ARTICLE IN PRESS

Physica A 360 (2006) 422–444

www.elsevier.com/locate/physa

Asymmetric responses of international stock

markets to trading volume
Richard Gerlacha, Cathy W.S. Chenb,,
Doris S.Y. Linb, Ming-Hsiang Huangc
a
School of Mathematical and Physical Sciences, University of Newcastle, Australia
b
Graduate Institute of Statistics and Actuarial Science, Feng Chia University, Taichung 407, Taiwan
c
Department of Business Administration, National Changhua University of Education, Taiwan
Received 28 February 2005; received in revised form 30 May 2005
Available online 18 July 2005

Abstract

The major goal of this paper is to examine the hypothesis that stock returns and return
volatility are asymmetric, threshold nonlinear, functions of change in trading volume. A minor
goal is to examine whether return spillover effects also display such asymmetry. Employing a
double-threshold GARCH model with trading volume as a threshold variable, we find strong
evidence supporting this hypothesis in five international market return series. Asymmetric
causality tests lend further support to our trading volume threshold model and conclusions.
Specifically, an increase in volume is positively associated, while decreasing volume is
negatively associated, with the major price index in four of the five markets. The volatility of
each series also displays an asymmetric reaction, four of the markets display higher volatility
following increases in trading volume. Using posterior odds ratio, the proposed threshold
model is strongly favored in three of the five markets, compared to a US news double
threshold GARCH model and a symmetric GARCH model. We also find significant nonlinear
asymmetric return spillover effects from the US market.
r 2005 Elsevier B.V. All rights reserved.

Keywords: Asymmetry; Double threshold GARCH; MCMC methods; Model selection; Trading volume
change

Corresponding author. Tel.: +886 4 24517250x4412; fax: +886 4 24517092.

E-mail address: chenws@fcu.edu.tw (C.W.S. Chen).

0378-4371/$ - see front matter r 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2005.06.045
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 423

1. Introduction

As a result of globalization, de-regulation and advances in information

technology, the modern theory of international markets has switched from the
traditional view of market segmentation to the concept of market integration. A
substantial amount of research emphasizes the co-movements of international stock
markets and explores the dynamics of return co-variances and spillover effects
between markets. For example, Jaffe and Westerfield [1] provide empirical evidence
of a significant direct spillover effect among some national stock markets, while Eun
and Shim [2] find return spillover effects among national markets and an influential
role of the US market on the cross-country market index series. Ross [3] argues
further that information from one stock market can be incorporated into the
volatility process of other stock markets. Hamao et al. [4], Theodossiou and Lee [5],
Chiang and Chiang [6] and Martens and Poon [7] subsequently find supporting
evidence for volatility spillover among major stock markets.
There is also substantial evidence in the literature that stock markets react
asymmetrically to market news results. This phenomenon was first discovered by
Black [8] and Christie [9] who discuss the leverage effect as the cause of higher
volatility following negative stock returns; similar to the market over-reaction
hypothesis discussed in Ref. [10]. There is also the volatility feedback hypothesis
which says that higher volatility causes stock prices to fall. Many models have been
developed to capture types of asymmetric behavior, most fit into the threshold
GARCH framework. Glosten et al. (1993) employ a threshold GJR-GARCH model
and find evidence that local negative market news causes increased market volatility
[47]. This finding is confirmed in studies by Koutmos [11], Nam et al. [10] and Brooks
[12], using double threshold models. These papers also find evidence of faster mean
reversion dynamics following local market bad news. More recently, Chen et al. [13]
employ a double-threshold GARCH model with a US market threshold variable, to
explore the dynamics of daily stock-index returns for six international markets from
1985 to 2001. Their results provide strong evidence supporting an asymmetric
nonlinear spillover effect from the US market to other markets in Europe and Asia.
The US market news transmits asymmetrically, around a threshold value, to each of
the national stock markets considered with average volatility in each national market
much higher following bad US news. Further, Chen and So [14] explore a range of
international markets to use as exogenous threshold values in a double threshold
GARCH model. They find that the Japanese market has little spillover or threshold
nonlinear effect on mean returns in Asian markets, in comparison with the US
market. These results are supported by Wang and Firth [15], who find the emerging
market of China does not exhibit significant spillover effects to other markets
worldwide, including those in Asia. The US return has thus evolved as the preferred
threshold variable in the examination of return spillover and nonlinear asymmetric
effects.
However, this previous work ignores the possible correlation between the stock
price or return and trading volume. Numerous financial studies have documented
this important relationship. Clark [16] and Epps and Epps [17] suggested that trading
 ARTICLE IN PRESS
424 R. Gerlach et al. / Physica A 360 (2006) 422–444

volume is a good proxy for information arrival from the capital market. The
hypothesis has been further supported by empirical evidence; Lamoureux and
Lastrapes [18], Kim and Kon [19], Andersen [20], Gallo and Pacini [21] found the
same effect for the US stock market; Omran and McKenzie [22] observed this effect
for the UK stock market; Bohl and Henke [23] reported similar evidence for the
Polish stock market.
Ying [24] was the first to provide strong empirical evidence supporting an
asymmetric relation between trading volume and price-change. By investigating six
series of daily data from NYSE, Ying made the following conclusions: a small
trading volume is usually accompanied by a fall in price; a large volume is usually
accompanied by a rise in price; and a large increase in volume is usually
accompanied by either a large rise in price or a large fall in price. These
propositions lay an important foundation for our nonlinear asymmetric
hypothesis and illustrate that a linear relationship between price return and
volume, and/or volatility and trading volume, may not be sufficient to capture the
true relationship. This hypothesis is also documented by Karpoff [25] in an
extensive survey of research into the relationship between stock–price change and
trading volume. Karpoff suggests several reasons why the volume–price change
relationship is important and provides evidence to support the asymmetric
volume–price change hypothesis. His asymmetric hypothesis implies that the
correlation between volume and price change is positive when the market trend is
going up, but that this correlation is negative when the market trend is
downwards. This is again important and highlights that we should not simply
add a linear exogenous volume term to the mean equation in a GARCH model for
stock returns. To capture the possible nonlinearity we will consider an asymmetric
piecewise linear relationship between price (return) and volume, as can be captured
by threshold models [26].
Departing from traditional work that focused on the contemporaneous relation
between return and trading volume, Chordia and Swaminathan [27] examine the
causal relationship and the predictive power of trading volume on the short-term
stock return. Their empirical evidence suggests that volume plays a substantial role
in the dissemination of national market-wide information. In a dynamic context, Lee
and Rui [28] utilize the GARCH(1,1) model to investigate the relationship between
stock returns and trading volume using the New York, Tokyo and London stock
markets. Their empirical results suggest that US financial market variables, in
particular US trading volume, have extensive predictive power in both the domestic
and cross-country markets, after the 1987 market crash. Moosa et al. [29] employ a
bivariate VAR model and find significant mean level asymmetry in the price–volume
relationship for the future market in crude oil prices; they did not consider a
heteroscedastic model and they enforced the threshold variable to be zero. The above
findings further enforce our belief that a consideration of trading volume as a
threshold variable might add to the understanding of capital market behavior in
general. The major objective of this study is thus to investigate whether stock
returns, volatility and international return spillover effects react in a threshold
nonlinear fashion to changes in trading volume, in five international markets. As far
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 425

