An Empirical Study on Stock Exchange Linkages between ...533177/FULLTEXT01.pdf · thank my thesis...

29
Master Essay Department of Statistics An Empirical Study on Stock Exchange Linkages between Chinese and Western Markets Heng Wang Supervisor: Anders Ågren

Transcript of An Empirical Study on Stock Exchange Linkages between ...533177/FULLTEXT01.pdf · thank my thesis...

  • Master Essay

    Department of Statistics

    An Empirical Study on Stock Exchange Linkages between Chinese and Western Markets

    Heng Wang

    Supervisor: Anders Ågren

  • An Empirical Study on Stock Exchange Linkages between

    Chinese and Western Markets

    Heng Wang

    Department of Statistics, Uppsala University

    May, 2012

    Abstract

    This paper examines the linkages between Chinese stock exchange

    markets and western developed markets. We use SHCOMP and SIASA

    representing Chinese exchanges, FTSE100 the European exchange, and

    S&P500 the U.S. exchange. We find evidence of bidirectional returns and

    volatility spillovers from western markets to Chinese markets by four bivariate

    GARCH-BEKK models. The volatility impulse response functions show the

    impact from U.S is normally stronger than that from Europe and SIASA is more

    sensitive to the spillovers than SHCOMP.

    Keywords: Stock market linkages, volatility spillovers, multivariate GARCH,

    BEKK, volatility impulse response functions

    E-mail: [email protected]

  • Dedication

    I dedicate this paper to my wonderful family. Particularly to my brave

    father, Junguo Cao, who has taught me to be strong and persistent, I wish he

    could conquer the disease and get healthy soon. I must also thank my loving

    mother, Xiufang Wang, who always supports me to move on, and my

    understanding sister, Ling Cao, who has helped me to keep patient for study

    and given her fullest encouragement. Finally, I dedicate this work to my

    girlfriend, Jiaman Yuan, who accompanied me in spirit during the tough days.

    I’m grateful for all of them, who always stand behind me and believe I will

    succeed.

  • Acknowledgement

    I am grateful for the guidance and support of my supervisor Anders Ågren. I

    thank my thesis opponent, Yumin Xiao, and many professors in Uppsala

    University, especially Lars Forsberg, for helpful suggestions on improving the

    paper. The encouragements from families and friends are heartily

    acknowledged.

  • 1

    1. Introduction

    This paper models the stock markets in Shanghai and Shenzhen, representatives of

    Chinese stock markets, focusing on the structure with global influences and studies

    whether these two Chinese markets are linked to the western developed markets in Britain

    or U.S. Under the development of economic globalization, there are more and more mutual

    trades and investment between China and the rest of the world. The European Union and

    the U.S are the two largest trading partners of China. Europe has been the largest exporting

    region of China since 2004. At the end of 2011, China also became the largest exporting

    country of Europe. Besides, the Chinese oversea investment in E.U. exceeded that in U.S. for

    the first time, making Europe the No. 1 position in Chinese international trading and

    economics. In the meanwhile, China is known as the country which owns most American

    national debts and dollar reserves. Therefore U.S. plays a very important role in the

    Chinese economy. As one component of the economy, the stock exchange reflects the status

    of the domestic economy. The relationship between Chinese and western economies is

    believed to produce the reasonable linkage between stock exchange markets. NYSE and

    NASDAQ, both involved with American and European companies, take the leading two

    positions of market capitalization in the world. And for individual European countries,

    British and German market capitalizations get the rank of 41 and 10. The overall

    capitalizations of Chinese markets, including Shanghai exchange (rank 5) and Shenzhen

    exchange (rank 12), would exceed Japan (rank 3) to generate the third largest market.

    Through the economic globalization, the large stock markets should not be isolated. The

    co-movements are frequently observed when big economic events strike the world

    economic system. A simple example is during the financial crisis in U.S around 2008, when

    most markets in different regions crashed with the disaster from the American market.

    Our study is based on the assumption that Chinese stock exchanges are associated with

    western markets.

    As far as Chinese markets are concerned, the study of the regional interaction of the

    Chinese stock markets begins very early. Brooks and Ragunathan (2003) and Wang et al.

