King 1990

Download as pdf or txt
Download as pdf or txt
You are on page 1of 30

The Society for Financial Studies

Transmission of Volatility between Stock Markets


Author(s): Mervyn A. King and Sushil Wadhwani
Reviewed work(s):
Source: The Review of Financial Studies, Vol. 3, No. 1, National Bureau of Economic Research
Conference: Stock Market Volatility and the Crash, Dorado Beach, March 16-18, 1989 (1990),
pp. 5-33
Published by: Oxford University Press. Sponsor: The Society for Financial Studies.
Stable URL: http://www.jstor.org/stable/2961954 .
Accessed: 02/06/2012 17:30

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Oxford University Press and The Society for Financial Studies are collaborating with JSTOR to digitize,
preserve and extend access to The Review of Financial Studies.

http://www.jstor.org
Transmission of Volatility
between Stock Markets
Mervyn A. King
Sushil Wadhwani
London School of Economics

Tbis article investigates why, in October 1987,


almost all stock marketsfell together despite widely
differing economic circumstances. We construct a
model in which "contagion" between markets
occurs as a result of attempts by rational agents
to infer information from price changes in other
markets. Tbis provides a channel through which
a "mistake" in one market can be transmitted to
other markets. We offer supporting evidence for
contagion effects using two different sources of
data.

The stock market crash of October 1987 generated a


large number of reports and commentaries. Most of
these concentrated on the alleged failure of market
mechanisms in particular countries, especially the
United States, and largely ignored the question of why
markets around the world fell simultaneously and with
surprising uniformity (Figure 1).
The fact that stock markets in different countries
are correlated is, of course, not surprising in itself.
Any standard asset pricing model, such as the inter-

This article was prepared for the NBER Conference on Stock Market Vola-
tility, March 16-19, 1989; an earlier version was presented to the LSE Finan-
cial Markets Group Conference on Stock Market Behavior in March 1988
and to the European Meeting of the Econometric Society at Bologna in
August 1988. Financial assistance was provided by the LSE Financial Markets
Group. The authors thank the International Stock Exchange, LIFFE, and
Nomura for their provision of data, and Victor Hung and Bertrand Kan for
excellent research assistance. They also thank Michael Brennan, the referee,
Fischer Black, John Campbell, Douglas Diamond, Rob Engle, Charles Good-
hart, Sandy Grossman, James Poterba, and Richard Roll for helpful comments
and suggestions. Address reprint requests to Prof. King, London School of
Economics, Houghton Street, London WC2A 2AE.

The Review of Financial Studies 1990 Volume 3, number 1, pp. 5-33


? 1990 The Review of Financial Studies 0893-9454/90/$1.50
The Review of Financial Studies/ v 3 n 1 1990

Index
110 .FT30

_- _ _Dow Jones
100 ......_.
Nkked

80

70

1 5 12 19 26
October
Figure 1
Stock market indexes, October 1987 (October 1, 1987 = 100)

national capital asset pricing model (ICAPM), would allow for such
a correlation. But to interpret the data solely within a Walrasian equi-
librium framework with fully informed agents seems inadequate for
two reasons. First, it is difficult to come up with a credible story that
links "fundamentals" to the crash; what could explain a fall of almost
23 percent, the largest one-day fall, on the NYSE? Moreover it is
extremely hard to imagine that any such explanation would be con-
sistent with the uniform decline in equity prices in different coun-
tries.' Second, the correlation coefficients between stock markets are
remarkably unstable over time [Brady Report (1988)].
In this article we examine a rational expectations price equilibrium
and model contagion between markets as the outcome of rational
attempts to use imperfect information about the events relevant to
equity values. Because investors (including market-makers) have
access to different sets of information they can infer valuable infor-
mation from price changes in other markets. Although published
news should affect all markets at the same time (albeit in different
ways because the significance of a piece of news may vary from country
to country), not all information, nor the ability to process it, is public.

ISome commentators saw the crash as the end of a speculative bubble. Although there may be some
truth to this story, there remains the puzzle of why the end of a bubble should have led to (1)
similar falls in markets that had behaved in very different ways prior to the crash and (2) subsequent
recoveries in markets, such as Tokyo, that had been most frequently cited as examples of speculative
bubbles.

6
Transmission of Volatility

Valuable information is contained in the prices that other traders are


willing to pay. Hence, an individual trading in London may feel that
information is revealed by the price changes in the New York and
Tokyo stock exchanges. Observed price changes are used to infer
other agents' information. In models of rational expectations equi-
librium with asymmetric information market prices reveal all relevant
information to agents, provided that the information structure is rel-
atively simple [Bray (1985), Green (1977), Grossman (1976, 1978,
1981)]. When this is the case markets are strongly informationally
efficient. Stock prices reflect fundamentals. But when the information
structure is more complex the mapping from signals (the information
observed by one agent that is relevant to others) to market prices will
not be invertible, and so the equilibrium will be non-fully-revealing.
In general, this will be true when the dimension of the signal space
exceeds the dimension of the price space, or when the number of
signals exceeds the number of markets [Jordan (1983)]. In a non-
fully-revealing equilibrium price changes in one market will, there-
fore, in a real sense depend on price changes in other countries
through structural contagion coefficients. Mistakes or idiosyncratic
changes in one market may be transmitted to other markets, thus
increasing volatility. It is this feature that appeals to us as an alter-
native to "news" as an explanation of the contemporaneous fall in
all major stock markets in October 1987. For example, a failure in
the market mechanism in one country that is not immediately rec-
ognized as such will be transmitted to other markets. Our principal
aim is to explore the empirical implications of the idea that a non-
fully-revealing equilibrium implies the possibility of contagion effects.
There is a fundamental identification problem in distinguishing
between the Walrasian efficient markets and the non-fully-revealing
rational expectations models. This is because, in the absence of any
prior restriction on the way in which the covariance structure of
returns may vary over time, any observed correlations between stock
markets can be said to be consistent with an asset pricing model that
satisfies the efficient markets hypothesis. Nevertheless, there are cer-
tain features of the data that throw light on the plausibility of the two
models. Stock markets are not open 'round the clock. In the non-
fully-revealing, but not the Walrasian equilibrium, model there is a
jump in the price in all other markets when one market reopens,
reflecting the information contained in the value of the opening price.
This provides one clear-cut test of the model, and we pay particular
attention to the modeling of price changes when there is time zone
trading.
As was mentioned above it is well known that the links between
stock markets vary over time, and we provide further supporting evi-

7
The Review of Financial Studies/ v 3 n 1 1990

dence below. The interesting question is whether the time-varying


covariance structure can be modeled in a plausible manner. We explore
the idea that the correlation between markets rises following an
increase in volatility. Using monthly data the Brady Report (1988)
found that annual covariances were not stable and did not exhibit
any clear trend. This they interpreted as evidence of the insignificance
of international transmission of price volatility during the 1987 crash.
With high-frequency data, however, we show that covariances are
related to volatility in a way that is consistent with both the contagion
model and also the observed low-frequency correlations. An impli-
cation of this result is that an increase in volatility could be self-
reinforcing and persist for longer than would otherwise be the case.
We conjecture that this might be one reason for the uniform fall in
stock markets during October 1987, despite their varying experience
both before and after that date. As volatility declines, market links
become weaker, and price changes are less closely tied together.
Sections 1 and 2 set out the theoretical framework of the article.
Estimates of the contagion model based on hourly data for London,
New York, and Tokyo over the period July 1987 to February 1988 are
described in Section 3. These suggest that the contagion coefficients
increased during the crash. Our conclusions are presented in Sec-
tion 4.

