An Examination of the Asian Crisis Part II:
Information Spillover, Herding, and Regime Shifts
Jarl G. Kallberg, Crocker H. Liu, and Paolo Pasquariello*
Stern School of Business - New York University Revision April 18, 2002
Abstract
We study the impact of capital flights on the interaction of currency and equity markets in Indonesia, the Philippines, Malaysia, South Korea, Taiwan and Thailand during the Asian crisis using the sequence of regime shifts in reduced-form linear relations between currency and equity returns, and between currency and equity return volatility estimated non-parametrically by Kallberg, Liu, and Pasquariello (2002). We relate those structural breaks to data on net monthly flow of funds from foreign investors in each of the equity markets in our sample. We test for whether the timing of those regime shifts implies a causal relationship across countries, and whether such relationship can be explained by common information shocks, portfolio rebalancing or “herding” phenomena. Our results suggest that portfolio rebalancing was a major channel for the transmission of information shocks across East Asia in 1997, but not in 1998. We provide evidence that information spillover and herding effects, rather than common information shocks, generate the propagation of the estimated structural shifts from a country to another during the Asian crisis. In particular, we find that the highest degree of herding occurred in 1998, when most of the breaks in the reduced-form return relations were observed. We also show that, when signed herding was more intense, i.e., when more equity markets were simultaneously experiencing funds’ withdrawals, the expected number of return and return volatility break events per period was bigger, and the clustering of regime shifts across countries was more likely. JEL classification: C51; G15
Keywords: Regime shifts; Asian crisis; Information spillover; Exchange rates; Herding
*Associate Professor of Finance, Associate Professor of Finance and Ph.D. candidate respectively at the Stern School of Business. We thank an anonymous referee for many suggestions that have substantially improved this paper We are also grateful to Viral Acharya, Edward Altman, Yakov Amihud, Robert Dittmar (our discussant in Bloomington), Tony Kao, Kanak Patel (our discussant in Maui), Bertrand Renaud, Vijay Singal (our discussant in New Orleans), and other seminar participants at the 1999 AREA/ AREUEA conference in Maui, the 1999 NYU Conference on Risk and Return Management for Insurance Companies, the 2000 Indiana University 6th Biennial Conference on Financial Crises, the 2001 AFA Annual Meeting in New Orleans, and the 2002 EFMA Annual Meeting in London. Please address comments to the authors at the Leonard N. Stern School of Business, New York University, Kaufman Management Education Center, Suite 9-190, 44 West 4th Street, New York, NY 10012-1126, or through email: [email protected], [email protected] and [email protected], respectively. The usual disclaimer applies.
2
An Examination of the Asian Crisis Part II:
Information Spillover, Herding and Regime Shifts
Abstract
We study the impact of capital flights on the interaction of currency and equity markets in Indonesia, the Philippines, Malaysia, South Korea, Taiwan and Thailand during the Asian crisis using the sequence of regime shifts in reduced-form linear relations between currency and equity returns, and between currency and equity return volatility estimated non-parametrically by Kallberg, Liu, and Pasquariello (2002). We relate those structural breaks to data on net monthly flow of funds from foreign investors in each of the equity markets in our sample. We test for whether the timing of those regime shifts implies a causal relationship across countries, and whether such relationship can be explained by common information shocks, portfolio rebalancing or “herding” phenomena. Our results suggest that portfolio rebalancing was a major channel for the transmission of information shocks across East Asia in 1997, but not in 1998. We provide evidence that information spillover and herding effects, rather than common information shocks, generate the propagation of the estimated structural shifts from a country to another during the Asian crisis. In particular, we find that the highest degree of herding occurred in 1998, when most of the breaks in the reduced-form return relations were observed. We also show that, when signed herding was more intense, i.e., when more equity markets were simultaneously experiencing funds’ withdrawals, the expected number of return and return volatility break events per period was bigger, and the clustering of regime shifts across countries was more likely.
JEL classification: C51; G15
Keywords: Regime shifts; Asian crisis; Information spillover; Exchange rates; Herding
3
I. Introduction
The Asian crisis was precipitated by large-scale capital outflows from Southeast Asian
economies, prompted by concerns about over-borrowing and mounting debt burdens. The
grounds of the crisis were laid mid-decade, when there was a substantial increase in the flow
of capital to liberalizing Asian markets without a commensurate improvement in their
capacity to manage these resources.
United Nations (1998; p.2)
The Asian currency crisis of 1997-1998 began when Thailand abandoned its fixed peg
currency regime and devalued the baht on July 5, 1997. The resulting shocks to Asian and
world economies have been the subject of an intense amount of discussion and analysis,
without many clear conclusions emerging. The reasons for its intensity, the role of foreign
investment, the manner in which the crisis propagated across the economies of Asia, and its
long-term impact continue to be discussed. However, as the above quote suggests, the most
frequently mentioned culprit for the crisis is the destabilizing effect of the flight of foreign
capital. In particular, much popular press blames the “excessive speculation” and rapid
withdrawal of capital by hedge funds for the events of 1997 and 1998. Although many
academic studies suggest that portfolio rebalancing may have represented a major channel
for transmitting information shocks across markets during the Asian crisis, the available
empirical evidence is inconclusive.
This paper attempts to clarify some important aspects of the crisis through a more formal
analysis of foreign exchange markets, equity markets and flows of capital in Indonesia,
Malaysia, the Philippines, South Korea, Taiwan and Thailand. We select these six countries
because they experienced the greatest extent of economic and financial turmoil during 1997
and 1998. Furthermore, their domestic monetary authorities maintained similar currency
management policies at the time of the crisis. Each of these countries was in fact pegging its
currency to the U.S. dollar (or to a basket of currencies in which the U.S. dollar dominated)
before the crisis swept through the region. This peg was fixed in the case of Thailand; the
other countries allowed the rate to fluctuate within pre-specified, relatively narrow bands.
The combination of such protection against currency devaluations and the rapid growth
experienced by these economies in the 1990s led to rapid inflows of foreign capital. These
flows of capital declined, first in 1997 during the Asian currency crisis, then again in 1998
4
with the devaluation of the ruble and the collapse of many major hedge funds, including
Long Term Capital Management (LTCM).
The financial devastation of those once thriving markets during the second half of 1997
was remarkable. Over the interval from July to December of 1997, the domestic equity
indexes for Indonesia, Malaysia, the Philippines, South Korea, Taiwan and Thailand dropped
by 37%, 52%, 41%, 42%, 12% and 56%, respectively. The corresponding local currencies
depreciated by 56%, 35%, 36%, 46%, 17% and 46%, respectively, versus the U.S. dollar.
After July 1997, most of these countries reacted swiftly in revising their currency regimes.
Thailand abandoned its fixed exchange rate regime on July 2, 1997. Indonesia gave up the
enlarged band of currency fluctuation on August 14, 1997. On October 17, 1997, Taiwan
decided to adopt a floating exchange rate regime. Korea abandoned the defense of its
currency in November of 1997, as did Malaysia.
A companion paper, Kallberg, Liu, and Pasquariello (2002), tests for regime shifts in
reduced-form linear relations between currency and equity returns and between currency and
equity return volatility in the above mentioned Asian markets. In that paper, using a non-
parametric technique developed by Bai, Lumsdaine, and Stock (1998), we identify and
characterize regime shifts in each of these countries in 1997 and 1998 and argue that the
behavior and interaction of those currency and equity markets during the Asian crisis
resulted from significant structural breaks, as opposed to being merely a manifestation of
natural dependencies. In our data, shocks to volatility lead shocks to returns; while regime
shifts in returns occur almost a year after the currency crisis, volatility breaks take place
around the crisis period. We further show that the sequential nature of breaks in the
structural relation between currency and equity market return and volatility market returns
and volatility is consistent with information spillover effects.
Figures 1 and 2 summarize the main findings in Kallberg et al. (2002) and motivate our
analysis. There we display net monthly flow of funds from foreign investors in each equity
market for Indonesia, Malaysia, the Philippines, South Korea, Taiwan and Thailand (Figures
1A and 2A), and confidence intervals at the 5% level around those estimated break dates �k ,
ranked in order of increasing statistical significance, over two time intervals, March to
December 1997 and January to December 1998, respectively. These two intervals
5
correspond to most of the observed break-dates for equity versus currency volatility (Figures
1C and 2C) and equity versus currency returns (Figures 1B and 2B).
The observed pattern in the flows of funds for 1997 (in Figure 1A) is striking. Intense
portfolio rebalancing activity took place during the selected time period. While Malaysia
suffered the most significant outflows, countries such as South Korea, the Philippines and
Thailand experienced inflows of capital for most of that summer. Portfolio rebalancing
appears to be a major channel by which information shocks were transmitted across markets
during 1997. The picture for 1998 (Figure 2A) suggests a different interpretation. Most of
the countries in our sample experienced significant contemporaneous inflows and outflows
of capital at that time, which is when we identify structural breaks in the hypothesized return
relations. In this case, Figure 2A seems to suggest that portfolio reallocation efforts affected
the entire region during 1998, although with a different intensity for each market, i.e., that
herding may have been a driving force for the events of that year.