    1 Data from “List of stock exchange” in Wikipedia: http://en.wikipedia.org/wiki/List_of_stock_exchanges

    http://en.wikipedia.org/wiki/List_of_stock_exchanges

  • 2

    (2004) use a GARCH model to investigate the interactions between Chinese A and B shares

    traded on the Shanghai and the Shenzhen stock exchanges. In the further studies, a

    conjecture of global interaction between the Chinese stock markets and the western stock

    markets has drawn a lot of attentions.. Chow (2003) studies the relationship between

    Shanghai and New York stock price indices by simple correlation and multiple regression

    methods. However the capital markets are found not to be integrated in his paper. Li (2007)

    examines the linkage between stock exchanges in China (including Shanghai, Shenzhen and

    Hong Kong) and in U.S. through a multivariate GARCH model. The return linkages between

    the stock exchange in mainland China and in U.S. are indirectly depending on the return

    linkage between the stock exchange in Hong Kong and the U.S. market. Hong Kong has

    acted as a go-between in the information flow. Diekmann (2011) analyzes the stock market

    integration of mainland China indices, Hong Kong indices and Dow Jones Industrial index

    from 1998 to 2006, proving an evidence of global and regional integration. In the newest

    research, Zhou et al. (2011) measure the dynamic volatility spillovers between Chinese

    stock indices and western indices by variance decompositions in a generalized VAR

    framework. Their study shows that from 1996 to 2005 the Chinese market was slightly

    affected by western markets and from 2005 to 2009, the Chinese stock market had a

    significant volatility spillover effect on western markets, indicating that the influence of the

    Chinese stock market was greatly enhanced during the years. They also show that the US

    market had dominant volatility impacts on the Chinese markets during the subprime

    mortgage crisis.

    The multivariate, mostly bivariate, BEKK models are widely used to investigate the

    interrelation between markets. Chou et al. (1999) use a bivariate BEKK model to examine

    the volatility linkage between the Taiwanese and the U.S. stock exchanges and prove that

    volatility spillovers from the developed market in the U.S. to the emerging market in

    Taiwan exist. Haroutounian and Price (2001) also find a volatility spillover from Poland to

    Hungary in a bivariate BEKK model. Worthington and Higgs (2004) use a nine-variable

    BEKK model to examine the transmission of equity returns and volatility among nine Asian

    developed and emerging markets. They find evidence of return spillovers from the

  • 3

    developed to emerging markets. Li and Majerowska (2007) test the linkages between

    emerging markets in Poland and Hungary and the established markets in Germany and the

    U.S under a multivariate BEKK approach. Through a four-variable BEKK model, Li (2011)

    investigates stock market linkages among China, Korea, Japan and the U.S. with particular

    attention to the impact of Chinese stock market reforms. In this paper, we will concentrate

    not only on the stock linkages between Chinese markets and western markets in the recent

    unstudied years, but also compare the spillovers from different regions, Europe and U.S, to

    different components of Chinese markets, the market in Shanghai and Shenzhen.

    We study the differences among linkages based on the impulse response function,

    proposed by Sims (1980) for the analysis of VAR models. The impulse response function

    has been generalized to nonlinear models and higher conditional moments. Hafner and

    Herwartz (1998) apply the impulse response function to the volatility of time series,

    defining the volatility impulse response function (VIRF). Hafner and Herwartz (2006)

    improve the VIRF to an intuitionistic one, which is employed by many researchers through

    years. In this paper, we use Hafner and Herwartz’s VIRF methodology to demonstrate the

    shocks in BEKK models. The linkages will be easily distinguished from the illustration of

    VIRF.

    2. Data and preliminary analysis

    There are two main stock markets in China, Shanghai and Shenzhen. In this paper, we

    use the Shanghai Stock Exchange Composite Index (SHCOMP) to represent the Shanghai

    market and the Shenzhen Stock Exchange Constituent A-Share Index (SIASA) the Shenzhen

    market. The stock markets in Britain (FTSE100) and the U.S. (S&P500) are considered to

    serve well as proxies for the western developed markets and their economic development.

    They are considered to have an influential impact on the Asian financial economies,

    especially on the Chinese economy.

    Following the previous literature, we use daily observations of the stock indexes of the

    markets in Shanghai (SHCOMP), Shenzhen (SIASA), London (FTSE100) and U.S. (S&P500)

    in this study. The stock indexes are computed from the closed stock prices in local markets.

