VJPS2.3qq
POLICY RESEARCH WORKING PAPER 2399
Managers, Investors, This study of an important
class of investors-U.S.
and Crises mutual funds-finds that
mutual funds do engage in
Mutual Fund Strategies momentum trading (buying
winners and selling losers).
in Emerging Markets They also engage in
contagion trading strategies
Graciela Kaminsky (selling assets from one
Richard Lyons country when asset prices fall
Sergio Schmukler in another).
The World Bank
Development Research Group
Macroeconomics and Growth
July 2000
POLICY RESEARCH WORKING PAPER 2399
Summary findings
Kaminsky, Lyons, and Schmukler address the trading Lagged momentum trading (buying past winners and
strategies of mutual funds in emerging markets. The data selling past losers) is stronger during noncrises, and
set they develop permits analyses of these strategies at stronger for fund managers.
the level of individual portfolios. Investors also engage in contagion trading-selling
A methodologically novel feature of their analysis: assets from one country when asset prices fall in another.
they disentangle the behavior of fund managers from that These findings are based on data about mutual funds
of investors. that represent only 10 percent of the market
For both managers and investors, they strongly reject capitalization in the countries considered. Were it a
the null hypothesis of no momentum trading. Funds' larger share of the market, finding counterparties for
momentum trading is positive: they systematically buy their trades (the investors who buy when they sell and
winners and sell losers. sell when they buy) would be difficult-and the premise
Contemporaneous momentum trading (buying current that funds respond to contemporaneous returns rather
winners and selling current losers) is stronger during than causing them would become tenuous.
crises, and stronger for fund investors than for fund
managers.
This paper-a product of Macroeconomics and Growth, Development Research Group-is part of a larger effort in the
group to understand capital flows to developing countries. The study was funded by the Bank's Research Support Budget
under the research project "Mutual Fund Investment in Developing Countries." Copies of the paper are available free from
the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Emily Khine, room MC3-347, telephone 202-
473-7471, fax 202-522-3518, email address kkhine@worldbank.org. Policy Research Working Papers are also posted on
the Web at www.worldbank.org/research/workingpapers. The authors may be contacted at graciela@gwu.edu,
lyons@haas.berkeley.edu, or sschmukler@worldbank.org. July 2000. (36 pages)
Phe Policy Research Working Paper Seies disseminates the findings of work in progress to encourage the exchange of ideas aiout
development issues. An objective of the series is to get the f ndings out quickly, even if the presentations are less than fully polished. The
papers carry the names of the authors and sbould be cited accordingly. The findings, interpretations, and conclusions expressed in this
paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the
countries they represent.
Produced by the Policy Research Dissemination Center
Managers, Investors, and Crises:
Mutual Fund Strategies in Emerging Markets
Graciela Kaminsky
Richard Lyons
Sergio Schmukler
JEL: F3, GI, G2.
Keywords: mutual funds; managers; investors; trading strategies; emerging markets;
momentum; feedback trading; crisis; contagion.
Respective affiliations are George Washington University, UC Berkeley and NBER, and the World
Bank. Correspondence to Sergio Schmukler, The World Bank, 1818 H Street NW, Washington, D.C.,
20433, Tel: 202-458-4167, sschmuklereworldbank.org. We thank the following for valuable comments:
Jeff Frankel, Mike Gavin, George Hoguet, Andrew Karolyi, Federico Sturzenegger, and participants at the
World Bank/Universidad Torcuato Di Tella conference on Integration and Contagion (June 1999), the
Cancun Meeting of the Econometric Society, the Latin American Economic Association, and the IDS-U.
of Sussex. For help with data we thank the World Bank (East Asia and Pacific Region), Erik Sirri from the
SEC, Konstantinos Tsatsaronis from the BIS, and Ian Wilson from Emerging Market Funds Research. For
excellent research assistance we thank Jon Tong, Sergio Kurlat, Cicilia Harun, Jose Pineda, and Allen
Cheung. (The efforts of Sergio Kurlat and especially Jon Tong were prodigious.) For financial support we
thank the NSF and the World Bank (Latin American Regional Studies Program and Research Support
Budget).
Managers, Investors, and Crises:
Mutual Fund Strategies in Emerging Markets
L Introduction
Financial crisis in 1997 engulfed not only Asia, it spread to countries as distant as
South Africa, the Czech Republic, and Brazil. To understand why, a literature has
developed that examines why the spreading of crisis might be due to financial links. There
is evidence that banks, for example, were important in spreading the 1997 crisis. The
transmission channel was lending: countries were exposed to the same banks (Kaminsky
and Reinhart 1999). Portfolio investors have also been scrutinized, particularly
institutions, such as hedge funds, pension funds, and mutual funds (Brown et al. 1998,
Eichengreen and Mathieson 1998, Kim and Wei 1999, Frankel and Schmukler 1998, among
many others). A common conclusion is that institutions sometimes panic, disregarding
fundamentals, and spreading crisis even to countries with strong fundamentals. The
literature notes that individuals, too, can contribute to this panic by fleeing from funds-
particularly mutual funds-forcing fund managers to sell when fundamentals do not warrant
selling.
This paper contributes to this literature on financial links by examining the trading
strategies of an important class of investor: U.S. mutual funds. Surprisingly, systematic
analysis of mutual funds' international strategies does not yet exist.' Consequently, our
results are of more general interest than our crisis motivation might suggest. At the same
time, the lack of systematic analysis of funds' behavior during crises warrants special
attention. Though there is some evidence that funds help crisis to spread, that evidence is
indirect and highly aggregative. Frankel and Schmukler (1998), for example, use closed-
end mutual funds to show that the Mexican crisis in 1994 was not transmitted to Asia
directly, but indirectly, via New York, where the funds are traded. The opposite view-
that funds do not spread crisis-also has some support in aggregate data. For example, net
IFunds' domestic (U.S.) strategies have been analyzed extensively, however. See Grinblatt et al. (1995),
Warther (1995), and Wermers (1999), among others.
1
redemption by mutual-fund investors during crisis periods is not large, and outflows that
occur tend to be small and short-lived (at least during Mexico's crisis-see Marcis et al.
1995 and Rea 1996). Froot et al. (1998) present a similar picture based on aggregated
flows that mix mutual funds with other types of international investor. They find that net
inflows during the Mexican and Asian crises decreased, but there is little evidence of net
outflows.2
Our paper departs from the more aggregated analysis above by effecting analysis at
the portfolio level. We develop a novel data set that includes individual portfolios, which
allows us to examine trading strategies at much higher resolution. The data include the
quarterly holdings of 13 mutual funds from April 1993 to January 1999. All 13 funds are
dedicated Latin America funds. (At year-end 1998, there were 25 Latin America funds; the
13 we track account for 88% of the value of these 25 funds.) We use these data to address
two sets of questions. The first set relates to whether funds engage in momentum trading-
systematically buying winning stocks and selling losing stocks (Jegadeesh and Titman
1993, Grinblatt et al. 1995). The second set of questions relates to whether funds engage in
contagion trading, by which we mean systematically selling stocks from one country when
stock prices are falling in another. In addressing this second set of questions, we establish
a first, direct empirical link between contagion and trading strategies.
The methodological contribution of the paper is our approach to attributing actions
to fund managers versus underlying investors. Despite a vast literature on the behavior of
domestic (i.e., U.S.) funds, to our knowledge we are the first to disentangle the two. In
effect, the trades of mutual funds reflect both institutional and individual decisions. To
understand those trades, particularly in the intemational context, ensuring that the decisions
are not commingled is an important step.
Our results show that emerging-market funds do indeed engage in momentum
trading. Their strategies exhibit positive momentum-they systematically buy winners and
sell losers. This is due to momentum trading at both the fund-manager level and the
investor level (through redemptions/inflows). We further distinguish between
2 Though a lovely data set, the Froot el al. (1998) data do not include transactions settled in foreign
currencies, e.g., ADR trades in New York and Brady bonds. These trades can be especially important in
times of crisis when local-market liquidity is at a minimum. Our fund-portfolio data include these trades.
2
contemporaneous momentum trading (buying current winners and selling current losers)
and lagged momentum trading (buying past winners and selling past losers).
Contemporaneous momentum trading is stronger during crises, and stronger for fund
investors than for fund managers. Lagged momentum trading, on the other hand, is stronger
during non-crisis periods, and stronger for managers. We also find that funds engage in
contagion trading, by which we mean that they systematically sell assets from one country
when asset prices fall in another. This contagion trading is due primarily to underlying
investors, not managers.
The paper is organized as follows. The next section outlines our approach to
measuring momentum trading and contagion trading. Section III describes our data. Section
IV presents our momentum and contagion results. Section V addresses whether return
autocorrelation within Latin America can rationalize our section-IV results. Section VI
concludes. The appendix provides some related regression-based analysis.
H. Strategies: Momentum Trading and Contagion Trading
This section presents our approach to testing whether funds employ momentum and
contagion trading strategies. Momentum trading-also called positive feedback trading-is
the systematic purchase of stocks that have performed well, and sale of stocks that have
performed poorly ("winners" and "losers"). Contagion trading is the selling of assets from
one country when asset prices are falling in another. Contagion trading is thus a cross-
country phenomenon, in contrast to momentum trading, which is a within-country
phenomenon. (This type of cross-country analysis is not possible using recent single-
country data sets, such as those of Kim and Wei 1999 and Choe, Kho, and Stulz 1999.)