as the authors are aware, this is the first time that trading volume has been employed
as a threshold variable in a dynamic heteroscedastic model of stock returns.
To address the above issue, we follow Chen et al. [13] and employ a double-
threshold autoregressive GARCH model, using a Bayesian estimation approach
through Markov chain Monte Carlo (MCMC) methods. This new model allows
threshold nonlinearity in both the mean and volatility processes to be driven by
lagged changes in trading volume. We compare this model to the US news threshold
model in Ref. [13] and a symmetric GARCH model. We utilize the posterior odds
ratio, the standard Bayesian model comparison technique, as in Ref. [30], to
determine which is the most favored model in each market. We also employ the
causality tests of Ref. [29] to motivate and validate our models. Building on the work
of Ying [24], Karpoff [25] and Moosa et al. [29], our empirical results provide strong
evidence supporting the return-volume nonlinear threshold relation. The stock
return, volatility and international return spillover do react asymmetrically, around
a threshold value of change in trading volume, in the five markets considered. Our
findings shed new light on the application of trading volume in the market
integration literature and develop a new avenue for asset pricing in a multi-market
framework.
The remainder of this study proceeds as follows. Section 2 describes the data used
in this study and presents some statistical properties of the stock returns in a
standard GARCH(1,1) specification. Section 3 discusses the Bayesian methods for
estimation and present the double threshold models considered. Section 4 discusses
the estimated results for each model, and compares the findings with the existing
literature. Section 5 discusses the Bayesian model comparison methods and discusses
the findings for each market. Section 6 contains concluding remarks.

2. Data and basic statistics

The analysis undertaken in this article is based on daily closing prices and trading
volume of five stock market indexes: the Korean Composite Price Index (South
Korea), Thailand SET index (Thailand), the Taiwan Stock Exchange Weighted
Stock Index (Taiwan), the CAC 40 (France), and the FTSE 100 (UK). We also
employ the US Dow Jones Industrial Average as a benchmark. The data, obtained
from Datastream International, run from January 1, 1994 to November 26, 2003.
The market return at time t is defined by Rt ¼ ðlnðpt Þ
lnðpt
1 ÞÞ  100%, where pt is
the price index at time t. Time series plots for each market return series are given in
Fig. 1. We also consider the series of logarithm of trading volume over the same time
period in each market. The variable change in volume at time t is defined as
V t ¼ c  ðlnðvt Þ
lnðvt
1 ÞÞ  100%, where vt is the trading volume at time t. The
multiple c is used here because the scales of the market return and percentage change
in volume series are very different. We thus scale the change in volume series by a
constant c, chosen to allow the series V to have a similar scale to the return series R.
In order to form a statistically adequate model, the variables should first be
checked as to what extent they can be considered stationary. The results of applying
 ARTICLE IN PRESS
426 R. Gerlach et al. / Physica A 360 (2006) 422–444

TWSI SETI

10
5
5
0
0

-5 -5

-10 -10
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500

KSll CAC
10
5
5

0 0
-5
-5
-10

0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500

FTSE DJIA
6 6
4 4
2 2
0 0
-2 -2
-4 -4
-6 -6
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500

Fig. 1. Time series plots for the five market returns and the returns of US Dow Jones Industrial Average
as a benchmark.

an augmented Dickey–Fuller (ADF) test to the logs of the price, returns, log-volume
and volume changes are shown in Table 1. We conclude from the p-values in this
table that the log-price and log-volume series contain unit roots, while the returns
and volume change series appear stationary in mean. To provide a general
understanding of the nature of each market’s returns, we summarize the daily return
and scaled volume change statistics in Table 2. Perhaps due to the Asian financial
crisis within this time period, all three Asian markets exhibit negative mean returns,
while the UK, France and the US have positive mean returns over this period. All six
return series exhibit the standard property of asset return data: they have fat-tailed
distributions, as indicated by the positive coefficient of excess kurtosis. This
characteristic is also shown by the highly significant Jarque–Bera normality test
statistics, a joint test for the absence of skewness and kurtosis. All series are
negatively skewed, except Thailand.
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 427

Table 1
ADF tests for a unit root on log-prices, return, log-volume, and log-volume changes

Index Taiwan Thailand South Korea France UK

weighted SET SE composite CAC 40 FTSE 100

ADF statistic
1.8593
1.8327
2.0664
1.1472
1.4719

(log-price)
ADF statistic
32.6736
32.5623
35.5427
35.9939
37.3928
(return) ðo0:01Þ ðo0:01Þ ðo0:01Þ ðo0:01Þ ðo0:01Þ
ADF statistic
3.1763
2.4134
1.6049
2.2453
2.1092
(log-volume)
ADF statistic
49.3402
23.3543
21.4808
43.8022
31.6039
(volume changea) ðo0:01Þ ðo0:01Þ ðo0:01Þ ðo0:01Þ ðo0:01Þ
a
V t ¼ c  ðlnðvt Þ
lnðvt
1 ÞÞ  100%, where vt is the trading volume at time t and c ¼ 0:1.

Table 2
Summary statistics: Stock Index Returns and volume change

Return Taiwan Thailand South Korea France UK US

weighted SET SE composite CAC 40 FTSE 100 DJI

Mean
0.0038
0.0422
0.0049 0.0162 0.0103 0.0384

Std. 1.7156 1.8673 2.1918 1.4830 1.1549 1.1220
Skewness
0.0824 0.4619
0.0647
0.0658
0.1574
0.2648
Kurtosis 1.9856 3.7963 3.0240 2.1205 2.3953 4.0236
Minimum
9.9360
10.0280
12.8047
7.6780
5.8853
7.4549
Maximum 8.5198 11.3495 10.0238 7.0023 5.9026 6.1547
Observations 2408 2421 2401 2491 2498 2491
J–B testa 398.3084 2039.7620 1298.5400 453.2870 714.1045 1957.0780
(o0.01) (o0.01) (o0.01) (o0.01) (o0.01) (o0.01)

Volume changeb
Mean
0.0015 0.00093 0.0133 0.0109 0.0078 0.0045
Std. 2.3952 3.7479 2.1108 3.8009 2.9139 2.4186
Skewness 0.1414 0.2297 0.3449 0.0234 0.0523 0.0889
Kurtosis 1.1420 1.3255 3.4734 2.7646 4.7098 5.6539
Minimum
12.8846
24.6453
9.7940
25.9523
20.0322
15.0145
Maximum 12.7297 18.1687 14.7527 23.5100 19.7815 18.3530
Observations 2408 2421 2401 2491 2498 2491
J–B testa 137.7552 197.0956 1247.7470 788.9356 2298.3250 3301.8310
(o0.01) (o0.01) (o0.01) (o0.01) (o0.01) (o0.01)
a
The Jarque–Bera normality test statistic and p-value are listed. The lower p-value indicates the null
hypothesis of normality can be rejected.
b
V t ¼ c  ðlnðvt Þ
lnðvt
1 ÞÞ  100%, where vt is the trading volume at time t and c ¼ 0:1.