1. An Example with Two Markets


For simplicity, especially when we come to model time zone trading,
we consider the case of risk-neutral investors. There is, however, a
cost to this assumption. With risk neutrality and arbitrage between
stock markets, all information is fully revealed. To prevent this we
assume that there is no trading in stocks across frontiers. There are
three reasons for making this assumption. First, it makes possible a
non-fully-revealing equilibrium with risk-neutral investors that per-
mits a linear structure for price changes. Second, in practice prices
are not determined by a Walrasian auctioneer, and uninformed (in
this case foreign) investors know that information will be revealed
to them by past transaction prices rather than notional prices trans-
mitted by an auctioneer. Third, even though developments in infor-
mation technology mean that market-makers and many large investors
in different countries now receive news simultaneously, the impli-
cation of screen-based news for prices is not costless to calculate.
There is a difference between "news" that arrives on screens and
"news" in the sense of unanticipated revaluations of asset prices. It
is costly to process the former to yield the latter. Some, perhaps many,
investors may find it less costly to infer valuations, albeit imperfectly,

8
Transmission of Volatility

from changes in market prices than to incur the direct costs of pro-
cessing information. As a modeling strategy, therefore, we want to
analyze a non-fully-revealing equilibrium. Although this is perfectly
compatible with international trading in stocks, the greater com-
plexity involved in modeling the behavior of risk-averse investors,
especially with time zone trading, adds little to the implications of
empirical work.
The model is best illustrated by considering the case of two coun-
tries, both with their own stock market. The general case is examined
in Section 3. Assume first that both markets are open 'round the clock.
The change in the stock market index over an hourly period, say, is
a function of the news released between the beginning and the end
of that hour. Information is of two types, systematic and idiosyncratic.
The former, denoted by u, is information that affects market values
in both countries. The latter, denoted by v, is relevant only to a specific
country. Both u and v are assumed to comprise two components,
corresponding to information that is observed in one country or the
other. Consider, first, a fully revealing equilibrium. If information
from both countries were fully revealed, then the process that would
generate changes in stock prices is assumed to be

AS(1) = u(') + a,2u(2) + V(l) (1)


AS(2) = a21u(') + u(2) + v(2) (2)

where ASO denotes the percentage return in country jbetween time


t - 1 and time t measured by the change in the logarithm of the
stock market price index. The superscripts on the information vari-
ables denote the country in which that information is observed. The
four information variables are assumed to be uncorrelated and to
follow a white-noise process. The only economic restriction implied
by this (and, in particular, the assumption that u(1) and u(2) are inde-
pendent) is that news which affects both countries is always revealed
(or interpreted) first in one country or the other, but never simulta-
neously. This assumption is made purely for convenience. The con-
sequences of relaxing it are minor and are discussed in the Appendix.
If information is not fully observable in both markets, then investors
and market-makers set prices according to

\S(') = u(l) + al2E1(u(2)) + VM'1 (3)


\S(2)= a21E2(u(')) + u(2) + v(2) (4)

where E1 and E2 denote the expectations operator conditional upon


information observed in markets 1 and 2, respectively. We assume

9
The Review of Financial Studies / v 3 n 1 1990

that the only information available to market 1 about the value of u(2)
is the contemporaneous price change in market 2. The unconditional
expectation of u(2) in market 1 is zero, but a nonzero realization of
ASt2)provides information to market 1 about the information that has
been observed in market 2. The message is contaminated by the fact
that some information which leads to price changes in market 2 is
idiosyncratic and irrelevant to market 1. Hence the equilibrium is not
fully revealing. In addition market 1 players realize that their coun-
terparts in market 2 are going through the same exercise in order to
infer information from price changes in market 1. We assume that the
distributions of the stochastic news processes and the parameters of
the model are common knowledge. Hence agents can solve the signal
extraction problem to find the minimum-variance estimator for the
value of the relevant news term that has been observed in the other
market. The solution to this problem is

El( U(2)) S2[AS2)- a (5)


E2(u()) = \1[AS(l)- al2El ( U(2)) (6)
where o2 denotes the variance of x and

i= 1, 2 (7)
c0-i) + t(o2

Substituting these expressions back into Equations (3) and (4) yields2

'AS( = ( - a12a21X1X2) (u(l) + V(1)) + a12X2 tS(2) (8)


S(2) = (1 - a12a2lXlX2) (U(2) + V(2)) + a21X1 AS(t) (9)

Because the a and X parameters cannot be separately identified we


define
ji3=
0aj\ i, j= 1, 2 (10)
Denote
7) = u'i) + vU) i= 1, 2 (11)

2
When u(I) and u(2) are correlated we obtain equations that are analogous to (8) and (9). These are
derived in the Appendix. We have also made the simplifying assumption that agents in one market
never learn subsequently about past realizations of the random variable u in other markets. When
agents learn in later periods about the currently unobservable news the mistakes in valuation are
corrected. This adds noise to the price changes but induces no serial correlation. The correlation
coefficient between markets is equal to its value in the fully revealing model. The implications for
price jumps as discussed below are unchanged. Details of this "catch-up" model are available on
request.

10
Transmission of Volatility

Table 1
Volatility in the two-market case

Variance of market 1 Variance of market 2 Covariance

Full information OT(I) + (a12)2o1(2) O7(2) + (a2l)2o2() a2lao(I) + a12)2)


Contagionmodel IT(2)+ X2(a2)2a (2) IT(2)+ X2(a2l)2 I,(2) a2o,(I) + a2o2)
No communication a2(I) O2(2) 0

Solving Equations (8) and (9) simultaneously we obtain

Aso) = 7(l) + f127(2) (12)


t = (2) + 021(tl) (13)
With 'round the clock trading the variances and covariances of stock
price changes are
var(ASM1)= o_2(1)+ (fl2)2 (2) (14)
var(ASt2)) = O72(2)+ (021)2 -(1) (15)
cov(AS(l), AS(2)) = 021o2(l) + f3120o2(2) (16)

We have concentrated on the case where the underlying variances


are constant. However, the analysis carries over to the case in which
the variances vary over time, and in our empirical work we will exam-
ine changes in 3ijthat might occur as a result of changes in Xj.
The covariance structure of stock price changes in this model may
be compared with that in two polar cases; first, the fully revealing
equilibrium described by Equations (1) and (2) in which all infor-
mation is either available at the same time in both countries or may
be inferred from prices, and, second, the other extreme in which
there is no communication at all between the markets and the price
change in market j is simply i7Q). Table 1 shows the values of the
variances and covariance in all three cases [using Equations (7), (10),
and (11)]. The variance of stock price changes in both markets is
higher in the fully revealing equilibrium than in the imperfectly
revealing equilibrium, which in turn exceeds the variance in the case
of no communication. These results follow from the assumption of
the rational use of available information. The covariance between the
markets is identical in the fully revealing and the imperfectly reveal-
ing equilibria, and hence the correlation coefficient is higher in the
latter case.
Consider the effect of an idiosyncratic shock in one market on prices
in-the other market. In both the fully revealing and no communica-
tions equilibria the impact of such a shock is zero. But in the non-
fully-revealing equilibrium the elasticity of the change in the price

11
The Review of Financial Studies / v 3 n 1 1990

in market i with respect to an idiosyncratic shock in market j is 3Oj.