First, we review the existing finance literature on momentum trading. Second, we
present our approach to testing for momentum trading, an approach that draws from this
earlier literature. Then we turn to contagion trading, presenting first a brief review of the
"contagion" literature, followed by our approach to testing for contagion trading. The
approach we adopt in testing for contagion trading is in the same spirit as our test for
momentum trading.
3
11.1. Introduction to Momentum Trading
The literature on momentum trading includes two lines of work, one based in asset
pricing and the other based in international finance. The asset-pricing line begins with the
finding that a strategy of buying past winners and selling past losers generates significant
positive returns over 3- to 12-month holding periods (Jegadeesh and Titman 1993, Asness
et al. 1997, Rouwenhorst 1998).3 Once established, this result inspired work on whether
investors actually follow momentum trading strategies. Grinblatt et al. (1995), for example,
examine the domestic strategies of U.S. mutual funds and find that they do systematically
buy past winners. They do not systematically sell past losers, however. They also find that
funds using momentum trading strategies realize significantly better performance.
Evaluation of performance is a central theme for all the papers in this asset-pricing line of
the literature.
The second line of work on momentum trading is based in international finance. Its
organizing theme is the link between returns and intemational capital flows. At the center
of this literature is the positive contemporaneous correlation between capital inflows and
returns. Early work establishes this correlation using data aggregated over both time and
types of market participant (Tesar and Werner 1994, Bohn and Tesar 1996). Later work
relaxes the aggregation over time to address whether the contemporaneous correlation in
quarterly data is truly contemporaneous (Froot et al. 1998, Choe et al. 1999, Kim and Wei
1999). Higher frequency data can distinguish three possibilities. Returns may precede
flows, indicating positive feedback trading (which is not necessarily irrational, per the
asset-pricing literature noted above). Returns and flows may be truly contemporaneous,
indicating that order flow itself may be driving prices.4 And returns may lag flows,
3
The return "continuations" that are implied by this result are not inconsistent with the return "reversals"
documented elsewhere in the literature. Horizon length is the key to understanding this: continuations
appear at mid-range horizons, 3 to 12 months. Return reversals, in contrast, appear at short horizons (up to
I month, see Jegadeesh 1990 and Lehmann 1990) and at long horizons (3 to 5 years, see De Bondt and
Thaler 1985). Reversals call for "contrarian" (or negative feedback) trading strategies. Parenthetically, all
these timne-series anomalies are distinct from the cross-sectional anomalies that have received much
attention in the asset-pricing literature recently (e.g., size and book-to-market effects).
4Microstructure finance provides three channels for truly contemporaneous price impact. The first is
information-if the buyer has superior information about a security's payoffs, then the purchase signals
that information, shifting expectations, and thereby increasing price. The second is incomplete risk-
sharing at the marketmaker level-the buyer's purchase temporarily disturbs the marketmaker's position,
4
indicating flows' ability to predict returns. Using high-frequency data aggregated across
types of market participant, Froot et al. (1998) find evidence of all three, with the first-
positive feedback trading-being the most important for explaining quarterly correlation.
Choe et al. (1999) and Kim and Wei (1999) use high-frequency data from Korea to
examine positive feedback trading around the 1997 currency crisis. Choe et al. find that
foreign investors as a group engage in positive feedback trading before the crisis, but
during the crisis feedback trading mostly disappears. Kim and Wei examine foreign
institutional investors separately and find that they engage in positive feedback trading at
all times-before, during, and after the crisis.
Our analysis is related to, and borrows from, both the intemational-finance and
asset-pricing lines of the literature. Like the work in international finance, we are more
concerned about international flows and crisis transmission than portfolio performance.
Like work in asset pricing, however, we maintain a direct link to investment strategy and
its measurement. In particular, we focus on a specific class of international investor-
mutual funds. A benefit of focusing on a specific investor class is that we can characterize
the evolution of actual portfolios, and how that evolution relates to returns in various
countries. Another benefit is that our data allow us to analyze jointly the behavior of fund
managers and their underlying investors. On the cost side, focusing on funds as a specific
investor class means that we lose resolution in terms of data frequency: our data are
quarterly.
which requires the buyer to pay compensation in the form of a higher price (so-called "inventory
effects"). The third is imperfect substitutability-the buyer's purchase may be a large enough portfolio
shift relative to the market as a whole that permanently higher price is required to clear the market (even
if it is common knowledge that the buyer does not have superior information about the security's
payoffs).
5
11.2. Measuring Momentum Trading
Our momentum-trading measure is akin to that used to analyze funds' domestic
strategies (e.g., Grinblatt et al. 1995). The measure captures the relation between security
transactions and returns. It is based on the mean of individual observations of the variable:5
ij,y = Q t (1)
where Qi,j,t is the holding by fund i of stock j (in shares) at time t, Qij t is (QiJ,t+Qii,t )/2,
and Rj,tk is the return on stock j from t-k-l to t-k. When k-0, this measure captures the
contemporaneous relation between trades and returns-referred to as lag-zero momentum
trading (LOM). When k=l, the measure captures the lagged response of trades to returns,
and is referred to as lag-one momentum trading (L1M). Parenthetically, notice the
implication of the j subscript: the mean of ,, measures the intensity of momentum trading
at the level of individual stocks. Testing the null of no momentum trading is a test of
whether the mean of Kj,, over all i, j, and t is zero.
This measure of momentum trading has two important advantages. First, it is not
contaminated by "passive price momentum." Passive price momentum arises in momentum
trading measures-like those of Grinblatt et al.-where the term in brackets is a change in
portfolio weight, rather than a percentage quantity adjustment. When using a portfolio
weight, a price increase in one stock (relative to prices of other holdings) produces a
positive relation between weights and returns that has nothing to do with trading strategy.
(A similar positive relation arises for losing stocks.) The second advantage of our measure
over one based on portfolio weights is that our measure is not contaminated by another
passive effect- "passive quantity momentum." When using portfolio weights, a large trade
in one stock can have substantial effects on the weights of holdings that involve no
transactions. Our main concem here-as in the rest of the international-finance-based
5This mean estimate does not value-weight the individual stock positions. This could make a difference if
the intensity of momentum trading differs depending on position value. After calculating it both ways, we
did not find any qualitative difference in the results.
6
literature on momentum trading-is the relation between returns and transaction flows.6
Accordingly, we want our realizations of MK,, to reflect actual transactions-the buying
and selling of winners and losers.
Separating Manager and Investor Momentum Trading
An important issue in the context of mutual-fund strategies is the effect of net
redemptions. Many funds experience substantial redemptions during crisis periods. If, on
average, funds sell shares to meet redemptions when Ri,t-k is negative, then our momentum
trading measures will be positive. This result is not spurious. But it does reflect strategies
of underlying investors, rather than strategies of the fund manager.
We control for this redemption effect by measuring the quantity transacted in each
stock relative to a fund-specific benchmark. This benchmark reflects the quantity that
would be transacted if a fund's net flows from investors produced proportional adjustment
in all stocks. Specifically, to isolate the manager's contribution to momentum trading we
calculate individual observations of:
Q _Q 2;~~I(Qi.X., -QQ,.t [)j, p
M.t = QiLj,QZt4 jEi ] RJ (2)
where Pj,, is the price of security j at time t, and Pj, is (Pj,,+Pj,, )/2. The second term in
brackets is a term that is fund-specific, i.e., for a given fund i and time t, it is invariant
across stocks j. It captures the percent increase in portfolio size due to net inflows. Here,
we use the notation je i to denote all those stocks j held by fund i. The overall momentum
trading measure in equation (2) therefore reflects the degree to which the manager of fund i
buys winners and sells losers beyond any average quantity adjustment due to fund
inflows/outflows. To understand why, note that the numerator of the second term in
6This contrasts with the asset-pricing-based literature on momentum, whose main concem is portfblio
performnance, in which case it is necessary to consider the return impact of all portfolio positions. Note
too that emerging-market funds are subject to large and rapid redemptions which, depending on liquidity
in specific markets, can produce significant passive quantity momentum.
7
brackets is the change in portfolio value due to inflows/outflows-using the P,, term
factors out capital gains/losses-and the denominator is the average portfolio value. (As
with our first momentum trading measure M;j,,, when k=O M', captures the
contemporaneous relation between trades and retums-LOM-and when k=l M,j
captures the lagged response of trades to retums-LIM). Under the null hypothesis of no
momentum trading at the manager level, the mean of the observations M'j, is zero.
We can also examine the investor-level term in isolation. That is, we can calculate
individual observations of
,ki,j,t - Qij,it-l )Pj5
j,f= ]5 Rj,t. (3)
Henceforth, we refer to momentum trading statistics calculated from equations (1)-(3) as
whole-fund, manager-only, and investor-only momentum, respectively.