We choose to employ autoregressive GARCH-type models for analysis. It is of

interest to examine whether the stock returns on price indexes in the advanced
markets are affected by international market return spillover. Possible choices of
exogenous factor here include the US, Japan and the emerging market in China.
 ARTICLE IN PRESS
428 R. Gerlach et al. / Physica A 360 (2006) 422–444

However, Chen and So [14] illustrate that the US market dominates Japan in terms
of return spillover effects in Asian markets. China’s stock market has recently
become more correlated with major international financial markets, but it is still
an emerging market in nature. Previous literature suggests that return spillover
effects are unidirectional, transmitted from advanced financial markets to less-
developed financial markets. This argument was further confirmed in recent work by
Wang and Firth [15]. Their findings suggest the existence of only unidirectional
return spillover effect from the US to China. Thus, for the time being it seems
inappropriate to consider China as an exogenous factor. We thus consider only the
US stock market for return spillover effects. It is important when doing this to note
that stock markets in different countries operate in different time zones with
subsequently different opening and closing times. To compare the realized returns
for international markets in a given calendar day in different real-time periods is
difficult. However, the stock trading at New York City (the Dow Jones Index in the
US market) is the last one to close among the international stock exchanges under
investigation. So the closing news in the US market at day ðt
1Þ will have sufficient
time to transmit to the Southeast Asian markets and various European markets,
see Ref. [13].
We employ standard Granger bi-directional causality tests to examine, using
a bivariate VAR model between the US and each market separately and
employing a regression F-statistic, whether each domestic market is signifi-
cantly affecting the US (in mean) and/or vice versa. The results from these tests
are presented in Tables 3. Clearly, the US is significantly affecting each
market’s return in the mean, while each market has a much weaker and mostly
insignificant effect on the US market return. For this reason, we restrict ourselves to
a univariate, as opposed to a bivariate, model and include a 1-day lagged cross-asset
return in the mean equation. We also assume the conditional variance is a

Table 3
Causality tests for stock returns in global markets

Causal direction Lag 1 p-value Lag 1–2 p-value

F test F test

US ! FR 171.99 0.0000 87.80 0.0000

US ! UK 228.45 0.0000 122.18 0.0000
US ! SK 142.99 0.0000 71.47 0.0000
US ! THAI 76.72 0.0000 38.48 0.0000
US ! TWN 112.36 0.0000 56.85 0.0000
FR ! US 3.70 0.054 3.134 0.044
UK ! US 2.712 0.100 2.561 0.077
SK ! US 0.037 0.847 2.945 0.053
THAI ! US 0.187 0.666 0.212 0.809
TWN ! US 0.309 0.578 0.517 0.596
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 429

GARCH(1,1) process
Model 1 Rit ¼ f0 þ f1 Rit
1 þ c1 Rjtþm
1 þ at ,
pffiffiffiffi
at ¼ ht e t ,
ht ¼ a0 þ a1 a2t
1 þ b1 ht
1 ; et tðnÞ , ð1Þ
where m is the time difference between i and j markets, Rit and Rjt are stock returns
from countries i (i ¼ Taiwan, Thailand, South Korea, France and the UK) and j
(the US), respectively; ht is the conditional variance; f0 , f1 , c1 , a0 , a1 and b1 are
unknown parameters; and at is a random error term following a standardized t-
distribution with n degrees of freedom, mean 0 and variance 1. As in Ref. [13] we
estimate the degrees of freedom parameter n as part of the Bayesian estimation, see
details in Appendix.

2.1. Bayesian estimation and preliminary results

We employ Bayesian MCMC methods to estimate the model parameters, as in

Ref. [13]; details are given in an Appendix. Briefly, MCMC is an iterative sampling
scheme that, in turn, samples parameter values as follows:

 Sample f0 ; f1 ; c1 jointly conditional on the data R and the other current

parameter values.
 Sample a0 ; a1 ; b1 jointly conditional on the data R and the other current
parameter values.
 Sample n conditional on the data R and the other current parameter values.

The Bayesian approach has the advantages over classical statistical methods of
simultaneous inference, and incorporation of any prior information, on all model
parameters. The priors used are described in Appendix, but are mostly chosen to be
uninformative over the standard region ensuring stationarity and finite, positive
variance in the model.
Estimates of Model 1, for each of the five series, are presented in Table 4. This
table contains the posterior means, together with standard errors, for the unknown
parameters (f0 , f1 , c1 , a0 , a1 , b1 ). Here we see that, as expected, indeed there are
significant spillover effects from the US market to each national stock index, as
measured by the significantly positive estimates of c1 . We note the standard
empirical result, high level of persistence in volatility in each market, with estimates
of a1 þ b1 between 0.96 and 0.99 for each market. There is significant negative first-
order persistence in mean in the UK and French markets, but significant positive
persistence in mean in the Thailand and South Korea markets, after accounting for
the exogenous US mean effect. Finally, the degree of freedom parameter is especially
low in the three Asian markets, indicating a higher level of kurtosis in the error
distribution compared to the UK and France. This justifies the use of the t-
distribution and indicates the tails of the conditional error distribution are
significantly fatter than a normal distribution.
 ARTICLE IN PRESS
430 R. Gerlach et al. / Physica A 360 (2006) 422–444

Table 4
Bayesian estimates for Model 1

Taiwan Thailand South Korea France UK

f0 0.0125
0.0445
0.0145 0.0266 0.0241

(0.0303) (0.0301) (0.0325) (0.0244) (0.0171)
f1 0.0108 0.1056 0.0675
0.0928
0.0910
(0.0207) (0.0214) (0.0204) (0.0224) (0.0213)
c1 0.3035 0.2234 0.3169 0.3416 0.2780
(0.0290) (0.0276) (0.0351) (0.0274) (0.0206)

a0 0.1202 0.0967 0.0312 0.0218 0.0112

(0.0463) (0.0365) (0.0125) (0.0078) (0.0037)
a1 0.0961 0.1164 0.0691 0.0636 0.0792
(0.0190) (0.0214) (0.0142) (0.0092) (0.0116)
b1 0.8659 0.8588 0.9253 0.9260 0.9124
(0.0304) (0.0284) (0.0152) (0.0108) (0.0125)
n 6.5699 7.1010 9.6063 20.4054 20.7016
(0.9262) (1.0312) (1.6805) (8.4478) (8.7777)

Figures in brackets represent posterior standard errors for each parameter.

3. Double TAR-GARCH models and Bayesian methods

In forming our models, we are motivated by the causality tests of Moosa et al. [29],
useful for vector autoregressive models in order to determine directions of causality
or dependence. While these tests make strict assumptions: the threshold variable is
exactly zero, volatility is constant and errors are Gaussian; they do allow asymmetric
relationships in the mean equation. They are thus useful as an initial guide in
asymmetric model choice, while being much less general than the results in this
paper. The results of these tests are presented in Table 5. Clearly, the results indi-
cate that both positive and negative US market returns significantly affect returns
in all five domestic markets. This motivates the choice of the US as both an
exogenous variable in the mean equation and a threshold variable. The model we
employ is a Double TAR-GARCH model, which generalizes the DT-ARCH model
of Li and Li [31] and Chen [32]. This model is motivated by several nonlinear
characteristics commonly observed in practice, such as asymmetry in declining and
rising patterns of a process. It uses piecewise linear models to obtain a better
approximation of the conditional mean and conditional volatility equations based
on a threshold variable.
We specify the first two-regime asymmetric heteroscedastic model as
8 ð1Þ ð1Þ j
< f0 þ fð1Þ i
1 Rt
1 þ c1 Rtþm
1 þ at ; Rjtþm
d pr1 ;
i
Model 2 Rt ¼
: fð2Þ þ fð2Þ Ri þ cð2Þ Rj Rjtþm
d 4r1 ;
0 1 t
1 1 tþm
1 þ at ;
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 431

Table 5
Causality tests without assuming symmetry for stock returns in global markets

Causal direction Lag 1 p-value Lag 1–2 p-value

F test F test

USþ ! FR 55.185 0.0000 27.103 0.0000

US
! FR 76.530 0.0000 41.704 0.0000
USþ ! UK 64.191 0.0000 35.989 0.0000
US
! UK 106.240 0.0000 56.563 0.0000
USþ ! SK 35.217 0.0000 14.397 0.0000
US
! SK 60.712 0.0000 35.548 0.0000
USþ ! THAI 19.508 0.0000 6.333 0.002
US
! THAI 63.147 0.0000 32.180 0.0000
USþ ! TWN 23.527 0.0000 12.634 0.0000
US
! TWN 55.225 0.0000 26.672 0.0000

The asymmetric variables, where RUS

t is the return on the US market, are defined as
( (
þ
RUS
t if RUS
t X0;
RUS
t t p0 ;
if RUS
USt ¼ USt ¼
0 otherwise; 0 otherwise :

pffiffiffiffi
at ¼ ht et ; et tðnÞ ,
8 ð1Þ
< a0 þ að1Þ 2 ð1Þ
1 at
1 þ b1 ht
1 ; Rjtþm
d pr1 ;
ht ¼ ð2Þ
: að2Þ þ að2Þ a2 þ bð2Þ Rjtþm
d 4r1 ;
0 1 t
1 1 ht
1 ;

where m is the time different between i and j markets. The definition of Rit and Rjt is
the same as Model 1. This is the model considered in Ref. [13] where the mean and
volatility equations can react asymmetrically to good or bad return news from the
US market. This model has the advantage that the threshold parameter, usually fixed
at r1 ¼ 0, and the delay parameter, usually fixed at d ¼ 1, are not specified but can be
estimated from the data simultaneously with the other model parameters, in a
Bayesian framework.

