Vous êtes sur la page 1sur 16

A Multivariate GARCH Model of International Transmissions of Stock Returns and Volatility: The Case of the United States and

Canada Author(s): G. Andrew Karolyi Reviewed work(s): Source: Journal of Business & Economic Statistics, Vol. 13, No. 1 (Jan., 1995), pp. 11-25 Published by: American Statistical Association Stable URL: http://www.jstor.org/stable/1392517 . Accessed: 15/03/2012 15:25
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.

American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of Business & Economic Statistics.

http://www.jstor.org

@ 1995 AmericanStatisticalAssociation

Journalof Business & EconomicStatistics,January1995, Vol. 13, No. 1

Multivariate GARCH
Returns

Model

of

of International Transmissions
Stock
The

and
the

Volatility:
United States and

Case

of

Canada

G. AndrewKAROLYI OH 43210 MaxM.Fisher Columbus, Collegeof Business,TheOhioStateUniversity,


Thisstudyexaminesthe short-run and for dynamicsof returns volatility stocks tradedon the New stock exchanges. The mainfinding thatinferencesaboutthe magnitude is York Toronto and and innovations originate eithermarket thattransmit the othermarthat in and to persistenceof return on ket depend importantly how the cross-market are dynamicsin volatility modeled. Moreover, and muchweakercross-market prevail duringlatersubperiodsand dynamicsin returns volatility stocks withshares duallylistedin NewYork.Implications international for especiallyforCanadian asset pricing,hedgingstrategies,and regulatory policyare discussed. KEYWORDS: Stock returns. Multivariate; ARCH; International;

of The growing international integration financialmarkets has prompted several recent empirical studies to examine the mechanismthroughwhich stock marketmovementsare transmittedaroundthe world. These studies evaluate how stock returnsin one nationalstock marketinfluencethose of anotherstock marketandtheirimplicationsfor pricingof securities within those markets,for hedging and other trading strategies, and for regulatorypolicies within their financial markets. These issues have been of heightened interest in crash that saw the wake of the October 1987 international large, correlatedprice movements across most stock markets. As a result, variousregulationsand institutionalrules were introducedto dampenthe cross-market impactof large stock-pricemovements. Roll (1989) surveyedthe regulatory policies that have been designed to mitigateagainst"excessive" marketvolatility, such as futures-market price limits, taxes, in 23 countries. marginrequirements,and transaction In this study, I examine the short-run dependencein price movementsfor stocks tradedon the TorontoStockExchange (TSE) and the New York Stock Exchange (NYSE). Specifically, I focus on the dynamic relationshipbetween daily stock returnsand stock-return and volatilityfor the Standard Poor (S&P) 500 andTSE 300 stock indexesfromApril 1981 throughDecember 1989. The North Americancontext has receivedattentionin severalrecentstudiesin the financeliterature,such as those of Booth andJohnston(1984), Jorionand Schwartz (1986), Alexander,Eun, and Janakiranan (1988), Mittoo (1993), and Foersterand Karolyi(1993), which have examined the extent to which these two marketsare financially integrated. Their motivationfor studyingthe case of the United States and Canadaarises from the similarityof marketstructuresand regulationand the absence of controls on the free flow of capital between the countries. Indeed,
11

these studies generally found, using internationalversions of the Capital Asset Pricing Model (CAPM) or Arbitrage Pricing Theory (APT), that Canadiansecurities are priced to reflect their risk exposure to the larger North American market ratherthanjust the Canadianmarket, especially in the 1980s. My interestin the North American context for examination of the short-run dynamicsof stock returnsand volatility is, however,furthermotivatedby three institutionalfeatures of these markets.First, uniqueto these two stock marketsis the fact that they representthe two largest nationalmarkets in terms of equity capitalizationwith perfectly synchronous tradinghours. The experiment in this article is then able to circumventthe problem of disentangling the confounding effects of nonsynchronoustradinghours and correlated price changes and volatility. This focus is importantbecause most studies of international dependence in stock returns (e.g., Eun and Shim 1989; Hamao, Masulis, and Ng 1990; Kochand Koch 1991) have determinedthatsystematic lagged cross-marketadjustmentsare typically short-term-that is, no longer than one day. Second, a vast majorityof all interlistedstocks on U.S. exchanges are Canadianbased, comprising42% of the total numberof foreign stocks listed on the NYSE and the American Stock Exchange (AMEX) and 53% of the total marketcapitalizationof foreign stocks, as of December1990 (AmericanStock Exchange 1991; New YorkStock Exchange 1991). Moreover,approximately133 of about 1,200 companies listed on the TSE are interlisted on U.S. exchanges, tradingof which comprises about 57% of the TSE's total-dollartradingvolume (TorontoStock Exchange 1990). Classifying Canadianstocks as interlistedor purely domestic allows my study to formulatea controlled to experimentof the impactof differenttypes of barriers inter-

12

Journal Business Economic of & 1995 Statistics, January between the two markets.Finally,I show that the dynamics of the impactof U.S.-basedstock-return innovationson portfolios of interlisted noninterlisted and Canadian-based stocks aremeasurably different.This evidence suggests thatinvestment barriersfaced by U.S. tradersinterestedin Canadian stocks--due to differences in disclosure requirements,restrictionson ownershipof foreign securities(e.g., 10%limit for Canadian pensionfunds),orperhapstax considerations-do indeedaffect the short-term dynamicsof lagged price adof those stocks to informationshocks that arise in justment the U.S. markets. Section 1 offers some preliminaryanalysis of the data. Section 2 presentsthe econometricapproach. Section 3 exhibits the main results. Section 4 assesses the differences in linkages across marketsfor interlistedand noninterlisted stocks. Section 5 summarizesthe main reCanadian-based sults and discusses theirimplications.

nationalinvestmenton the dynamicrelationsbetween price movements'acrossnationalstock markets. Finally, through daily news reports,marketregulators,traders,and the general investing public in Canadahave become sensitized to marketmovements in the United States and their impacton Canadianmarkets(e.g., Financial Post headlines "TSE300 lags Dow on the Week,"December 12, 1992; "U.S.Bull only a Steer in Canada,"August 3, 1992; "Sun Shines on Bay Street:MarketHealthyDespite Chilly Blast fromU.S.,"July 6, 1992). My methodology is also different from the previous effortsjust cited in thatI examine the dynamicrelationship between the U.S. and Canadiandaily stock-market returnsand returnvolatilities using a bivariategeneralized autoregressive conditionalheteroscedastic(GARCH)model, a family of statistical models originally developed by Engle (1982) and Bollerslev (1986). With this model, I test not only how innovationsoriginatingin the U.S. and rapidly stock-return Canadianmarketstransmitto the othermarketbut also how rapidly the volatility of these innovationstransmitsto the othermarketby simulatingthe impulseresponsesof the estiGARCHmodel. Earlierstudiesexaminedthe matedbivariate marcorrelationof asset-price changes across international Shim (1989), von Furstenkets. Hilliard (1979), Eun and bergandJeon (1989), Roll (1988, 1989), and Kochand Koch and (1991) all focused on the contemporaneous laggedcorreacrossmajorstock marlationin daily closing-pricechanges kets, usually includingCanadaand the United States. These studiesignored,however,the changingconditionalvolatility of stock price changes and, more importantly,the international spillovers of these price-changevolatilitiesthatmight be occurringat the same time. In a study similarto this one, Hamao et al. (1990) used autoregressive conditionallyheteroscedastic models to demonstratethe differentdynamics for spillovereffects in price changes and volatilitiesbetween the United States and Japanand foundthatshocks thatoriginate in the United States are largerand more persistent.Lin, Engle, and Ito (in press) used similar techniques to study such spillover effects, but they explicitly modeled the nonsynchronoustradinghoursin these marketsand foundmuch weakerresults. I confirm that the bivariateGARCH model is a reasonmoveablerepresentation the linkagesbetweenstock-price of in the U.S. and Canadianmarkets.In fact, inferences ments about the magnitude and persistence of the stock-returns shocks that originate in either marketand that transmitto on the othermarketare shown to dependimportantly how we model the cross-market dynamicsin the conditionalvolatilities of the respective markets. Specifically,tests using multivariateGARCH models indicatethat the effects of shocks from S&P 500 index returnsfor the TSE 300 index returns and volatility are smallerand less persistentthanthose measured with traditionalvector autoregressive(VAR) models. Second, using a subperiodanalysis, I show that the effects of S&P 500 returnsshocks on Canadianstocks have dissipatedover the decade of the 1980s, a findingconsistentwith the evidence in otherstudies of growingfinancialintegration