The timing information on the volatility break events contained in Figures 1C and 2C is
especially significant for this paper. Such evidence suggests not only that breaks in volatility
anticipated breaks in returns but also that volatility shocks in one country may have affected
the posited relations between equity and currency markets in other countries during the
crisis. To test for this possibility, we estimate the impact of the break event in country i at
time t on the reduced-form volatility relation of other countries. If information spillover
were at least partially responsible for the propagation of the event from a country to another,
we would expect the statistical power of such structural relations to improve from the
inclusion of dummies for break dates in other countries. We find some evidence of
information spillover across Indonesia, the Philippines and South Korea between 1997 and
1998, when past regime shifts in the linear relation between their domestic equity and
currency return volatility appear to have induced a statistically and economically significant
decline in equity returns.
This paper links those estimated regime shifts to the movements of foreign capital. In
particular, we seek to gain some insight into how the actions of international investors,
specifically capital flight, influenced the structure of Asian currency and equity markets
around the time of the crisis. We start by asking whether the observed sequence of regime
breaks described in Figures 1 and 2 implies a causal relation across countries. We then test
6
for whether the observed sequence of structural breaks can be explained by portfolio
rebalancing or the “herding” behavior of foreign investors. Finally, we develop a Poisson
regression model to analyze whether changes in the flows of funds were responsible for the
observed clustering in regime shifts.
Our analysis of flows of funds data provides further evidence that information spillover
or herding effects generate the observed return and volatility shocks among countries. We
show that portfolio rebalancing was a major channel for transmitting information shocks
across markets during 1997, but not in 1998 (when the highest degree of herding occurred,
and when most of the breaks in returns were instead observed). We also demonstrate that
the expected number of return and return volatility break events per period was greater, and
the clustering of those break events more likely, when signed herding was more intense (i.e.,
when more equity markets were simultaneously experiencing withdrawals).
The organization of the paper is as follows: Section II briefly reviews the main economic
explanations for the turmoil of 1997 and 1998. Section III illustrates the statistical
methodology. Section IV describes the data set, summarizes the findings of Kallberg et al.
(2002) that we use in this paper, and presents our first set of results. In Section V we study
the relation between herding behavior and cross-country spillover. Section VI concludes.
II. The economics of the Asian crisis
… any country that adopts a fixed exchange rate for stabilization reasons needs to
have a well-thought-out exit strategy. In the crisis year of 1997, the widely held implicit
strategy of waiting until a speculative attack forced a devaluation caused an overshooting of
currency depreciation and significantly lowered stock returns.
Grier and Grier (2001; p. 146)
While there is general agreement among academics and practitioners that the Asian crisis
was triggered by the devaluation of the Thai baht in July of 1997, there is less agreement
about the relative significance of the many factors that contributed to its severity.1 In this
section we briefly overview these explanations and the related economic literature. The
major contributing factors discussed here are: (i) the role of currency regimes, (ii) the
1For a further discussion of the causes of the Asian crisis, refer to Gangnes (1998) and Wu (1998).
7
interaction of foreign investment and domestic current accounts and (iii) macroeconomic
factors.
(i) Currency regimes
A major factor in the turmoil that affected the six countries in our sample was the
adoption by each of them of a pegged currency regime. This effectively virtually eliminated
the currency risk for foreign investors, but at the same time left these countries more
exposed to economic shocks or the actions of speculators than countries adopting a floating
exchange rate regime.2 The literature has argued that such circumstance may help explain the
occurrence of financial crises. Supportive empirical evidence on the adverse effects of
pegging the domestic currency can be found in Grier and Grier (2001) for 25 developing
countries in 1997, and in an earlier study of African economies by Tornell and Velasco
(1995).3 In particular, Grier and Grier (2001) use a macroeconomic model to show that the
nations in their sample that adopted pegged currency regimes suffered more than what could
have been anticipated from their fundamentals.
(ii) Foreign investment and current account balance
These implicit guarantees on the value of the local currency and the remarkable growth
rates this region experienced in the first half of the 1990s drew vast amounts of foreign
capital into the region. In 1996, the amount of foreign investment in the developing markets
of Asia totaled $56 billion.4 This pool of capital encouraged over-investment. Investment in
South East Asia countries (measured as fixed capital formation as a percent of GDP) grew
more quickly than GDP, inducing persistent deficits on the current account. In Thailand,
investment accounted for 40% of GDP from 1990 to 1996, while GDP increased by 6.4%
over this period. Over-investment in speculative real property (particularly in Thailand,
Malaysia, and Indonesia) put stress on the domestic financial systems, given the fast pace of
Asian financial market liberalization, weak regulation of the financial sector and insufficient
quality of lending oversight. Corsetti, Pesenti, and Roubini (1998) report that Thailand,
2 For example, in the first nine months of 1997, the Bank of Thailand spent an estimated $10 billion in foreign exchange reserves in an effort to defend the baht (Time, 8/25/1997). 3 Kaminsky, Lizondo, and Reinhart (1998) survey the empirical literature relating to currency crises. 4 See Van Wincoop and Yi (2001) for detailed figures and a discussion of the impact on the U.S. and Europe of foreign capital leaving Asia in 1997 and 1998.
8
Indonesia, and the Philippines experienced growth rates in bank credit in excess of 20% per
annum. Growth rates in Korea and Malaysia were only slightly lower. At the same time,
estimates of non-performing loans prior to the crisis were significant, ranging from 14% in
the Philippines to 19% in Thailand.
Since in many of these countries the foreign reserves were inadequate, foreign capital was
eventually used to fund deficits on their current account balances. Compounding these
problems was the fact that much of the foreign capital was short term in nature. This “hot”
capital is very sensitive to changes in the international economic environment. Hence, its
sudden withdrawal, as we will explore later, can disrupt an economy, especially when these
funds were used to finance long-term projects.
Furthermore, the analysis in many academic studies (e.g., Fleming, Kirby, and Ostdiek
(1998), Kaminsky, Lyons, and Schmukler (2001), and Kodres and Pritsker (2002)) also
suggests that portfolio rebalancing may have represented a major channel for transmitting
information shocks across markets during the Asian crisis. The paper by Fleming et al.
(1998) is particularly relevant for this research, as it investigates how volatility linkages of
two distinct types, a common information shock or information spillover, may result in
investors rebalancing their holdings in other markets in response to a change in the hedging
component of their demand for one market. Brown, Goetzmann, and Park (2000) address
the issue of whether the withdrawal of hedge fund capital has played an important role in the
Asian crisis, by analyzing the investment decisions of ten major currency funds during that
period. They show that changes in those funds’ positions were not highly correlated with
fluctuations in exchange rates. Lin and Kuo (2000) argue that the actions of international
hedge funds, although central to the Thai baht crisis, were not a major force in the currency
turmoil involving Indonesia, Malaysia and the Philippines. These results thus suggest that the
importance of hedge funds’ activity as a catalyst of the crisis may have been overstated.
(iii) Macroeconomic factors
As interest rates rose in the U.S. and the outlook for European markets improved, the
flow of foreign capital began to contract. Additionally, almost all project returns were
denominated in local currency, while most of the debt financing those investments was
dollar-denominated. These mismatches, together with limited forward markets, weakened
9
the domestic banking sectors. The gradual rise in the dollar in mid-1995 also reduced
exports, since the region’s exchange rates were pegged to the U.S. dollar, albeit to varying
degrees. Such currency arrangements also led to severe losses of foreign reserves, as in
Taiwan between June and October of 1997, in the Philippines in July 1997, and in South
Korea later that year. The stress from these factors, combined with the high debt burdens
culminated in the second half of 1997.5 The high interest payments now become
unsustainable as interest rates rose, the value of the domestic currency plummeted and the
values of real assets tumbled.
Kallberg et al. (2002) identify and characterize such potential interaction between equity
and currency returns, and equity and currency return volatility around the time of the Asian
crisis. In the remainder of this paper, we use their chronology of regime shifts in reduced-
form relations between currency and equity markets to test for whether the resulting
sequence of structural breaks implies a causal relationship across countries, and for whether
such regime shifts and relationships can be explained by the dynamics of flows of funds
across countries, i.e., by portfolio rebalancing or “herding” by foreign investors.
III. Methodology
Testing for regime breaks
In this section we briefly describe the empirical method used in Kallberg et al. (2002) to
test for the existence, timing and intensity of shocks in a set of linear reduced-form structural
relations between equity and currency returns, and equity and currency volatility for each of
the countries in our sample. Kallberg et al. adopted a statistical methodology originally
devised by Bai, Lumsdaine, and Stock (1998) to make statistical inference about regime
breaks, including interval estimation of the break date, with minimal restrictions on the
underlying data generation process (DPG). In particular, Bai et al.’s non-parametric
technique searches for a single break in univariate or multivariate time series models (with or
without stationary regressors) assuming only that the DPG is a stationary VAR before and
after the break, and specifies asymptotic confidence intervals for the estimated break point.