3.1. Threshold trading volume DT-GARCH model

Model 2 above ignores any potential relationship between trading volume and
price movements on the stock market. Financial theory suggests that volume and
stock return are strongly associated and many authors have found empirical results
to back this up, see Refs. [27,28] among others. To motivate our model, we again
employ the Moosa et al. [29] causality tests, without assuming symmetry, applied to
each market’s volume-change and return series together in a bivariate VAR model.
Results are presented in Table 6. Here we see, assuming a zero threshold, constant
volatility and Gaussian errors (these assumptions are violated for these markets, see
Table 8) that results are somewhat mixed, however, this test is meant as an initial
guide only. Both positive and negative changes in trading volume significantly affect
 ARTICLE IN PRESS
432 R. Gerlach et al. / Physica A 360 (2006) 422–444

returns in South Korea (5% level), while Taiwan is affected strongly by only positive
volume-change at 5% level. Also, at the 10% level, decreasing volume affects returns
on the French and Taiwanese markets. However, the UK and Thailand seem not be
significantly affected in the mean equation by trading volume-change. We also
observed the standard causality tests, not presented here to save space, which
illustrated that volume was not a significant exogenous factor in these markets.
These results suggest that volume may potentially affect market returns asymme-
trically and motivates the choice of volume as a threshold variable, but not as an
exogenous variable in the mean equation.
We thus develop a new model in this paper, motivated by the work of Ying [24]
and Karpoff [25], attempting to capture the potentially asymmetric reaction of
stock prices to changes in trading volume. We introduce a Double TAR-GARCH
model using change in trading volume as the threshold variable. This is the first
time that trading volume has been employed as a threshold variable in an
asymmetric heteroscedastic dynamic model of stock returns. The model is posited
as follows:
8 ð1Þ ð1Þ j
< f0 þ fð1Þ i
1 Rt
1 þ c1 Rtþm
1 þ at ; V it
d pr2 ;
i
Model 3 Rt ¼
: fð2Þ þ fð2Þ Ri þ cð2Þ Rj V it
d 4r2 ;
0 1 t
1 1 tþm
1 þ at ;
pffiffiffiffi
at ¼ ht et ; et tðnÞ ,
8 ð1Þ
< a0 þ að1Þ 2 ð1Þ
1 at
1 þ b1 ht
1 ; V it
d pr2 ;
ht ¼ ð3Þ
: að2Þ þ að2Þ a2 þ bð2Þ ht
1 ; V i 4r2 ;
0 1 t
1 1 t
d

where m is the time difference between markets i and j. Here, V it ¼ c  ðlnðvit Þ

lnðvit
1 ÞÞ  100% is the change in trading volume, where vit is the trading volume of
market i at time t and c ¼ 0:1. This model will allow us to test the major hypothesis
about asymmetries between mean and volatility of returns in reaction to changes in
trading volume, and will allow us to empirically confirm the findings and theories of
Ying [24] and Karpoff [25] in a GARCH-type dynamic volatility setting. The model
allows trading volume to drive or influence stock returns in a piecewise linear
fashion, allowing different return behavior following increasing or decreasing
volume. Again, the threshold change in volume level and delay lag parameter are
estimated simultaneously with the unknown model parameters, a significant
advantage of the Bayesian approach to estimation.
We employ a similar MCMC method to that for Model 1, as in Ref. [13], to
estimate both Models 2 and 3. See Appendix for some details of this approach.
Briefly, this involves the same iterative sampling scheme as for Model 1 (the two
steps above are performed separately for each regime) with the addition of the steps:

 Sample rk conditional on the data R and the other current parameter values.
 Sample d from it’s discrete posterior distribution conditional on the data R and
the other current parameter values.
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 433

The prior distributions on model parameters here are described in Appendix, but are
mostly chosen to be uninformative over the region ensuring stationarity and finite
variance in each model.

4. Results for threshold models

Results from Model 2, with a US news threshold parameter, are presented in

Table 7. These illustrate a clear nonlinear asymmetric reaction to US return news in
each of the five markets, similar to results found in Ref. [13]. Firstly in each market,
the mean reversion is faster in the negative regime, following bad news from the US
market, i.e., fð1Þ ð2Þ
1 of1 , with these parameters being significantly different to zero in
all five markets following good or positive US return news. Again, this local
persistence in return is positive in the three Asian markets, but negative in UK and
France. In addition, the spillover effect from the US market is significantly greater
following bad US news, i.e., cð1Þ ð2Þ
1 4c1 in each market. We note that the spillover
effect from the US market is significantly stronger than the local return persistence in
each of the five markets, as indicated by jc1 j4jf1 j and jc1 j4jf0 j in each regime.
Also, the persistence in volatility, as measured by a1ðjÞ þ bðjÞ
1 is higher following bad
news in each market.
Table 9 highlights the asymmetric reactions in the five markets. Here, we see
that volatility level increases by as little as 7.5 times and as much as 68 times
following bad US news compared to that following good US news. Thailand is the
exception, with volatility level appearing stable regardless of the news from the US.
In fact, the asymmetric reaction in volatility appears to be much stronger in the UK
and France, compared to the three Asian markets. We also see that the mean return
estimate in each market has decreased from the positive to negative threshold, i.e.,
following bad news from the US. This mean return decrease is from positive return
following good US news to negative return following bad US news, in three of
the markets.
We note that the threshold parameters in Table 7 are estimated to be negative and
significant in four of the five markets. This illustrates that small negative US returns
do not induce an asymmetric reaction and are not perceived as ‘bad news’ by these
five national markets. However, significant negative US returns, that is below each
threshold estimate, are needed to induce the asymmetric reaction in each market.
Thailand is the exception to this rule, although the positive estimated threshold
parameter is not significantly different to 0 in this case. We also note in Table 9 that
the posterior mode for the delay parameter d is 1 in four of the five markets, but is
two days for the Taiwan market. The degrees of freedom estimates are similar to
those obtained for Model 1.
Results for Model 3 are presented in Tables 8 and 9. Results here illustrate a clear
nonlinear asymmetric reaction to change in trading volume in each of the five
markets. Firstly in each market, the mean reversion is faster following negative
changes or decreases in trading volume, i.e., fð1Þ ð2Þ
1 of1 in four of the five markets,
excepting Taiwan. The international spillover effect from the US market is greater in
 ARTICLE IN PRESS
434 R. Gerlach et al. / Physica A 360 (2006) 422–444