1. PRELIMINARY ANALYSIS 1.1 Data


The data used in the study consist of time series of daily indexes at the close of the markets(4:00 p.m. stock-market in terms of local currency,for the S&P 500 and Eastern), TSE 300. Both marketaggregatesare value-weightedcomposite indexes of a large cross-section of listed stocks. The stocks that have been TSE 300 comprises Canadian-owned listed for at least threeyears and thathave total continuously marketvalues of over $3 million and annual dollar trading volume in excess of $1 million. Care is taken to eliminate control blocks of more than 20% of outstandingshares in the calculationof marketvalue weights (Hatch and Robinson 1989). These data are obtainedfrom Reuters Datalink and the Index Section of the TorontoStock Exchangefor the periodApril 1, 1981, throughDecember29, 1989. I consider daily returnsfor which the respectiveclose-to-close trading periodsin New Yorkand Torontoareperfectlyaligned, generating a total of 2,133 observations. When multiple-day returnsarise as a result of weekends or holidays observed by either market, these observationsare accounted for in my estimationproceduresby means of a dummy variable. My analysis is performedwith both own currencyand U.S. wherethe latteremploy the daily dollar-denominated returns, 4:15 p.m. (Eastern)New York marketmidpoint quotes for the Canadiandollar, also obtainedfrom Reuters. Note that the results report only those experiments performed with U.S. dollar-denominated returns,althoughresults from experimentsusing own-currencyreturnsare obtainablefrom me. Finally, I eliminate from the sample the four influential daily returns-October 16, 19, 20, and 21-around of the marketcrashwhen, it appears,the distribution returns differs dramaticallyfrom the distributionon the other days in the sample. The model estimates and inference tests in this study are also influencedby these four observations. I chose to focus on the long-rundynamicpatternsin stock rereadersspecifically turnsandvolatilityfor these two markets;

A Modelof International Transmissions Karolyi: GARCH

13

interestedin the global marketlinkages aroundthe episode of October 1987 can consult Roll (1989) and Bennett and Kelleher(1988). The stock-market indexesin this studydo not double-count stocks interlistedon either marketso that, for example, no interlistedstocks of the TSE 300 index are Canadian-based membersof the S&P 500 index. Any measurable interdependence between the marketscannot be attributed multiple to listings of stocks.

additionalstructurefor the conditional volatility of returns, as in the multivariate GARCH(M-GARCH)framework proposed in this study.

2. THEECONOMETRIC APPROACH
In this section, I introducethe bivariateGARCH model, describethe numericalmaximumlikelihood techniquesused to estimate the system, and illustratethe impulse response analysis. The tests in this article are based on the ARCH family of models developed by Engle (1982) and generalized (GARCH) by Bollerslev (1986). These models have been empiricallyshown to capturereasonablywell the time variationin the volatility of daily and monthly stock returns (Bollerslev, Chou, and Kroner1992). Moreover,in its multivariateform, it has been applied successfully for asset returnsby Bollerslev,Engle, and Wooldridge(1988), Schwert and Seguin (1990), Lin et al. (in press), Chan, Chan, and Karolyi(1991), Chan, Karolyi,and Stulz (1992), and Engle and Susmel (1993).

1.2 Some Diagnostics


Table 1 presents a wide range of descriptivestatisticsfor the daily stock-indexreturnsof the S&P 500 andTSE 300 for the full sampleperiodandfor two equallydividedsubperiods and the post-October-1987period. The sample momentsfor all returnsseries indicateempiricaldistributions with heavy tails relative to the normaldistribution.There is some negative skewness, especially in the S&P 500 returns,and zero excess kurtosisis confidentlyrejectedfor all series and in all subperiodsat the 95%level. The overallstatisticsarenotably influencedby the post-October-1987 period,even withoutthe omittedreturnsaroundthe crashperiod itself. The sample autocorrelation functions for the daily rawreturnsseries and their respective squared-returns series up to two lags with Ljung-Box (LB) statisticsup to 6 and 12 lags are also exhibited. As seen in previousstudies, the autocorrelations in the stock indexes are statisticallydifferentfrom 0 for the firstandpossibly higherlags. Due to the largesamcriteriafor statistical ple size of this analysis, the appropriate significance for sample statistics and estimatedcoefficients areunclear. I highlightthroughout text andtablescritical the valuesat the 5%significancelevel butcautionthereadersthat a moreconservativecutoff may be appropriate (Zellner1984, returnssechap. 3). The LB statisticsfor the rawandsquared ries rejectthe null hypothesisof white noise easily. Note that the magnitudeof the first-order autocorrelation coefficientfor the TSE 300 index is larger,which, given the lower trading volume and frequencyin the TSE 300 stocks, suggests that the presence of asynchronoustradingof componentstocks in the index may be spuriously inducing the serial correlation in the series (Muthuswamy1989). The autocorrelation for the squareddaily returnsmay be evidence of nonlinear dependence in the returnsseries possibly due to changing conditionalvolatility overtime. The ARCHmodels of Engle (1982) have been developed to capturesecond-ordernonlinear dependence in which the first and second moments of returnsdepend on past values. Finally,the lead and lag correlationsfor the raw and squaredreturnsseries for the S&P 500 and TSE 300 confirm a significantlypositive and large associationandin the one-periodleadfrom contemporaneous the S&P 500 to the TSE 300. The apparentasymmetryof the relationwith strongertransmissionsof movementsfrom New Yorkto Torontoconcurs with the findings of Eun and Shim (1989) and others. Although a crude diagnostic, this effect indicates that cross-market influencescan be revealed in the volatility of the returnsas well as the mean returns themselves. This confirms the importanceof building the

2.1 The Bivariate GARCHModel


The following bivariateGARCH model is posited for the for joint processesgoverningthe daily ratesof return the S&P 500 and TSE 300 stock indexes:
P

r, = a + E
p=1

pr,_, + DIHOL, + D2WKND,+ e,,

e, I Rand H, =
K L

N(0, H,)

(1)

'T+
k=1

FkH,_kFk+
1=1

G'e,_- e,G

+ V'VHOL, + V2V2WKND,, (2) where the daily index returnsare the logarithmof the stock price relativefor the S&P 500 (ri,) and the TSE 300 (r2t) and the returnsvector is denoted by r' = [ri,, r2t]. The residual vector is given by e' = [Lt, 2t], with its correspondingconditional covariancematrix {H,}i = hi,,. e, is representedas a columnvectorof forecasterrorsof the best linearpredictor of r, conditionalon past information,denoted by ft, 1, and vectors including the P lagged values of r,. The parameter and matricesof the mean returnsequation(1) are defined as a' = [al,a2] for the constantand {.Ip}i = di, for the matrix of coefficientswith the p lagged returns.The parameter matrices for the varianceequation(2) are defined as {r}0 = y, which is restrictedto be uppertriangular, and free matrices =j;j,k and {Gi}i = go, for lags k and I, respectively. {Fk}1 The dummy-variablespecification for mean returnsand variancesin this model follows the approachof Baillie and Bollerslev (1989) to isolate the daily spillovers of stock returnsand volatilityfor the marketswhen close-to-close trading hoursareperfectlyaligned. This is in contrastto Cheung and Kwan (1992), who comparedthe average volatility for TSE 300 stock-indexreturnson days in which New Yorkis

14

Journalof Business & EconomicStatistics,January1995 Table 1. SummaryStatisticsforDailyReturnson the Standard Poor500 and Toronto & Stock Exchange300 Stock Indexes period Overall 1-Dec/89 Apr/8 (2,127 obs.) Apr/81-Jun/84 (793 obs.) S&P500 returns Mean Std. dev. (Tstat of mean) Skewness Kurtosis Rawreturnscorrelations Rho (lag = 1) Rho (lag = 2) LB(6) LB(12) Squaredreturnscorrelations Rho (lag = 1) Rho(lag = 2) LB(6) LB(12) .054 1.012 2.463* -.230* 8.780* .062* -.016 34.232* 37.541* .176* .100* 675.633* 753.015* .016 .948 .487 .483* 2.415* .102* .005 9.545 13.684 .032 .003 56.323* 89.252* TSE300 returns Mean Std. dev. (Tstat of mean) Skewness Kurtosis Rawreturnscorrelations Rho (lag = 1) Rho (lag = 2) LB(6) LB(12) Squaredreturnscorrelations Rho (lag = 1) Rho (lag = 2) LB(6) LB(12) .030 .923 1.519 -.242* 12.934* .130* .080* 61.060* 64.646* .339* .176* 708.123* 784.990* -.023 1.019 -.639 .280* 2.765* .251* .042 59.519* 59.681* .102* .141* 42.042* 59.683* .075 .693 3.035* .039 2.538* .267* .064* 62.878* 66.873* .153* .055* 32.413* 41.551* .045 1.062 .979 -.921* 24.629* -.125* .136* 45.340* 54.011* .394* .181* 241.015* 263.189* .094 .862 3.070 -.237* 2.279* .100* -.030 1.819 18.480 .086* -.014 11.422 17.093 .054 1.276 .978 -.626* 11.533* .000 -.025 41.591* 46.881* .194* .112* 222.571* 239.819* Subperiods Jul/84-Oct/87 (793 obs.) Nov/87-Dec/89 (541 obs.)