    The Shanghai stock index, SHCOMP is an overall market index, computed from all the listed

  • 4

    companies in this market. The Shenzhen stock index, SIASA is a selected index that consists

    of the 40 top companies’ A-shares on the Shenzhen Stock Exchange. The reason I choose

    different kinds of indexes in these two markets comes from the behavior habit of Chinese

    market investors. SCHCOMP represents most large listed companies and SIASA represents

    medium and small size companies. The two indexes cover a general situation of Chinese

    listed companies, making a quality summary for Chinese stock markets. The London Stock

    Exchange is the fourth-largest stock exchange in the world and the largest in Europe, so the

    index FTSE100, a share index of the stocks of the top 100 companies listed on the London

    Stock Exchange having the highest market capitalization, typically reflects the financial

    situation in Europe. The S&P500, known as the combination of the two largest stock

    exchanges in U.S. as well as in the world, is the most accurate reflection of the U.S. stock

    market. The Chinese data in the paper are downloaded from the stock trading software

    Great Wisdom2. The data of FTSE100 and S&P500 are from Yahoo Finance.

    Figure 1. Stock indices from June 2006 to March 2012

    2 Official website: http://www.gw.com.cn/.

    http://www.gw.com.cn/

  • 5

    The period of the data is from 6 December 2005 to 1 March 2012. Our chosen period

    includes the big events in recent years, the bull market increase in China (2005.12 ~

    2008.11), American financial crisis (2008.08 ~ 2008.10) and European sovereign debates

    crisis (2009.11 ~ now). These events have huge influences on the regional and global

    markets, determining the trend of stock exchange. Based on the different stock trends

    under the events, we divide the whole period into three separate intervals: Increasing

    Period I (2005/12/6 ~ 2007/11/12), Decreasing Period II (2007/11/13 ~ 2008/11/10)

    and Depression Period III (2008/11/11 ~ 2012/3/1). Comparison of the results from

    different markets in different periods can show if there is any market linkage in the

    economic integration age and how it varies if existing.

    Figure 1 presents the time plots of the series. The vertical dash lines split out the three

    periods. It is impressive that the two Chinese markets follow a similar movement while

    FTSE100 and S&P500 have a similar trend. The Chinese indices grew rapidly in the first

    period due to the domestic product development. But in the following period, the global

    financial crisis made the indices fall down quite a lot to the past level two years ago. Even

    with the domestic economic encouragement in the recent three years, 2009, 2010 and

    2011, the stock markets couldn’t recover. Moreover, the European sovereign debates crisis

    led to a new wave of depressed market. On the contrast, the British and American markets

    are more defensive to the economic turnovers. When FTSE100 and S&P500 got to the

    lowest after the crisis in 2009, they reacted quickly to get saved, regained the markets

    confidence and rebounded to the acceptable level.

    Figure 2 presents the returns of stock exchanges within the three periods. We make an

    adjustment to extend the natural logarithm a 1000 times to increase returns to an easily

    visual level. The Chinese markets have quite high volatility from 2007 to 2009 while the

    western markets only have shocks during 2009. Compared to Britain and U.S., as Harvey

    (1995) pointed out, emerging markets in Asia have high-expected returns and high

    volatility. The cluster phenomenon is also observed in Figure 2. Small volatility is more

    likely followed by small volatility and large volatility followed by large volatility, which

    implies a regular system in the structure.

  • 6

    Figure 2. Stock returns from June 2006 to March 2012

    Table 1 shows the summary results of returns. During Period I and Period III, the

    larger means and standard errors of the Chinese series prove our statement in the last

    paragraph. But in Period II, when the Chinese markets face the strong strike of financial

    crisis, there are smaller negative returns and larger volatilities in China than in the

    developed markets. These facts prove the higher risk in Chinese markets, especially in the

    decreasing time. During Period I, Chinese markets have negative skewness, smaller than

    the skewness of FTSE100 and S&P500. It means the Chinese stock returns are more likely

    positive and the possibilities of negative returns are larger than those in Britain and U.S. In

    Period II and Period II, the Chinese skewness keeps different sign from the mean. For the

    western markets, FTSE100 gets a negative skewness but S&P500 is positive-skewed. In

    Period III, it is the opposite. In an overall view, all the series are slightly skewed in Period II

    due to a small absolute value of skewness. Except SHCOMP in Period I, the kurtosises of the

    Chinese stock exchanges are smaller than 3, i.e. in most cases they have significant thinner

    tails and shorter peaks, which is not common in the world. On the contrast, FSTE100 and

  • 7

    S&P500 have skewness larger than 3 in Period II&III, with fatter tails and higher peaks.