the positive regime, following increases in trading volume, i.e., cð2Þ ð1Þ
1 4c1 in each
market, except Thailand. We note that the spillover effect from the US market is
again significantly stronger than the local return persistence in each of the five
markets, as indicated by jc1 j4jf1 j and jc1 j4jf0 j in each regime. In this case, the
persistence in volatility (a1 þ b1 ) is mostly higher following positive trading volume
changes.
Table 9 again highlights the asymmetric nonlinear reactions to trading volume
change in the five markets. Here, we see that average volatility increases by as little as
1.5 times (South Korea) up to 100 times (UK, France) following increases in trading
volume, compared to that following declining trading volume. This is consistent with
the hypothesis of Ying [24] who suggested that large increases in volume lead to large
increases or decreases in price and hence to volatility. Taiwan is the notable
exception to this rule, with volatility level significantly decreasing as trading volume
increases. This distinctive empirical result in the Taiwanese market is interesting in
light of the under and over reaction hypothesis, in the modern theory of behavior
finance, proposed by Daniel et al. [33] and Barberis et al. [34]. Their hypothesis states
that individual investors tend to be over-confident and over-reliant on their own
private information, in general, resulting in over-reaction to public information, on
the basis that they perceive this to confirm their individual beliefs. Proposition 3 in
Ref. [33] states that ‘overconfidence can increase or decrease volatility’ around public
information events. The result for Taiwan thus agrees with this theory, as indeed the
volatility has changed following news on trading volume, it just changes in a
different way to that in the other four markets. The particular Taiwan market
behavior has been well documented in previous studies of price–volume relationships
with the most plausible explanation falling in the realm of market imperfection and
investor irrationality [35,36]. The Taiwanese stock market’s prolonged behavioral
characteristic is very short-run speculation, which may also contribute to the
inherent political jitters in Taiwan, due to the Chinese Missiles incident, and
economic crisis in 1997–1998. In addition, the majority (approximately over 92%) of
investors are uninformed traders and individual investors. Thus, we suspect that this
speculative behavior, incorporated with investor irrationality, caused the particular
information flow pattern and price–volume relationship in the Taiwan stock market.
On close examination of the mean and volatility of returns under Model 3 in Table 6,
the evidence suggests that in regime 1 investors over-react, proxied by the relatively
high volatility and lower mean return, to the negative volume change. On the other
hand, the result in regime 2 suggests an under-reaction, with lower volatility, but in a
bullish fashion, with increasing volume of trade.
The estimates of average return in each market have increased following increases
in trading volume, excepting Thailand. The greatest increases in mean return are in
South Korea and Taiwan. The particular finding in the Thailand stock market is also
of interest. The turmoil of the Asian currency crisis in July 1997 led the economy of
Thailand into a well-documented seven-year stagnation, falling into a long sluggish
bear market until June of 2003, except for a short rebound between August 1998 and
June 1999. This phenomenon is reflected by the largest, in magnitude, negative mean
return ð
0:0422Þ among our five sample markets in Table 2 and also a large negative
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 435

Table 6
Causality tests without assuming symmetry for domestic returns and volume changes

Causal direction Lag 1 p-value Lag 1–2 p-value

F test F test

Vþ ! FR 0.226 0.634 0.363 0.696

V
! FR 3.375 0.066 0.432 0.239
Vþ ! UK 0.247 0.619 0.103 0.902
V
! UK 1.263 0.261 0.732 0.481
Vþ ! SK 4.743 0.029 3.152 0.043
V
! SK 4.712 0.030 1.296 0.274
Vþ ! THAI 1.752 0.186 0.914 0.401
V
! THAI 0.538 0.463 0.422 0.656
Vþ ! TWN 6.607 0.010 5.453 0.004
V
! TWN 2.925 0.087 3.372 0.034

The asymmetric variables are defined, where V t is volume-change, as

( 
V t if V t X0; V t if V t p0 ;
Vþt ¼ V

t ¼
0 otherwise; 0 otherwise :

volume threshold value of
1:5375 in Table 7. Financial theory suggests that trading
volume tends to expand in the direction of the major market trend. If the observed
stock market is up in trend, the volume should increase with price increase, and the
volume should diminish as price decreases. Conversely, if the major trend of the
stock market is down, the volume should diminish as price rallies up but expand as
price dips down. The relatively large magnitude of negative mean return ð
0:0385Þ of
regime 2 under Model 3 in Table 6 is consistent with these findings in the literature, a
large increase in volume is usually accompanied by either a large rise in price or a
large dip in price, the dip here associated with a period of poor market return
performance in Thailand.
These results suggest very clear and significant asymmetric and nonlinear reactions
in each market to changes in trading volume and support the original work by Ying
[24] and Karpoff [25]. The results suggest that the mean return in each market is
higher (except Thailand) and the volatility level is much higher (except Taiwan) when
trading volume increases:
We note that the threshold parameters in Table 8 are estimated to be positive and
significant in three of the five markets. This illustrates that small increases in trading
volume do not induce an asymmetric reaction and are not perceived as significant by
these five national markets. However, significant positive changes in volume, that is
above each positive threshold estimate, are needed to induce the asymmetric reaction
in Taiwan, South Korea and France. Thailand is again the exception to this rule, as
explained above. We also note in Table 9 that the posterior mode for the delay
parameter d is 1 day in all five markets. The degrees of freedom estimates are similar
to those obtained for Model 2.
 ARTICLE IN PRESS
436 R. Gerlach et al. / Physica A 360 (2006) 422–444

Table 7
Bayesian estimates for Model 2

Taiwan Thailand South Korea France UK

f0ð1Þ
0.0051
0.0707 0.2828 0.0561 0.1072

(0.0697) (0.0621) (0.1372) (0.0756) (0.0570)
f1ð1Þ 0.0780 0.0981 0.0153
0.1119
0.1198
(0.0377) (0.0298) (0.0418) (0.0394) (0.0367)
c1ð1Þ 0.3030 0.2497 0.5464 0.404 0.3827
(0.0502) (0.0537) (0.1027) (0.0691) (0.0497)

f0ð2Þ 0.0864 0.083
0.0435 0.0414 0.0091

(0.0410) (0.1075) (0.0455) (0.0400) (0.0302)
f1ð2Þ 0.1429 0.1034 0.0895
0.0814
0.0824
(0.0257) (0.0366) (0.0244) (0.0269) (0.0279)
c1ð2Þ 0.0890 0.1808 0.3224 0.2855 0.2623
(0.0392) (0.0751) (0.0550) (0.0502) (0.0365)

a0ð1Þ 0.0869 0.0892 0.1614 0.1234 0.0514

(0.0965) (0.0568) (0.0734) (0.0315) (0.0152)
a1ð1Þ 0.1134 0.1129 0.1036 0.0939 0.1066
(0.0284) (0.0282) (0.0248) (0.0197) (0.0231)
b1ð1Þ 0.8608 0.8639 0.8811 0.8985 0.8871
(0.0400) (0.0364) (0.0296) (0.0213) (0.0240)

a0ð2Þ 0.1897 0.2018 0.019 0.0156 0.0051

(0.0524) (0.1027) (0.0133) (0.0117) (0.0040)
a1ð2Þ 0.1424 0.1464 0.0734 0.043 0.0542
(0.0234) (0.0342) (0.0197) (0.0121) (0.0159)
b1ð2Þ 0.8129 0.801 0.9149 0.9114 0.9117
(0.0374) (0.0481) (0.0219) (0.0161) (0.0192)

r1
0.2410 0.2697
0.4208
0.2269
0.1735

(0.1332) (0.3178) (0.1197) (0.0697) (0.0722)
n 6.6773 7.2158 9.8639 21.9762 25.3133
(0.9553) (1.0223) (1.8114) (13.9071) (18.6993)

Figures in brackets represent posterior standard errors for each parameter. r1 : threshold value for US
returns.