Statistics

S&P500 and TSE300 returnscross-correlations Rawreturns Rho (lag = -2) Rho (lag = -1) Rho (lag = 0) Rho (lag = +2) Squaredreturns Rho (lag = -2) Rho (lag = -1) Rho (lag = 0)
Rho (lag = +1) Rho (lag = +1)

.055* .198* .661*


-.049*

.041 .271* .712*


.032

.008 .236* .599*


.069*

.099* .097* .661*


- .214*

.011 .155* .196* .764*


.301*

-.007 .024 .076* .632*


.052*

-.020 -.023 .130* .438*


.064*

.048* .186* .218* .817*


.360*

Rho (lag = +2)

.119*

.063*

-.027

.133"

NOTE: * indicates significance at the 5% level. LB(n) denotes the Ljung-Box test of significance of autocorrelations of n lags. Rho (lag = j) denotes a lag correlation coefficient between the TSE 300 return and the jth lag of the S&P 500 return (j less than 0 is leading correlation from S&P 500 to TSE 300).

open and closed. I, therefore,define

D,=

dl2J(3)

for HOL,, a dummy variableequal to 1 for days thatfollow holidays in either market,and 0 otherwise. D2 is similarly

defined for a Monday seasonal dummy,WKND,, to isolate effects. In the varianceequation,theparamday-of-the-week to etermatrices,V1 andV2, correspond the dummyvariables for holidays HOL,and weekendsWKND,, respectively,and are restrictedto be uppertriangular, similar in construction to the constantmatrixr.

A Modelof International Transmissions Karolyi: GARCH

15

Equation(1) models the index returnsas a VAR process, such as were employed to study the international transmissions of stock-marketmovementsby Eun and Shim (1989), von Furstenburg Jeon (1989), andKochandKoch(1991). and allows us to measurethe effects of The multivariate structure an innovationin the stock returnsof one marketon its own lagged return and that of the other market. For example, the i,jth component of P,,measuresthe direct effect that a change in the returnto thejth marketwould have on the ith marketin p days. Because of the complicatedcross-equation feedbacks built into such an autoregressivesystem, measuring the system's responseto a typical randomshock requires us to traceoutthe appropriate moving averagerepresentation, as shown by Sims (1980). More discussion of the impulse response analysis follows in Subsection2.4. Conditional on this dependence structurein the mean returns,the residual vector is bivariatenormallydistributed with conditionalcovariancematrixH,. Equation(2) models the dynamic process of H, as a linear function of its own K past values, Ht-k, as well as L past values of squaredinnovations,e,_te'_,, both of which allow for own-marketand cross-marketinfluences in the conditional variances. This model was originally proposed by Baba, Engle, Kraft, and Kroner(1989) (BEKK). The important featureof this specification is that it builds in sufficientgenerality,allowing the conditionalvariancesand covariancesof the two stock markets to influence each other, and, at the same time, it does not requireestimation of many parameters [eight for the bivariatesystem with (L, K) = (1, 1)]. Even more importantly, perhaps,the BEKK process guaranteesby constructionthat the covariancematricesin the system are positive definite. Sequential testing proceduresare used to determinethe order,P, of the VAR process, as well as for the conditional varianceequation,L andK in a BEKKprocess. In this article, Ijustify the specificationfor the conditionalmeanreturns and conditionalvolatility equationsusing the optimallag-length algorithmwith the Akaike informationcriterion.

and
h12,t =

p[hi,

th22, t1/2,

(5)

where A' = [ala2], {Bk}ij = bij,k, and {CI}1 = cij,1,for lags k and 1, respectively. This model imposes an assumption of a constantcorrelationmatrixof returns(p) over time. This parameterization ensures that H, is positive semidefinite and also offers a parsimoniousrepresentationfor the tests that follow. Note also that this specificationallows for own-marketand cross-marketinfluences in the conditional variancesthroughthe off-diagonalcoefficients in each of the Bk and C, matrices. We denote this model the GARCH-CC process to contrastwith the GARCH-BEKKprocess. Finally,I considera specificationin whichthe cross-market influences in the conditionalvarianceprocess are restricted to be 0. That is, I estimate a univariateGARCH process for the conditionalvariances, while at the same time allowing for spillovers in the returnsequation. This is similar to the GARCH-CCprocess, except that the correlationcoefficient, p, and the off-diagonalcoefficients in each of the Bk and Ct matricesare constrainedto be 0. I denote this the univariate GARCHmodel and use it to gauge how importantmodeling of the cross-marketinfluences in the conditional variance process will be for measuringthe short-rundependence in the mean returns.

2.3 Estimation
Given a sample of T observationsof the returnsvector, of r,, the parameters the bivariatesystems are estimatedby the conditional log-likelihood function for each computing time period as L,(O) = - log 27rand L()
=

-t logI H,j

e()H-

1()e,(()

(6)
T

EL,(8),
t=l

(7)

2.2 OtherConditional Variance Processes


To measure the sensitivity of the analysis to the specification of the BEKK conditionalvarianceprocess, the study also employs other specifications. The simplest specification correspondsto thatof a traditional VARsystem in which the residual vector, e,, is white noise with a constant covariancematrix, H, = H. This process is nested within the BEKK model by restrictingthe F and G matricesto be 0. This comparisonrepresentsan important check of the stability of a full informationmaximumlikelihood system such as M-GARCH.Moreover,it affordsus an opportunity benchto markthe findings of earlierstudies of international transmissions of stock returnsthat employ differentspecifications. I also explore an alternativebivariateGARCH specification employed previously by Baillie and Bollerslev (1987), Schwertand Seguin (1990), and Chanet al. (1991):
h1 ,= C2
k=l , t-k

where 8 is the vector of all parameters. Numerical maximizationof the log-likelihoodfunctionfollowing the Berndt, Hall, Hall, and Hausman(1974) algorithmyields the maximumlikelihoodestimatesandassociatedasymptoticstandard errors. The standard errorsand associated t values reported in this articleare, however,those calculatedusing the quasimaximumlikelihoodmethodsof Bollerslev and Wooldridge (1992), which are robustto the density function underlying the residuals. To performresidualdiagnostics, standardization is based on a Choleskydecompositionof the conditional covariancematrixat each period t.

2.4

Impulse Response Analysis

1
2, t-lI

L h22,t

=1

To illustratethe dynamicsof the bivariateGARCHsystem for the conditionalmean returnsof U.S. and Canadianstock markets,I solve for the impulse response functions implied by the system. The impulse response coefficients can be obtained for the conditional mean returns as in traditional

"16

Journalof Business & EconomicStatistics,January1995

Table2. EstimatesFromDynamicModelsof DailyReturnson the S&P500 and TSE300 Stock IndexesfromApril1981 to December 1989, VAR Modeland M-GARCH BEKK Model VAR model 2 Equation1 Equation S&P500 returns TSE300 returns Parameters S&P500 Lag -1 -2 -3 -4 -5 -1 -2 -3 -4 -5 Coef. .1686 -.0617 -.0825 -.0851 .0049 -.1658 .1067 -.0183 .1024 -.0522 .0837 -.1510 .1312 .0349 1.9973 -.2662* 4.9442* .0973 4.2034 342.1167* 432.2920* 11.5644 9.8361 14,941.3342 (.00) (.00) t value [5.79]* [-2.09]* [-2.81]* [-2.89]* [.17] [-5.12]* [3.28]* [-.56] [3.14]* [-1.64] [3.40]* [-2.69]* [1.31] Coef. .1990 -.0424 -.0212 -.0580 -.0009 -.0136 .1292 -.0305 .1175 -.0289 .0726 -.2974 .2202 .0692 2.0026 -.3143* 1.0554* .1180 4.2191 659.4262* 762.9053* 13.8256 7.7861 (.00) (.00) t value [7.64]* [-1.61] [-.81] [-2.20]* [-.03] [-.47] [4.43]* [-1.04] [4.03]* [-1.01] [3.30]* [-5.91]* [2.45]* M-GARCH BEKK model 2 Equation1 Equation S&P500 returns TSE300 returns Coef. .0573 .0180 -.0663 -.0047 -.0478 .0510 -.0225 -.0225 .0804 -.1749 .1284 .0103 1.9975 -.1332 2.7001 1.6599 4.7732 12.0861 13.6191 1.6410 1.4825 15,204.8753 (.07) (.12) t value [2.64]* [.81] [-2.99]* [-.21] [-2.10]* [2.23]* [-.98] [-.98] [3.23]* [-3.07]* [1.26] Coef. .1985 .0270 -.0145 .0260 .1213 .0961 -.0230 .0085 .0494 -.3332 .1986 .0807 2.0007 -.3092* 5.2469* 5.2902 13.0350 36.5527* 38.1322* 8.9172 5.4894 (.00) (.00) t value [9.60]* [1.28] [-.69] [1.23] [5.60]* [4.42]* [-1.05] [.39] [2.09]* [-6.15]* [2.05]*

TSE 300

Constant WKND dummy HLDY dummy AdjR2 DW Skewness Kurtosis Returnsresiduals LB(6) LB(12) Squaredreturnsresiduals LB(6) LB(12) Ftests of blockof lags S&P500 TSE 300 Log-likelihood

for are based on Bollerslev Wooldridge and likelihood values (in brackets) t NOTE: Quasi-maximum (1992) are employed. Skewness, kurtosismeasures,and autocorrelations computed and at standardized residuals. F statistics(withp values in parentheses)denotetests of exogeneityof blockof lags. A * indicatessignificance the 5% level. The specifications estimation for procedures each modelare describedin the text.