A Second Investor-Level Measure
Before moving on, it is important to recognize what our investor-only measure is
capturing, and what it is not capturing. What our investor-only measure does capture is
investor effects on our whole-fund measure; that is, the sum of the investor-only and
manager-only measures equals the whole-fund measure. (This is not quite true in our
reported results because we omit some outlier observations for robustness, as described
below in section IV.) Though this investor-only measure is certainly an object of interest, it
does not recognize that investors' decisions are made at the level of the fund, not at the
level of individual stocks. (Manager decisions, in contrast, are made at the level of
individual stocks). To capture this, we also estimate an investor-only measure at the fund
level. Specifically, we estimate the mean of the statistic:
(Q,' -Q,j,tQ.i. l)Fj,Rj, t-
M ji (4)
mi't I U~~~iJ,)it j,t
jei
8
Clearly, this reduces the number of observations-we lose the stock dimension-but it
better corresponds to the decision that investors actually face.
Conditional Momentum Trading
In addition to the momentum measures LOM and LiM, we are also interested in
conditional momentum trading. Specifically, we split our sample into sub-periods: crisis
and non-crisis. The crisis portion of our full sample (April 1993 to January 1999) includes
four sub-periods: December 1994 to June 1995 (Mexico), July 1997 to March 1998
(Asia), August 1998 to October 1998 (Russia), and January 1999 (Brazil).7
Statistical Inference
Several inference issues deserve further attention. First, the percentage quantity
changes-the term in brackets in equations (1) through (4)-may have fund-specific
volatilities. Two factors could account for differing volatilities at the fund level. Factor
one is the considerable cross-sectional difference in fund size-size can affect trading
strategies. Factor two is fund differences that are distinct from size, such as turnover ratios,
redemption penalties, and other factors. Below, we test for heteroskedasticity across funds
i, and after finding it, we correct for it.5
While the first inference issue pertained to heterogeneity across funds, a second
inference issue pertains to dependence across observations within funds. Specifically,
individual observations of our various momentum trading statistics, Mj-,,, are unlikely to be
independent across stocks within a given fund. Our mean estimate should account for this
7We also examined a second conditional momentum measure by splitting our sample into buys and sells
(as in Grinblatt et al. 1995). Buying past winners and seUling past losers need not be symmetric. We
found, however, that our results were extremely sensitive to the specification of expected returns, an
adjustment that is necessary when splitting buys from sells (see Grinblatt et al., page 1091). We do not
report those results due to their fragility.
8 Because our heteroskedasticity correction affects only standard errors, each observation of Mij,, gets
equal weight in the calculation of a momentum measure's mean. Our correction for heteroskedasticity
therefore does not alter the fact that funds with more observations have more effective weight.
Regrettably, we have little statistical power to explore whether funds differ appreciably in the intensity of
their momentum trading. As for heteroskedasticity in the time-series dimension, our sample partition into
crisis and non-crisis periods accounts for the most obvious correction.
9
cross-stock, within-fumd correlation. Our estimates of the mean cluster observations within
funds, and allow the weights assigned to individual observations to vary with the
covariance structure.
A third inference issue that warrants attention is the possibility that our momentum
trading measures might be biased due to high return volatility, which is clearly a feature of
our crisis-ridden sample (see Forbes and Rigobon 1998). In fact, we are not exposed to
this bias under our null, because under our null the statistics we report in Tables 1-5 are
equal to zero. In this case the bias is not problematic.9
113. Introduction to Contagion
The financial crises of the 1990s in Europe, Mexico, Asia, Russia, and Brazil
spread rapidly across countries, including countries with diverse market fundamentals.10
These events spawned a literature to make sense of the seeming "contagion." The term
contagion is used quite differently by different authors, however, so let us be more
specific. From the outset, however, it was clear that authors use that term quite differently.
Presently, the literature on contagion identifies three types: fundamental-spillover
contagion, common-cause contagion, and non-fimdamental contagion. Fundamental-
spillover contagion occurs when an inside disturbance is rapidly transmitted to multiple,
economically interdependent countries. Common-cause contagion occurs when an outside
disturbance is rapidly transmitted to multiple countries (e.g., a fall in commodity prices, or
learning about common fundamental factors). Fundamental disturbances underlie both of
these first two types. The third type-non-fundamental contagion-can stem from any kind
9 Under the alternative hypothesis of non-zero measures, however, precise statistical comparisons across
crisis and non-crisis sub-samples would require adjustments for the volatility-specific nature of the
sample split. This type of comparison is not central to our paper. Nevertheless, we did re-estimate our
main comparative results with a Forbes-Rigobon correction (in this case, a correction to estimated
covariance, rather than correlation), and found no qualitative change in the results.
10 Witness Indonesia in 1997. Nobody can disagree that there were signs of weakness in the Indonesian
economy at the outset of the Asian crisis: the banking sector was fragile, the economy was not growing,
and there was a current account deficit. Still, these problems were not insurmountable. Kaminsky (1998),
for example, estimates that the probabilities of crisis in Indonesia by June 1997 amnounted to only 20
percent. This probability stands in sharp contrast to the likelihood of a currency crisis in Thailand, which
skyrocketed to 100 percent at the beginning of 1997. Still, the Indonesian rupiah collapsed only weeks
after the floating of the Thai baht.
10
of disturbance; the defining characteristic is that the rapid transmission to multiple
countries is beyond what is warranted by fundamentals (i.e., controlling for fundamentals
cannot account for it). This third type is sometimes referred to as pure or true contagion.
Many authors focus on the first two types of contagion, those driven by
fundamentals. For example, Eichengreen, Rose, and Wyplosz (1996) examine whether
contagion is more prevalent among countries with either important trade links or similar
market fundamentals. In the first case, devaluation in one country reduces competitiveness
in partner-countries, prompting devaluations to restore competitiveness (fundamental-
spillover contagion). In the second case, devaluation acts like a wake-up call: investors
seeing one country collapsing learn about the fragility of "similar" countries, and speculate
against those countries' currencies (common-cause contagion). The Eichengreen et al.
evidence points in the direction of trade links rather than similar fundamentals. Corsetti et
al. (1998) also claim that trade links drive the strong spillovers during the Asian crisis.
Kaminsky and Reinhart (1999) focus instead on financial-sector links. In particular, they
examine the role of common bank lenders and the effect of cross-market hedging (a type of
conmnon-cause contagion). They find that common lenders were central to the spreading of
the Asian crisis (as they were to the spreading of the Debt Crisis of the 1980s).
The non-fundamental category of contagion has attracted more attention than the two
fundamentals-driven categories. Theoretical work on non-fundamental contagion focuses
on rational herding. For example, in the model of Calvo and Mendoza (1998), the costs of
gathering country-specific information induce rational investors to follow the herd. In the
model of Calvo (1999), uninformed investors replicate selling by liquidity-squeezed
informed investors because the uninformed mistakenly (but rationally) believe these sales
are signaling worsening fundamentals. Kodres and Pritsker (1999) focus on investors who
engage in cross-market hedging of macroeconomic risks. In that paper, international market
comovement can occur in the absence of any relevant information, and even in the absence
of direct common factors across countries. For example, a negative shock to one country
can lead informed investors to sell that country's assets and buy assets of another country,
increasing their exposure to the idiosyncratic factor of the second country. Investors then
-hedge this new position by selling the assets of a third country, completing the chain of
contagion from the first country to the third.
11
The literature on non-fundamental contagion also has an empirical branch.
Kaminsky and Schmukler (1999) find that spillover effects unrelated to market
fundamentals are quite common, and spread quickly across countries within a region.
Valdes (1998) examines the degree to which comovement of Brady-bond prices is
unexplained by fundamentals. Interestingly, contagion in his paper is symmetric, applying
both on the downside during crises and on the upside during periods of rapid capital
inflow. A different line of empirical work on non-fundamental contagion examines whether
crises are spread by particular investor groups. For example, Choe, Kho, and Stulz (1998)
use transaction data in the Korean equity market to examine whether foreign investors
destabilize prices. They find evidence of herding by foreign investors before Korea's
economic crisis in late 1997, but these effects disappear during the peak of the crisis, and
there is no evidence of destabilization. Since their data include only transactions on the
Korean Stock Exchange, these authors cannot examine the transmission of crisis across
countries.
11.4. Measuring Contagion Trading
Our approach to testing for contagion is different from the literature reviewed
above. Data on individual portfolios allow us to address contagion in a new way-from
the trading-strategy perspective. We will use the term contagion trading to mean the
systematic selling (buying) of stocks in one country when the stock market falls (rises) in
another."1
To do this we introduce a new measure-a contagion trading measure. Our
contagion trading measure is based on the methodology outlined above for measuring
momentum trading. Like the momentum measures, we present contagion trading measures at
three different levels: whole-fund contagion trading (C), manager-only contagion trading
11Notice that this definition does not take account of the fundamental-versus-non-fundamental distinction
introduced above. The appendix introduces a regression-based approach that allows us to test for
contagion with controls for various fundamnental factors.