5. Model comparison

We use model selection to determine whether the threshold trading volume

model, Model 3, adds any value over the simple and DTAR-GARCH
models, Models 1 and 2. To compare models pairwise we use the posterior
odds ratio which is the ratio of integrated likelihoods for each model, multi-
plied by the prior odds ratio. This is a Bayesian decision rule where for any
two models A and B the posterior odds ratio, showing evidence in favor of A
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 437

Table 8
Bayesian estimates for Model 3

Taiwan Thailand South Korea France UK

f0ð1Þ
0.0690
0.0369
0.2520 0.0193 0.0124

(0.0380) (0.0549) (0.0432) (0.0303) (0.0247)
f1ð1Þ
0.0027 0.0160 0.0275
0.1293
0.1447
(0.0290) (0.0560) (0.0242) (0.0288) (0.0316)
c1ð1Þ 0.2988 0.2537 0.2642 0.2722 0.2286
(0.0344) (0.0481) (0.0396) (0.0360) (0.0340)

f0ð2Þ 0.2261
0.0712 0.6071 0.0441 0.0346

(0.0667) (0.0367) (0.0786) (0.0411) (0.0245)
f1ð2Þ
0.0431 0.1309 0.0910 0.0207
0.0266
(0.0335) (0.0248) (0.0375) (0.0436) (0.0372)
c1ð2Þ 0.3147 0.2103 0.3338 0.4280 0.3058
(0.0546) (0.0343) (0.0657) (0.0529) (0.0278)

a0ð1Þ 0.1538 0.1810 0.0244 0.0082 0.0067

(0.0562) (0.1169) (0.0153) (0.0085) (0.0069)
a1ð1Þ 0.1348 0.0905 0.0785 0.0633 0.0938
(0.0284) (0.0508) (0.0169) (0.0158) (0.0194)
b1ð1Þ 0.8438 0.8344 0.9150 0.8982 0.8703
(0.0353) (0.0539) (0.0179) (0.0186) (0.0258)

a0ð2Þ 0.1265 0.0733 0.0732 0.1263 0.0323

(0.0941) (0.0416) (0.0452) (0.0385) (0.0109)
a1ð2Þ 0.0534 0.1171 0.0806 0.1020 0.0825
(0.0223) (0.0218) (0.0230) (0.0238) (0.0193)
b1ð2Þ 0.8714 0.8686 0.9075 0.8868 0.9119
(0.0522) (0.0253) (0.0263) (0.0270) (0.0203)

r2 1.1448
1.5375 1.0179 1.0832
0.1654

(0.3979) (0.9330) (0.1962) (0.6038) (0.4197)
n 6.6030 7.2002 8.5599 22.3516 21.4837
(0.9211) (1.0382) (1.3779) (12.9051) (10.0179)

Figures in brackets represent posterior standard errors for each parameter. r2 : threshold value for volume-
change.

over B is
pðAjRÞ pðRjAÞPrðAÞ
ORA;B ¼ ¼ .
pðBjRÞ pðRjBÞPrðBÞ
The decision rule is to choose Model A if ORA;B 41, otherwise choose model B. See
Ref. [37] for a complete discussion of Bayesian decision rules. We use the integrated
or marginal likelihood function pðRjAÞ here, so that we may integrate over the
possible values of unknown parameters based on the observed sample data, and thus
 ARTICLE IN PRESS
438 R. Gerlach et al. / Physica A 360 (2006) 422–444

Table 9
The estimates of means and unconditional variance for each regime

Taiwan Thailand South Korea France UK

Model 1
m 0.0243
0.0402
0.0024 0.0380 0.0337
a0 3.1627 3.9258 5.6165 2.0900 1.3417
1
a1
b1
Model 2
mð1Þ
0.0768
0.2108
0.3954
0.2939
0.2101
mð2Þ 0.0777 0.1824 0.1484 0.1985 0.1601
a0ð1Þ 9.5368 3.2780 13.4063 19.0026 11.1588
1
a1ð1Þ
b1ð1Þ
a0ð2Þ 1.3355 4.2612 1.4863 0.2911 0.1235
1
a1ð2Þ
b1ð2Þ
d1 2 1 1 1 1
Model 3
mð1Þ
0.0614
0.0318
0.2629 0.0223 0.0088
mð2Þ 0.2279
0.0385 0.6763 0.0721 0.0568
a0ð1Þ 7.6983 2.3874 4.1567 0.1696 0.1010
1
a1ð1Þ
b1ð1Þ
a0ð2Þ 1.4935 5.7459 6.0943 14.7160 8.0672
1
a1ð2Þ
b1ð2Þ
d1 1 1 1 1 1

account for parameter uncertainty. We assume prior model ignorance so that the
prior ratio is 1, equivalently PrðAÞ ¼ PrðBÞ ¼ 0:5. To estimate the integrated
likelihood for each model can be quite difficult, with many Bayesian approaches
suggested in the literature, see Refs. [38–40] for examples. We employ the method
suggested in Ref. [30], employing importance sampling techniques. These methods
are implemented in Ref. [14], who give a detailed exposition of this work applied to
extended threshold GARCH models. See Appendix for details.
Table 10 contains the logarithm of the odds ratio comparing each pair of Models
1, 2 and 3. The first listed model is preferred over the second listed model if the
number in the table is positive (and hence the corresponding odds ratio is 41),
otherwise the second model is preferred. The boxed number is for the most preferred
model in each market, confirmed by the final line of the table. It is interesting to note
that the model with trading volume as a threshold variable, Model 3, is strongly
preferred over both Models 2 and 1 in the three Asian markets: Taiwan, Thailand
and South Korea. Using the Bayesian odds ratio interpretation in Ref. [39], this is
‘very strong’ to ‘decisive’ evidence in favor of the volume threshold asymmetric
model. In contrast, however, for the European markets of the UK and France, the
asymmetric model with lagged US news as the threshold variable is preferred over
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 439

Table 10
Logarithm of the posterior odds ratio

Taiwan Thailand South Korea France UK

Model 2 vs Model 1
2.1363 3.4860
8.4749 1:7713 1:5733

Model 3 vs Model 1 6:3943 8:9871 3:6360
0.4313
5.8874
Model 3 vs Model 2 8:5306 5:5011 12:1109
2.2026
7.4607
Preferred model Model 3 Model 3 Model 3 Model 2 Model 2

Models 1 and 3. This represents ‘substantial’ evidence in favor of this model for these
two markets. Clearly, the asymmetric effect from trading volume is comparatively
strongest in the Asian markets considered, while the US trading news dominates
trading volume news in the UK and France markets. Table 10 also shows that the
symmetric GARCH model with leptokurtic innovations, Model 1, ranked second in
four of the five markets. This result is not surprising, and confirms previous findings,
such as those in Ref. [41], that a simple GARCH formulation with t-errors can
sometimes outperform more complex models for real market returns.
It is hard to make the usual good news–bad news financial argument with this
trading volume threshold model, as asymmetric increases in volatility also coincide
with increases in mean return in each market; following an increase in trading
volume. However, this is a common finding in many financial studies; an increased
level of return is associated with, and often caused by, an increased level of risk
(volatility) [42]. The traditional bad news argument can only be applied in Taiwan,
where investors seem to perceive decreases in trading volume as bad news, as this
result is associated with lower negative mean returns and higher return volatility.