VAR models by successively substitutingon the right side of Equation(1) to generatea moving averagerepresentation (Sims 1980) as follows:
00

r, =
s=

R,e,_,

(8)

which representsrt as a linear combinationof currentand past one-step-aheadforecasterrorsor innovations.The i,jth componentof R, shows the responseof the ith marketin s peunit riodsaftera standardized randomshock in thejth market and none in other markets. To observe the distinctresponse to patternsof the system, the errorsaretransformed orthogoin nalizethe innovationsusing a Choleskyfactorization which a lower triangular matrixV is selected so thatV-'HV -' = I and used to compute new innovations,vt = V-'et. Making to an orthogonalizedtransformation v,, Equation(8) can be as rewritten

estimatesof the moving averagecoefficientsin Equation(1), erthoughthey are normalizedby dividingby their standard rors to control for differentvariationsin returnsacross the two stock markets. The impulse response functions for the bivariateand univariateGARCH processes apply the same to residualsseries, which have procedures theirstandardized been correctedfor time-varyingconditionalheteroscedasticity, as impliedby theirrespectivemodels. One caveatabout impulseresponseanalysesderivesfromthe difficultyin comstandard-error bands (Runkle 1987) and puting appropriate from their sensitivity to the underlyingmodel specification (Braun and Mittnik in press). These studies imply that, if these errorbandsarelarge,inferencesaboutthe relativemagnitudeof responsesacross models will be problematic.

3. RESULTS 3.1 PrimaryResults


Tables2 and3 reportthe resultsof fittingthe VAR,univariateGARCHandtheBEKKandCCbivariate GARCHmodels to the S&P 500 and TSE 300 index returns. In each case, I reportthe coefficient estimates for the various models with associatedrobustt values, adjustedR2 measures,likelihood-

r, =

s=O

RsVv,_,=

s=O

QSv_,

(9)

where the i,jth componentof Q, representsthe impulse responseof the ithmarketins periodsto a shockof one standard errorin thejth market.The responsescan be thoughtof as the

Transmissions Modelof International A Karolyi: GARCH

17

Table 3. EstimatesFromDynamicModelsof DailyReturnson the S&P500 and TSE300 Stock IndexesfromApril1981 to December 1989, Model CC GARCH M-GARCH Modeland Univariate CC M-GARCH model Equation1 S&P500 returns Parameter S&P500 Lag -1 -2 -3 -4 -5 -1 -2 -3 -4 -5 Coef. .0600 .0191 -.0638 -.0032 -.0544 .0466 -.0316 -.0228 .0812 -.1706 .1310 .0116 1.9978 -.2714* 5.4505* .0703 1.8102 91.4459* 94.1430* 1.6329 1.8206 15,196.4238 (.08) (.04) t value [2.76]* [.86] [-2.89]* [-.15] [-2.48]* [2.12]* [-1.43] [-1.04] [3.24]* [-2.98]* [1.28] 2 Equation TSE300 returns Coef. .1884 .0334 -.0196 .0312 .1012 .0832 - .0274 .0289 .0542 -.3155 .2037 .0662 2.0040 .1548 11.1014* 2.2342 5.9626 147.7939* 148.5949* 7.8140 4.0736 (.00) (.00) t value [8.75]* [1.52] [-.90] [1.42] [4.66]* [3.83]* [-1.26] [1.33] [2.18]* [-5.57]* [2.01]* model GARCH Univariate Equation1 S&P500 returns Coef. .1079 -.0095 -.0294 -.0642 .0181 .0164 .0819 -0.1578 .1138 .0049 1.9982 -.1505 4.8471* .2585 2.6846 17.6840* 19.5341 1.9922 1.2201 15,186.5012 (.04) (.28) t value [3.56]* [-.31] [-.97] [-2.12]* [.59] [.54] (3.31]* [-2.79]* [1.12] 2 Equation TSE300 returns Coef. .1304 -.0382 -.0171 .0713 .0687 .0207 .0959 - .3229 .2105 .0470 2.0008 .3392* 15.2530* .0927 4.7869 224.8783* 245.3134* 2.6289 1.8175 (.01) (.06) t value [4.32]* [-1.27] [-.57] [2.36]* [2.27]* [.68] [3.89]* [-5.73]* [2.08]*

TSE 300

Constant WKND dummy HLDY dummy AdjR 2 DW Skewness Kurtosis

Returns residuals residuals returns Squared


LB(6) LB(12) Ftests of blockof lags S&P500 TSE 300 Log-likelihood LB(6) LB(12)

for are and based on Bollerslev Wooldridge t likelihood values (in brackets) NOTE: Quasi-maximum (1992) are employed. Skewness, kurtosismeasures, and autocorrelations computed at and residuals. F statistics(withp values in parentheses)denote tests of exogeneityof blockof lags. A * indicatessignificance the 5% level. The specifications estimation standardized for procedures each modelare describedin the text.

function values, and various residual diagnostics. I do not reportthe coefficient estimates for the conditionalvariance dynamicsto conserve space, thoughthey are availableon request. Finally, I compute tests of the joint significanceof a block of lags (block exogeneity test) for the returnsof one marketin termsof F statistics,reportedwith associatedp values. Note that these F statistics are not computedusing the likelihoodestimateof the coefficientcovariquasi-maximum coefficients ance matrixbut with restrictedand unrestricted VAR studies. as of determination, in traditional For each model, I estimate lag lengths that correspondto the optimal number as determinedby the Akaike criterion. This resultedin five lags for the VAR,fourlags for the BEKK andCC bivariateGARCHmodels, andthreelags only for the univariateGARCH model. Similarly,though not reported, the Akaike criterion dictated that the conditional variance process be governed by a lag (L, K) structureof (1, 3) for the CC model (16 parameters)and the univariatemodel (8 and parameters) of (1, 1) for the BEKKmodel (8 parameters). The patternsin the coefficientsof the matrices4 thatcapturethe dependenceof the indexreturns theirlaggedvalues on are differentacross models. In general,the estimatesfor the

first-order arequitedifferentfor the VARmodel thanfor lags any of the GARCHmodels. Forexample,the coefficientvalues for the own-market lags of the S&P 500 rangefrom .1686 for the VAR model to .0573 for the BEKK GARCH model, thoughall are statisticallysignificant. On the otherhand,for the TSE 300 returns,the coefficient estimates for the ownmarketlagged returnsrange from an insignificant -.0136 in the VAR model to .1213 for the BEKK GARCH model. More interesting,however, is the variabilityacross models in inferences aboutthe cross-marketdependencein returns. Whereasthe VAR and GARCHmodels agree on a large and significantvalue of about .1990 for the dependenceof TSE 300 returnson lagged S&P 500 returns,the VAR models appear to overestimatethe dependenceof S&P 500 returnson lagged TSE 300 returnswith a significantvalue of -.1658 as comparedwith the GARCHestimates all around-.0500. This findingis consistentwith VARmodels estimatedby Eun andShim (1989) in which they determinedthatin accounting for the total forecast errorvariancefor U.S. marketreturns, a surprisinglylarge 33% of the foreign (eight differentmarkets in total) marketcomponent derives from the Canadian returns.Finally,higher-order lags of the VAR system appear

18 0.25

Journalof Business & EconomicStatistics,January1995 0.25 ModelA - VAR ModelB - M-GARCH BEKK

0.1 .............................................................................................................. 0 . ................................................................................................................ ------------------------5 0.1------?--------?--0,1

0.1--------------------............................................................................................................... ................................................................................................................0................ 0.1-----------------------------------

---I---- --------0,1- --------?-------------

0.15

---------------0.15 1
0.25

---- 1------3 4 5 6 7 11 12

-0.1 05 1
O.L-

10

11

12

Model C - M-GARCH CC

Model D - Univarlate GARCH

0,15

-------------?

------------------

0-1

0,1

-----------------------------

0.1
.

........

O,5. o .----................---------------------------0*
.I
.

......................... ............................
0.05-........
O

...............................................-0.1 --------------------------0.15- 11

-01 ...........................................

...................