It is because of this effect, and the resulting higher correlation coef-
ficient between the markets, that we call the imperfectly revealing
equilibrium the contagion model.
The contagion model described by Equations (14) to (16) is not
fully identified because there are four parameters and only three
pieces of information from the data. As we now show, however, the
fact that markets operate in different time zones anad are closed for
part of the day may help us to identify the contagion coefficients.
Each of the three markets examined in our empirical work is closed
for a significant part of the day. The length of time for which markets
operate has been increasing, but the indices are computed and pub-
lished only for a certain number of hours each day. Only London and
New York have overlapping trading hours. During the sample period
used in our empirical work both the United States and United King-
dom changed from daylight saving (or summer) time to winter time
on October 25.
When a market is closed there is no explicit index of prices. But
we may define the shadow index as the price that would clear the
market if trading were to take place conditional upon the information
that is available when it is closed. Although the shadow index is
unobservable, the concept plays an important role in our model.

Case 1: overlapping trading hours. In the two-market case there


are, in general, four regimes in which trading may occur. Without
loss of generality market 1 is defined to be that which opens first.
* Regime 1: Both markets are open and price changes are described
by Equations (12) and (13).
* Regime 2: Market 1 is closed but market 2 remains open. Inves-
tors in market 2 can no longer use information from market 1 to form
conditional expectations about the value of u(1). Because the uncon-
ditional expectation is zero, then from Equation (4) the price change
in market 2 in regime 2 is given by
AS(2) = (2) (7
In market 1 the shadow index incorporates information directly
observable in market 1 as well as the information that is inferred from
the actual price change in market 2. Hence (using the superscript s
to denote the shadow index) we have3
ASols) = 1 + S(12AS(2) (18)

3 If u(I) and u(2)are correlated, then Equations (17) and (18) are no longer valid. The coefficient in
(18) is no longer equal to the corresponding coefficient in (8) when both markets are open. This
implies that the estimates of the contagion coefficients from overlapping trading hours and open-
to-close regressions need not coincide.

12
Transmission of Volatility

* Regime 3: Both markets are closed and the shadow price changes
are given by
s(tis) = nt i= 1, 2 (19)
* Regime 4: Market 2 is closed and market 1 is open. This situation
is obviously the mirror image of regime 2 and price changes are
described by Equations (17) and (18) with superscripts 1 and 2 inter-
changed.
It is necessary to examine also the jumps in price that take place
when switching from one regime to another. Such jumps occur when-
ever a market reopens and are a unique feature of the imperfectly
revealing equilibrium model. There are two cases to examine. First,
when market 1 reopens the shadow index in market 2 jumps to reflect
the information that is contained in the opening price in market 1.
Denote by to, and tc, the times at which market i opens and closes,
and by St3$and SQ)the logarithm of the stock price in market j at the
time when market i opens and closes, respectively.4 The change in
price between the close of trading on one day and the opening of
trading on the next day, the "close-to-open" (CO) price change, is
defined (for market j) by
CO (j)-=S( - S(p) (20)

Denote bypJi) the jump in the (actual or shadow) price in market


jwhen market i reopens. When market 1 reopens market 2 is closed
and so there is a jump in the shadow price in market 2. This is equal
to the inferred value of the relevant information contained in the
opening price in market 1 allowing for the fact that market 1 itself is
reacting to information revealed by the previous day's change in
market 2 after market 1 had closed.
J12) = 21[CO(l)- 012(sf2 - Scl)] (21)
When trading commences in market 1, the opening price will reflect
the market's reaction to the price changes in market 2 that occurred
when market 1 was shut. The close-to-open price change in market
1 is the sum of the changes in the shadow price from regime 2 (while
market 2 remains open) and from regime 3 (while both markets are
closed). Hence
tol

CO(1) (
n1) + f12(S'c - S(2)) (22)
t-= tcI

4 To simplify notation we omit the day to which these values refer. This should be obvious from the
context.

13
The Review of Financial Studies/ v 3 n 1 1990

Combining Equations (21) and (22) yields the price jump as


tol

J12) = :217 t (23)


t= tcI

The opening price in market 1 reveals to market 2 the accumulated


value of the total news terms q(') since the market closed on the
previous day. This revelation property of the opening price is a gen-
eral result. Given their assumed knowledge of the structure of the
problem, agents in all other markets can infer the accumulated value
of the total news term in any market that has just reopened.
The second case is when market 2 reopens, and market 1 is open
so that there is a jump in the actual price in market 1 given by

J 21) = 012[CO(2) - f2l(S(ol2) - S(c))] + (l12)2121(S() - Scl) (24)


The jump is equal to the inferred value of the information in the
opening price in market 2 taking into account the reaction of market
2 to (1) the price changes earlier in the day in market 1 and (2) the
opening price in market 1, which in turn reflects price changes in
market 2 on the previous day after market 1 had closed.
The close-to-open price change in market 2 is the sum of the changes
in the shadow price in regime 3 when both markets are closed [Equa-
tion (19)], the jump in the shadow price when market 1 reopens
[Equation (21)], and the changes in the shadow price in regime 4
while market 1 is trading [Equation (18)]. Summing over this set of
changes yields the close-to-open price change in market 2 as

to2

C2= 7t + f21(S(2) - S(1)) - fl120l21( - Sc)l (25)


t= tc2

Combining Equations (24) and (25), the jump in the price in mar-
ket 1 when market 2 reopens is
to2

J=
(1 (2) (26)
t= tc2

The complete description of price changes with time zone trading


consists of the equations for each of the four regimes [Equations (12),
(13), (17), (18), and (19)] together with the equations for the jumps
in price at the switch points that link regimes [Equations (23) and
(26)]. The change in price over any finite period is the sum of the
changes in either the actual or shadow prices over that period and is
obtained by summing over the equations for the relevant regimes and
switch points. Equations (22), (23), (25), and (26) can be regarded