5 See Furman and Stiglitz (1998) for an analysis of the role of external debt during the Asian crisis.
10
More specifically, the reduced-form model for returns uses month t equity index returns
denominated in local currency as the dependent variable. The independent variables are the
currency returns in each of the five months t – 2 to t + 2 (to account for possible lead-lag
relations) and the equity return in month t=− 1 (to account for first-order autocorrelation in
returns). We have the following equation:
ti i
itittititt xykdxbAyy εβαλµ +�
��
�+++++=
= =−+−−+−
5
1
5
13131 )( , ( 1 )
with dt(k) equal to 1 if t is greater or equal to k, and zero otherwise. Here, k is a potential
break date, yt is the equity index return for a certain country in month t, and xt is the
corresponding currency return in month t.
Kallberg et al. (2002) also focuses on the second moments of returns, by extending the
basic model of Eq. (1) to the case of equity and currency volatility. We in fact search for
breaks in a structural relation between monthly rolled volatility for the available series of
equity returns and the corresponding monthly rolled volatility series for currency returns. We
calculate the rolled volatility as a moving standard deviation of 12 monthly returns for the
equity (σy) and currency (σx) index returns. The proposed model is then
( )[ ] txttxtytyt kdbA εβσλσσµσ +++++= −1 . ( 2 )
Many of the approaches proposed in the literature to test for a break in any of the
parameters in Eqs. (1) and (2) are based on Wald statistics. Bai et al.’s test, which is similar to
that of Quandt (1958, 1960), identifies a break date �k as the value of k that maximizes an F-
statistic � ( )F k . The date �k is then statistically significant if )ˆ(ˆ kF is greater than the
corresponding critical value (for the selected significance level). This technique also allows us
to construct confidence intervals for the estimated break dates assuming only that the
disturbances form a sequence of martingale differences.
In Kallberg et al. (2002), we used two monthly time series, currency returns and equity
returns for Indonesia, Malaysia, the Philippines, South Korea, Taiwan and Thailand, and the
specification of our model described above, to estimate break dates for volatility and return
shocks, i.e., for breaks in the linear reduced-form models of Eqs. (1) and (2). We employ the
results of that analysis in this research to examine whether the ensuing chronology of break
11
dates provides enough evidence of causal relationships across countries around the time of
the Asian crisis, i.e., whether information spillover from one country to another was
occurring during 1997 and 1998.
Herding measure
In this paper we investigate the potential causes for the information effects observed
during the Asian crisis, and described in Kallberg et al. (2002). One popular explanation is
the herding behavior of foreign investors. Capital flight, in particular the widespread
withdrawal of foreign funds from emerging equity markets, appears to have played an
important role in the Asian crisis, especially in creating pressure on the domestic currencies
and interest rates.6
The concept of herding comes into play when different markets simultaneously (or with
only a slight lag) experience flows of foreign capital in the same direction. Herding
intensifies the transmission of single-market information shocks through geographically or
fundamentally heterogeneous markets. In our analysis, inter-country herding can arise
because of distinct, yet related reasons, as in Wermers (1999). First, managers might mimic
each other’s behavior, disregarding their private information, to avoid jeopardizing their
reputation by trading differently from other managers who are acting in apparently
interrelated markets. Second, as suggested earlier, managers might make country allocation
decisions simultaneously. This could happen either because the original information shock
affecting a single market is correlated to the information sets of each of the other markets, or
because a common information shock affects all the information sets at the same time.
Third, managers could infer private information from observing trading in geographically
heterogeneous markets. Finally, institutional investors trading in different markets might
need to rebalance their aggregate holdings in these markets as a result of global portfolio
reallocation. Measuring the existence and intensity of the herding effect during the Asian
crisis is another important objective of this study.
We use a measure developed by Lakonishok, Shleifer and Vishny (1992) and adopted by
Wermers (1999). Choe, Kho, and Stulz (1999) use a similar criterion to analyze herding
6 In 1997, the Prime Minister of Malaysia, Dr. Mahathir Mohamed, called hedge fund managers the “highwaymen” of the global economy (Chancellor (2000)).
12
behavior in the Korean stock market in 1997. The measure assumes a null hypothesis that
observed herding behavior is due purely to chance. The herding measure for a given month
t, H(t), is defined as
})]([)({)]([)()( **1 tbEtbEtbEtbtH t −−−= − , ( 3 )
where b(t) is the number of countries in which we observe negative flows of funds divided
by the total number of countries for which flows of funds were available at time t.
Eq. (3) implies that either sign of H(t) is equally likely. The proxy for Et-1[b(t)], the
expected proportion of outflows during a given month, is the simple average of the observed
b(t) during the past quarter. We calculate the adjustment term E{|b*(t) − E[b*(t)]|} under
the null hypothesis that herding is only observed as a result of random chance. Hence, b*(t)
is the proportion of negative flows that we would observe by drawing xt, the number of
observed outflows at time t, from a binomial distribution B(xt, nt), where the probability of a
single sell outcome is one half, and nt is the number of countries in the sample for which
flows of funds data were available at time t. Since xt follows a binomial distribution with
probability b*(t) of success, E{|b*(t) − E[b*(t)]|}, is then easily calculated given b*(t) and nt.
Our final measures of herding are generated as averages of H(t) over selected time intervals.
When we analyze the herding measure for the chosen sample sub-periods, we can test for
systematic patterns in the aggregate behavior of fund managers who invest in the Far East.
Any significant value of H(t) signals that international investors tend to trade in different
markets together, at the same time, and in the same direction, more often than would be
expected by random and independent trading.
IV. Empirical results
Data
We use three monthly time series: currency returns, equity returns, and flow of funds, in
our analysis of Indonesia, Malaysia, the Philippines, South Korea, Taiwan, and Thailand.
We assume that the exchange rate of each local currency relative to the dollar is the key
exchange rate variable. Therefore, for the time series of monthly currency returns, we use
the spot rates that correspond to the noon buying rate for cable transfers payable in foreign
13
currencies, as recorded by the Federal Reserve Bank of New York every business day. For
each of the six markets, we use the major equity index in local currency to calculate a
monthly time series of local returns. The six equity indexes are JCI Jakarta Composite Index,
EMAS Equity Index, PSE Index, Kospi 200, the Thailand Stock Exchange Index and the
TWSE Stock Index. Monthly equity indexes for each country are from Bloomberg. Table 1A
presents summary statistics for each country. Sample periods are of different length, due to
data availability constraints, but they all end in March 1999.
To measure the extent to which international portfolio managers reallocate their wealth
across the region, we use time series data on the net monthly flow of funds in each equity
market. The flow is the difference (in millions of U.S. dollars) between total foreign
purchases and sales of domestic equities in each month. Flow of funds data are collected by
local stock market authorities, and were made available to us by a major investment bank.
Sample periods in most cases overlap the ones for equity and currency data, and again end
all in March 1999. The flow of funds data in Table 1B shows great variation across the
sample, although the means are positive for each country.
Indonesia, Malaysia, the Philippines, South Korea, and Thailand arguably suffered the
greatest degree of economic, financial, and political turmoil during 1997 and 1998. More
importantly, those countries also shared a similar exchange rate regime at or around the time
of the Asian crisis, being all unilaterally pegged to the U.S. dollar by relatively tight bands of
fluctuations. Thailand abandoned its fixed exchange rate regime on July 2, 1997. Indonesia
gave up the enlarged band of currency fluctuation on August 14, 1997. The Indonesian
rupiah crashed soon afterwards. South Korea renounced battling the increasing selling
pressure on the won in November of 1997, and so did Malaysia.
The Philippines and Taiwan were also hit by the crisis, although less severely. In the
Philippines, the Central Bank was forced to relax its previously successful band of
fluctuations for the exchange rate by the end of July 1997. Taiwan, previously known as one
of the East Asian Tigers, was able at first to fend off expectations of increased depreciation,
through its Central Bank’s policy of active foreign exchange intervention. Therefore, the
country initially appeared to be unaffected by the economic instability undermining the rest
of the region. However, on October 17, 1997, Taiwan decided to adopt a floating exchange
rate regime, and the New Taiwan Dollar subsequently fell by 30% versus the U.S. dollar.
14
Many analysts and economists seem to agree that Taiwan’s economy paid a very high price
for the comparative resilience shown throughout 1997 and 1998, in particular in the
domestic financial markets, where the steep rise of short term interest rates and the loss of
foreign reserves were eventually accompanied by a severe decline of the stock market.