10 11 12

-0.15

11

1 - I 10 11 12

for ReturnsApril1981 to December 1989, up to 12 Lags Following Figure 1. ImpulseResponse Coefficients DailyS&P500 Stock-Index Shock Originating S&P 500 (domestic,solid line) and TSE300 (foreign,dottedline) Markets.Estimatesare fromVAR in Unit-Return and models of Tables2 and 3. M-GARCH

to be statistically significant in contrastwith the GARCH models. For example, though the GARCH models rarely yield statisticallysignificantcoefficientestimatesfor higher thantwo lags either for own-market cross-market or returns, those of the VAR model are almost all significantup to the fourthlag. TheresidualdiagnosticsindicatethattheM-GARCH models obtain a better fit to the returnsprocess. The distributionalcoefficientsrevealstill significantnonnormality the for VARmodel with largenegativeskewnessandexcess kurtosis. These models do appearto absorbthe dependencein the autocorrelationsof the residualsfor both series, but higher-order dependencein the squaredresidualsremains. In contrast,the residualsfrom the bivariateGARCH models approachnormality, with insignificantexcess skewness and only slightly positive excess kurtosis. The BEKKmodel also capturesthe structurein the raw and squaredresidualsautocorrelations, with LB X2 statistics unable to reject white noise at the 5% significancelevel for the S&P 500 returns,thoughnot for the squaredTSE 300 returnsresiduals. Waldtests (not reported in the table) were conductedof the null hypothesisthat the

conditionalvariancedynamicsin the bivariateand univariate GARCHmodels are0. Thatis, the VARmodel was proposed as a restricted to alternative the more generalGARCHmodels. The X2 statistics (degrees of freedom) yielded values of 526.9 (8), 510.2 (16), and 490.3 (8) for the BEKK, CC, and univariateGARCH models, respectively, all of which p represent values less that .0001. Tables 2 and 3 also present results on zero exclusion or block exogeneity tests of the null hypothesis that no crossmarketspilloversof stock-return innovationsexist. I implement these tests using the F statistic with degrees of freedom equal to the number of restrictions imposed on the system and the numberof parametersin the unconstrained system. The VAR model indicates that the null hypothesis of zero cross-market spillovers can be easily rejectedin both directions. That is, not only are the five lags of S&P 500 returnsstatisticallyimportant predictorsof TSE 300 returns, but so are the five lags of TSE 300 returnsfor the S&P 500 returns. This is counterintuitive likely due to and model misspecification,as seems to be the case when considering the bivariateand even univariateGARCH models.

A Transmissions Modelof International Karolyi: GARCH


25Model - VAR AModel A VAR 25Model B - M-GARCHBEKK

19

.............................................................................................................. 0 .5 ................................................................... ............................................. .a 0.1


,

.........................

................................................

-..................................................................................
01

.... 0.Q5

...............

.................. -....

-G151........................................................

-Q

1
a025 0.2

7 8 9 10 112 1'
--------

-0-10, 5 '

12

"'

3456

,..

7 8 9101112

0.25 ModelD - Untvariate GARCH 2-..........................................

Model C - M-GARCH CC .......................................---

Q .....................15 ...................................................

.-- \---.------------'-'-'-"""" "'--. ri1

................................................

......................................................... 0015-..................................... .1

0,1

-Q 5 0 -0 5 .................................................................................................................................................................................................................................
_Q

1-Q...................................................... II--------.-1 15-L---1-11--

a-aI 5 -Q...............................

.........

..

1 2

5 6

8 9 10 11 12

1 2

4 5 6 7

8 9 10 11 12

for ReturnsApril1981 to December 1989, up to 12 Lags Following Figure2. ImpulseResponse Coefficients DailyTSE300 Stock-Index in and Shock Originating TSE 300 (domestic,solid line) and S&P 500 (foreign,dottedline) Markets.Estimatesare fromVAR Unit-Return M-GARCH models of Tables2 and 3.

For example, in the BEKKprocess, the F statisticassociated with the null of zero cross-market spilloversfromlaggedTSE 300 returnsto futureS&P 500 returnscannotbe rejected(p value of .1233). Interestingly,the statisticalimportanceof the lagged S&P 500 returnsfor future TSE 300 returnsis confirmedfor all four models estimated,thoughthe magnitudeof the F values areconsiderablysmallerfor the GARCH models. The evidence in Tables2 and 3 suggests thatthe dynamics of the conditionalvolatilityprocesses buildsimportant structure into the data. As shown in the residualdiagnosticsand block exogeneity tests, inferencesaboutthe statisticalimportance of cross-marketspilloversof returnsbetween the S&P 500 and TSE 300 marketreturnsdo appearto be sensitiveto the conditionalvolatility dynamics.

3.2 Impulse Response Analysis


To illustratethe dynamics of the VAR and bivariateand univariateGARCH models for S&P 500 and TSE 300 market returns,we computedthe simulatedresponsesof the esti-

matedsystems to innovationsoriginatingin each marketusing an impulse response analysis. The normalizedresponse functionsimpliedby the four models are plotted in Figure 1 for the S&P 500 marketreturnsand in Figure 2 for the TSE 300 marketreturns. Both figures indicate that the innovations in these marketsarerapidlytransmitted acrossmarkets with the most dramaticresponsesusually on day 1 and later responsestaperingoff. One contrastbetween S&P 500 and TSE 300 responsefunctionsis the absoluteandrelativemagnitudeof the domestic and foreign responses at the first and even subsequent lags. For the S&P 500, the averageresponse is generallywithin a band of .10 around0 for both domestic and foreign shocks, which are likely to be insignificant (Runkle 1987). Moreover,the magnitudeof responsesto domestic shocks are on averagelargerthan for foreign shocks, consistentwithfindingsof Eun andShim (1989). FortheTSE 300 returns, responsesarelargeratthe firstlag, in general, the andespecially those fromthe foreignshocks thatoriginatein the S&P500 returns.This findingis consistentacrossall four models, although the sizes of the response coefficients are smallerfor the GARCHCC and univariate GARCHmodels.

20

Journalof Business & EconomicStatistics,January1995 Table 4. SubperiodEstimatesFromDynamic Modelsof DailyReturnson the S&P500 and TSE300 Stock Indexes Apr/81-Jun/84 793 obs. JuV8 4-Oct/87 793 obs. Nov/87-Dec/89 541 obs.

2 2 2 Equation1 Equation Equation1 Equation Equation1 Equation S&P500 returns TSE300 returns S&P500 returns TSE300 returns S&P500 Returns TSE300 Returns

Parameters
S&P500

Lag Coef.
-1 -2 -3 -4 -1 -2 -3 -4

t value Coef. t value

Coef.

t value Coef. t value

Coef.
-.0349 .0293 -.1813 -.0728 -.0791 .0906 -.0634 .0287 .0561 -.0585 .1770

t value Coef. t value


.0992 [2.35]* [-.80] [.66] .0669 [1.58] [-4.18]* -.1355 [-3.24]* .0460 [1.08] [-1.64] .0703 [1.61] [-1.74] [1.99]* .1204 [2.74]* .0048 J.11] [-1.39] [.63] -.0683 [-1.56] [1.05] .0607 [1.18] [-.47] -.2089 [-1.73] [.71] .3095 [1.30]

TSE 300

Constant WKND dummy HLDY dummy

.1128 [3.15]* .2748 [8.08]* .0257 [.69] -.0168 [-.48] -.0137 [-.37] .0645 [1.82] -.0021 [-.06] -.0137 [-.39] -.0592 [-1.58] .1647 [4.63]* .0204 [.54] .0625 [1.74] .0474 [1.26] .0083 [.23] -.0365 [-.97] .0141 [.40] .0421 [1.05] .0520 [1.37] -.2245 [-2.45]* -.4524 [-5.21]* 0783 [.50] .1021 [.69]

.1002 [2.78]* .2231 [6.57]* -.0447 [-1.20] -.0111 [-.32] .0036 [.10] .0318 [.91] -.0224 [-.60] .0033 [.09] .0175 [.46] .1654 [4.61]* .0086 [.22] .0636 [1.75] -.0231 [-.60] -.0424[-1.17] -.0770 [-2.00]* .0396 [1.09] .1302 [3.33]* .0368 [1.00] -.1752 [-2.00]* -.2724 [-3.30]* .1178 [.76] .2857 [1.95]

AdjR2 DW
Skewness Kurtosis

.0123 1.9975
.3203* 1.5222 1.2811 1.1207 20,13.6487

.1346 1.9909
-.0355 1.9533 (.24) 6.7345 (.34) 3.3304 (.00) (.00)

.0106 1.9979
-.1652 1.8439 1.4950 1.0922 38,422.9087

.0848 2.0105
-.0072 .8381 (.14) 5.2378 (.37) 3.1943 (.00) (.00)

.0400 1.9950
-.8751* 9.0747* 2.4424 1.7827 24,266.8386

.0455 1.9723
-.1141 12.6153 (.02) 2.6568 (.06) 2.8837 (.01) (.01)

Ftests of block lags of


S&P500 TSE 300 Log-likelihood

and NOTE: See Tables2 and 3 fornotation symbols.