12
(C), and investor-only contagion trading (C"). These three measures are the sample
averages of the variables:
CQ = (Qt.i4 - R (5)
QQj . Qij.t-I - rt JR (6)
Q i,j ,t Qf,jJ-)p-J,t
C,TJ,, = jEtQ J tPiv RfJ (7)
Instead of testing for a relation between quantity changes and own-stock returns, our
contagion trading measure tests for a relation between quantity changes and foreign-country
equity returns. In effect, we are testing for what might be called "cross-country momentum
trading." Here, R*1 is the return on the foreign-country index f from t-l to t. For each of the
three measures above (C, C', and C"), we consider five different contagion trading
measures, each one constructed from a different foreign equity index. Those foreign equity
indexes include Brazil, Mexico, Asia, Russia, and the U.S. Naturally, when calculating the
contagion trading measure when f-Brazil, we do not include observations where stock j is
from Brazil (similarly for Mexico). Under the null hypothesis of no contagion trading, the
mean of the observations C is zero.
Our contagion trading measure in equations (5)-(7) allows us to address many of
the issues we address with our momentum trading measure. For example, we examine
crisis versus non-crisis sub-samples, and we partition the crisis sub-sample further to
isolate the effects of particular crises. We do not offer a contagion-trading analogue to
13
equation (4}-investor-only at the fund level-only because the results we shall find for
that measure are, in the end, similar to the investor-only results from equation (3)
m. Data
Our data on mutual-fund holdings come from two sources. The first source is the
U.S. Securities and Exchange Commission (SEC). Mutual funds are required to report
holdings to the SEC twice a year. The second source is Morningstar. Morningstar conducts
surveys of mutual fund holdings at a higher frequency: quarterly surveys are the norm for
most funds. For our purposes, quarterly data are available from Momingstar for about 50%
of the funds we examine. In those instances where our measure of Mi, is based on
portfolio holdings that are not measured three months apart, these observations of AQj,, are
multiplied by 3/x, where x is the number of months between Qij,t and Qijt-l.
Our sample includes the holdings of 13 Latin America equity funds (open-end) from
April 1993 to January 1999 (24 quarters). Those funds are (1) Fidelity Latin America, (2)
Morgan Stanley Dean Witter Institutional Latin America, (3) Van Kampen Latin America
(formerly Morgan Stanley), (4) BT Investment Latin America Equity, (5) TCW Galileo
Latin America Equity, (6) TCW/Dean Witter Latin America Growth, (7) Excelsior Latin
America, (8) Govett Latin America, (9) Ivy South America, (10) Scudder Latin America,
(11) T. Rowe Price Latin America, (12) Merrill Lynch Latin America, and (13) Templeton
Latin America. Not all of these funds existed from the beginning of our sample; on average
we have about 10 quarters of data (out of a possible 24) per fund.
Our third source of data is Bloomberg and the Intemational Finance Corporation
(IFC). Bloomberg provides monthly price series for all equities held by the 13 funds,
including ADRs. (The need for monthly price data arises in our analysis of lag-one
momentum trading.) These price series are corrected for splits and dividends. The IFC
provides information on stock market indexes, which we need for our contagion trading
analysis. Our contagion trading analysis uses the IFC Latin America Stock Market index,
the IFC Asia Stock Market index, and several IFC country stock market indexes. The U.S.
equity return is the S&P 500 return. All return data are expressed in percent
14
IV. Results: Momentum and Contagion Trading
We present our results in four parts. First, we present aggregate evidence on the
trades of mutual funds in times of crisis. Then, we present results on within-country
momentum trading (equations 1-4). We follow these with cross-country contagion trading
results (equations 5-7). In the appendix, we also present some regression-based results
relating momentum and contagion trading with other determinants of trading strategy.
IV.1. Aggregate Evidence on our Sample of Funds During Crisis
Though our data set does include individual portfolios, let us first consider
evidence based on the aggregation of those portfolios. We focus this aggregate evidence on
funds' experience with investor inflows and outflows. During the fourth quarter of 1997-
the peak of the Asian crisis-Latin American funds suffered large outflows (Figure 1).12
The reversal from inflows to outflows during the Asian and Russian crises is more severe
than that during the Mexican crisis in December 1994. In the Mexican crisis, funds tended
to pull out of Mexico, Argentina, and Brazil, all of which are relatively liquid; funds
tended not to pull out from more illiquid markets, such as Colombia. Moreover, the
Mexico-induced pullout was temporary-by the third quarter of 1995 fund inflows to Latin
America had resumed (consistent with the findings of Marcis et al. 1995 and Rea 1996).
Relative to the Mexican crisis, the Asian and Russian crises of 1997 and 1998 were more
broad-based and persistent. In those crises the retreat from Latin America was more
indiscriminate, with heavy sales reaching even the most illiquid markets. On average, net
sales in 1998 were about 32 percent. This result differs from that of Froot et al. (1998),
who find little evidence of net outflows during the Asian crisis. A possible explanation is
that the aggregated data used by Froot et al. include institution types that counteract the
clear net selling by mutual funds (hedge funds?). Another possible explanation is that the
Froot el al. data do not include transactions settled in dollars, euros, or yen, e.g., ADR
12 Net selling in Figure I is calculated as the change in number of shares-as a percentage of average
shares held during the quarter-valued at the beginning-of-quarter price. The average shares held during
the quarter is the mean of the beginning- and end-of-quarter holdings.
15
trades in New York and dollar denominated bonds. This is very important in Latin
America. Our data set includes all these trades.
One technique available to managers is using "cash" (e.g., liquid money-market
instruments such as U.S. Treasury bills) to buffer their portfolios from redernptions.
Holding cash allows managers to meet redemptions without the need to sell less-liquid
assets. In principle, this can mute the effect of investor outflows on the underlying stocks.
However, managers can also reinforce investors' actions if they increase their liquid
positions in times of investor retrenchment. For our whole sample, funds kept an average of
4.4 percent of their net asset value in cash. We then split our sample into two sub-samples,
one where on average these funds received inflows, and one where on average these funds
suffered outflows. In the inflows sub-sample we find an average cash position of 4.6
percent, whereas in the outflow sample we find an average cash position of 4.3 percent.
Average cash positions are remarkably stable. Managers' choice of cash position does not
appear to either mute or reinforce investor actions.13
IV.2. Momentum Trading Results
In our full sample, we find strong evidence of lag-zero momentum trading at all
three levels: whole-fund, manager-only, and investor-only (Table 1, column 1).
Interestingly, contemporaneous momentum trading is especially strong during crises. In
terms of attribution, it is investors that account for the lion's share of the contemporaneous
momentum trading at the whole-fund level. Significant lag-one momentum trading is present
only in the non-crisis portion of our sample, and it is concentrated at the manager level.'4
For robustness, we estimate each cell based only on observations of Mi,, within three
standard deviations of its mean. This is the reason why, within any column of Table 1, the
13 A natural question is whether these cash positions are stable because managers face some kind of
constraint. The reality is that funds are far less constrained than our cash-holding results might indicate in
any de jure sense. De facto, however, managers are sensitive about departing too much from their
benchmarks. The classic example is the hapless manager at Fidelity's Magellan Fund in the late 90s who
felt that the stock market was over-valued, switched heavily into cash, watched the market rise further, and
was fired for the decision.
14 In our estimation, L1M always relates the transacted quantities between t-1 and t with the return over
the month preceding t-l. Increasing the length of the period over which lagged returns are measured
diminishes explanatory power, in general.
16
manager-only and investor-only estimates do not sum to the whole-fund estimate exactly. 15
To interpret the size of the coefficients, consider the whole-fund LOM estimate of
2.36. Given the units of our data, an LOM estimate of 2.5 implies that on average the
product of (AQ,,,,/ Q j,) and P,t over a quarter is 2.5 percent (a representative example
would be a return of -10% and a position reduction of 0.25, or 2.5%).16
Table 2 presents estimates of our investor-only measure at the fund level, rather
than at the stock level as in Table 1. Recall that this fund-level variant of the investor-only
measure recognizes that investors' decisions are made at the level of the fund, not at the
level of individual stocks. Despite fewer observations from losing the stock dimension, our
results are sharpened in terms of statistical significance, though the overall pattern remains
the same. The only notable change in the pattem is the significance of LlM at the investor
leve: it is now significant at the 1 percent level, whereas it was insignificant in Table 1.
Table 3 presents momentum trading measures for three crisis-period sub-samples:
the Mexican Crisis (December 1994 to June 1995), the Asian Crisis (July 1997 to March
1998), and the Russian Crisis (August 1998 to October 1998). The interesting question
here is whether momentum trading is equally strong across different crises. The answer is
no. Within our Latin American sample, we find that positive momentum trading was
strongest during the 1994 Mexican Crisis.
IV.3. Contagion Trading Results
Tables 4 and 5 present our contagion trading results. Table 4 presents the all-
sample results, as well as the crisis versus non-crisis sub-samples. Table 5 splits the crisis
sub-sample further into the Mexican, Asian, and Russian crises. In Table 4, we find more
significance at the investor level than at the manager level. Thus, investors clearly engage
in contagion trading, but managers are less apt. Of the five different return benchmarks
(Brazil, Mexico, Asia, Russia, and the U.S.), Russia clearly has the strongest effects-
15 Using all observations tends to increase both point estimates and t-statistics.
16 Returns are measured in percent. The quantity-adjustment term in momentum is untransformed (e.g.,
the 0.25 in the example). Note that the quantity-adjustment term uses the average quantity in the
denominator, so that the position reduction in our parenthetical example is only approximate. Note too
that our LIM measures below are based on monthly returns, not quarterly retums as in our LOM
measures, so their size is correspondingly smaller.