    Furthermore in Table 1, the Shapiro-Wilk test3 rejects all the normally distributed

    hypothesis of the four series except SIASA in Period II. For all the above, GARCH type

    models are capable of modeling data with such features.

    Table 1 Summary statistics of returns during December 2005 and March 2012

    SHCOMP SIASA FTSE100 S&P500

    Mean P1 4.0781 4.5881 0.3741 0.3394

    P2 -3.8464 -3.2481 -1.3668 -1.8473

    P3 0.6798 1.0465 0.4331 0.5387

    S.D P1 17.8163 19.8520 8.9983 7.7514

    P2 27.6310 29.5603 21.8485 22.1755

    P3 15.4937 18.0416 13.6861 15.9025

    Skewness P1 -1.1552 -0.8120 -0.4343 -0.5945

    P2 0.3175 0.0634 -0.1323 0.0736

    P3 -0.4776 -0.4446 0.1485 -0.3551

    Kurtosis P1 3.8517 2.8669 2.4370 2.8407

    P2 1.0698 0.4154 3.5722 5.9377

    P3 2.0277 1.8958 4.1056 4.2110

    Shapiro P1 0.9273

    (p=0.0000)

    0.9551

    (p=0.0000)

    0.9673

    (p=0.0000)

    0.9502

    (p=0.0000)

    P2 0.9813

    (p=0.0017)

    0.9941

    (p=0.4105)

    0.9374

    (p=0.0000)

    0.9034

    (p=0.0000)

    P3 0.9706

    (p=0.0000)

    0.9746

    (p=0.0000)

    0.9567

    (p=0.0000)

    0.9327

    (p=0.0000)

    Note: P1, P2 and P3 stand for Period I, Period II and Period III, mentioned above.

    3. Methodology

    3.1 The AR-GARCH process

    As the first principal aim of this paper is to find out the behavior of Chinese stock

    returns by exploring the short-run volatility either in a separate scenario or a cross-market

    setting, the univariate GARCH model is an appropriate approach for a first trial. Also in the

    previous literature, the GARCH model is widely used for financial returns.

    3 Shapiro-Wilk test, by Samuel Shapiro and Martin Wilk, tests the null hypothesis that a sample came from a normally

    distributed population. We may reject the null hypothesis if the statistic is too small.

  • 8

    Engle (1982) introduced the ARCH (Autoregressive Conditional Heteroskedastic)

    process instead of the classical assumption on time series and econometric models, which

    assumes the variance to be constant. The volatility of econometric series usually shows

    some random characteristics on its innovation, indicating a possibility assuming a white

    noise process. Besides, the clustering of high level volatility also implies autoregressive

    possibilities. The GARCH model, developed from the ARCH process by Bollerslev (1986),

    combines the lagged noises and conditional variances together to achieve a general

    approach.

    Let 𝜖𝑡 denote a discrete-time stochastic process, and 𝐼𝑡 the information set at time t

    through the past. The GARCH(p,q) process is given as:

    ϵt|𝐼𝑡−1~𝑁(0, ℎ𝑡)

    ht = 𝜔 +∑𝛼𝑖𝜖𝑡−𝑖2

    𝑞

    𝑖=1

    +∑𝛽𝑖ℎ𝑡−𝑖

    𝑝

    𝑖=1

    where

    p > 0, q > 0

    ω > 0, 𝛼𝑖 ≥ 0, 𝑖 = 1,… , 𝑞,

    βi ≥ 0, 𝑖 = 1,… , 𝑝

    In this equation, a conditional normal distribution of ϵt with zero mean and variance

    of ht is considered. In a special case, when p=0, the process is an ARCH(q) process

    without its own lag effect. And when p=q=0, ϵt is a white noise with variance α0. If the

    white noise could be generated from a linear equation, then the new GARCH (p,q) model

    has an additional structure with a mean equation, which is presented as.