3.3 SubperiodAnalysis
Studies by Booth and Johnston (1984), Jorion and Schwartz (1986), Alexander et al. (1988), Mittoo (1993), and Foersterand Karolyi (1993) empiricallyexamined the or extentof integration segmentationof the financialmarkets for in Canada the UnitedStates. Theirmotivation considand erationof the NorthAmericancase arose fromthe similarity and of marketstructures regulation,and from the absence of many controls on capital flows. The tests took the form of empirical investigationsof how domestic and foreign market risk is priced for portfolios of stocks tradedon the TSE versionsof assetandNYSE in accordancewith international pricing models such as the CAPM or APT. The evidence is begenerallyconsistentwith a growingdegreeof integration tween the financial markets over the decade of the 1980s (Mittoo 1993). An important hypothesisthat supplementary I test in this section is whetherthis recent phenomenoncan be observedin the short-term dynamicsof the S&P 500 and TSE 300 returnsand volatility. Table4 presents subperiodresults of the BEKK GARCH modelforthe S&P500 andTSE300 returns series.The model is estimatedwith four lagged returnsfor each marketand usof ing an (L, K) lag structure (1,1) in the conditionalvolatility process, as in Section 2. The subperiodscorrespondto two equally dividedperiods before the October1987 market crash(1981-1984 and 1984-1987) andone post-crashperiod (1987-1989). The coefficient estimates associatedwith the

first lagged returnsshow an attenuationover time. For the S&P 500 returns,the coefficient for the own-marketlagged returnfalls from a value of .1128 to a statisticallyinsignificant -.0349. Less dramatically,the first own-marketlag for TSE 300 returnsfalls from .1647 to .1204. The more interestingpatternacross subperiods,however,is in termsof cross-market dependencein which the coefficient of future TSE 300 returnson the lag-one S&P 500 returnsfalls from influenceof TSE 300 .2748 to only .0992. The cross-market returnsfor subsequentS&P 500 returnsis, as expected,weak over the entirehorizon. This diminishing cross-marketdependence of TSE 300 returnson the U.S. marketshocks is furtherdemonstrated in the block exogeneity tests, shown at the bottom of Table 4. Whereas the F statistics associated with the null hypothesis that the block of lagged S&P 500 returnshave no predictivepower for TSE 300 returnscan easily be rejected in the earlier two subperiods,this is not the case in the post-crashperiod, at least at the 1% significance level. Figures 3 and 4 plot the impulse response functions for the S&P 500 and TSE 300 returns, respectively, in response returnsshocks. The reto own-marketand foreign-market sponse function for the S&P 500 returnsshow no dramatic changes over time and, overall,are well within a band of .1 around0. These responses are likely to be statistically insignificant. By contrast, the TSE 300 response functions show larger changes over the three subperiods, with the first day lagged response of around .30 in the 1981-1984

A Modelof International Transmissions Karolyi: GARCH

21

a3 Subperlod: Apr/81- Jun/84 a 25 ................................................................................................................ a 2 ................................................................................................................. 02 a 15 ................................................................................................................ ai 015

Q3

Subperlod: Apr/81- Jun/84

- i1 a3 JuV/84 Oct/87 Subperiod: a 25 ............................................................................................................... 0,25Q2a 2 . ................................................................................................................


02

10

11

12 Jul/84- Oct/87 Subpe-rod:

Q ................................................................................................................ i15

-015

03 ................ -,O5 ... ............... ....................................

-Qi -Qi ...................... ................................ .. ...............................................


-al
1

1 1 2 3 4 5 6 7 8 9 10 11 12 a3Subperlod:Nov/87 - Dec/89

3 1

6 1

10

11

12

Q3

025 ?--?------------0 (3a5 ....................................o............................................... 1 0........................... 02

Subpedod:Nov/87 - Dec/89 25 l 02 ...........................................................................................

. . 5...... '............-............
01
0

10

11

12

Figure 3. Impulse Response Coefficients for Daily S&P 500 Stock-IndexReturns, SubperiodResults, up to 12 Lags Following Shock Originating S&P500 (domestic,solid line)and in Unit-Return TSE 300 (foreign,dotted line) Markets.Subperiodestimates correspond to those of Table4.

Figure 4. Impulse Response Coefficients for Daily TSE 300


Stock-Index Returns, Subperiod Results, up to 12 Lags Following Unit-Return Shock Originating in TSE 300 (domestic, solid line) and spond to those of Table 4.

S&P500 (foreign,dottedline) Markets.Subperiodestimates corre-

22

Journalof Business & EconomicStatistics,January1995

subperiodfalling to only .10 in the most recent 1987-1989 subperiod. Overall,the short-run dependenceof returnson Canadian stocks to innovationsor shocks that arise in New Yorkapthis pearsto havediminishedovertime. I interpret subperiod evidence as consistent with the findings in other studies of the NorthAmericancapitalmarketthatdemonstrate growing financialmarketintegrationduringthe 1980s.

4. INTERLISTED AND NONINTERLISTED STOCKS CANADIAN


Many Canadianfirms have chosen to simultaneouslylist their stocks on the major U.S. marketexchanges in addition to Canadianmarketsover the past two decades. The usual arguments forwardby firmsrelateto the enhanced put of marketability their securities, betteraccess to new funds at lower cost, and a greaterprofile for Canadianproducts sold in the U.S. market. Switzer (1986), Mittoo (1992), and Foersterand Karolyi (1993) have examinedthe benefitsand costs of listing Canadianstocks in U.S. markets.Two critical facts are important. First, of all foreign-basedcompanies firms represent listed on U.S. exchanges, Canadian-based the largest foreign-stock contingent. Second, a large fraction (as of December 1990, about 133 of 1,200 companies) on of all listed securitieson the TSE areinterlisted theNYSE, AMEX, or NASDAQ exchanges. Moreover,these comprise some of the largest TSE-listed companiesin terms of capitalizationand total-dollartradingvolume. In fact, the TSE

hasestablisheddirectelectronictradinglinkswith theAMEX thatallowsfortwo-waytrading their30 Canadian-based in interlistedstocks. Orderscan be routedeitherway betweenthe exchangesusing the MarketOrderTradingSystem (MOST) of theTSE or the PostExecutionReporting(AUTOPER) system of AMEX. The TSE has a similarthoughless important tradinglink with the Midwest Stock Exchangein Chicago. The patternof spilloversin returnsand volatilities across the U.S. and Canadianmarketsand globally can arise from many differentbarriersto investmentsacross nationalmarkets such as the quality of financialreportingdue to differential accountingdisclosure standards,tax considerations, foreign ownershiprestrictions,or just the difficulty of obaboutforeign stocks. For U.S. investors, taininginformation interlistedCanadian stocks, however,aresubjectto the same andareas easy to tradeas domesticU.S. listingrequirements of stocks. One way to gauge the economic importance these is investmentbarriers to examinethe patternsof spilloversof and returns volatilitiesfromU.S. marketsto interlistedversus stocks. In the studiesby Jorionand purelydomesticCanadian andMittoo(1993) strongerevidenceof segSchwartz(1986) mentation(versus integration)was uncoveredin the pricing of purely domestic versus interlistedCanadianstocks relative to a NorthAmericanmarket.A second reasonto examTSE 300 stocks ine separatelyinterlistedand noninterlisted arises from the study by Bertero and Mayer (1989), which across the world's proposedthatthe degreeof correlatedness marketsis related to the trading of overseas securistock the ties and, in particular, fractionof cross-listed securities.

of and Table 5. EstimatesFromDynamicModelsof DailyReturnson the S&P500 Stock Indexand Portfolios Interlisted Domestic TSE300 Stocks TSEstocks Interlisted Equation1 S&P500 returns Parameters S&P500 Lag -1 -2 -3 -4 -1 -2 -3 -4 Coef. .1287 -.0566 -.0292 .0113 -.0627 .0595 -.0144 -.0097 .0880 -.1638 .1314 .0116 1.9971 -.0372 3.0689* 4.9807 1.7090 62,206.4209 (.00) (.13) t value [4.44]* [-1.94] [-1.00] [.39] [-2.19]* [2.06]* [-.50] [-.34] [3.62]* [-3.00]* [1.17] 2 Equation Portfolio returns Coef. .1163 -.0594 .0031 .0275 .0576 .0790 -.0159 -.0072 .0481 -.1726 .2747 .0286 1.9982 .0351 1.4663 2.0497 2.5073 (.02) (.03) t value [3.97]* [-2.02]* [.10] [.93] [1.99]* [2.71]* [-.54] [-.25] [1.96]* [-3.13]* [2.43]* Domestic TSEstocks Equation1 S&P500 returns Coef. .1147 -.0308 -.0412 -.0001 -.0604 .0395 .0030 .0111 .0842 -.1533 .1399 .0116 1.9977 -.0729 3.3361* 5.4090 1.8572 36,472.1287 (.00) (.10) t value [4.63]* [-1.21] [-1.63] [-.01] [-2.46]* [1.61] [.12] [.45] [3.51]* [-2.84]* [1.27] 2 Equation returns Portfolio Coef. .2513 -.0151 .0072 .0287 -.0634 .1027 .0166 .0380 .0692 - .2520 .2956 .0648 2.0009 .1801" 3.5270* 2.6023 5.7375 (.00) (.00) t value [1.09]* [-.59] [.28] [1.13] [-2.57]* [4.16]* [.67] [1.54] [2.87]* [-4.64]* [2.66]*

TSE 300

Constant WKND dummy HLDY dummy AdjR2 DW Skewness Kurtosis F tests of blockof lags S&P500 TSE 300 Log-likelihood

and NOTE: See Tables2 and 3 fornotation symbols.