17
funds are systematically buying Latin American equities when Russia's returns are high,
and vice versa. This is especially true during the Russian Crisis, which squares with
informal accounts of the extraordinarily intense contagion at that time. Even during the
Russian Crisis, however, fund managers remained cool-headed: there is no evidence they
engaged in contagion trading. The conternporaneous relation with U.S. equity returns is the
only one of the five return benchmarks that is concentrated at the manager level. It is also
the only significant effect that is negative. This negative LOC statistic for the U.S. return
implies that fund managers systematically buy Latin American equities when U.S. returns
are low (controlling for fund inflows/redemptions). Though past work has shown clear
links between emerging-market returns and U.S. interest rates, this is the first evidence of
which we are aware that links actual portfolio shifts to U.S. equity returns.
Table 5 focuses on contagion trading during three specific crises: the Mexican, the
Asian, and the Russian. The reaction of investors to Russian equity returns during the
Russian crisis was particularly strong: investors systematically sold Latin American
equities when Russian equity returns were low. Note, though, that this link to Russia is not
operative at the manager level. In the case of the Mexican crisis, the effect is smaller, but
still significant, and there is some evidence that managers were involved in that case. In the
case of the Asian crisis, there is no discernable link to the trading of Latin American
equities. The last three columns show the link to U.S. market returns during each of these
three crises. Given the proximity to Mexico, and the importance of economic links between
the two countries, it is not surprising that the link between Latin-American portfolios and
U.S. returns is strongest during the Mexican crisis. Interestingly, the contagion trading
statistic is negative, and is significant at both the manager and investor levels. This
suggests that, during the Mexican crisis, managers and investors tended to sell Latin
American equities when U.S. returns were high, and vice versa. One interpretation is that
strong U.S. returns in the face of Mexico's crisis bodes well for Mexican equities, which
induces a portfolio shift away from the rest of Latin America.
In closing this section on contagion trading, it is worthwhile re-emphasizing the
qualitative difference between the results above and the existing contagion literature. The
difference is that we measure quantities, as well as prices, and address their joint
behavior, whereas much of the literature focuses on correlation in prices only.
18
V. Rationalizing Momentum Trading: Return Autocorrelation?
In an environment with positively autocorrelated returns, momentum trading is a
natural response. The previous section presented evidence of positive LOM and, at least
during non-crisis periods, positive LIM. This raises the question of whether returns within
Latin America exhibit positive autocorrelation. One common way to test for return
autocorrelation is using variance ratios. If returns follow a random walk, then return
variance is a linear function of horizon length. That is, the variance of returns over k
periods is k times the variance of returns over one-period. If instead returns are positively
autocorrelated, the variance of k-period returns is larger than the sum of one-period
returns-variances grow faster than linearly. Thus, variance ratios larger than one are
consistent with rational positive momentum trading. Alternatively, when returns are
negatively autocorrelated, the variance of k-period returns is smaller than k times the
variance of one-period returns. Variance ratios smaller than one would call for negative
momentum (or contrarian) trading.
Table 6 reports the values of the variance-ratio test statistic at different horizons,
together with p-values, for seven Latin American countries. For comparison we also
provide results for the U.S. stock market.17 Interestingly, stock returns in several Latin
American markets are highly persistent (variance-ratio statistics larger than one), even at
three and four-year horizons. In contrast, U.S. returns show no persistence at any horizon.
Though certainly not proof that the positive momentum trading we fmd in Latin America is
rational-after all, this persistence in returns is at the index level-these results do point to
the possibility of rationalizing our momentum results, at least for some countries (e.g.,
Mexico, Chile, Colombia, and Venezuela).
It is important to note, however, that while positive autocorrelation is necessary for
rationalizing positive LIM, it is certainly not necessary for rationalizing positive LOM. As
noted in Section 11. 1, returns and trades may be truly contemporaneous if order flow itself
is driving prices. This is possible where fund transactions are "large" relative to liquidity
17 See Campbell, Lo, and MacKinlay (1997) for the asymptotic distribution of the variance-ratio test.
19
in the market (the imperfect substitutability channel noted in footnote 4), or when fund
managers' trades are perceived as containing superior information.
VI. Conclusion
Discriminating among the various ways that financial markets can spread crisis
requires a sharper picture of actual behavior. Who is doing the trading? What are their
trading strategies? In this paper we examine portfolios of an important class of
international investor-US mutual funds. We address two sets of questions. The first
relates to whether and when these funds engage in momentum trading-systematically
buying winning stocks and selling losing stocks. We find that international funds do engage
in momentum trading. Their trading exhibits positive momentum, due to momentum at two
levels: the fund manager level and the investor level (through redemptions/inflows). Funds
also engage in momentum trading in both crisis and non-crisis periods. Contemporaneous
momentum trading is stronger during crises, and stronger for fund investors than for fund
managers. Lagged momentum trading, on the other hand, is stronger during non-crisis
periods, and stronger for managers.
The second set of questions we address relates to funds' use of contagion trading
strategies-selling assets from one country when asset prices fall in another. We find that
funds do engage in contagion trading. Per the appendix, this result is robust to controlling
for own-stock returns, the local-market factor, and the US-market factor. Strictly speaking,
while these controls have a sound theoretical basis, they are not sufficient to conclude that
this contagion trading is non-fundamental (or pure) contagion trading. In any event, we have
uncovered several stylized facts that are useful for evaluating hypotheses about the
emerging-market crises and their transmission.
Beyond these stylized facts, this paper includes several methodological
innovations. For example, the distinction between momentum trading at the manager and
investor levels is new to the literature, as is our method for distinguishing the two. Our
method of measuring contagion trading via transaction quantities is also new. Finally, our
regression-based approach to controlling for systematic return factors in measuring
momentum and contagion trading provides a valuable check on the bilateral measures'
robustness.
20
An important question we have not addressed is, Who takes the other side of these
momentum and contagion trades? Someone certainly must. This question is, unfortunately,
beyond the feasible scope of our analysis. We can offer some parting thoughts however.
Consider for example the following question: If the model in our managers' and investors'
heads is one of undershooting prices, followed by positively autocorrelated returns, then
must it be that their counter-parties believe the opposite model? No, this is not necessary.
The literature in microstructure finance-which we touch on in section II.1-provides
many models of liquidity providers who do not have opposite models or views, they
simply require compensation for providing liquidity in the form of transaction costs
(revenues from their perspective). It is also appropriate to keep in mind that, together, the
mutual funds we examine own only about 10 percent of the market capitalization of the
countries we consider. If they were a more substantial fraction, then finding counter-parties
for their trades would be much more difficult. Indeed, the premise that funds respond to
contemporaneous returns rather than causing them would be become rather tenuous.
21
Appendix: A Regression-Based Approach
The bivariate relations examined via equations (1)-(7) draw from, and therefore
allow direct comparison with, past empirical work on momentum trading. But these
bivariate relations provide no means of testing joint significance. Is lag-one momentum
trading still significant after controlling for lag-zero momentum trading (i.e., after
controlling for contemporaneous price effects)? Is cross-country contagion trading still
significant after controlling for own-price effects via lag-zero and lag-one momentum
trading? Are these relations robust to including local-market index returns and the US-
market index return?
A regression-based approach provides a natural framework for addressing these
questions. At the whole-fund level, the questions of the previous paragraph can be
addressed by estimating:
(QjJ fQ-QHJ )1 = + P1R, -+ 2R- t-l + 3RLA,t + 34RLM,t + 5Rus, +£, ,j,. (Al)
Here, N.,t and Rj,t l are own-stock returns, as before. These variables capture lag-zero and
lag-one momentum trading, respectively. The variable FqA,, is the contemporaneous return
on a Latin American equity index. 18 This variable captures cross-country contagion trading.
The fourth variable, Ru6,t, is the local-market index return. This variable does not enter the
analysis introduced in the previous sections, and is intended here as a control for country-
level systematic factors. The last variable, Rus,t, is the US-market index return. This
variable also does not enter in the previous sections, and is intended here as a control for
systematic U.S. factors, which have well established effects on emerging equity markets.
At the nimanager-only and investor-only levels, the dependent variable in equation
(Al) is replaced with:
18 We do not attempt to remove the own-country portion of the broader Latin American index. Note,
thought, that the own-country index is also in the regression, and our results are able to distinguish quite
sharply between them. In fact, the own-country index is never significant, so it is highly unlikely the
effects are confounded.
22
(Q1 Q __________ X(Q,iJ - Q'-J j,t
Manager-only _ : ) (A2)
jF;
I(Qijit -Qij.t-l)Pj.,
Investor-only : j -l (A3)
jci
This follows the separation of the manager-only and investor-only levels in our analysis of
bivariate momentum and contagion trading.
