denotes tails thatincrease in correlation as the thresholds are moved away from the mean.

Positive Negative Difference T-stat Positive Negative Difference T-stat Positive Negative Difference T-statDeveloped Markets

Australia 0.228 0.304 0.076 0.599Japan 0.157 0.308 + 0.151 1.215Hong Kong 0.296 0.450 0.154 1.186Singapore 0.527 0.625 + 0.098 0.795

Belgium 0.199 0.137 -0.062 -0.527France 0.303 + 0.225 + -0.078 -0.595Germany 0.177 0.221 0.044 0.363Italy 0.202 0.215 0.013 0.105Netherlands 0.265 0.252 -0.013 -0.099Spain 0.278 + 0.221 -0.057 -0.465Sweden 0.260 0.121 -0.139 -1.148Switzerland 0.209 0.272 + 0.063 0.497UK 0.265 0.122 -0.143 -1.201

Canada 0.188 0.280 0.092 0.713US 0.167 0.331 + 0.164 1.352 *

Emerging Markets:Argentina 0.000 0.134 0.134 1.654 ** 0.000 0.131 0.131 1.623 * 0.013 0.029 0.016 0.196Brazil 0.001 0.178 0.177 1.785 ** 0.000 0.184 0.184 2.198 ** 0.000 0.124 0.124 1.722 **Chile 0.107 0.185 + 0.078 0.711 0.097 0.185 + 0.088 0.815 0.025 0.198 + 0.173 1.680 **Colombia 0.019 0.062 0.043 0.412 0.028 0.032 0.004 0.040 0.000 0.000 0.000 0.002Mexico 0.073 0.291 + 0.218 1.836 ** 0.032 0.287 + 0.255 2.288 *** 0.047 0.130 + 0.083 0.843Peru 0.064 0.286 + 0.222 1.536 * 0.061 0.276 + 0.215 1.533 * 0.003 0.045 0.042 0.508Venezuela 0.161 0.215 + 0.054 0.436 0.096 0.082 + -0.013 -0.112 0.004 0.155 + 0.151 1.685 **

China 0.000 0.000 0.000 -0.870 0.075 0.337 0.262 3.559 *** 0.003 0.000 -0.003 -0.173India 0.084 0.265 + 0.181 1.534 * 0.067 0.236 0.169 1.255 * 0.061 0.241 0.180 1.292 *Indonesia 0.663 0.442 -0.221 -2.453 0.638 + 0.414 -0.224 -1.556 0.552 + 0.384 + -0.168 -1.110Korea 0.396 0.485 + 0.089 0.649 0.467 + 0.451 + -0.016 -0.104 0.389 + 0.500 + 0.111 0.701Malaysia 0.427 0.548 + 0.121 0.875 0.344 0.553 + 0.209 1.525 * 0.422 0.469 + 0.047 0.331Pakistan 0.028 0.221 + 0.193 1.685 ** 0.024 0.227 + 0.203 1.744 ** 0.004 0.173 + 0.169 1.826 **Philippines 0.351 0.506 + 0.155 1.178 0.283 0.397 + 0.114 0.850 0.386 0.497 + 0.111 0.843Sri Lanka 0.062 0.078 0.016 0.124 0.072 0.094 0.023 0.171 0.028 0.134 + 0.106 1.004Taiwan 0.226 0.312 + 0.086 0.500 0.236 0.299 + 0.063 0.371 0.169 0.311 + 0.142 0.851

Czech Republic 0.091 0.182 + 0.091 0.574 0.076 0.136 0.060 0.409 0.094 0.013 -0.081 -0.981Greece 0.153 0.286 + 0.133 1.071 0.146 0.261 0.115 0.921 0.000 0.100 0.099 >100 ***Hungary 0.163 0.282 + 0.119 0.783 0.141 0.255 0.114 0.757 0.002 0.186 0.184 1.422 *Poland 0.000 0.096 0.096 0.990 0.027 0.098 0.072 0.531 0.009 0.003 -0.005 -Portugal 0.339 0.293 + -0.046 -0.307 0.343 + 0.272 + -0.071 -0.483 0.255 + 0.328 + 0.073 0.482Russia 0.269 0.192 -0.077 -0.417 0.226 0.000 -0.226 -1.580 0.267 0.077 -0.190 -0.957Slovakia 0.000 0.064 0.064 0.618 0.000 0.086 0.086 0.616 0.000 0.004 0.003 0.476

Egypt 0.002 0.000 -0.002 -0.024 0.004 0.041 0.037 0.185 0.014 0.000 -0.014 -0.197Israel 0.273 0.118 -0.155 -1.348 0.272 0.118 -0.154 -0.799 0.002 + 0.052 0.049 3.806 ***Jordan 0.039 0.039 0.000 -0.002 0.025 0.035 0.009 0.121 0.051 0.024 -0.027 -0.322Morocco 0.000 0.000 0.000 -0.003 0.000 0.000 0.000 - 0.002 0.000 -0.002 -0.131South Africa 0.368 0.418 0.050 0.321 0.368 0.417 0.049 0.314 0.239 0.215 -0.024 -0.108Turkey 0.054 + 0.244 0.190 1.584 * 0.054 0.262 0.209 1.694 ** - 0.843 - -Zimbabwe 0.042 0.243 0.201 1.473 * 0.043 0.212 0.169 1.598 * 0.030 0.226 0.196 1.335 *

Total Number of Countries Estimated: 44 29 29Countries with Positive Difference: 31 23 20Sign Test (P-value) 0.002 0.000 0.012

Table 9BTest 3: Portfolio Balancing vs. Wealth Constraint (Exceedance Correlation)

Crisis Country: Thailand

Investable Index Return Non-Investable Index ReturnDifferenceDifference

Total Index ReturnDifferenceCorrelation Correlation Correlation

This figure shows weekly index returns for Hong Kong, Thailand, Indonesia and Malaysia from January to December 1997.The total index return data is from Datastream.

Figure 1

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3JA

NFE

BMAR

APR MAYJU

NJU

LAUG

SEP

OCTNOV

DEC

Hong Kong

Thailand

Indonesia

Malaysia

The figures show the time-series probabilities of the volatile regime, from the estimated regime-switching model, for the periodJanuary 1996 to January 2000. Figure 2A shows probabilities from the regime-switching model estimated using Hong Kong stockindex and the world index. Figure 2B shows probabilities from the regime-switching model estimated using Thailand stock indexand the world index.

Figure 2: Regime Probabilities

Figure 2A: Hong Kong and World Index

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Jan-

96

Mar

-96

May

-96

Jul-9

6

Sep-

96

Nov

-96

Jan-

97

Mar

-97

May

-97

Jul-9

7

Sep-

97

Nov

-97

Jan-

98

Mar

-98

May

-98

Jul-9

8

Sep-

98

Nov

-98

Jan-

99

Mar

-99

May

-99

Jul-9

9

Sep-

99

Nov

-99

Jan-

00

Figure 2B: Thailand and World Index

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Jan-

96

Mar

-96

May

-96

Jul-9

6

Sep-

96

Nov

-96

Jan-

97

Mar

-97

May

-97

Jul-9

7

Sep-

97

Nov

-97

Jan-

98

Mar

-98

May

-98

Jul-9

8

Sep-

98

Nov

-98

Jan-

99

Mar

-99

May

-99

Jul-9

9

Sep-

99

Nov

-99

Jan-

00