A Modelof International Transmissions Karolyi: GARCH


Q3-

23

the conditionalreturnsand volatility dynamicsuncoveredin the previous subsection are simply statistical artifactsof a lagged responseby Canadianstocks to shocks originatingin U.S. marketsbecause of the many small TSE 300 stocks that tradeinfrequently. ............................................................................................................................. QL25i Table 5 reports the estimates of the bivariateGARCHBEKK model for the interlistedand noninterlisted portfolios 015 with the S&P 500. The main contrastingfeature of the interlisted and noninterlistedportfolios is the significant and ................................... much largerpositive dependence of the portfolio returnon 00 .............. the one-day lagged S&P 500 returns. The coefficient value of .2513 is more than double that of the interlistedportfolio of .1163. The block exogeneity tests at the bottom of Table 5 also confirm the greaterimportanceof the lagged S&P 500 returnsfor subsequentreturnsof the domesticTSE stocks (F statisticof 20.6023) thanfor those of the interlisted TSE stocks (F statistic of 2.0497). Finally, Figure 5 displays the impulse response coefficients for the two portfoTSE Non-antedisted 300 Stocks lios of TSE 300 stocks in response to both own-marketand OL25 ................................................................................................................ foreign-marketshocks. Whereas for the interlisted stocks the foreign marketshock generatesa response equivalentin 2 0Q............................................................................................................................ magnitudeat the first and subsequentlags to that of a domestic shock, the foreign-market shock has a much greater ......... .................................................................................... ............. 0Q 15 immediateimpactfor the noninterlistedstocks. Despite po3 4 5 6 11 7 8 12 - 1 1-0 2 bands that surroundthese imtentially large standard-error the economic impact of the U.S. pulse response functions, 1 1 12 2O534...7..10 shocks appearsto be demonstrablylargerfor noninterlisted Canadianstocks. In sum, our results suggest that the magnitude and persistence of the innovationsoriginatingin S&P 500 stock returnsthathave an impacton subsequentreturnsof interlisted stocks are smallerthan those of noninterlistedstocks in the Canadianmarkets. Furthermore, find evidence that the we 1 2 3 4 5 1 10 1 12 of the spillovers in the returnsuncovered cannot dynamics Figure 5. Impulse Response Coefficientsfor Daily Returns on be directly attributed the phenomenon of asynchronous to of and Portfolios Interlisted Domestic TSE300 Stocks up to 12 Lags of portfolio securitiesbecause our results obtainfor in TSE300 (domestic)and Unit-Return Shock Originating Following trading S&P500 (foreign) Markets.Estimatescorrespond those of Table 5. to returnson portfoliosof large and actively tradedTSE stocks as well as for broad-based market-index returns.
TSE 300 Stocks Intertlsted
7 8

Barclay,Litzenberger,and Warner(1990) examinedthe importanceof this element for averagemarketvolatility in the New Yorkand Tokyo markets. In this section, I repeat the experimentsof Section 3 for two portfolios of large, actively tradedTSE 300 stocks that are interlistedor "purelydomestic."The interlistedportfolio comprises Alcan Aluminum, CanadianPacific, Moore Corporation, Inco, NorthernTelecom, and Placer Dome, each of which tradeon the NYSE. They represent,as of December 31, 1989, 15.92% of the total marketcapitalizationof the TSE 300. The matchedsample of large, actively traded, purely domestic stocks includes CanadianTire,Imasco, Noranda, Power Corporation,Stelco, and Thomson Corporation, which total 6.88% of the capitalizationof the TSE 300. This supplementarytest is also importantbecause I control for the effects of asynchronoustradingof componentsecurities in either the aggregateTSE 300 index or its subportfolios. Studying the model estimates for two portfolios of actively tradedTSE stocks will help to determinewhether

5. CONCLUSIONS 5.1 Main Findings


This article examines the structureof short-rundynamics of returnsand volatility for stocks traded on the TSE and the NYSE for the period from 1981 to 1989. The experimentis designed to account for the fact that these markets tradesimultaneouslyso thatour measuresof returnand volatilityspilloversbetweenmarketsarenot affectedby measurementproblemsdue to nonsynchronoustradinghours in the two markets. The study employs multivariateGARCH techniquesto capturethe mechanismby which stock-returns innovationsin one markethave an impact on not only the conditional marketreturnsbut also the conditional market volatility of the other market. I also exploit the fact that manyCanadianstocks interliston U.S. exchanges to explore the importance differenttypes of linkagesbetweenmarkets of

24

Journal Business Economic of & 1995 Statistics, January S&P 500 marketreturns,future research can examine the viability of cross-hedgingtradingstrategiesusing the liquid, U.S.-based derivativecontractsfor traderswith exposure to Canadianmarketrisk. Finally,ourfindingthatthe magnitudeof shocks thatoriginate on the NYSE and that spill over to the TSE has been previously overstatedin several studies may provide useful guidelines for regulatorypolicy in the Canadiansecurities industry. For example, the adoptionby the NYSE in 1988 of limits on large negativedaily price movements,known as "circuit led of breakers," to the introduction similarmeasures on marketsaroundthe world, includingthe TSE. The Brady Commission'sReportof the PresidentialTaskForceon Market Mechanisms(1988) originally recommendeddownside price limits in the stock-index futures and options markets and tradinghalts in all marketsat price levels equivalentto Av250- and400-point declines in the Dow Jones Industrial eragefromthe previousday's close. The benefitsandcosts of thesepolicies havebeen debatedin variouscommitteereports sponsoredby the CommodityFuturesTradingCommission (Kuhn, Kuserk, and Locke 1990) and the NYSE's Market and Panel Report(1990). VarVolatility InvestorConfidence ious U.S. securities,futures,and options exchanges adopted in these recommendations October 1988, followed the next year by the Boardof Governorsof the TorontoStock andFuturesExchanges,which adoptedcircuitbreakerrules identical to thoseof theNYSE (Toronto Futures ExchangeNotice to MembersTS89-21,August 15, 1989). Interestingly, though, the TSE's circuit breakersare triggered, similarly to those of the NYSE, by down moves of the Dow Jones Industrial Averageand not of any TSE-basedaggregate. Because the that evidence in this study demonstrates the influenceon the financialmarketsof U.S.-based stockvolatilityof Canadian price movementsis weakerthan previously understoodand has diminishedover time, the rationalebehind these regulatory policies should be seriouslyreexamined.

that interlistedand noninterlistedstocks face. The main results can be summarizedas follows: 1. A bivariateGARCH model is a useful representation of the joint process governing S&P 500 and TSE 300 returns. Inferencesaboutthe magnitudeandpersistenceof the returninnovations that originate in either market and that transmitto the other marketdepend importantly how the on cross-marketdynamics in the conditionalvolatilities of the respectivemarketsare modeled. 2. The cross-marketpatternsin S&P 500 and TSE 300 returnsand volatilities have changed over time. During the latterpartof the 1980s, the magnitudeof shocks originating in New York have had a diminished impact on subsequent TSE 300 returns. 3. The impact of S&P 500 stock-return innovationson of interlistedversusnoninterlisted TSE 300 stocks portfolios are distinctly different. The magnitudeand persistence of S&P 500 shocks are greaterfor subsequentreturnsof nonreinterlistedstocks. This suggests that investmentbarriers latedto differentialaccountingdisclosurestandards, foreignmay be imporownershiprestrictions,andtax considerations the tantfor understanding dynamicsof comovementsin stock aroundthe world. prices 4. Spurious autocorrelation the returnsseries for the in S&P 500 and TSE 300 aggregateindexes due to asynlarge chronoustradingof componentsecurities likely cannot exof dependencein conditional plainthe structure cross-market mean returnsor volatility.