Results
Tables Al-A3 present OLS estimates of the models in equations (Al), (A2), and
(A3). At the whole-fund level (Table Al), the full sample generates significant positive
coefficients on all of the first three variables. Thus, momentum and contagion trading are
robust to moving from bivariate measures to multivariate measures, and including controls
for the overall local and U.S. markets. Interestingly, the local-market control is never
significant. The U.S.-market control, in contrast, is quite significant, and negative. This
squares with past empirical work showing that U.S. investors tend to chase emerging
markets when returns at home are low. When the sample is split into crisis and non-crisis
sub-periods, we find that contagion trading is largely a crisis-period phenomenon.
Tables A2 and A3 present results for the manager-only and investor-only
regressions, respectively. At the manager level, we find significant positive momentum
trading (both lag zero and lag one), and significant contagion trading with respect to the
U.S. market, but no evidence of contagion trading with respect to other Latin American
markets (3), except in times of crisis. Our investor-level results tell a distinctly different
story. Once we control for the local index return, we find that investors do not engage in
stock-specific momentum trading. This is not surprising: one would not expect investors to
respond to individual stocks, but to the market as a whole. They do respond strongly,
however, to the contemporaneous local-index return. And they also respond strongly within
23
the quarter to other Latin American markets per the significant positive coefficient f3.
Note, though, that these latter two effects are concentrated in the non-crisis periods.
24
Figure 1: Mutual Funds' Net Buying/Selling of Stocks in Latin American Countries
Argentina Mexico
60 -
40 4
20 ~~~~~~~~~~~~~~20
Brazil Pr
60 0
-20 \.AL I 2(|1 1 |_ wl - E;8
20 440
0 205| s O_ >
-60. ~~~~~~~~~~~~~~~~-60:
-80. ~~~~~~~~~~~~~~~~-80
-100 -100
25 t
40 0
-80
60 40-
40 I20
20 0
20 ~~~~~ ~ ~~~~~~~~~~-20 -W0
0 WI ~~~~~~~~~~~~~~40
-20 ~~~~~~~~~~~~~~~~-60
-80
Coloxmbia Latin America
60 -3
-20.
-30-
-60 -40
C, ~~ ~~~0 C, 00 0% C .70%00 0
Source: Our data set. Net Buying/Selling is equal to the value-weighted percentage change in quaLrterly holdings of all funds in each
country, where the value weighting uses the beginning-of-period share price. All figures are in percent. However, since quarterly
change in the nurmber of shares is divided by the mean number of shares (at the beginning and end-of-period), changes can be greater
than 100 percent.
2 5
Table 1
L_ Iag-O and Lag-I Momentum Trading
All Sample Non-Crisis Crisis
Whole-Fund Momentum
LOM 2.36*** 0.98*** 5.13***
T-statistic 5.63 3.19 4.55
Observations 4924 3288 1636
LIM 0.20 0.25** 0.11
T-statistic 153 2.35 0.40
Observations 4852 3214 1638
MAnaver-Oly Momentu
LOM 0.868** 0.29 2.01***
T-statistic 2.90 1.27 2.68
Observations 4929 3287 1642
LIM 0.16 0.18** 0.11
T-statistic 1.58 2.11 0.61
Observations 4849 3210 1639
Investor-ngv Mintentu
LOM 1.70*** 0.81*** 3.46***
T-statistic 6.12 3.45 4.09
Observations 4954 3292 1662
LIM 0.08 0.05 0.16
T-statistic 1.06 0.75 0.75
Observations 4854 3221 1633
LOM is the point estimate for the mean lag-0 momentum trading measure. LIM is the point estimate for the mean lag-i momentum
trading measure (measured from return over the previous month). Whole-Fund momentum tests whether the mean of (AQu/Q i,j )P,,-kis
zero, per equation (1). Manager-Only momentum controls for investor redemption effects as in equation (2). Investor-Only momentum
reflects only investor redemption effects as in equation (3). All t-statistics are corrected for heteroskedasticity across flunds. Full
sample: quarterly data from April 1993 to January 1999. The crisis portion of the sarnple is December 1994-June 1995, July 1997-March
1998, August 1998-October 1998, and January 1999. The non-crisis portion is the rest of the sample. The total of roughly 4400
observations is 13 funds times an average of about 35 stocks per fund, times an average of about 10 quarters of available data per fund.
For robustness, results in each cell are based only on observations within tbree standard deviations of the mean.
* Statistically Significant at the 10-percent level
* Statistically Significant at the 5-percent level
** Statisticaly Significant at the 1-percent level
26
Table 2
Investor-Only Momentum at the Fund Level
All Sample Non-Crisis Crisis
Investor-Only: Fund Le''
LOM 1.99*** 0.97*** 3.78***
T-statistic 5.49 3.55 4.30
Observations 127 81 46
LIM 0,54** 0.49* 0.63**
T-statistic 2.83 1.79 2.09
Observations 115 72 43
LOM is the point estimate for the mean lag-0 momentum trading measure. LIM is the point estimate for the mean lag-I momentum
trading measure (measured from return over the previous month). Investor-Only: Fund Level reflects only investor redemption effects
at the fund level as in equation (4). All t-statistics are corrected for heteroskedasticity across funds. Full sample: quarterly data from
April 1993 to January 1999. The crisis portion of the sample is December 1994-June 1995, July 1997-March 1998, August 1998-October
1998, and January 1999. The non-crisis portion is the rest of the sample. The total of roughly 127 observations is 13 funds times an
average of about 10 quarters of available data per fund For robustness, results in each cell are based only on observations within three
standard deviations of the mean.
Statistically Significant at the 10-percent level
** Statistically Significant at the 5-percent level
'# Statistically Significant at the I-percent level
27
Table 3
Momentum Trading Results by Crisis
Mexican Criss Asian Crisis Rassian Crisis
Whole-Fund Mometum
LOM 12.11**# 1.69*** 8.26**"
T-statistic 3.45 2.97 4.24
Observations 268 920 417
LIM 1.00* -025 0.22
T-statistic 1.82 -0.69 0.57
Observations 297 898 413
LOM 6.56* 0.99** 1.04
T-statistic 2.16 2.32 0.90
Observations 279 920 412
LIM 1.00*** -0.17 -0.04
T-statistic 2.71 -0.74 -021
Observations 297 898 414
h^n lyto-oi Mvments
LOM 7.56** 0.71** 6.86***
T-statistic 2.38 2.30 5.84
Observations 284 921 426
LIM 0.12 0.00 0.64
T-statistic 0.34 -0.02 130
Observations 294 910 398
LOM is the point estimate for the mean lag-0 momentum trading measure. LIM is the point estimate for the mean lag-I momentum
trading measure (measured from return over the previous month). Whole-Fund momentum tests whether the mean of (AQqij, )RA,., is
zero, per equation (1). Manager-Only momentum controls for investor redemption effects as in equation (2). Investor-Only momentum
reflects only investor redemption effects as in equation (3). All t-statistics are corrected for heteroskedasticity across funds. The
Mexican Crisis portion of the sample is December 1994-June 1995. The Asian Crisis portion of the sample is July 1997-March 1998. The
Russian Crisis portion of the sample is August 1998-October 1998. For robustness, results in each cell are based Dnly on observations
within three standard deviations of the mean.
* Statistically Significant at the 10-percent level
** Statistically Significant at the 5-percent level
Statisticaly Significant at the 1-percent level
28
Table 4
Contagion Trading Results
County/Regioua Index
Brazil Mexico Asia Russia U.S.
Statistics All Non- Crisis All Non- C(isis All Non- Crisis Al Non- Crisis Al Non- crisis
Swmpic Crisis Sanplc G isis Samplc Crisis Sapkl Crisis Sampic Crisis
Whole Fund
LOC 1.80*** 0.63 4.15*** 0.83*** 0.70 1.10* 0.72** 0.39*** 1.38 3.91*** 2A8 6.18*** -0.58**' -0.25 -1.26***
T-statistic 3.45 1.06 3.00 2.56 1.63 1.68 2.23 2.79 1.38 3.05 138 2.73 -2.81 -1.08 -3.08
Man&=r Oni,
LOC 0.09 -0.63 1.52*** 0.02 -0.13 032 0.53** 0.12 1.36** -0.59 -1.26 0.46 -0.50*s* -0.50+** 4.51***
T-statistic 0.22 -0.99 3.17 0.08 -0.43 0.63 2.21 0.93 2.15 -0.66 -1.05 0.29 -3.61 -2.90 -2.84
Inestor Onl
LOC 1.89+** 1.69*5* 2.30** 1 .12* 0.92*** 1.52*5* 0.62*** 0A5*** 0.96 5.87** 4.36*** 8.28** -0.02 0.21 -0.48
T-statistic 3.95 3.04 3.64 4.63 3.23 4.09 2.58 3.59 1.34 4.70 2.88 443 -0.09 1.15 -1.27
LOC dcnotcs lag-0 contagion trading. Whole-Fund contagion tests whether the mean of (AqI,/ ,)RKis zero, where R, is the return on foreign index f from t-l tot, with fe (Bsazil. Mexico, Asia, Russia,
U.S. }, per equation (4). Manager-Only contagion controls for investor redemption effects as in equation (5). lnvestor-Only contagion reflects only investor redemption effects as in equation (6). All t-
statistics are corrected fbrheteroskedasticity across funds. Full sample: April 1993 to January 1999. The crisis portion ofthe sample is December 1994-June 1995. July 1997-March 1998, August 1998-
October 1998, and January 1999. The non-crisis portion is the rest of the sample. Asia is the IFC Asia Stock Market Index. Note that Braziliani equities are excluded from the calculation of LOC for Brazil
(similarly for Mexico).