Despite the major devaluations of their domestic currencies, during this time both the
Philippines and Taiwan experienced significant, albeit extremely volatile portfolio
rebalancing activity. For instance, the average absolute flows of funds in and out of Taiwan’s
domestic equity market over the interval January 1997-March 1999 was second only to South
Korea. While many other Asian equity markets were plummeting, The TWSE Stock Index
was instead rising by 13.15%, to then fall by more than 21% in 1998. Investigating the
potential role of portfolio rebalancing in explaining these disparities and the occurrence of
potential “information spillover” events from a market to another is one of the objectives of
this research.
Analyzing regime shifts
Using the database described above, in a companion paper we estimate breaks in the
assumed structural relations between local equity returns and local currency returns (Eq. (1),
in Table 2A), and between the volatility of local equity returns and the volatility of local
currency returns (Eq. (2), in Table 2B).
Tables 2A and 2B reveal that, for three of the countries in our sample, Indonesia, the
Philippines and South Korea, breaks in the hypothesized reduced-form relations described in
Eqs. (1) and (2) do occur in 1997 and 1998, and are statistically significant at less than 5%
for both returns and return volatility. Indonesia and the Philippines, where the domestic
central banks strictly controlled the corresponding currencies and market regulations were
poor or nonexistent, experienced probably the greatest extent of political turmoil ensuing
from the Asian crisis. The troubles of South Korea surprised many investors and
commentators, in light of the enormous economic and social progress made by the country
in the previous fifty years. Regime shifts in the lead-lag relation between equity and currency
returns were registered for Malaysia, Thailand, and Taiwan as well. Nonetheless, for these
three countries breaks in return volatility cluster in the fall of 1994, during the Mexican Peso
crisis.
15
Furthermore, breaks in the hypothesized structural relations between equity and currency
return volatility appear to consistently anticipate the corresponding break event in the
returns’ relations for most of the countries in our sample. More specifically, in our sample of
Asian countries, past regime shifts in the linear relation between equity and currency return
volatility appear to have induced a statistically and economically significant decline in equity
returns, even after controlling for the occurrence of a break in the relation between equity
and currency returns of Eq. (1). This evidence is consistent with the results of French et al.
(1987) for the U.S. markets, and shows that, in Asia, volatility shocks influence future
returns.
Regime shifts and information spillover across countries
In our companion paper we argued not only that, during the events of 1997 and 1998, a
negative relationship between volatility and returns characterized many of the Asian
countries under investigations but also that the observed sequence of breaks (in Table 2)
appeared to imply a causal relationship from volatility to returns. An intuitive explanation of
this phenomenon is that any information shock that affects the volatility of a single market’s
equity index returns, the volatility of the exchange rate, or the relation between the two, may
eventually induce a shock in returns through the adjustment of the predicted component of
the index volatility itself.
Significant cross-country correlations could channel those information shocks from one
country into other markets, with market frictions, differences in economic fundamentals,
and constraints to hedging influencing the intensity of the spillover phenomenon.7 The
sequential nature of those estimated structural breaks suggests that information spillover
effects created return and volatility linkages between Asian markets. Information alters
expectations in one market and affects returns and volatility in other markets through
changes in hedging demand. The effect of these changes is protracted over time and delayed
by the existence of market frictions. The existence of such linkages in East Asia has often
been related to the availability of hedging and portfolio rebalancing opportunities among
7 The available empirical evidence on the state of the economies involved in the Asian crisis (e.g., Eichengreen, Rose, and Wyplosz (1996), Corsetti et al. (1998), or Barro (2001)) leads us to exclude the possibility that idiosyncratic shocks occurred in each of the countries for which the statistic F is statistically significant could explain the chronology of regime shifts in the reduced-form relations between equity and currency markets.
16
local equity markets around the time of the Asian turmoil. Indeed, the previously mentioned
studies of Fleming et al. (1998) and Kodres and Pritsker (2002) have argued that portfolio
rebalancing represented a major channel for transmitting information shocks across markets
during 1997 and 1998. In the remainder of this paper, we address the closely related issues
of whether the sequence of breaks reported in Tables 2A and 2B implies a causal
relationship across countries, and of whether cross-country spillover at or around the time of
the Asian turmoil can be attributed to herding behavior by foreign investors.
We start by using the timing information on the volatility break events contained in Table
2B to explain whether the observed sequence of structural breaks implies their transmission
across countries. Such evidence suggests that breaks in volatility anticipated breaks in
returns, and that volatility shocks affected future returns for most of the nations that have
experienced regime shifts in the posited relations between equity and currency markets
during the crisis. The timing of the structural breaks in volatility will allow us to test for the
existence of information spillover by estimating the impact of the break event in country i at
time t on the structural relation itself.8 From Table 2B we identify two sets of countries that
experienced a (quasi) contemporaneous volatility break event. The first set consists of
Thailand, Malaysia, and Taiwan in the fall of 1994. The second set comprises Indonesia,
South Korea and the Philippines in 1997-98. We then estimate the regression
( )[ ] ( ) tkdxtktdxtbytAytK
ityii i
εσαβσλσσµσ +++++−+==1
ˆˆ1 . ( 4 )
Eq. (4) tests whether the statistical power of Eq. (2) can be improved by including a
dummy for break dates in the other countries in the group. Thus, d( �k i) is equal to zero when
t is below the lower limit of the confidence interval for the volatility break registered in
country i, and one otherwise. If information spillover were at least partially responsible for
the propagation of the event from a country to another, we would expect the estimated αs
to be statistically significant. Table 3 reports the results of our analysis. To correct for
heteroskedasticity and autocorrelation in the residuals of the above regression, statistical
significance of the coefficients is measured using Newey-West standard errors. The
clustering of volatility breaks in the fall of 1994 does not seem to be characterized by
8 We are grateful to Robert Dittmar for suggesting this analysis of the break chronology.
17
information spillover. Indeed, all the coefficients for break events in other countries are
statistically indistinguishable from zero, except for the parameter measuring the propagation
of a quasi-contemporaneous volatility shock from Malaysia to Thailand (α1 in Table 4). The
results for the second set of countries are strikingly different. Many of the coefficients for
other countries’ volatility-adjusted break event dummies are significant at the 10% level or
less, in particular suggesting the occurrence of spillover from Indonesia to the reduced-form
volatility relations of the Philippines and (more weakly) of South Korea. Therefore, the
cross-country propagation variables in Table 3 appear to help explaining the estimated
sequence of volatility breaks occurring in Asia during 1997 and 1998 and the ensuing
statistically and economically significant declines in equity returns.
Prima facie, these findings seem to suggest that, after controlling for past regime shifts in
the linear relation between their domestic equity and currency return volatility, Malaysia,
Taiwan and Thailand in the fall of 1994 did not experience spillover of information from
other markets in the region, while volatility shocks to Indonesia affected the Philippines and
South Korea during the Asian crisis. However, the simultaneity of many of the spillover
dummies prevents us from being able to distinguish between a rapid propagation of
information innovations from a country to another and a common information shock, and
reduces the statistical power of the analysis of Eq. (4). Moreover, such dummies do not
explicitly account for one of the possible sources of spillover of information across
countries, portfolio rebalancing, nor do they distinguish it from the widespread withdrawal
of foreign funds that would accompany a common information shock. Therefore, in the
remainder of the paper we concentrate our attention on the analysis of flows of funds and
explore more directly their relationship to the observed sequence of regime shifts between
equity and currency markets across the Asian countries in our sample.
Analysis of flows of funds
To relate the timing of regime breaks in Asia to portfolio rebalancing activity, we employ
our flow of funds data and the estimated break dates in Tables 3 and 4.9 Figures 1 and 2
9 A word of caution on the use of such data is necessary. The difference between total foreign purchases and sales of domestic equities in each month accounts only for the trading activity of foreign investors in the domestic spot equity markets. Speculative forces can nonetheless deeply affect financial markets even while
18
display our monthly flows of funds data for Indonesia, Malaysia, the Philippines, South
Korea, Taiwan, and Thailand, and confidence intervals at the 5% level around those
estimated break dates �k , ranked in order of increasing statistical significance, over two time
intervals, March to December 1997 and January to October 1998 respectively. These two
intervals correspond to most of the observed break-dates for equity versus currency volatility
(Figures 1C and 2C) and equity versus currency returns (Figures 1B and 2B).
The observed pattern in the flows of funds for 1997 (in Figure 1A) is striking. Intense
portfolio rebalancing activity occurred during the selected time period. While Malaysia
suffered the most significant outflows, countries such as South Korea, the Philippines and
even Thailand experienced inflows of capital for most of that summer. These findings,
coupled with the evidence that most of the volatility shocks accompanying the Asian crisis
(in Table 2B) occurred between August and October 1997, are consistent with information
spillover effects. Portfolio rebalancing appears to be a major channel by which information
shocks were transmitted across markets during 1997. The picture for 1998 (Figure 2A) is not
as clear. Most of the countries in our sample experienced significant contemporaneous
inflows and outflows of capital at that time, with Malaysia again acting as the bellwether.