5.2 Implications
These results have importantimplicationsfor the global pricingof securities,for hedgingandothertradingstrategies, andforregulatory policies withinthesetwo financialmarkets. If one accepts that the short-run cross-market dependencein the security returnsand volatilities are significant,then we need to assess the impactof these international spilloversfor of our understanding the degree of integrationor segmentation in the pricingof CanadianandU.S. securitieswithinthe global NorthAmericancontext. Most of the tests of internationalasset-pricingmodels ignorethe time-varying natureof conditionalexpectedreturnsandtheircovarianceor variance risks, including those that focus on the U.S. and Canadian markets(JorionandSchwartz1986;Mittoo 1993). Future researchneeds to evaluatehow sensitiveare conclusionsabout the extent of integrationor segmentationin these marketsin the context of models thatallow for the dynamicsuncovered in this study. We also need to understandthe dynamics of conditional volatilities in these marketsand how they influencehedging behaviorusing, for example, Canadianderivativesecurities, such as the TSE 35 index options and futurescontractsand the more popularTIPS (index participation certificates),all tradingon the TSE. For example, one of the difficultieswith these contractsis their low liquidity,particularly contrast in with that of the Chicago MercantileExchange's S&P 500 index futures and futures options contracts. With a better of understanding the short-termdynamics of TSE 300 and

ACKNOWLEDGMENTS
Financial support was provided by a seed grant from The Ohio State University,a Faculty Research Grantfrom the Embassy of Canada, and the Dice Center for Financial Economics at Ohio State University. I am grateful for comments from both Editors John Geweke and George Tauchen, an anonymous associate editor, two anonymous referees, WarrenBailey, Stephen Foerster,David Johnson, Eric Kirzner, FrancisLongstaff,MarlenePuffer,Rena Stulz, JustinWood, and participants the September1992 Northat ern Finance Association meetings and at Wilfrid Laurier University. Discussions with Keith Boast, Director of the Toronto Stock Exchange's Market Listings Division, and Craig Hurl of the DerivativeMarkets group were helpful. Bong Chan Kho providedcapable researchassistance. Remaining errors are my own responsibility. All correspondence to G. Andrew Karolyi, Fisher College of Business, The Ohio State University,318 HagertyHall, 1775 College Road, Columbus,OH 43210-1309. Phone: (614) 292 -1875.
[ReceivedNovember1991. RevisedJune 1994. ]

A Modelof International Transmissions Karolyi: GARCH

25

REFERENCES
G., Alexander, Eun,C., andJanakirananan, (1988), "International S. Listings and Stock Returns:Some EmpiricalEvidence,"Journalof Financialand Quantitative Analysis, 23, 135-151. AmericanStock Exchange (1991), Fact Book, New York:Author. SiBaba, Y, Engle, R., Kraft, D., and Kroner,K. (1989), "Multivariate multaneousGeneralizedARCH,"unpublishedmanuscript, Universityof Californiaat San Diego, Dept. of Economics. GeneralizedARCH Baillie, R., and Bollerslev, T. (1987), "A Multivariate Approachto Modelling Risk Premiain ForwardForeignExchangeRate Journalof InternationalMoneyand Finance, 9, 309-324. Markets," (1989), "The Message in Daily Exchange Rates: A Conditional VarianceTale,"Journal of Business & EconomicStatistics,7, 297-305. J. R., Information, Barclay,M., Litzenberger, and Warner, (1990), "Private Reviewof Financial StudVariances," TradingVolumeand Stock-Return ies, 3, 233-253. Transmissionof Bennett, P., and Kelleher, J. (1988), "The International Stock Price Disruptionin October 1987,"Federal Reserve Bank of New York QuarterlyReview, 13, 17-33. Berndt, E., Hall, B., Hall, R., and Hausman, J. (1974), "Estimationand Inferencein NonlinearStructural Models,"AnnalsofEconomicandSocial Measurement,3, 653-665. and Performance: Global Bertero, E., and Mayer, C. (1989), "Structure of Interdependence Stock MarketsAroundthe Crashof October 1987," Centrefor Economic Policy Research,London. unpublishedmanuscript, Bollerslev, T. (1986), "Generalized Autoregressive Conditional HetJournal of Econometrics,31, 307-327. eroscedasticity," Bollerslev, T., Chou, R., and Kroner,K. (1992), "ARCHModeling in Finance: Theory and EmpiricalEvidence,"Journal of Econometrics,37, 231-356. J. Bollerslev,T., Engle, R., andWooldridge, (1988), "ACapitalAsset Pricing Journal of Political Economy, Model With Time-varyingCovariances," 96, 116-131. LikelihoodEsBollerslev,T., and Wooldridge,J. (1992), "Quasi-Maximum Covaritimationand Inferencein Dynamic Models With Time-Varying ances,"EconometricReviews, 11, 143-172. Booth, L., and Johnston, D. (1984), "The Ex-dividendDay Behaviourof CanadianStock Prices: Tax Changes and Clientele Effects,"Journal of Finance, 39, 457-476. in Braun,P.,andMittnik,S. (in press), "Misspecification VectorAutoregressions and Their Effect on Impulse Responses and VarianceDecompositions,"Journal of Econometrics,39. Chan, K., Chan, K. C., and Karolyi, G. A. (1991), "Intraday Volatilityin Reviewof Financial the Stock Index and Stock Index FuturesMarkets," Studies,4, 657-683. Chan,K. C., Karolyi,G. A., andStulz, R. (1992), "GlobalFinancialMarkets and the Risk Premiumon U.S. Equity," Journalof Financial Economics, 32, 137-167. Cheung, C. S., and Kwan, C. C. Y (1992), "A Note on the Transmission of Public InformationAcross International Stock Markets,"Journal of Bankingand Finance, 16, 831-837. ConditionalHeteroscedasticity With EsEngle, R. (1982), "Autoregressive timates of the Varianceof U.K. Inflation," Econometrica,50, 987-1008. Engle, R., and Susmel, R. (1993), "CommonVolatilityin International EqJournal of Business & EconomicStatistics, 11, 167-176. uity Markets,"

Eun, C., and Shim, S. (1989), "International Transmissionsof Stock Market Movements,"Journal of Financial and QuantitativeAnalysis, 24, 241-256. Foerster,S., and Karolyi, G. A. (1993), "The Impact of CanadianStock Listing in the United States,"Journal of InternationalBusiness Studies, 24, 763-784. in Hamao, Y., Masulis, R., and Ng, V. (1990), "Correlations Price Changes and VolatilityAcross International Stock Markets," Review of Financial Studies, 3, 281-307. Hatch, J., and Robinson, M. (1989), InvestmentManagementin Canada, Toronto:Prentice-Hall. Hilliard, J. (1979), "The RelationshipBetween Equity Indices on World Exchanges,"Journalof Finance, 34, 103-114. vs. Jorion, P., and Schwartz,E. (1986), "Integration Segmentationin the CanadianStock Market," Journalof Finance, 41, 603-616. Koch, P., and Koch, R. (1991), "Evolutionin Dynamic Linkages Across Daily National Stock Indexes,"Journal of InternationalMoney and Finance, 10, 231-251. Kuhn, B., Kuserk,B., and Locke, P. (1990), "Do Circuit BreakersModerate Volatility? Evidence From October 1989," unpublishedmanuscript, CommodityFuturesTradingCommission,Washington. Lin, W., Engle, R., and Ito, T. (in press), "Do Bulls and Bears Move Across Borders? International Transmissionof Stock Returnsand Volatilityas the WorldTurns," Reviewof Financial Studies, 7. Mittoo, U. (1992), "ManagerialPerceptions of the Net Benefits of Foreign Listing: CanadianEvidence,"unpublishedmanuscript,University of Manitoba,Dept. of Finance. Evidence on Integration the CanadianStock in (1993), "Additional Journalof Finance, 47, 2035-2053. Market," J. Muthuswamy, (1989), "Nonsynchronous Tradingand the Index AutocorrelationProblem," Ph.D.dissertation,Universityof Chicago, unpublished Graduate School of Business. New YorkStock Exchange(1991), Fact Book, New York:Author. Reportfor Market Volatilityand Investor ConfidencePanel (1990), New York:New YorkStock Exchange. Reportof the PresidentialTaskForceon MarketMechanisms(1988), Washington: U.S. GovernmentPrintingOffice. Crashof October, 1987,"Financial AnRoll, R. (1988), "TheInternational 34-54. alysts Journal,September/October, MarketLinks, and TheirIm(1989), "PriceVolatility,International plications for RegulatoryPolicies," Journal of Financial Services Research, 3, 211-246. Runkle, D. (1987), "VectorAutoregressionsand Reality,"Journal of Business & EconomicStatistics,5, 437-442. in Schwert,W., and Seguin, P. (1990), "Heteroscedasticity Stock Returns," Journal of Finance,45, 1129-1156. and Sims, C. A, (1980), "Macroeconomics Reality," Econometrica,48, 1-48. Switzer, L. (1986), "The Benefits and Costs of Listing CanadianStocks in in U.S. Markets," CorporateStructure,Finance and Operations,ed. L. Sarna,Toronto:CarswellCompany,pp. 241-254. TorontoStock Exchange(1990), TSEReview,December, Toronto:Author. von Furstenberg, and Jeon, B. (1989), "International Stock Price MoveG., ments: Links andMessages,"BrookingsPapers on EconomicActivity,1, 125-179. Zellner, A. (1984), Basic Issues in Econometrics, Chicago: Universityof Chicago Press.

Vous aimerez peut-être aussi