14
Transmission of Volatility

as a simultaneous system for the close-to-open price changes and the


price jumps in the two markets.
In our empirical work we will use Equation (26) to examine the
effect on London prices of the opening price in New York. In principle
OLS estimation of Equation (22) yields a consistent estimator for the
contagion coefficient ,12 because there is no problem of simultaneity.
This is because market 2 is closed when market 1 reopens, so that
there is no feedback from market 1 to market 2.5
Case 2: nonoverlapping trading hours. When trading hours do not
overlap (London and Tokyo, for example) the outcome is symmetric.
The equations governing price changes within regimes and at the
jump points are as described above. In both markets the jump in the
shadow price equals the informational content of the opening price
of the other market, allowing in turn for that market's reaction to the
earlier price change in its own market. Because the two markets do
not overlap means that there is a recursive structure to the price jumps.
For this reason it is convenient to examine changes in prices over
the 24-hour period from the close of trading on one day to the close
of trading on the next. Denote the "close-to-close" (CC) price change
on day d in market j by CCU/),and the cumulative value of the total
news term x77i)during the close-to-close period that ends on day d
by Md ). Let market 1 be the market that both opens and closes before
market 2 opens. Summing over the price changes in regimes 3, 2,
and 4, Equations (19), (18), and (17), respectively, and at the jump
point [Equation (26)], yields the close-to-close change in market 1 as

to2

CC(d =
NdJ) + 1(S c2 S02) + n(2)J (27)
L ~~~~~~~t--
4c2 J

Substituting from Equation (17)


CC(1)= NAdl) + f12NdT1 (28)

Similarly
CC(2) = NAJ2 + fl1) (29)
Equations (28) and (29) are natural generalizations of Equations
(12) and (13). The lag in Equation (28) but not in Equation (29)
reflects the fact that market 1 opens and closes before market 2 trades.
These two equations may be solved to yield

CC(d1) =
f12CCd-1 + (1 12 021,L)N) (30)

CCd = f21CCdl + (1 -
012fl2,L)d) (31)

5When u(") and u(2) are correlated IV estimation is required.

15
The Review of Financial Studies / v 3 n 1 1990

In both cases the close-to-close price change is linearly related to


the previous close-to-close price change in the other market and a
first-order moving average error process. These equations represent
the reaction of market 1 to market 2, taking into account the fact that
market 2 reacted to the close-to-close change in market 1 on the
previous day (CCOL)),which, in turn, had reacted to CC(2), and so on.
This chain of reactions may be represented as a moving average pro-
cess.

2. The Many-Markets Model


The model described above for the case of two markets may be gen-
eralized to any number of markets, although, as we have seen, esti-
mation of the model with time zone trading introduces a number of
complications. When markets overlap fully the equation describing
price changes for the general case of J markets is
AS= +Ae (32)
where
AS = a J x 1 vector of price changes
71 = a J x 1 vector of news terms
A = a J x J matrix of the aiqcoefficients (ai1 = 0, V])
e = a J x 1 vector of expectations of u held by agents in other
markets
The solution to the signal extraction problem is
e = A(AS - Ae) (33)
where A is a J x J diagonal matrix with X. as the jth element of the
leading diagonal.
Combining Equations (32) and (33) yields
AS = (I + B) n (34)
where B = AA, and the ijth element, /,3j, is the response of market i
to changes in the price in market j.
B is the matrix of contagion coefficients. As we saw in the case of
two markets, the contagion model has the property that an idiosyn-
cratic shock (such as a market breakdown) in one market may have
a multiplier effect on markets elsewhere. The matrix formulation
provides tests of two interesting, albeit rather extreme, hypotheses.
The first is that there are multiple equilibria so that the rate of change
of prices is indeterminate. This occurs if the matrix (I + B) is singular
so that there is no unique solution for the rate of change of market
prices. In conditions of a crash, for example, the fi coefficients might

16
Transmission of Volatility

rise to a level at which the matrix became singular. Second, if the


matrix is decomposable, then there is a hierarchy of influence of
markets on each other, which can be thought of as a leader-follower
relationship.
The existence of time zone trading in the case of Jmarkets means
that there are 2J possible regimes, consisting of all possible combi-
nations of markets being either open or closed. The model describing
price changes within regimes is a switching regressions model with
exogenous switching. The form of the equations governing price
changes in any given regime is similar to Equation (34) with B replaced
by the submatrix formed by deleting the rows and columns corre-
sponding to the markets that are closed. The number and sequence
of regimes are exogenous to the model and are determined by time
zone differences and local hours of trading.6 In addition there are up
to J jump points when markets reopen, and hence J(J - 1) jumps in
actual or shadow prices. When any market reopens the accumulated
value of the total news observed in that market is revealed to all other
markets. When market j reopens the jump in market i is equal to

to'

J(!) = ij r(
2;
77i#] V i ?# j(5 (35)
t-tcj

One possible regression model is to take the close-to-open price


change as the dependent variable with the changes during the trading
day in the other markets prior to the opening of the dependent market
as the independent variables. This procedure yields consistent esti-
mates of the contagion coefficients. But more efficient estimates could
be obtained by using the information contained in the opening prices
of the independent markets that open while the dependent market
is closed, as in the two-market analysis of the previous section. The
implications for estimation in this case depend upon the degree of
overlap of trading in the various markets.

3. Empirical Results
In this section we provide some empirical tests of the contagion
model using high-frequency data from the stock markets in London,
New York, and Tokyo for an eight-month period around the crash,
July 1987 to February 1988. Together these three markets account for
80 percent of total world market capitalization.

6
Local exchanges may choose their hours of trading in the light of experience of the price movements
determined endogenously within the model and, if this is the case, then the regimes become
endogenous to the model. The issue of the optimal length of the trading day is beyond the scope
of this article.

17
The Review of Financial Studies/ v 3 n 1 1990

3.1 Tests for price jumps


One of the features that distinguishes the contagion model from any
fully revealing equilibrium model are the price jumps that occur in
all markets whenever one market reopens. For example, when New
York opens there is a jump in the London price reflecting the infor-
mation contained in the New York opening price [Equation (35)]. In
fact, for the three main financial centers, the effect of the New York
open on London is the only example of an observable jump in the
price of a market that is open. All other jumps are of shadow prices
in markets that are closed. In practice such jumps may be attenuated
for a variety of reasons. An S&P 500 futures contract is traded (albeit
in a thin market) in Amsterdam prior to the opening of New York;
some U.S. stocks are traded in London; and information about the
state of the order books of specialists on the NYSE may be available
to some market-makers in London. As a result the jumps in the London
price when New York opens may not be as clear-cut in practice as
they appear in the theoretical model.
The empirical test of such jumps is that, ceteris paribus, the vola-
tility of prices in London should rise when New York opens. Using
data on the FTSE-100 Index in London we computed the volatility of
half-hourly returns at 15-minute intervals throughout the trading day.
Since there are considerable changes in the average level of volatility
over the sample period, intraday volatility was calculated for three
different subperiods as shown in Figures 2 to 4.7 Figure 2 shows the
intraday volatility during the precrash period (July 1 to October 13-
75 trading days).8 There are three times at which volatility is signif-
icantly higher (at the 10 percent level) than during the rest of the
day: (1) 9:15-9:45 A.M., (2) 11:15-11:45 A.M., and (3) 2:45-3:15 P.M.
The first of these periods is just after the opening of trading during
which the market is incorporating overnight news. The second is the
half-hour around 11:30 A.M., which is the time at which all official
economic statistics for the United Kingdom are announced. The third
is just after the New York opening (which is at 2:30 P.M. London time).
Note that this period is not one during which official U.S. economic
statistics are released. This occurs at 1:30 P.M. London time, and it is
striking that there is a significant decline in volatility in London around
this time suggesting that London reacts more to New York's assess-

7 The three subperiods were selected on the basis of differences in the average level of volatility.
The results are not sensitive to the precise dates that were used and, in particular, to the choice
of October 13 rather than October 16 for the end of the precrash subperiod.
8 We have excluded the observation for 11:30 A.M.-12 noon on August 20, as it represents by far the
largest single change in the subperiod (following an unexpected announcement of changes in
interest rates) and distorts the graph. Including this observation increases volatility by 50 percent
for that 30-minute period.