This is when our empirical model identifies structural breaks in the hypothesized return
relations. In this case, Figure 2A seems to suggest that portfolio reallocation efforts affected
the entire region during 1998, although with a different intensity for each market, i.e., that
herding may have been a driving force for the events of that year.
Table 4 reports summary results on the herding measure H(t) described in Eq. (3) for four
different time intervals. Over the longest time period, from May 1993 (when we have flows
of funds for at least three countries) to March 1999, the figures are significant at the 1%
level, indicating the presence of herding behavior. Breaking the time period into three sub-
periods shows that herding occurred primarily during 1998 and early 1999, when the t-
statistic for H(t) is equal to 2.12 and significant at (slightly more than) the 5% level. In the
inducing little or no change in those flows of funds. For example, our dataset ignores the flow of loans from foreign banks to domestic borrowers that, as previously mentioned, characterized many East Asian economies during the early 1990s. Furthermore, if speculators choose to trade in derivatives (like swaps) to take a negative stance on a particular country, thus exercising a downward pressure on the corresponding domestic markets, their actions would cause only marginal fluctuations in the flows of funds to that country. Unfortunately, evidence on such activity is too scant, and often just anecdotal, to be employed in the statistical analysis that follows. We thank an anonymous referee for these observations.
19
interval surrounding the Mexican Peso crisis, between 1993 and the end of 1995, evidence of
herding in our sample of Asian markets is weaker. In 1996 and 1997, the herding measure is
instead barely different from zero, as is the corresponding t-test. These results suggest that,
consistently with the analysis of Figures 1 and 2, herding was strongest in 1998, i.e., in the
period during which most regime shifts in returns occurred.
V. Herding and clustering of break events
One of the most puzzling characteristics of the financial crises of the past decade is the
speed at which they appear to move from one country to the other, as the results of our
investigation into how currency and equity markets interact during stress periods seem to
confirm. As previously mentioned, and as evident from the bottom panels of Figures 1 and
2, breaks in the relation between currency and equity returns and return volatility are
concentrated in relatively short periods of time, even after we account for the statistical
uncertainty surrounding our estimates. Figures 1 and 2, and the analysis of Table 4, make it
apparent that capital-flow reversals from international investors played a significant role in
the clustering of estimated breaks in the reduced-form relations of Eq. (1) and (2) across
Asian countries during 1997 and 1998. This should not be surprising, given the large
holdings of emerging markets’ publicly available equity by international investors and the
mounting evidence of institutional panic and herding around the time when the crises
occurred. Most significantly for our paper, Kaminsky et al. (2001) observe that mutual fund
investments were very responsive and volatile during most of the crises of the 1990s. In
particular, they argue that the events of 1997 and 1998, although initiated in a single country
(Thailand and Russia respectively), rapidly propagated to other markets, due to the large and
widespread withdrawals of funds from Asia, Latin America, and Eastern Europe, thus
providing support for cross-country spillover effects.
In this section we explore this issue in more depth. In the analysis we have presented so
far, breaks in the estimated relations between equity and currency return volatility at or
around the time of the Asian crisis cluster over a few months of 1997. We also found that,
during the same time interval, the herding measure H(t) was barely different from zero.
Breaks in the reduced-form relation between equity and currency returns were instead
concentrated in 1998, when H(t) was large and significant. This evidence seems to suggest
20
that herding played a role in the clustering of return breaks, but less so in the clustering of
volatility breaks.
To translate this intuition into a more rigorous statistical investigation, we define some
new variables. The first is It , the number of countries that at time t are experiencing a
statistically significant break in their reduced-form relation between equity and currency
markets; more specifically It is the number of countries for which t falls in the corresponding
break confidence interval, as specified in Tables 2A and 2B. Hence, It measures the extent of
concentration of break events across the countries in our sample for each period t. We
compute It for both the sequence of return breaks and the sequence of return volatility
breaks. In the remainder of this section, we develop an empirical model to test whether
herding explains the variability of these series, which proxy for clustering of volatility or
return break events.
To that end, we use the flows of funds data described in Section IV to derive a measure
of the intensity of the herding behavior of international investors at each point in time t. As
previously mentioned, our dataset covers net purchases (or sales) by foreign investors in the
domestic spot equity markets for each of the countries in the sample. We define HM1t as the
number of countries that at time t were experiencing a net outflow of funds from
international investors.10 HM1t is a measure of signed herding, therefore focusing on the
extent of net withdrawals across Asia at each point in time, which Kaminsky et al. (2001) have
already identified as a likely culprit in the events of 1997 and 1998.
We intend to verify whether the herding proxy HM1t can explain the clustering of the
break events reported in Tables 2A and 2B. The easiest approach would be to use linear
regressions. However, because the break is, by construction, a unique event, and because
these events are concentrated over relatively few dates, the discrete variables It for returns
and volatility are characterized by a preponderance of zeros and small values. Hence, such
data appear to be better analyzed by a specification that accounts for those properties. In
particular, the econometric literature suggests the use of the Poisson regression model.
10 As previously mentioned, our dataset covers net purchases (or sales) by foreign investors in the domestic spot equity markets for each of the six countries in our sample. However, although such data starts on January 1992, both equity return and flows data are simultaneously available for all of those six countries only starting from March of 1996. To accommodate these limitations, the variables It and HM1t are computed using only the subset of countries for which both flows of funds data and regime shift information were available at time t.
21
According to this approach, each variable It is assumed to be drawn from a Poisson
distribution with parameter λ t. The primary equation of the model is then
( ) �,2,1,0!
Pr ===−
ii
eitt I
t Itλλ
, ( 5 )
where i is the number of countries experiencing a break event at time t. The variable λ t is
related to the signed herding regressor HM1t in terms of a log-linear model, i.e., lnλ t = B’Rt ,
where B = [constant, β] and Rt = [1, HM1t].11 We estimate the model of Eq. (5) by maximum
likelihood for the full sample 01/1992-03/1999. Our parameter of interest is clearly β. If
estimates of β are positive and significant, then the expected number of break events per
period E[It|HM1t] is bigger, and the clustering of break events more likely, when herding is
more intense, i.e., when HM1t is higher. It is easy to show that the partial derivative of the
conditional expected number of break events per period E[It|HM1t] with respect to the
corresponding herding measure is given by λ tβ. In Table 5 we report the results of the
estimation of Eq. (5) for It computed with respect to the breaks in the reduced-form
equation (2) for return volatility (listed in Table 2B) and with respect to the breaks in the
reduced-form equation (1) for return (listed in Table 2A).
The results from Table 5 support our initial hypothesis that signed herding was a relevant
factor in explaining the estimated clustering of return break events during 1998. In the third
column of Table 5, the coefficient β is positive and statistically significant at the 1%
confidence level, as is the chi-squared statistic. Interestingly, signed herding appears to have
played an important role in the clustering of volatility break events observed in 1994 (as
hinted by Table 4) and 1997 as well. Indeed, the estimated coefficient β is again positive and
significant, although the evidence for the model’s fit is much less compelling, suggesting
lower explanatory power for HM1t in the model for volatility.
We perform a parallel analysis using an alternative, more general measure of herding,
HM2t , defined as the maximum of HM1t and the number of countries that at time t were
experiencing a net inflow of funds from international investors. HM2t is a proxy for unsigned
herding because it captures the extent by which international investors in different countries
11 For more on this topic, see Greene (1997), especially Chapter 19.
22
act together independently of the direction of their flows. We estimate Eq. (5) using HM2t for
both return- and volatility-based It . Unsigned herding appears to explain the concentration
of regime shifts in returns, but not the clustering of return volatility breaks during both the
Mexican Peso and the Asian crises. Indeed, the model for volatility (in the second column of
Table 5) is clearly rejected by the data. Nonetheless, unsigned herding significantly improves
the fit of Eq. (5) with respect to the clustering of return breaks, generating two different R2s
between 40% and 50%.12 Furthermore, using the resulting estimates for β in the fourth
column of Table 5 to compute λ tβ over the mean values for Rt, we find that, on average,
about one additional country suffers a regime shift in the reduced-form return relation of
Eq. (1) when the number of countries experiencing flows of funds of the same sign increases
by 3 units.13
To provide more intuition for these results, in Figure 3 we plot the probabilities for
various outcomes of the clustering variable It , in particular for It = 0, 1, 2, and 3, computed
according to the following recursion:
( )
( ) .3,2,1Pr
0Pr0
=���
�===
=== −
ii
iP
eP
tit
tt
PI
I
-1itt
t
λ
λ
( 6 )
The probabilities in Eq. (6) are estimated using the coefficients from the first and third
columns of Table 5, for both the measures of volatility break clustering (Figure 3A) and
return break clustering (Figure 3B), together with the corresponding dependent variable
itself and the signed herding measure HM1t. The correlation between those Its and HM1t is
evident. As a consequence of the good fit provided by the model of Eq. (5), the probability
of no regime shifts drops, the expected number of countries suffering a break increases, and
the clustering probabilities P2t and P3
t rise significantly when signed herding is more intense,
i.e., when the percentage of funds’ outflows across Asian equity markets is higher.