* Statistically Significant at the 10-percent levdl
* Statistically Significant at the 5-percent level
Statistically Significant at the I -pcrcent level
29
Table 5
Contagion Trading Results by Individual Crisis
Mexico During Asia During Russia During U.S. During U.S. During U.S. Daring
Mcxican Crisis Asian Crisis Russian Crisis Mcxican Crisis Asian Crisis Russian Crisis
Statistic
Whole Fund
L0C 3.899 0.27 22.0"*' -3.59*** -0.90** .22
T-statistic 1.88 0.40 3.55 -3.71 -2.23 -0.33
MnA_ .ul
LOC 1.89* 1.66 6.03 -1.73*** -0.41 0.10
T-statistic 1.80 135 1.18 -3.91 -1A8 0.86
in tr Oly
LOC 3.23*" 0.42 24.0*"' -1.86* -0.17 0.17
T-statistic 2.00 0.47 5.53 -2.44 -0.42 0.30
1~ ~ ~ ~
LOC denotes lag-0 contagion trading. Wholc-Fund contagion tests whethcr the mean of(AQ Qijt )Rf,is zero, where Rftis tdreturn on forcign imdcx f from t- Ito t, with fe (Mcxico, Asia. Russia, U.S.. pcr
equation (4). Manager-Only contagion controls for investor redemption effects as in equation (5). Investor-Only contagion reflects only investor redemption effects as in equation (6). All t-statistics are
corrected for heteroskedasticity across funds. The Mexican Crisis portion of the sample is December 1994-June 1995. The Asian Crisis portion of the sample is July 1997-March 1998. The Russian Crisis
portion of the samplc is August 1998-October 1998. Asia is the IFC Asia Stock Market Index. Note that Mcxican cquities arc cxcluded from the calculation of LOC for Mcxico.
* Slatisticaly Significant at the 10-percent level
** Statistically Significant at the 5-percent level
"' Statistically Significant at the I -percent level
30
Table 6: Variance Ratio Test of Stock Returns
Horizon
COUNTRY 3-months 12-months 24-months 36-months 48-months 60-months
Argentina 1.02 0.88 0.70 0.62 0.60 0.60
(0.84) (0.59) (0.35) (0.35) (0.39) (0.45)
Brazil 1.01 0.99 0.82 0.71 0.75 0.83
(0.93) (0.95) (0.57) (0.46) (0.59) (0.75)
Chile 1.34 1.94 2.50 2.88 3.16 2.97
(0.00) (0. 00) (0.00) (0.00) (0.00) (0.00)
Colombia 1.43 2.22 2.40 2.63 2.76 2.81
(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Mexico 1.31 1.50 1.61 1.74 1.85 1.84
(0.00) (0.02) (0.06) (0.06) (0.07) (0.11)
Peru 1.07 0.80 0.64 0.70 0.82 0.58
(0.45) (0.37) (0.27) (0.45) (0.69) (0.42)
Venezuela 1.15 1.59 1.53 1.10 0.97 0.88
(0.09) (0.01) (0.10) (0.80) (0.95) (0.82)
USA 0.96 0.91 0.83 0.90 0.91 0.94
(0.64) (0.69) (0.60) (0.81) (0.84) (0.91)
P-valucs shown in parenthcscs for the null hypothcsis that the variance ratio cquals 1, where thc nuncrator is the variancc of k-month retums and thc denominator is k timcs thc variancc of 1-month
rcturns. If-returns follow a random walk (i.e., no rctum autocorrelation), thcn return variance is a linear function of horizon lcngth: thc varianrc of returns overk periods is k times the variance ofreturns
ovcr one-period. If returns arc positively autocorrclated, the variance of k-period rcturns is larger than thc sum of onc-period returns-variances grow faster than lincarly. Thus, variance ratios larger than
onc are consistent with rational positivc momcntun trading. Alternatively, when returns arc negatively autocorrclated, the variancc of k-period rcturns is smaller than k times thc variancc of onc-period
returns. Variance ratios smaller than onc would call for ncgative momcntum (or contrarian) trading. Samplc: monthly index returns from January 1975 to October 1998.
31
Table Al
Regression Results: Whole Fund
( Q )J = a + PRj,, + flRj,, + P3RLA, + /34RLM J + (35RUSJ + -, jf
Independent Variables All Sample Non-Crisis Crisis
Own Return (P,) 0.0021*** 0.0029*** 0.00150*t
T-statistic 5.34 4.04 2.76
Own Return Lagged (,&) 0.0029+** 0.00420** 0.0002
T-statistic 3.15 5.501 0.138
Latin America Retum (03) 0.0035*5* 0.0021 0.0041***
T-statistic 3.20 1.49 2.62
Local Index Return (0) 0.0000 -0.0003 0.0006
T-statistic -0.01 -032 0.44
US Return (a) -0.0065*** -0.0041w .0096%**
T-statistic -6.24 -1.95 4.54
Constant -0.0048 -0.0086 0.0026
T-statistic -024 -030 0.05
Observations 4,842 3,223 1,619
Adjusted R-squared 0.05 0.03 0.06
These results are "Whole Fund" in that they include no control for investor redemption effects as in equation (7). T-statistics are
contected for heteroskedasticity across funds. Full sample: April 1993 to January 1999. The crisis portion of the sample is December
1994-June 1995, July 1997-March 1998, August 1998-October 1998, and January 1999. Thenon-crisis portion is the rest of the sample.
For robustness, results in each cell are based only on observations within three standard deviations of the mean.
* Statistically Significant at the 10-percent level
** Statistically Significant at the 5-percent level
** Statistically Significant at the 1-percent level
32
Table A2
Regression Results: Manager Only
(Q _Q ~~~(Qiij,- Qi,j,,l )Tj,,
Ql iJ,t JIA i,l j ] = - 11Rj,, + P,Rjtt + 02Rj,t I33RLA,t + 4RLM,t + I5RUs,t + i,i,t
ji
Independent Variabks AD Sample Non-Crisis Crisis
Own Return (P,) 0.0042*** 0.0052*** 0.0033***
T-scatistic 5.56 3.76 2.76
Own Return Lagged (N) 0.0049*** 0.0071** 0.0004
T-statistic 2.78 4.04 0.17
Latin America Return (N) 0.0014 -0.0001 0.004#**
T-statistic 1.17 -.44 3.46
Local Index Return (34) -0.0004 -0.0011 0.0011
T-statistic -0.328 -0.55 0.48
US Return (Ps) -0.0115*## -0.0063**# -0.0196***
T-statistic -6.54 -2.90 -7.018
Constant 0.0108 -0.0155 0.0970**
T-statistic 0.326 -0.40 2.53
Observations 4,942 3,274 1,668
Adjusted R-squared 0.03 0.02 0.05
These results are "Manger Only" in that they control for investor redenption effects as in equation (8). T-statistics are corrected for
heteroskedasticity across funds. FuU sample: April 1993 to January 1999. The crisis portion of the sample is December 1994-June 1995,
July 1997-March 1998, August 1998-October 1998, and January 1999. The non-crisis portion is the rest of the sample. For robustness,
results in each cell are based only on observations within three standard deviations of the mean.
* Statistically Significant at the 10-percent level
** Statistically Significant at the 5-percent level
" Statistically Significant at the 1-percent level
33
Table A3
Regression Results: Investor Only
S(QiJ,t - i,j, I [0i
S Qi,x,tPjJ = a-+ ,Rj,, + 132R1,-1 + fi3RLA,+ + 04RL I+ -fsRus5R + Fejj/
jei
Indendent Variables Al Sample Non-Crisis Crisis
Own RetLurn (,) -0.0001 0.0001 -0.0001
T-statistic -0.84 1.13 -OA5
Own Return Lagged (,) 0.0007 0.0004 0.0018
T-statistic 1.36 0.76 1.69
Latin America Return (f3) 0.0037** 0.0039*** 0.0022*
T-statistic 6.62 4A45 1.66
Local Index Retum s 04) 0.0011"** 0.0012*** 0.0005
T-statistic 3.83 5.02 0.94
US Return (Ps) -0.0021 -0.0034 0.0013
T-statistic -1.26 -1A8 0.33
Constant 0.0094 0.0247 -0.0562
T-statistic 0.52 0.94 -1.18
Observations 4790 3241 1549
Adjusted R-squared 0.29 0.21 0.19
These results are "Investor Only" in that they reflect only investor redermption effects as in equation (9). T-statistics are corrected for
heteroskedasticity across funds. Full sample: April 1993 to January 1999. The crisis portion of the sample is December 1994-June 1995,
July 1997-March 1998, August 1998-October 1998, and January 1999. The non-crisis portion is the rest of the sample. For robustness,
results in each cell are based only on observations within three standard deviations of the mean.