18
Transmission of Volatility

0.0022

i3 0.X2 A 01 3
3lxStd
* 0.002

i 0.00128 ----------------------------- .--._._._._.-._

0.0016
C 0.0014

0.0012
co

0.001
o '- r'2 >
'-4 '-4 '-4
~ 4 a

Local Time

Figure 2
Standard deviation of 30-minute returns using the FTSE-100 index, computed for 15-minute
intervals, July 1-October 13, 1987

ment of the statistics than to the news itself. Figure 2 appears to


support the contagion model.
Figure 3 shows volatility in the subperiod from December 1 to the
end of February (61 trading days) following the crash and its imme-
diate aftermath.9 There appear to be two peaks in volatility, the first
at the start of trading and the second for the first hour after the opening
in New York (although this is not statistically significant). Again there
is some support for the idea that London reacts to the opening price
in New York (i.e., 2:00-2:30 P.M.). Figure 4 relates to the subperiod
including the crash and its aftermath (34 trading days). Here the local
peak in volatility comes just before the official opening in New York.
Anecdotal evidence suggests that there was much greater commu-
nication between traders in London and New York regarding the size
of the latters' order books immediately after the crash. This would
have the effect of bringing forward in time the effect of New York on
London, as observed in Figure 4. These results are broadly supportive
of the notion that the time around the New York opening is associated
with unusually high volatility in London, although the response is
more diffused than would be predicted by the simple theoretical
model examined above.

3.2 Contemporaneous correlation between markets


To identify the contagion coefficients we estimate the model on hourly
data for stock price changes in New York, Tokyo, and London for the
9 In this subperiod we have excluded the observations for 1:30-2:00 P.M. on December 10 andJanuary
15, which followed the announcement of U.S. trade figures and are clear outliers.

19
The Review of Financial Studies / v 3 n 1 1990

0.003

1 .3lxStd
CC 0.0025- ......

E ~~~~~~~~~~~~~~~~
.3lxStd

0.002
0

o0.0015 ,

0.0005
O '-4 r'2 > N n

Local Time

Figure 3
Standard deviation of 30-minute returns using the FTSE-100 index, computed for 15-minute
intervals, December 1, 1987-February 26, 1988

period September to November 1987. For New York we use the Dow
Jones Index, for London the Financial Times 30 Share Index, and for
Tokyo the Nikkei-Dow Index.
One hypothesis to which we will pay particular attention is that
the contagion coefficients increased during and immediately after the
crash in response to the rise in volatility but then declined as volatility
decreased. Nothing in the model implies that the contagion coeffi-
cients are necessarily constant. The variances of the information vari-
ables may change over time. Suppose that investors do not know the
true variances of the information variables. With Bayesian updating
of beliefs about variances a common shock to all markets, such as
the crash, would result in an increase in the perceived variances of
the common news terms. In turn this would lead to a rise in the
contagion coefficients. Note that "common news" here refers to either
fundamentals in the conventional sense or other sources of changes
in equity values. If periods of high volatility exhibit little increase in
economic "news" (e.g., episodes like the crash), then in such periods
there is likely to have been a change in the underlying demand or
"taste" for equity. As in the Keynesian beauty contest parable, changes
in the average investor's taste for equity are important in determining
the demands of individual investors, who, therefore, in times of
increased volatility will rationally place greater weight on price
changes elsewhere. In our sample period, therefore, we might expect
that the contagion coefficients would be an increasing function of
volatility.

20
Transmission of Volatility

0.02

0.018
1..3lxStd
cc 0.016- - --

EG -1 .3lxStd
. . .
0.014 - - - - - - - -- - - - - - - - - - - - - - - -

o 0.012-
c

0.0
0.008
v.

0.006

0.004
co
0.002

Local Time
Figure 4
Standard deviation of 30-minute returns using the FTSE-100 index, computed for 15-minute
intervals, October 14-November 30, 1987

of the three marketsthat we consider only London and New York


have overlapping tradinghours. Denote Londonas market1 and New
York as market2. The model that describes changes in stock prices
when both marketsare open is [from Equations (8) and (9)]

ASt)= fl12ZS + (1 -tl 12 (36a)


ZtS = fl21ZAS(tl+ (1 -1 02)t(36 b)
As we showed in Section 1, the contagion coefficients fl12and#f21
arenot identified from estimation of the model for overlappingtrading
hours because of the simultaneity involved. The imposition of iden-
tifying restrictions that would enable instruments to be constructed
is discussed below.
When both markets are open it is difficult to distinguish the con-
tagion model from a fully revealing equilibrium model such as the
international market model (IMM). To see this suppose that price
changes satisfy the IMMs0 that
A()= ai+o Sw+' i= 1, 2 (37)

where
Aw= the percentage change in the world index
oi= the normalized covariance with the world index
Ei)=the idiosyncratic component of the return

21
The Review of Financial Studies / v 3 n 1 1990

In the two-market case


lASW = W1 A S(') + W2 A St2) (38)
where w, is the share of market i in the world portfolio (w1 + w2 =
1). It is also true by construction that
w1f1 + w2f2= 1 (39)
From these equations it follows that

(1) -
S= S(2) + a' (40a)
1-
2
, S(2) t L S()t + AZ + t
4b
(40b)
1 ~1 - 32w2

The difference between Equations (36) and (40) is that the IMM
implies a nonlinear restriction on the regression coefficients. (In the
two-market case the restriction is that the product of the coefficients
is unity.) This restriction is rejected in our data. In the remaining
empirical tests we focus on the change in the coefficients over time
and in particular on their relationship with volatility. Whether these
changes are more plausibly explained within a contagion model or
in the IMM framework is a judgment that we leave to the reader.
First, however, we examine the correlation between the markets
when both are open. Table 2 reports the correlation coefficient between
London and New York for hourly price changes during overlapping
trading hours (13:30 to 16:00 GMT), both before and after the crash.
The correlation coefficient is positive, which is consistent with the
idea of the contagion model. Using the published market index there
is some evidence of an increase in the correlation between the two
markets after the crash; the coefficient rises from 0.27 to 0.38. How-
ever, the published data for the United States may be misleading
because, during the week of the crash, the Dow Jones Index often
deviated from the "true" market-clearing price. For example, one
hour after the opening bell on October 19 more than one-third of
the stocks in the DowJones Index had failed to commence trading.
In contrast, the futures price is more likely to reflect market-clearing
levels (although the futures market itself shut down for a short period
on October 20). It seems highly plausible that the observation that
the futures price was often at a substantial discount to the cash price
reflects the presence of "stale quotes" in the cash index.10 In London