In short, our analysis suggests that the likelihood of observing a concentration of regime
shifts in the reduced-form models of interaction between equity and currency returns and
12 More precise definitions of the two different R2s used to measure the fit of the log-linear model are in the note to Table 5. 13 In Table 5, none of these measures is however statistically significant, even though β is, because of the standard errors of the mean regressors.
23
return volatility posited in this paper in more than one country over a certain period of time
is an increasing function of signed herding, i.e., of widespread withdrawals from the
corresponding equity markets. Nonetheless, as in Table 4, unsigned herding appears to be
less successful in explaining the clustering of return volatility breaks during 1994 and 1997.
These findings are consistent with our previous assertion that, as a result of herding behavior
by international investors, the Asian financial markets experienced cross-country spillover
during the recent crisis events.
VI. Conclusions
In this study we investigate the role of portfolio rebalancing and herding in the Asian
crisis. In a companion paper, we specified reduced-form linear relations between equity
index returns and lead-and-lag currency returns, and tested whether any of those relations in
six Asian countries experienced a regime shift during the events of 1997 and 1998. The
resulting estimated sequence of structural breaks in the posited relations between currency
and equity markets allowed us distinguish whether market linkages between countries are
due to common information shocks or to information spillover effects. Most of the
estimated breaks in the hypothesized return and return volatility structural relations are not
simultaneous, even when the uncertainty surrounding our estimates is taken into account.
The sequential nature of these regime shifts indicates that information spillover effects
created linkages across the Asian markets during 1997 and 1998.
In this paper, we address the issue of whether the observed sequence of breaks implies a
causal relationship across countries, and of whether cross-country spillover at or around the
time of the Asian turmoil can be attributed to herding behavior by foreign investors. Our
analysis shows no evidence of information spillover for Malaysia, Taiwan, and Thailand in
1994, when they all experience a simultaneous volatility break event. A common information
shock seems a more likely explanation for the set of volatility breaks that occurred in those
countries. We find instead some evidence of information spillover across Indonesia, the
Philippines, and South Korea between 1997 and 1998.
Our analysis suggests that portfolio rebalancing was a major channel for transmitting
information shocks across markets during 1997, but not in 1998. Over the entire time
period, the herding measure used in our analysis is significant at the 1% level, revealing the
24
presence of herding behavior. We nonetheless find that the lowest degree of herding occurs
between 1996 and 1997, when most of the breaks in volatility were observed, but that
herding was particularly intense in 1998, i.e., around the time when most of the breaks in the
posited reduced-form return relations were detected.
Finally, using a Poisson regression model, we provide additional support for the
hypothesis that the Asian markets were affected by cross-country spillover at or around the
time of the crisis, by showing that the expected number of return and return volatility break
events per period was bigger, and the clustering of those break events more likely, when
signed herding was more intense, i.e., when more equity markets were simultaneously
experiencing funds’ withdrawals.
25
References
Bai, J., Lumsdaine, R., Stock, J., 1998. Testing for and dating common breaks in multivariate time series. Review of Economic Studies 65, 395-432.
Banerjee, A., Lumsdaine, R., Stock, J., 1992. Recursive and sequential tests of the unit root and trend break hypothesis. Journal of Business and Economic Statistics 10, 271-288.
Barro, R.., 2001. Economic growth in East Asia before and after the financial crisis. Working paper 8330. NBER, Cambridge, MA.
Bekaert, G., Harvey, C., Lumsdaine, R., 1998. Dating the integration of world equity markets. Working paper. Columbia University, New York, NY.
Bekaert, G., Harvey, C., Lumsdaine, R., 2002. The dynamics of emerging market equity flows. Journal of International Money and Finance, forthcoming.
Brown, S., Goetzmann, W., Park, J., 2000. Hedge funds and the Asian currency crisis of 1997. Journal of Portfolio Management, 26, 95-101.
Chancellor, E., 2000. Devil Take the Hindmost: a History of Financial Speculation. Second Edition. Farrar, Straus & Giroux Inc., New York, NY.
Choe, H., Kho, B., Stulz, R., 1999. Do foreign investors destabilize stock markets? The Korean experience in 1997. Working paper 6661. NBER, Cambridge, MA.
Corsetti, G., Pesenti, P., Roubini, N., 1998. What caused the Asian currency and financial crisis? Working paper. Yale University, New Haven, CT.
Eichengreen, B., Rose, A., Wyplosz, C., 1996. Contagious currency crises: first tests. Scandinavian Journal of Economics 98, 463-484.
Fleming, J., Kirby, C., Ostdiek, B., 1998. Information and volatility linkages in the stock, bond and money markets. Journal of Financial Economics 49, 111-137.
French, K., Schwert, W., Stambaugh, R., 1987. Expected stock returns and volatility. Journal of Financial Economics 19, 3-29.
Furman, J., Stiglitz, J., 1998. Economic crises: evidence and insight from East Asia. Brookings Papers on Economic Activity, 2, 1-135.
Gangnes, B., 1998. The Asian Crisis: toward recovery and reform. Report on a United Nations Expert Group Meeting, Department of Economic and Social Affairs, New York, March 17-18, 1998.
Greene, W., 1997. Econometric analysis. Third Edition. Prentice-Hall, New Jersey. Grier, K., Grier, R., 2001. Exchange rate regimes and the cross-country distribution of the
1997 financial crisis. Economic Inquiry 39, 139-147. Hamao, Y., Masulis, R., Ng, V., 1990. Correlations in price changes and volatility across
international stock markets. Review of Financial Studies 3, 281-307. Harvey, C., 1995. Predictable risk and returns in emerging markets. Review of Financial
Studies, 8, 773-816. Haugen, R., Talmor, E., Torous, W., 1991. The effect of volatility changes on the level of
stock prices and subsequent expected returns. Journal of Finance 46, 985-1007. Hirshleifer, D., Subrahmanyam, A., Titman, S., 1994. Security analysis and trading patterns
when some investors receive information before others. Journal of Finance 49, 1665-1698.
26
Kallberg, J., Liu, C., Pasquariello, P., 2002. An examination of the Asian Crisis part I: regime shifts in currency and equity markets. Working paper. Stern School of Business, NYU, New York, NY.
Kaminsky, G., Lizondo, S., Reinhart, C., 1998. Leading Indicators of Currency Crises. IMF Staff Papers, March, 1-48.
Kaminsky, G., Lyons, R., Schmukler, S., 2001. Mutual fund investment in emerging markets: an overview. Working paper. George Washington University, Washington, DC.
Kaminsky, G., Reinhart, C., 2001. Financial markets in times of stress. Working paper 8569. NBER, Cambridge, MA.
Kodres, L. E., Pritsker, M., 2002. A rational expectations model of financial contagion. Journal of Finance, 57, 769-799.
Lakonishok, J., Shleifer, A., Vishny, R., 1992. The impact of institutional trading on stock prices. Journal of Financial Economics 32, 23-44.
Lee, S., Kim, K., 1993. Does the October 1987 crash strengthen co-movements among international stock markets? Review of Financial Economics 3, 89-102.
Lin, W., Kuo, A., 2000. An overview of the East Asian financial crisis. Review of Pacific Basin Financial Markets and Policies 3, 107-120.
Merton, R., 1980. On estimating the expected return on the market: an exploratory investigation. Journal of Financial Economics 8, 323-361.
Officer, R., 1973. The variability of the market factor of New York Stock Exchange. Journal of Business 46, 434-453.
Park, Y., Jong-Wha L., 2001. Recovery and sustainability in East Asia. Working paper 8373. NBER, Cambridge, MA.
Picard D., 1985. Testing and estimating change-points in time series. Advances in Applied Probability 176, 841-867.
Quandt, R., 1958. The estimation of the parameters of a linear regression system obeying two separate regimes. Journal of the American Statistical Association 53, 873-880.
Quandt, R., 1960. Tests of the hypothesis that a linear regression system obeys two separate regimes. Journal of the American Statistical Association 55, 324-330.
Time, Too much too fast, Volume 150, No. 8, August 25, 1997. Tornell, A., Velasco A., 1995. Fixed versus flexible Exchange rates: which provides more
fiscal discipline? Working Paper 5108. NBER, Cambridge, MA. Van Wincoop, E., Yi, Kei-Mu., 2001. Asian crisis post-mortem: where did the money go and
did the United States benefit? Federal Reserve Bank of New York Working paper, New York, NY.
Wermers, R., 1999. Mutual fund herding and the impact on stock prices. Journal of Finance 54, 581-622.