6. Conclusions

In this paper, we have thoroughly examined the empirical dynamic relationship

between stock return, return volatility, international return spillover and change in
trading volume for five international stock markets. Conforming to well-established
empirical results, stock returns and return volatility display a certain degree of
persistence and consistent with a meteor-shower hypothesis, our empirical results
suggest that the stock-return news developed from US is transmitted significantly to
each of the five national stock markets. In particular, the US return-news is
positively correlated with the stock price of each national market but this correlation
is asymmetric around a threshold US return value and/or a value of trading volume
change. This confirms the minor goal of this paper. The major issue of this study has
been whether stock returns and return volatility react asymmetrically, in a threshold
nonlinear fashion, to trading volume change. By employing a double-threshold
GARCH model to capture the nature of market reaction to trading volume change,
we find that a large increase in volume is accompanied by a large rise in average
 ARTICLE IN PRESS
440 R. Gerlach et al. / Physica A 360 (2006) 422–444

stock price, while a decrease in volume is associated with lower average returns, with
the exception of Thailand. Moreover, the return volatility of all the stock-markets
also displayed an asymmetric reaction to volume change, around a threshold level.
With the exception of Taiwan, the average magnitude of return volatility following
an increase in trading volume was much larger than that following decreases in
volume. Our model comparison results revealed that the asymmetric nonlinear
model, with change in volume as a threshold, was decisively favored in the three
Asian markets: Taiwan, Thailand and South Korea. However, the DT-GARCH
model of Chen et al. [13], with US return news as threshold variable, was favored in
UK and France.
In summary, our findings are consistent with, and add to, the arguments of Ying
[24], Karpoff [25] and Gallant et al. [43]: that the joint study of stock price and
trading volume, as an asymmetric nonlinear relationship, leads to a better
understanding of capital market behavior. These findings shed new light on the
application of trading volume in the market integration literature and develop a new
avenue for asset pricing in a multi-market framework.

Acknowledgements

The authors thank two anonymous referees and the Editor, H. Eugene Stanley,
whose comments improved the paper. C.W.S. Chen is supported by National Science
Council (NSC) of Taiwan grants NSC93-2118-M-035-003. R. Gerlach was
supported by the Mathematical Research Promotion Center of the NSC of Taiwan,
Feng Chia University and the University of Newcastle, via an ECA networking
grant and the School of MAPS.

Appendix A

This appendix gives some details for the Bayesian MCMC sampling scheme used
to estimate the GARCH models in this paper and of the method used to estimate the
marginal likelihood required in the posterior odds ratio used for model comparison.
For further details of MCMC sampling see Ref. [44] or [14].

A.1. MCMC methods

Firstly, let

 /ðjÞ ¼ ðfðjÞ ðjÞ

0 ; f1 ; c1 Þ;
ðjÞ
j ¼ 1; 2,
 aðjÞ ¼ ða0ðjÞ ; aðjÞ
1 ; bðjÞ
1 Þ; j ¼ 1; 2,
 ð1Þ ð2Þ ð1Þ ð2Þ
H ¼ ð/ ; / ; a ; a ; rk ; dÞ.
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 441

The likelihood function pðRjHÞ is given by

(  
ðnþ1Þ=2 )
Yn X2
Gððn þ 1Þ=2Þ 1 ðRt
mt Þ2
2;n
pðR jHÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffi 1 þ I jt ,
t¼2 j¼1 Gðn=2Þ ðn
2Þp
ht ðn
2Þht

where I jt is the indicator variable IðRjtþm
d pr1 Þ, Gð:Þ is the Gamma function and n is
the degrees of freedom. For Bayesian inference, we need to set prior distributions on
all model parameters. We use the prior pðHÞ / Iðað1Þ ð1Þ ð1Þ ð2Þ
0 40; a1 þ b1 o1ÞIða0 40;
ð2Þ ð2Þ
a1 þ b1 o1ÞIðq1 ork oq3 ÞIðd 2 1; 2; 3ÞIðt 2 ½0; 0:25Þ, where IðÞ is an indicator
function, t ¼ 1=n and q1 and q3 are the first and third quantiles, respectively, of
the required threshold variable. In threshold modelling, it is important to set a
minimum amount of observations in each regime, so there is sufficient sample size to
generate meaningful inference results, driving our prior choice for rk here. We
restrict n44 so that the variance of
t is finite, while ensuring that the kurtosis is
greater than 3. Finally, for the mean equation parameters we assume the normal
prior /j Nð/j0 ; V j Þ, where usually we set /j0 ¼ 0 and V
1 j to be a large number,
such as 1, to ensure a reasonably flat prior.
Incorporating the likelihood above and the priors, using Bayes rule, leads to the
conditional posterior distributions for the parameter groupings we use in the
MCMC sampling scheme. We use the following iterative sampling scheme to
construct the desired posterior sample:

1. Draw the vector /ðjÞ jR; H
fj for j ¼ 1; 2, using the random walk MH algorithm.
2. Draw the vector aðjÞ jR; H
aj for j ¼ 1; 2 using the random walk MH algorithm.
3. Draw the parameter rk jR; H
r using the random walk MH algorithm.
4. Draw djR; H
d by noting that d is discrete valued and using it’s discrete posterior.
5. Draw t using the random walk MH algorithm.

These posterior distributions are in general not of a standard form and require us to
employ techniques such as the Metropolis random walk method [45,46] to achieve
the desired sample, as detailed now. Consider a general parameter vector q, a subset
of H. The posterior distribution for this parameter vector is evaluated by
pðqjR2;n ; H
q Þ / pðR2;n jHÞpðqÞ .
Details of the random walk MH algorithm are:
Step 1: Generate initial values q½0 from the prior distribution for this parameter
vector
Step 2: At iteration i, generate a point q from the kernel density,
q Nðq½i
1 ; aOÞ ,
where q½i
1 is the ði
1Þth iterate of q.
Step 3: Accept q as q½i with the probability p ¼ minf1; pðq jR2;n ; H
r Þ=
pðq½i
1 jR2;n ; H
r Þg. Otherwise, set q½i ¼ q½i
1 .
To yield good convergence properties, the choices of O and a for each parameter
vector are important. These choices can be made to ensure good coverage for each
 ARTICLE IN PRESS
442 R. Gerlach et al. / Physica A 360 (2006) 422–444

conditional posterior distribution, with O usually chosen with sufficiently large

diagonal terms, while the coverage and sampling efficiency can also be tuned using
the acceptance rate, in the burn-in period of the MCMC sample; see Ref. [14] for
details. A suitable value of a, with good convergence properties, can usually be
selected by having an acceptance probability of between 25% and 50% in the burn-
in period.
Parameter estimates of any functions of parameters are obtained as posterior
means by averaging the function over all sample iterates after the burn-in period.

A.2. Marginal likelihood estimation

We follow Gerlach
Q et al. [30] in estimating the marginal likelihood term for each
model pðRi Þ ¼ nt¼1 pðRit jR1;t
1 Þ by estimating the respective marginal term at each
time t using
PD
pðRi jR1;t
1 ; Yk½i Þ=pðRt;k jR1;t
1 ; Y½i
kÞ
pðR i
^ t jR 1;t
1
Þ ¼ i¼1 PtD t;k 1;t
1 ½i
,
i¼1 1=pðR jR ; Yk Þ
Y n
^ iÞ ¼
pðR ^ it jR1;t
1 Þ ,
pðR
t¼1

where Y½i
k is the ith MCMC iterate from the posterior distribution of the unknown
parameter vector YjR1;k , conditional on the first k observations only. Following
Gerlach et al. [30], we choose a maximum value for k
t ¼ 300 and allow
^ it jR1;t
1 Þ for t ¼ k

k ¼ 300; 600; . . . ; n. For each k, we simultaneously compute pðR
300; . . . ; k. This means we need to run the MCMC sampling scheme multiple times
(in fact n=300 integer rounded times) for each model. While this process can be more
time consuming than 1 single MCMC run, we consider the gain from including
parameter uncertainty in model comparison is sufficient, from a Bayesian viewpoint,
to justify this procedure.