10 Further discussion of this issue is contained in Miller et al. (1987).

22
Transmission of Volatility

Table 2
Correlation between London and New York stock markets (hourly data for overlapping
trading hours)

Correlation Number of
Sample period coefficient observations

July 1-October 16, 1987 .270 191


October 19-November 30, 1987 .379 86
October 19-November 30, 19871 .478 84
December 1-February 28, 1988 .194 175
Crash week' .649 10
Crash week2 .750 10

1
Uses data on the futures index for the United States during the crash week, as cash price data were
unreliable.
2
Uses data on the futures price index for both markets.

it has been argued by the International Stock Exchange that the official
index "moved closely in step" with actual transactions prices, although
there were occasions when the futures index traded at a discount."
For these reasons we recalculated the correlation coefficient using
the percentage change in (1) the S&P futures price (quoted on the
Chicago Mercantile Exchange) instead of the change in the Dow
Jones Index and (2) the FTSEfutures index (quoted on LIFFE) instead
of the FT-30 Index for observations in the week of the crash. As can
be seen from Table 2, this produces a significantly higher correlation
coefficient of 0.48 during the period October 16 to the end of Novem-
ber, implying a substantial rise relative to the period before the crash
and even higher values during the crash week itself. All the empirical
results reported below are based on use of the futures index rather
than the spot index for the United States during the crash week.12 It
is also striking that the correlation between the two markets had fallen
to approximately its precrash level by the beginning of 1988.
There was a significant increase in actual volatility during the week
of the crash (see Figure 5) in both London and New York. Measures
of implied (or expected) volatility derived from observed option
prices [using data from Franks and Schwartz (1988)] rose less than
actual volatility, although the time pattern is similar (see Figure 6,
which shows actual and implied standard deviations of hourly price
changes in London during each week of the sample period). After
the crash volatility fell, though it was not until February that it returned
to its precrash level.

11 Quality of Markets Quarterly, Winter 1987-1988, International Stock Exchange, London, p. 19.
12
The futures data used during the crash week refer to the S&P Index, whereas we use the Dow
Jones Index before and after the crash week. However, since the official Dow Jones Index during
the crash week was unreliable, we believe that the change in the price of an S&P futures contract
is likely to be a better proxy for the "true" change in the Dow Jones.

23
The Review of Financial Studies / v 3 n 1 1990

0.04 NowYork

London

0.03

>, 0.02

10

C
co
.... .~~~~~~~~~~~~~~~~~.......
0
)_ 0 H > H z
Z
-Y z Q >

Date
Figure 5
Standard deviation of hourly returns computed over the previous week using the Dow
Jones and FT-30 indexes, respectively

Using the implied volatility measure we may test formally the


hypothesis that the contagion coefficients are an increasing function
of perceived volatility. When the product of the price change in the
other market and the value of implied volatility is added to the regres-
sion model implied by Equation (36), there is striking evidence that
the links between the two markets have indeed varied with changes
in volatility. This may be seen in Table 3 (rows 1 and 2), where the
columns headed VOL contain the regression coefficients of the vari-
ables measuring the interaction between implied volatility and the
change in the other market. These coefficients are highly significant.
They suggest that at the precrash mean of volatility (around 0.2), the
response of both London and New York to a 1 percent change in the
other was around 0.2-0.25 percentage points. During the five-week
period starting from the crash week (during which volatility averaged
about 0.5), London's response to changes in New York rose to around
0.5, and the coefficient on New York's response to London rose to an
(implausibly high) point estimate of 1.3.
These results may not, of course, reflect contagion because regres-
sions of hourly price changes in one market on changes in the other
market are subject to simultaneity bias. But we explore the possibility
of identifying the contagion coefficients through the use of instru-
mental variables (IV) estimation. If stock prices follow a martingale,
as implied by the efficient markets hypothesis, then there are no
observable variables that could be used as instruments. But there is
now some evidence that stock returns are serially correlated [Fama

24
Transmission of Volatility

Actual volatility

\Implied volatility

I~~~ ,

--- -- - - - - - - - - - - - - - - - - - ---------------------- --- 1 -

July Aug Sept Oct Nov Dec Jan Feb

Month
Figure 6
Alternative measures of volatility, London

and French (1986) and Poterba and Summers (1987)]. This may arise
through either variations in expected returns or the activities of "noise"
traders. With serially correlated returns, estimation of the following
augmented version of Equations (40) would provide a way of iden-
tifying the contagion coefficients:

AS(tl) = 012 tS(2) + fr tS(tI + (10-12I321)i(tl) (41 a)

= f21
AS( AS2+ '2 AS(2)1 + (1-(2) (41b)

Hence lagged stock returns can be used as instruments in the esti-


mation of Equations (41) . When these equations were estimated using
IV for the first and third subperiods the coefficients were poorly deter-
mined. This is probably because the serial correlation in hourly returns
was greater during the crash week. Table 3, therefore, presents both
OLSand IV estimates of Equations (41) for the subperiod that included
the crash. The OLSestimates of the coefficients on lagged price changes

25
The Review of Financial Studies/ v 3 n 1 1990

The New New New


Price Table
London London London 3
most York York York Estimates
(analogous
change of
in
general the
equation

Oct. Oct. Oct. Oct. July July


equation
defined Sample contagion
for that 1-Feb. 1-Feb.
is 19-Nov.19-Nov.19-Nov.19-Nov.
28 28
30 30 30 30 period
market
1). coefficients
estimated

IV IV
takes OLS OLS OLS OLS
method between
the Estimation

form
t2) - - - - - - #12
London
= 0.41
(6.43) 0.36
(9.22) -0.07
(-0.68)
#21 and

AS," New
+
- - - - - -
42 -0.08 -0.09 0.01 York
(-0.98) (-1.25) (0.35)

5,2,
+ using
#1
- - - - 21
2.16
(6.35) 1.56
(9.14)
-0.46
(-1.82)

overlapping
#"(VOL,)(AS^,)
+ - - 0.16 - - 0.65 t2
0.09
(1.02) (2.05) (5.90) data
(2)

- - - - (t-values
3.54 VOL-#
0.94(7.86)
(4.05) in

- - - - -
-1.06 VOL.42
(-5.00)
parentheses)

60 60 60 60 316 316 of
obser-Number
vations

R2
0.630 0.568 0.628 0.644 0.516 0.588

1.74 1.74 1.80 1.81 1.98 1.93 D.W.

26
Transmission of Volatility

Table 4
Estimates of the contagion coefficients between London and New York with time zone
trading (t-values in parentheses)

Number
of
obser-
Sample period 12 21 vations R2 D.W.