Wu, R., 1998. Taiwan’s Role in the Asian Financial Crisis. Working paper. Taiwan Institute of Economic research. Available at the following URL: http://www.mofa.gov.tw/emofa/TwRAFCidx.htm.
27
Table 1 Descriptive statistics This table displays descriptive statistics (mean and standard deviation) for time series of equity indexes and local currency returns (Table 1A), and flows of funds (Table 1B) for each of the countries included in the study. Equity returns are computed from local equity indexes’ monthly time series obtained from Bloomberg. Currency returns are calculated from the exchange rate versus the U.S. dollar. The minus sign represents devaluation of the local currency. Currency data are from the Federal Reserve Bank of New York. Flows of funds are in millions of U.S. dollars and are calculated as the difference between the total foreign purchases and sales of domestic stocks in each month of the sample.
1A) Equity and currency data
Equity returns Currency returns Countries Mean Stdev Mean Stdev Sample Size Indonesia 0.06% 11.88% −5.56% 25.05% Jan 96 - Mar 99 Malaysia 0.19% 10.19% −0.62% 3.40% Nov 92 - Mar 99 Philippines −0.29% 10.06% −0.91% 3.77% Nov 94 - Mar 99 South Korea 0.21% 10.52% −0.55% 5.01% Jan 90 - Mar 99 Taiwan 1.02% 9.38% −0.34% 1.99% Apr 93 - Mar 99 Thailand −0.67% 11.84% −0.62% 5.23% Mar 93 - Mar 99
1B) Flows of funds data
Flows of funds Countries Mean Stdev Sample Size Indonesia 60.67 146.88 Mar 96 - Mar 99Malaysia 86.52 648.09 Jan 93 - Mar 99 Philippines 154.15 184.61 Jan 95 - Mar 99 South Korea 229.77 445.79 Jan 92 - Mar 99 Taiwan 100.78 335.04 Jan 96 - Mar 99 Thailand 64.55 432.00 May 93 - Mar 99
28
Table 2 Analysis of structural breaks This table, from Kallberg et al. (2002), reports estimated break dates k^ for the structural relation between equity and currency returns (Table 2A)
ti i
ititttiitt xykdxbAyy εβαλµ +�
��
�+++++=
= =−+−−−
5
1
5
13131 )ˆ( , ( 1 )
and between equity and currency return volatility (Table 2B) ( )[ ] txttxtytyt kdbA εβσλσσµσ +++++= −1 , ( 2 )
respectively. The Median column in the Tables shows the estimated break date �k , computed using the Wald statistic F described in Kallberg et al. (2002). For both Eqs. (1) and (2) we test the null hypothesis that the post-break coefficient changes are not significantly different from zero, i.e., that no break occurred in the sample period, by comparing the maximum value in the estimated time series )(ˆ kF to the 5% quantile of its limiting distribution. The null hypothesis is rejected when the maximum value for )(ˆ kF , reported in the Max-Wald column, is lower than the critical value for the selected significance level. Bekaert et al. (1998) compute a table with critical values for max )(ˆ kF (for dimensions up to 68), approximating the limiting distribution of the F process with partial sums of normal random variables for each possible dimension of the test statistic )(ˆ kF . From such table we use the asymptotic 1% (5%) critical value of 27.02 (22.21) for Eq. (1) (where 7 parameters are allowed to break) and of 16.37 for Eq. (2) (where 2 parameters are allowed to break). The 2.5th and 97.5th Percentile columns display estimated lower and upper bands, respectively, for the confidence intervals for the “true” break dates, computed with quantiles of the Picard (1985) distribution.
2A) Analysis of equity indexes monthly returns Country 2.5th Percentile Median 97.5th Percentile Max-Wald p-value Indonesia May-98 Jun-98a Jun-98 24.19 < 0.05 Malaysia Oct-97 Feb-98b May-98 35.64 < 0.01 Philippines Feb-98 Apr-98b May-98 47.33 < 0.01 South Korea Dec-97 Mar-98a May-98 26.52 < 0.05 Taiwan Jan-94 Feb-94a Feb-94 23.93 < 0.05 Thailand Jan-98 Mar-98b Apr-98 27.45 < 0.01 2B) Analysis of equity indexes monthly return volatility Country 2.5th Percentile Median 97.5th Percentile Max-Wald p-value Indonesia May-97 Jun-97c Jun-97 70.70 < 0.01 Malaysia Jun-94 Oct-94c Jan-95 23.56 < 0.01 Philippines Jul-98 Aug-98c Aug-98 63.61 < 0.01 South Korea Sep-97 Nov-97c Dec-97 17.67 < 0.01 Taiwan Sep-94 Oct-94c Oct-94 32.55 < 0.01 Thailand Sep-94 Oct-94c Oct-94 21.22 < 0.01 a Significant at the 5% level (for a critical value of 22.21). b Significant at the 1% level (for a critical value of 27.02) or less. c Significant at the 1% level (for a critical value of 16.37) or less.
Table 3 Evaluation of information spillover: volatility This table displays the results of tests for chronologically sequential relations between equity volatility breaks. From Table 2B we identify two sets of countries that experienced a contemporaneous volatility break. The first set consists of Malaysia, Taiwan and Thailand in mid-1994. The second set comprises Indonesia, the Philippines and South Korea in 1997-98. We then estimate the following regression:
( )[ ] ( ) tkdxtktdxtbytAytK
ityii i
εσαβσλσσµσ +++++−+==1
ˆˆ1 , ( 4 )
where d( �k i) is equal to zero when t is earlier than the lower limit of the confidence interval for the volatility break registered in country i, and one otherwise. The first column of the table shows the regression adjusted R2, R2a , for the structural relationship of Eq. (4) for each of the countries in the sample. We identify the break dates �k and �k i through the Wald statistic described in Eq. (5). Break dates in our sample are collected in Table 2B. Country i is where a break contiguous to �k has been recorded, according to the results reported in Table 2B. The Contiguous breaks columns displays, for each country in the sample, the two markets where a contiguous break was recorded. We use S if the break is simultaneous, P if country i breaks before �k and D if country i breaks after �k . The last three columns report the corresponding estimated coefficients for the information spillover dummies. To correct for heteroskedasticity and autocorrelation in the residuals of the above regression, statistical significance of the coefficients is measured using Newey-West standard errors.
Contiguous breaks Spillover dummy coefficients Country R2
a i = 1 i = 2 α1 α2
Malaysia 96.63% Thailand (S) Taiwan (S) 0.022 −0.007Taiwan 82.46% Malaysia (S) Thailand (S) −0.036 0.057 Thailand 95.82% Malaysia (S) Taiwan (S) 0.188c 0.050 Indonesia 92.45% Philippines (D) South Korea (D) 0.042b 0.262b Philippines 92.11% Indonesia (P) South Korea (P) 0.125c −0.021 South Korea 94.33% Philippines (D) Indonesia (P) 0.003 0.090a
a Significant at the 15% level. b Significant at the 10% level. c Significant at the 5% level or less.
Table 4 Analysis of the average herding measure for selected periods This table displays the results of tests for the null hypothesis that herding is observed in selected time periods only by random chance, i.e., that the average herding measure H(t) developed by Lakonishok et al. (1992) has mean equal to zero across the six countries in our sample, Indonesia, Malaysia, the Philippines, South Korea, Taiwan and Thailand. H(t) is computed as
})]([)({)]([)()( **1 tbEtbEtbEtbtH t −−−= − , ( 3 )
where b(t) is the number of countries in which we observe negative flows of funds over the total number of flows of funds available at time t. Flows of funds are in millions of U.S. dollars and are calculated as the difference between the total foreign purchases and sales of domestic stocks in each month of the sample. The proxy for Et-1[b(t)], the expected proportion of sellers during a given month, is the simple average of the observed b(t) during the past quarter. The adjustment term E{|b*(t)-E[b*(t)]|} is calculated under the null hypothesis that herding is only observed as a result of random chance. Hence, b*(t) is the proportion of negative flows that we would observe by drawing xt, the number of observed outflows at time t, from a binomial distribution B(xt, nt), where the probability of a single sell outcome is one half, and nt is the number of countries in the sample for which flows of funds data were available at time t. Since xt follows a binomial distribution with probability b*(t) of success, E{|b*(t)-E[b*(t)]|} easily follows given b*(t) and nt. Our final measures of herding are generated as averages of H(t) over selected time intervals.
Analysis of the herding measure H(t) Periods Mean H(t) Median H(t) Stdev H(t) t-statistic
05/93 – 03/99 6.55% 6.25% 19.07% 2.83* 05/93 – 12/95 6.99% 6.25% 19.85% 1.90a 01/96 – 12/97 2.05% −4.51% 13.95% 0.72 01/98 – 03/99 12.89% 12.15% 23.56% 2.12a
a Significant at the 10% level. * Significant at the 1% level or less.