    ϵt = 𝑦𝑡 − 𝑥𝑡′𝑏,

    where yt and xt are observed variables and b is a vector of unknown parameters.

    For the long-term autoregressive series, an AR process is somehow a better approach to

    describe the first order trend. Thus, we adjust the mean equation as

    ϵt = 𝑟𝑡 −∑𝑏𝑖𝑟𝑡−𝑖

    𝑠

    𝑖=1

    − 𝜇

    where b = (b1, … , bs) is the vector of autoregressive coefficients.

  • 9

    In independent univariate GARCH models, the conditional variance of each series

    could be estimated so that we can capture the trend and volatility of the series. The moving

    average model is also a proper approach to illustrate short-term volatility, making it clear

    to distinguish the models.

    3.2 The VAR-GARCH-BEKK process

    Our second task is to detect the relationship between Chinese stock markets and

    western markets, requiring a new model to combine separated series together. The cross

    market impact will be in two forms. One is an instant impact which means that a change

    from one market effects another immediately and lasts no more. The other one is a

    continuous impact which will exist during a period. In a univariate GARCH model, the

    mean equation gives an instant impact to the returns and the variance equation results in a

    continuous indirect impact in volatility of returns. Both impacts should be analyzed in a

    cross-market setting, with cross effects both from series and volatilities. So the series will

    be dependent series in a multivariate case. Then, dynamic covariance or correlation

    becomes important. Since cross-effects are not considered in a univariate model, a

    higher-dimension model is needed for examining cross-market effects.

    The BEKK model, proposed by Engle and Kroner (1995), is a multivariate GARCH

    model. Specifically, the following model presents the BEKK process, a joint process

    covering the multivariate financial return series to be studied.

    Yt = 𝛼 + Γ𝑌𝑡−1 + 𝜖𝑡

    ϵt|𝐼𝑡−1~𝑁(0,𝐻𝑡),

    where Yt is a N × 1 vector of multivariate series at time t, α is a N × 1 vector of

    intercepts and Γ is a N × N coefficient matrix associated with lagged own effects. The

    diagonal elements in Γ measure the autoregressive effects and the off-diagonal elements

    describe spillovers across the markets. The N × 1 vector of noise, or the random errors,

    ϵt is the innovation for every market at time t. In this multivariate model, the conditional

    distribution of ϵt based on the former information It−1, is in a multivariate normal form

    with a conditional variance-covariance matrix, Ht.

    Bollerslev et al. (1988) assume that Ht is a linear function of cross products of errors

  • 10

    and related to its lagged value, Ht−1. So one kind of the BEKK model is given as:

    vech(Ht) = c +∑𝐴𝑖𝑣𝑒𝑐ℎ(𝜖𝑡−1𝜖𝑡−1′ )

    𝑞

    𝑖=1

    +∑𝐺𝑖𝑣𝑒𝑐ℎ(𝐻𝑡−𝑖)

    𝑝

    𝑖=1

    ,

    where vech is an operator transforming the lower/upper triangular part of a symmetric

    matrix into a vector. Since the number of parameters to be estimated is large, results

    through the formula, which is also the conditional covariance matrix, may not be

    guaranteed to be positive definite, making it unreasonable. Years later, Engle and Kroner

    (1995) modified the model to a general decomposition one, the well-known BEKK model,

    explained better in Ht than its vech form:

    Ht = 𝐶′𝐶 + 𝐴′𝜖𝑡−1

    ′ 𝜖𝑡−1𝐴 + 𝐺′𝐻𝑡−1𝐺

    In the BEKK model, 𝐶 is a 𝑁 × 𝑁 lower triangular matrix of constants. 𝐴 and 𝐺

    are 𝑁 × 𝑁 parameter matrixes. The diagonal elements in 𝐴 and 𝐺 represent the lagged

    effects from single series and its volatility on the conditional variance while off-diagonal

    elements measure the spillovers’ effect on the conditional variance as well. Compared to

    the univariate GARCH model, the BEKK model has considered all cross effects from a single

    series and its volatility. If every matrix in a BEKK model is a diagonal matrix, then the

    BEKK model becomes a set of univariate models. It is convenient to model multi series

    with a BEKK model, since fixing different elements in the parameter matrices means

    different assumption of the effects. At the same time, the quadratic forms ensure the

    positive definiteness of the equation. The parameter matrices correspond to specific parts

    in the 𝑣𝑒𝑐ℎ formula, but in a very complicated relation.