United Kingdom July 1-Oct. 13 0.21 - 59 0.159 1.85


[Equation (22)] (3.28)
Oct. 14-Nov. 30 0.39 29 0.576 2.80
(6.06)
Dec. 1-Feb. 28 0.36 - 56 0.486 1.89
(7.14)
United States July 1-Oct. 13 -0.09 0.18 59 0.122 1.96
[Equation (25)] (-0.23) (2.69)
Oct. 14-Nov. 30 0.57 1.11 29 0.494 2.44
(4.23) (4.81)
Dec. 1-Feb. 28 0.53 0.44 56 0.425 2.02
(3.85) (6.00)

The equations estimated are

,0,
CO(l) = q(1) + #2 (Sli, - S2)

and

t?2
c02 = v (2) + #21 (S2 -
S())) -1221 (S(2-S)-
7

where 1 denotes the United Kingdom and 2 denotes the United States.

are significantly different from zero at the 5 percent level, as are the
IV estimates. This suggests that there is a true interrelationship between
the two markets and that the positive correlation coefficient is not
explicable in terms of the same news arriving in both markets simul-
taneously. It is somewhat surprising, however, that the estimated
contagion coefficients are slightly higher using IV than OLS estima-
tion, implying that the innovations in news in the two markets were
negatively correlated.

3.3 Time zone trading regression models


The model determining the close-to-open price changes incorporat-
ing interactions between London and New York is described by Equa-
tions (22) and (25). Table 4 shows estimates of the contagion coef-
ficients based on this model. Rows 1, 2, and 3 show the estimates of
Equation (22) for the United Kingdom for the three subperiods. This
is the set of results obtained by regressing the close-to-open price
change in London on the change in the New York price on the

27
The Review of Financial Studies / v 3 n 1 1990

Table 5
Estimates of the contagion coefficients between Japan, United Kingdom, and United States
with time zone trading (t-values in parentheses)

Number of
observa-
Regression Sample iJfiji 3 tions

London on Japan July 1-Oct. 13 0.19 0.01 67


(3.31) (0.04)
Oct. 14-Nov. 30 0.58 -0.46 31
(7.33) (-2.88)
Dec. 1-Feb. 28 0.06 -0.06 58
(0.92) (0.46)
Japan on London July 1-Oct. 13 0.17 0.12 66
(1.26) (1.00)
Oct. 14-Nov. 30 0.26 -0.49 31
(1.60) (-3.23)
Dec. 1-Feb. 28 0.16 -0.34 58
(1.22) (-2.73)
United States on Japan July 1-Oct. 13 0.08 -0.19 71
(1.38) (-1.58)
Oct. 14-Nov. 30 0.16 -0.12 32
(1.28) (-0.67)
Dec. 1-Feb. 28 0.04 0.09 59
(0.60) (0.66)
Japan on United States July 1-Oct. 13 0.20 0.25 70
(1.93) (2.13)
Oct. 14-Nov. 30 0.49 -0.56 32
(7.13) (-3.72)
Dec. 1-Feb. 28 0.11 -0.20 59
(1.12) (-1.56)

The equations estimated take the form


C') = t ,jCC 1, + (1 ij,fiL)N(d

previous day from the close of London to the close of New York.
Rows 4, 5, and 6 show estimates of Equation (25) for the United States
for each of the subperiods. This second set of estimates involves
regressing the close-to-open price change in New York on the change
in the London price from its previous close and on the change in the
New York market on the previous day during the period when London
was closed.
The message from the results shown in Table 4 is clear. They
suggest a statistically significant association between the two markets,
and, moreover, one that increases during mid-October and then
declines after the end of November. The contagion coefficient mea-
suring the effect of New York on London rose from an average of
about 0.2 before the crash to about 0.4 after the crash. This is very
much in line with the estimates based on hourly changes during
contemporaneous trading, which show a rise from 0.20 to 0.38. As far

28
Transmission of Volatility

as the effect of London on New York is concerned, the point estimates


of the contagion coefficient in Table 5 imply a rise from approximately
0.2 before the crash to around unity in the immediate aftermath of
the crash (a slightly smaller rise than was obtained with the contem-
poraneous data), and then a fall to about 0.4 in the period December-
February.
The contagion coefficients between Japan and both London and
New York are obtained by estimating Equations (30) and (31), which
represent a regression of the close-to-close price change in market 1
on the previous close-to-close change in market 2 with a moving
average error process. These estimates are shown in Table 5. The
contagion coefficient measuring the effect of market j on market i is
denoted by f,ij. The dependent variable is the change in the price in
country i. The estimated moving average error process gives an esti-
mate of the product of the contagion coefficients. All of the point
estimates are consistent with the view that the contagion coefficients
rose in the period during and immediately after the crash and then
fell to previous levels in the third of our subperiods. The pattern of
correlations between markets that is revealed by the data seems easier
to reconcile with the contagion model than with a fully revealing or
purely "fundamentals" model.

3.4 A direct test of contagion


The essence of the contagion model is that the trading of stocks in
one market per se affects share prices in other markets, that is, share
prices respond both to public information about economic funda-
mentals and to share-price changes elsewhere. In a related context
French and Roll (1986) examined whether trading contributed to
higher volatility by exploiting the fact that U.S. stock markets were
closed on Wednesdays during the second half of 1968 in order to
clear the settlement backlog. Our contagion model predicts that dur-
ing these U.S. exchange holidays share prices in London would be
less volatile than on other days. In contrast, a fully revealing equilib-
rium model would predict no change in volatility because there was
no interruption in the flow of official and other news.
To discriminate between the two hypotheses we computed the
variance of daily returns in London for various subperiods during
1968-1971. The results are shown in Table 6 and are striking. Although
there was no difference between daily return volatility on Wednesdays
and on other days of the week in the period 1969-1971, volatility in
London on Wednesdays during the second half of 1968 was only two-
thirds its average daily level on other days. This evidence is supportive
of the contagion hypothesis.

29
The Review of Financial Studies / v 3 n 1 1990

Table 6
Variances of daily returns In the United Kingdom

Wednesday Other days

July-December 1968 57.2 83.2


January 1969-December 1971 124.5 123.9

3.5 Related work


The idea that there may be volatility spillovers across markets also
has been examined by Hamao, Masulis, and Ng (1989). They used a
GARCH-based model of volatility and found that higher lagged vol-
atility in both own and other markets was associated with higher
current volatility. This is consistent with a contagion model but could
also be rationalized within a fully revealing equilibrium framework.
In a study of the Tokyo Stock Exchange, Barclay, Litzenberger, and
Warner (1989) concluded that volatility was caused by private infor-
mation revealed through trading. Neumark, Tinsley, and Tosini (1988)
examined the prices of stocks quoted both in New York and either
London or Tokyo. They found that price differentials narrowed during
the crash period. But this provides evidence only on the size of
transactions costs that may limit arbitrage and throws no light on the
signal extraction mechanism that underlies the contagion model.