31
Table 5 Herding and spillover effect This table displays the results of the estimation of the Poisson regression model for the number of simultaneous break events I in each period t computed using the confidence intervals for breaks reported in Table 2B (for equity versus currency return volatility) and Table 2A (for equity versus currency returns). The probability of i such events in period t is given by
( ) �,2,1,0!
Pr =−
== ii
tItte
i tIobλ
λ. ( 5 )
The variable λt is formulated in terms of a log-linear model, i.e., lnλ t = B’Rt , where the coefficient vector B = [constant, β], HM1t and HM2t are the measures of signed and unsigned herding, respectively, described in the text, and where Rt = [1, HM1t] or Rt = [1, HM2t]. The model of Eq. (5) is estimated via maximum likelihood. For each of the estimated coefficients (the constant and the herding parameter β) we report the corresponding t-statistic. The two R2 statistics help assess the improvement of the fit resulting from using λ t instead of the mean value for y to predict It. The Pearson R2
P is computed as R2P = 1 – [Σt(It - λ t)2 / λ t] / [Σt(It – Ibar)2 / Ibar], where Ibar is the mean of I
over the sample. The Deviance R2D is instead obtained as R2
D = 1 – [ΣtItln(It / λ t)] / [ΣtItln(It / Ibar)]. The goodness-of-fit statistics χ2 tests the hypothesis that the coefficients of the regression in Eq. (5) are all zero. We also report the partial derivative of the conditional expected number of break events per period E[It|HM1t] with respect to the corresponding herding measure, HM1t. This measure, given by λ tβ, is computed at the mean of HM1t. The clustering measure is computed using the six countries in our sample, Indonesia, Malaysia, the Philippines, South Korea, Taiwan and Thailand. The model is estimated over the interval from 01/1992 to 03/1999.
Poisson model for clustering of breaks Volatility breaks Return breaks
Statistics HM1t HM2t HM1t HM2t Constant −2.0315* −1.4831* −1.8561* −7.3703*
(t-stat) (-5.516) (-2.657) (-5.649) (-5.526) β 0.2544a 0.0289 0.3043* 1.3438*
(t-stat) (1.948) (−0.081) (2.757) (5.522)
Pearson R2P −7.38% 0.49% −16.17% 39.69%
Deviance R2D 5.50% 0.05% 5.89% 49.71%
χ2 3.426a 0.033 6.803a 57.440* (p-value) (0.0642) (0.8563) (0.0107) (0.0000)
∂E[I|HM]/∂HM 0.0526 −0.0060 0.0839 0.3707 (t-stat) (1.248) (-0.180) (1.504) (1.242)
a Significant at the 10% level. * Significant at the 1% level or less.
Figure 1 Flows of funds, return, and volatility breaks in East Asia: March 1997 to January 1998 Figure 1A reports monthly flows of funds data for the six countries in our sample, Indonesia, Malaysia, the Philippines, South Korea, Taiwan and Thailand. Flows of funds are in millions of U.S. dollars and are calculated as the difference between the total foreign purchases and sales of domestic stocks in each month of the sample. Figures 1B and 1C display confidence intervals at the 5% significance level around the estimated break date �k , i.e., the one that maximizes the Wald statistic )(ˆ kF over the sample interval, for the structural relations between equity and currency return and volatility respectively. The confidence intervals are computed according to Bai et al. (1998). The confidence interval measure of 6 (left axis) is associated with the country for which we measure the most significant )ˆ(ˆ kF in the sample. The confidence interval measure
of 1 (left axis) is associated to the country for which we measure the least significant )ˆ(ˆ kF in the sample. A B C
- 2 0 0 0
- 1 5 0 0
- 1 0 0 0
- 5 0 0
0
5 0 0
1 0 0 0
1 5 0 0
$ M
ill.
M a r - 9 7 A p r - 9 7 M a y - 9 7 J u n - 9 7 J u l - 9 7 A u g - 9 7 S e p - 9 7 O c t - 9 7 N o v - 9 7 D e c - 9 7T i m e
M a l a y s i a I n d o n e s i a
T h a i l a n d P h i l i p p i n e s
T a i w a n S o u t h K o r e a
0
1
2
3
4
5
6
7
M a r - 9 7 A p r - 9 7 M a y - 9 7 J u n - 9 7 J u l - 9 7 A u g - 9 7 S e p - 9 7 O c t - 9 7 N o v - 9 7 D e c - 9 7T i m e
S o u t h K o r e a
I n d o n e s i a
0
1
2
3
4
5
6
7
M a r - 9 7 A p r - 9 7 M a y - 9 7 J u n - 9 7 J u l - 9 7 A u g - 9 7 S e p - 9 7 O c t - 9 7 N o v - 9 7 D e c - 9 7T i m e
T h a i l a n d
S o u t h K o r e a
M a l a y s i a
33
Figure 2 Flows of funds, return, and volatility breaks in East Asia: April to December 1998 Figure 2A reports monthly flows of funds data for the six countries in our sample, Indonesia, Malaysia, the Philippines, South Korea, Taiwan and Thailand. Flows of funds are in millions of U.S. dollars and are calculated as the difference between the total foreign purchases and sales of domestic stocks in each month of the sample. Figures 2B and 2C display confidence intervals at the 5% significance level around the estimated break date �k , i.e., the one that maximizes the Wald statistic )(ˆ kF over the sample interval, for the structural relations between equity and currency return and volatility respectively. The confidence intervals are computed according to Bai et al. (1998). The confidence interval measure of 6 (left axis) is associated with the country for which we measure the most significant )ˆ(ˆ kF in the sample. The confidence interval measure
of 1 (left axis) is associated to the country for which we measure the least significant )ˆ(ˆ kF in the sample. A B C
- 6 0 0
- 4 0 0
- 2 0 0
0
2 0 0
4 0 0
6 0 0
8 0 0
1 0 0 0
1 2 0 0
1 4 0 0
$ M
ill.
J a n - 9 8 F e b - 9 8 M a r - 9 8 A p r - 9 8 M a y - 9 8 J u n - 9 8 J u l - 9 8 A u g - 9 8 S e p - 9 8 O c t - 9 8T i m e
M a l a y s i a I n d o n e s i a
T h a i l a n d P h i l i p p i n e s
T a i w a n S o u t h K o r e a
0
1
2
3
4
5
6
7
J a n - 9 8 F e b - 9 8 M a r - 9 8 A p r - 9 8 M a y - 9 8 J u n - 9 8 J u l - 9 8 A u g - 9 8 S e p - 9 8 O c t - 9 8T i m e
P h i l i p p i n e s
S o u t h K o r e a
0
1
2
3
4
5
6
7
J a n - 9 8 F e b - 9 8 M a r - 9 8 A p r - 9 8 M a y - 9 8 J u n - 9 8 J u l - 9 8 A u g - 9 8 S e p - 9 8 O c t - 9 8T i m e
P h i l i p p i n e s
T h a i l a n d
S o u t h K o r e a
I n d o n e s i a
M a l a y s i a
Figure 3 Probabilities of clustering of break events: Volatility and return Figure 3 displays the time series of probabilities for the following outcomes of the clustering variable It = 0, 1, 2, and 3, computed according to the recursion
( )
( ) ,3,2,1Pr
0Pr0
=���
�===
=== −
ii
iP
eP
tit
tt
PI
I
-1itt
t
λ
λ
( 6 )
for both the series of volatility breaks (3A) and the series of breaks in returns (3B), using the estimates for the constant term and the herding coefficient β for HM1t from Table 5 and the log-linear model lnλ t = B’Rt where B = [constant, β] and Rt = [1, HM1t]. On the left axis, we plot (dark line) the number of simultaneous breaks It, and (thin line) the explanatory variable HM1t , while on the right axis we plot Pt0, Pt1, Pt2, and Pt3. A) Volatility Breaks Clustering B) Return Breaks Clustering
0
1
2
3
4
5
6
Jan-92Apr-92Jul-92Oct-92Jan-93Apr-93Jul-93Oct-93Jan-94Apr-94Jul-94Oct-94Jan-95Apr-95Jul-95Oct-95Jan-96Apr-96Jul-96Oct-96Jan-97Apr-97Jul-97Oct-97Jan-98Apr-98Jul-98Oct-98Jan-99
0.00
10.0
20.0
30.0
40.0
50.0
60.0
70.0
80.0
90.0
100.
IP1P3HM1P2P0
0
1
2
3
4
5
6
Jan-92Apr-92Jul-92Oct-92Jan-93Apr-93Jul-93Oct-93Jan-94Apr-94Jul-94Oct-94Jan-95Apr-95Jul-95Oct-95Jan-96Apr-96Jul-96Oct-96Jan-97Apr-97Jul-97Oct-97Jan-98Apr-98Jul-98Oct-98Jan-99
0.00%
10.00%
20.00%
30.00%
40.00%
50.00%
60.00%
70.00%
80.00%
90.00%
100.00%
IP1P3HM1P2P0