Statistically Significant at the 10-percent level
** Statistically Significant at the 5-percent level
Statistically Significant at the 1-percent level
34
References
Asness, C., J. Liew, and R. Stevens, 1997, "Parallels Between the Cross-Sectional
Predictability of Stock and Country Retums," Journal of Portfolio Management, 3:
79-87.
Bohn, H., and L. Tesar, 1996, "US Equity Investment in Foreign Markets: Portfolio
Rebalancing or Return Chasing?" American Economic Review, 86: 77-81.
Brown, S., W. Goetzmann, and J. Park, 1998, "Hedge Funds and the Asian Currency Crisis
of 1997," NBER Working Paper 6427, February.
Calvo, G., 1999, "Contagion in Emerging Markets: When Wall Street Is a Carrier,"
University of Maryland working paper.
Calvo, G., and E. Mendoza, 1998, "Rational Herd Behavior and the Globalization of
Securities Markets," University of Maryland working paper.
Campbell, J., A. Lo, and C. MacKinlay, 1997, The Econometrics of Financial Markets,
Princeton University Press, Princeton, New Jersey.
Choe, H., B. Kho, and R. Stulz, 1999, "Do Foreign Investors Destabilize Stock Markets?
The Korean Experience in 1997," typescript, Ohio State University, January.
Corsetti, G., P. Pesenti, and N. Roubini, 1998, "What Caused the Asian Currency and
Financial Crisis?" typescript, New York University, March.
Corsetti, G., P. Pesenti, N. Roubini, and C. Tille, 1998, "Structural Links and Contagion
Effects in the Asian Crisis: A Welfare Based Approach," New York University
working paper.
De Bondt, W., and R. Thaler, 1985, "Does the Stock Market Overreact?" Journal of
Finance, 40: 793-805.
Eichengreen, B., and D. Mathieson, 1998, "Hedge Funds and Financial Market Dynamics,"
Occasional Paper No. 166.
Eichengreen, B, A. Rose, and C. Wyplosz, 1996, "Contagious Currency Crises," NBER
working paper No. 5681.
Forbes, K., and R. Rigobon, 1998, "No Contagion, Only Interdependence: Measuring Stock
Market Co-movements," NBER Working Paper #7267, September.
Frankel, J., and S. Schtmukler, 1998, "Crisis, Contagion, and Country Funds," in R. Glick,
ed., Managing Capital Flows and Exchange Rates (Cambridge University Press).
Froot, K., P. O'Connell, and M. Seasholes, 1998, "The Portfolio Flows of International
Investors, I," typescript, April.
Glick, R., and A. Rose, 1998, "Contagion and Trade: Why Are Currency Crises
Regional?" NBER Working Paper 6806, November.
Grinblatt, M., S. Titman, and R. Wenmers, 1995, "Momentum Investment Strategies,
Portfolio Performance, and Herding: A Study of Mutual Fund Behavior," American
Economic Review, 85, 1088-1105.
Jegadeesh, N., 1990, "Evidence of Predictable Behavior of Security Retums," Journal of
Finance, 45: 881-898.
Jegadeesh, N. and S. Titman, 1993, "Returns to Buying Winners and Selling Losers:
Implications for Stock Market Efficiency," Journal of Finance, 48, 1.
35
Kaminsky, G., 1998, "Currency and Banking Crises: The Early Warnings of Distress,"
International Finance Discussion Paper No. 629, Board of Governors of the Federal
Reserve System.
Kaminsky, G. and C. Reinhart, 1999, "On Crises, Contagion, and Confusion," George
Washington University Working Paper, forthcoming Journal of International
Economics.
Kaminsky, G. and S. Schmukler, 1999, "What Triggers Market Jitters? A Chronicle of the
Asian Crisis," Journal of International Money and Finance, Vol. 18, No. 4.
Kim, W., and S. Wei, 1999, "Foreign Portfolio Investors Before and During a Crisis,"
NBER Working Paper No. 6968, February.
Kodres, L., and M. Pritsker, 1999, "A Rational Expectations Model of Financial
Contagion," typescript; International Monetary Fund and Board of Governors of the
Federal Reserve System, May.
Lehmann, B., 1990, "Fads, Martingales, and Market Efficiency," Quarterly Journal of
Economics, 105: 1-28.
Marcis, R., S. West, and V. Leonard-Chambers, 1995, "Mutual Fund Shareholder
Response to Market Disruptions," Perspective, Investment Company Institute, Vol. 1,
No. 1.
Rea, J., 1996, "U.S. Emerging Market Funds: Hot Money or Stable Source of Investment
Capital?" Perspective, Investment Company Institute, Vol. 2, No. 6.
Rouwenhorst, Geert, 1998, "International Momentum Strategies," Journal of Finance, 53,
267-284.
Tesar, L., and I. Werner, 1994, "International Equity Transactions and U.S. Portfolio
Choice," in J. Frankel (ed.), The Internationalization of Equity Markets, University
of Chicago Press, 185-220.
Valdes, Rodrigo, 1998, "Emerging Markets Contagion: Evidence and Theory," Banco
Central de Chile, typescript, May.
Warther, V., 1995, "Aggregate Mutual Fund Flows and Security Returns," Journal of
Financial Economics, 39: 209-235.
Wermers, R., 1999, "Mutual Fund Herding and the Inpact on Stock Prices," Journal of
Finance, 54: 581-622.
36
Policy Research Working Paper Series
Contact
Title Aulhor Date for paper
WPS2378 Disintegration and Trade Flows: Simeon Djankov June 2000 R. Vo
Evidence from the Former Soviet Caroline Freund 33722
Union
WPS2379 India and the Multilateral Trading Aaditya Mattoo June 2000 L. Tabada
System after Seattle: Toward a Arvind Subramanian 36896
Proactive Role
WPS2380 Trade Policies for Electronic Aaditya Mattoo June 2000 L. Tabada
Commerce Ludger Schuknecht 36896
WPS2381 Savings and the Terms of Trade Pierre-Richard Agenor June 2000 T. Loftus
under Borrowing Constraints Joshua Aizenman 36317
WPS2382 Impediments to the Development and Thorsten Beck June 2000 E. Mekhova
Efficiency of Financial Intermediation 85984
in Brazil
WPS2383 New Firm Formation and Industry Thorsten Beck June 2000 E. Mekhova
Growth: Does Having a Market- or Ross Levine 85984
Bank-Based System Matter?
WPS2384 Are Cost Models Useful for Telecoms Daniel A. Benitez July 2000 G. Chenet-Smith
Regulators in Developing Countries? Antonio Estache 36370
D. Mark Kennet
Christian A. Ruzzier
WPS2385 The Rise, the Fall, and ... the Antonio Estache July 2000 G. Chenet-Smith
Emerging Recovery of Project John Strong 36370
Finance in Transport
WPS2386 Regulators and the Poor: Lessons Richard Green July 2000 G. Chenet-Smith
from the United Kingdom 36370
WPS2387 The Long and Winding Path to Private Antonio Estache July 2000 G. Chenet-Smith
Financing and Regulation of Toll Manuel Romero 36370
Roads John Strong
WPS2388 The Role of Special and Differential Constantine Michalopoulos July 2000 L. Tabada
Treatment for Developing Countries in 36896
GATT and the World Trade Organization
WPS2389 Vietnam: On the Road to Labor- Patrick Belser July 2000 H. Sutrisna
Intensive Growth? 88032
WPS 2390 The Social Rate of Return on David Canning July 2000 H. Sladovich
Infrastructure Investments Esra Bennathan 37698
Policy Research Working Paper Series
Contact
Title Author Date for paper
WPS2391 Are the Poor Protected from Budget Martin Ravallion July 2000 P. Sader
Cuts? Theory and Evidence for 33902
Argentina
WPS2392 What Factors Appear to Drive Private Dipak Dasgupta July 2000 S. Crow
Capital Flows to Developing Countries? Dilip Ratha 30763
And How Does Official Lending
Respond?
WPS2393 Will the Euro Trigger More Monetary Patrick Honohan July 2000 A. Yaptenco
Unions in Africa? Philip R. Lane 31823
WPS2394 Tax Evasion, Corruption, and the Waly Wane July 2000 H. Sladovich
Remuneration of Heterogeneous 37658
Inspectors
WPS2395 Decentralizing the Provision of Health William Jack July 2000 H. Sladovich
Services: An Incomplete Contracts 37698
Approach
WPS2396 Aid Dependence and the Quality of Stephen Knack July 2000 P. Sintim-Aboagye
Governance: A Cross-Country 38526
Empirical Analysis
WPS2397 Verifying Exchange Rate Regimes Jeffrey Frankel July 2000 E. Khine
Eduardo Fajnzylber 37471
Sergio Schmukler
Luis Serven
WPS2398 Determinants of Current Account Cesar Calder6n July 2000 H. Vargas
Deficits in Developing Countries Alberto Chong 38546
Norman Loayza