4. Conclusions
A world in which investors infer information from price changes in
other countries is also one in which a "mistake" in one market can
be transmitted to other markets. If, for example, a failure in the market
mechanism in the United States exacerbated the crash (and we remain
agnostic about that), then this would have transmitted itself to other
markets. Moreover, the empirical evidence suggests that an increase
in volatility leads in turn to an increase in the size of the contagion
effects. The rise in the correlation between markets just after the crash
is evidence of this. Were this result to prove robust, it would have
the important implication that volatility can, in part, be self-sustain-
ing.
The starting point of this article was the uniformity of the fall in
world stock markets during the October 1987 crash, despite important
differences in economic prospects, market mechanisms, and their
prior "degree of overvaluation." We believe that our story might
provide a part of the explanation. The evidence on price jumps with
time zone trading supports the contagion model and merits further

30
Transmission of Volatility

research with data from other markets. The evidence that volatility
in London was lower when the New York market was closed on some
Wednesdays in 1968 provides support for the contagion model. The
role of contagion should not be dismissed on the grounds that there
has been no historical trend increase in correlations between markets.
Nothing in our argument requires there to have been such an increase.
Rather, it is the volatility-related increase in contagion effects that is
the feature of the transmission mechanism.
The possibility of contagion means that one cannot assert that,
because markets without formal portfolio insurance fell as much as
the U.S. market, portfolio insurance could not have been responsible
for the crash."3It is possible-though we take no position on this-
that a U.S. price decline exacerbated by portfolio insurance could
have spread to other markets.

Appendix: Extensions of the Basic Model

The contagion model when ul, u2 are correlated


Suppose that

AS(') = u(t) + al2u 2) + V(2) (Al)


AS(= a21u(1) + U(2) + V(2) (A2)

as in the text, but instead of u(1) and u(2) being independent, we


assume that
(t = Z + (l)(A3)
U(2) = Zt + W(2) (A4)

where zt represents the common component in u(4) and U42). (w(1)


and w(2)are independent.)
In this case, the solution to the signal extraction problems yields

El(ut) = X2[ St2-aE2(ut"))]


- + 02utl1 (A5)
and

E2 ( Utl)) = Xl[-Sal)-al2Ej(u(2))] + O1U(2) (A6)


where in contrast to the text, agents infer something about u(I) from
the observed ut(). Note that
Xi= Xi- 6,0n> 0
13 Fora discussion of the plausibility of events in the U.S. market as the cause of the crash see Roll
(1989) and Wadhwani (1989).

31
The Review of Financial Studies / v 3 n 1 1990

where
+ o4O)
Xi = 2o(t)/(Of22(,)
as before,
-= 4/(G2 + c7A(,)+ a_Vo)

and
(1 - XJ(1 - 6,)o + X16i?oJ1)- (1 - Xi)6i?oW()
=
(1 - ,)c42 + -2W,) + 62f2,) + 6P at,)

Using (A5) and (A6), we can obtain

As(i) t = [a12X2
a22 t10 + (M1 + a 202)u(1) + Mlv(1)]
AS(2) (A7)
M + Ca12aX21X1

and

t A/2 + aI2a21x2 [a21X1 AS(1) + (M2 + a2101)u(2) + M2v 2)] (A8)

where
M2 = 1 - XX2a21Ia12 + X2a2181 > 0

Ml= 1 - X1X2a21a12 + x1a1282 > 0

Therefore, (A7) and (A8) are just generalizations of Equations (8)


and (9). The essence of the contagion model (i.e., a positive rela-
tionship between share-price changes, where the coefficient of inter-
relationship may vary as the ratios of different variances changes) is
preserved.
The equations for the case of overlapping trading hours are also
readily generalized. When both markets are closed, it is still possible
to infer something about u(i) from u(i). So,
AS(ls) = (1 + a12El) U(1) + VM1) (A9)
AS(2s) = (1 + a212) U(2) + V(2) (AlO)

where

2i= 2/(2 + O2(,))

If market 1 is closed but market 2 is open

,ASo) 12X2 [AS(2) + (1 + a12O2)u(1) + vt ]


K1

where

32
Transmission of Volatility

K1 = 1/(1 + a21X2of),

which is analogous to Equation (18).


Notice, however, that the coefficient of AS(2) differs according to
whether markets are open. This implies that our equations for "jumps"
also will have to be correspondingly modified. Once again, though,
the essential structure of the contagion model is unchanged.

References
Barclay, M. J., R. H. Litzenberger, and J. B. Warner, 1989, "Private Information, Trading Volume,
and Stock Return Variances," paper presented to the NBER Conference on Stock Market Volatility
and the Crash, March 1989.

Brady, N. F., 1988, "Report of the Presidential Task Force on Market Mechanisms," U.S. Government
Printing Office, Washington, D.C.

Bray, M., 1985, "Rational Expectations, Information and Asset Markets: An Introduction," Oxford
Economic Papers, 37, 161-195.

Fama, E. F., and K. French, 1986, "Permanent and Transitory Components of Stock Prices," CRSP
Working Paper 178, mimeo, Chicago.

Franks, J. R., and E. S. Schwartz, 1988, "The Stochastic Behaviour of Market Variance Implied in
the Prices of Index Options: Leverage, Volume, and Other Effects," mimeo, London Business School.

French, K. R., and R. Roll, 1986, "Stock Return Variances: The Arrival of Information and the
Reaction of Traders," Journal of Financial Economics, 17, 5-26.

Green, J., 1977, "The Non-Existence of Informational Equilibria," Review ofEconomic Studies, 44,
451-463.

Grossman, S. J., 1976, "On the Efficiency of Competitive Stock Markets Where Traders Have Diverse
Information," Journal of Finance, 31, 573-585.

Grossman, S. J., 1978, "Further Results on the Informational Efficiency of Competitive Stock Mar-
kets," Journal of Economic Theory, 13, 81-101.

Grossman, S. J., 1981, "An Introduction to the Theory of Rational Expectations Under Asymmetric
Information," Review of Economic Studies, 48, 541-560.

Hamao, Y., R. W. Masulis, and V. Ng, 1989, "Correlations in Price Changes and Volatility Across
International Stock Markets," mimeo, University of California, San Diego.

Jordan, J. S., 1983, "On the Efficient Markets Hypothesis," Econometrica, 51, 1325-1343.

Miller, M. H., J. D. Hawke, B. Malkiel, and M. Scholes, 1987, "Preliminary Report of the Committee
of Inquiry Appointed by the Chicago Mercantile Exchange to Examine the Events Surrounding
October 19, 1987."

Neumark, D., P. A. Tinsley, and S. Tosini, 1988, "After-Hours Stock Prices and Post-Crash Hang-
overs," Finance and Economics Discussion Paper 50, Federal Reserve Board, Washington, D.C.

Poterba,J., and L. H. Summers, 1987, "Mean Reversion in Stock Returns: Evidence and Implications,"
Journal of Financial Economics, 22: 27-60.

Roll, R., 1989, "Price Volatility, International Market Links and Their Implications for Regulatory
Policies," Journal of Financial Services Research, 3, 211-236.

Wadhwani, S., 1989, "Comment on Roll," Journal of Financial Services Research, 3, 255-260.

33

You might also like