References

[1] J. Jaffe, R. Westerfield, The weekend effect in common stock returns: the international evidence,
J. Finance 40 (1985) 433–454.
[2] C. Eun, S. Shim, International transmission of stock market movements, J. Financial Quant. Anal. 24
(1989) 241–256.
[3] S.A. Ross, Information and volatility: the no-arbitrage martingale approach to timing and resolution
irrelevancy, J. Finance 44 (1989) 1–17.
[4] Y. Hamao, R.W. Masulis, V. Ng, Correlations in price changes and volatility across international
stock markets, Rev. Financ. Stud. 3 (1990) 281–307.
[5] P. Theodossiou, Lee Unro, Relationship between volatility and expected returns across international
stock markets, J. Business Finance Account. 22 (1995) 289–300.
[6] T.C. Chiang, J. Chiang, Dynamic analysis of stock return volatility in an integrated international
capital market, Rev. Quant. Finance Account. 6 (1996) 5–17.
[7] M. Martens, S.H. Poon, Returns synchronization and daily correlation dynamics between
international stock markets, J. Bank. Finance 25 (2001) 1805–1827.
 ARTICLE IN PRESS
R. Gerlach et al. / Physica A 360 (2006) 422–444 443

[8] F. Black, Studies of stock price volatility changes, Proceedings of the 1976 Meetings of the Business
and Economics Section, American Statistical Association, Alexandria, VA, pp. 177–181.
[9] A.A. Christie, The stochastic behavior of common stock variances—value, leverage and interest rate
effects, J. Financ. Econ. 10 (1982) 407–432.
[10] K. Nam, C.S. Pyun, S.L. Avard, Asymmetric reverting behavior of short-horizon stock returns: an
evidence of stock market overreaction, J. Bank. Finance 25 (2001) 807–824.
[11] G. Koutmos, Asymmetric price and volatility adjustments in emerging Asian stock markets,
J. Business Finance Account. 26 (1999) 83–101.
[12] C. Brooks, A double-threshold GARCH model for the French Franc/Deutschmark exchange rate,
J. Forecasting 20 (2001) 135–143.
[13] C.W.S. Chen, T.C. Chiang, M.K.P. So, Asymmetrical reaction to US stock-return news: evidence
from major stock markets based on a double-threshold model, J. Econ. Business (Special issue on
Globalization in the New Millennium: Evidence on Financial and Economic Integration) 55 (2003)
487–502.
[14] C.W.S. Chen, M.K.P. So, On a threshold heteroscedastic model, Int. J. Forecasting, 2005, to appear.
[15] S.S. Wang, M. Firth, Do bears and bulls swim across oceans? Market information transmission
between greater China and the rest of the world, J. Int. Financ. Markets Institutions Money 14 (2004)
235–254.
[16] P.K. Clark, A subordinated stochastic process model with finite variance for speculative prices,
Econometrica 41 (1973) 135–155.
[17] T.W. Epps, M.L. Epps, The stochastic dependence of security price changes and transaction volumes:
implications for the mixture distributions hypothesis, Econometrica 44 (1976) 305–321.
[18] C.G. Lamoureux, W.D. Lastrapes, Heteroskedasticity in stock return data: volume versus GARCH
effects, J. Finance 1 (1990) 221–229.
[19] D. Kim, S. Kon, Alternative models for the conditional heteroscedasticity of stock returns,
J. Business 67 (1994) 563–588.
[20] T.G. Andersen, Return volatility and trading volume: an information flow interpretation of
stochastic volatility, J. Finance 51 (1996) 169–204.
[21] G. Gallo, B. Pacini, The effects of trading activity on market volatility, Eur. J. Finance 6 (2000)
163–175.
[22] M. Omran, E. McKenzie, Heteroscedasticity in stock returns data revisited: volume versus GARCH
effects, Appl. Financ. Econ. 10 (2000) 553–560.
[23] M.T. Bohl, H. Henke, Trading volume and stock market volatility: the Polish case, Int. Rev. Financ.
Anal. 12 (2003) 513–525.
[24] C.C. Ying, Stock market prices and volumes of sales, Econometrica 34 (1966) 676–685.
[25] J.M. Karpoff, The relation between price changes and trading volume: a survey, J. Financ. Quant.
Anal. 22 (1987) 109–126.
[26] H. Tong, in: K. Krickegerg (Ed.), Threshold Models in Non-linear Time Series Analysis, vol. 21,
Lecture Notes in Statistics, Springer, New York, 1983.
[27] T. Chordia, B. Swaminathan, Trading volume and cross-autocorrelations in stock returns, J. Finance
55 (2000) 913–935.
[28] B.S. Lee, O.M. Rui, The dynamic relationship between stock returns and trading volume: domestic
and cross-country evidence, J. Bank. Finance 26 (2002) 51–78.
[29] I. Moosa, P. Silvapulle, M. Silvapulle, Testing for temporal asymmetry in the price–volume
relationship, Bull. Econ. Res. 4 (2003) 0307–3378.
[30] R. Gerlach, C. Carter, R. Kohn, Diagnostics for time series analysis, J. Time Ser. Anal. 20 (1999)
309–330.
[31] C.W. Li, W.K. Li, On a double-threshold autoregressive heteroscedastic time series model, J. Appl.
Econ. 11 (1996) 253–274.
[32] C.W.S. Chen, A Bayesian analysis of generalized threshold autoregressive models, Stat. Probab. Lett.
40 (1998) 15–22.
[33] K. Daniel, D. Hirshleifer, A. Subrahmanyam, Investor psychology and security market under-and
over-reactions, J. Finance 6 (1998) 1839–1885.
 ARTICLE IN PRESS
444 R. Gerlach et al. / Physica A 360 (2006) 422–444

[34] N. Barberis, A. Shleifer, R. Vishny, A model of investor sentiment, J. Financ. Econ. 3 (1998) 307–343.
[35] C.H. Shen, L.R. Wang, Daily serial correlation, trading volume and price limits: evidence from the
Taiwan stock market, Pacific-Basin Financ. J. 6 (1998) 251–273.
[36] B.N. Huang, C.W. Yang, An empirical investigation of trading volume and return volatility of the
Taiwan stock market, Global Finance J. 12 (2001) 55–77.
[37] J.O. Berger, Statistical Decision Theory and Bayesian Analysis, second ed., Springer, New York,
1985.
[38] P.J. Green, Reversible jump MCMC computation and Bayesian model determination, Biometrika 82
(1995) 711–732.
[39] R. Kass, A. Raftery, Bayes factors, J. Am. Stat. Assoc. 90 (1995) 773–795.
[40] S. Kim, N. Shephard, S. Chib, Stochastic volatility: likelihood inference and comparison with ARCH
models, Rev. Econ. Stud. 65 (1998) 361–393.
[41] R. Gerlach, F. Tuyl, MCMC methods for comparing stochastic volatility and GARCH models, Int.
J. Forecasting, 2005, to appear.
[42] E. Fama, K. French, The cross section of expected stock returns, J. Finance 47 (1992) 427–466.
[43] R. Gallant, P. Rossi, G. Tauchen, Stock prices and volume, Rev. Financ. Stud. 5 (1992) 199–242.
[44] S. Chib, E. Greenberg, Understanding the Metropolis–Hastings algorithm, Am. Stat. 49 (1995)
327–335.
[45] N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, E. Teller, Equations of state calculations by fast
computing machines, J. Chem. Phys. 21 (1953) 1087–1091.
[46] W.K. Hastings, Monte-Carlo sampling methods using Markov chains and their applications,
Biometrika 57 (1970) 97–109.
[47] L.R. Glosten, R. Jagannanthan, D. Runkle, Relationship between the expected value and the
volatility of the nominal excess return on stocks, J. Finance 48 (1993) 1779–1802.