    In our paper, to study the interaction among regional markets and distinguish the

    differences, we choose the BEKK (1,1) model with one Chinese series and two western

    series. One key point is to reduce the computational processes. On the other hand, to

    remove the interference within highly related Chinese markets is also important. Thus, we

    get two BEKK models, one with SHCOMP, FTSE100 and S&P500, another one with SIASA,

    FTSE100 and S&P500.

  • 11

    3.3 Volatility Impulse Response Functions

    In Section 3.2, we know that for every BEKK model there is a unique vec model

    specifying the same structure. And for the BEKK(1,1) model, the equivalent vec

    specification is

    vech(Ht) = c + 𝐴1𝑣𝑒𝑐ℎ(𝜖𝑡−1𝜖𝑡−1′ ) + 𝐺1𝑣𝑒𝑐ℎ(𝐻𝑡−𝑖).

    When we get the long-term expectation of the model, the unconditional covariance matrix

    Σ could be found from the formula:

    vech(Σ) = (I − A1 − G1)−1𝑐.

    If Ht reaches to Σ, in the stationary condition, the volatility will remain the same

    even when the time goes on. But in this case, any change in Ht will break the current

    balance. If the system is stationary, it will recover from the interruption. If it is not, the

    strike from interruption will last forever. As Σ is a multivariate term, associated with

    dependent series, any change in one series will influence others. Assume that the volatility

    is stable at time t = 0, Σ0 = Σ. We introduce a shock ξ0, a vector related to each series.

    One component of ξ0 is set to be 1 and the others are set to be 0, which means the series

    with shock 1 will have effects on the system. Here the Volatility Impulse Response Function,

    VIRF, is defined as the expectation of the volatility conditional on the initial shock ξ0 and

    the initial volatility Σ0. The formula of VIRF is given by

    Vt(𝜉0) = 𝐸[𝑣𝑒𝑐ℎ(Σ𝑡)|𝜉0, Σ0 = Σ].

    Start with t = 1,

    V1(𝜉0) = 𝑐 + 𝐴1𝑣𝑒𝑐ℎ (Σ12𝜉0𝜉0

    ′Σ12) + G1𝑣𝑒𝑐ℎ(Σ).

    For t ≥ 2,

    Vt(𝜉0) = 𝑐 + (𝐴1 + 𝐺1)𝑉𝑡−1(𝜉0).

    In the formula above, Vt(𝜉0) is a 3 dimensional vector. For every component in Σ,

    there is one corresponding component in Vt(𝜉0), which measures the change of the

    covariance of the series due to the shock at time t. In our paper, we focus on the change of

    the Chinese indexes, SHCOMP and SIASA, influenced by shocks in western markets.

    Therefore we set ξ0 to be (0,1) and pick up the first element in Vt(𝜉0), Vt,1(𝜉0) as the

    Chinese response. To make clear how the responses differ, we calculate the differences

  • 12

    between Vt,1(𝜉0) and V0,1(𝜉0), DVt,1(𝜉0), to study the situation of Chinese series in the

    whole system. Generally, in the stationary system, DVt,1(𝜉0) will go to zero after a period

    of shocks.

    3.4 Estimation

    We adopt a maximum likelihood framework to estimate the BEKK models. The

    log-likelihood function of the joint distribution is calculated from the conditional

    multi-normal distribution. Denote 𝐿𝑡 as the log-likelihood of observation at time 𝑡, 𝑛 as

    the number of series and 𝐿 as the sum of log-likelihood of all time. The calculation is given

    as:

    𝐿 = ∑𝐿𝑡

    𝑇

    𝑡=1

    𝐿𝑡 =𝑛

    2ln(2𝜋) −

    1

    2ln|𝐻𝑡| −

    1

    2𝜖𝑡′𝐻𝑡

    −1𝜖𝑡

    Usually, by the BHHH4 method we use first order and second order derivatives to

    iterate and get the estimated parameters after achieving a required tolerance. However, in

    the BEKK model, the number of parameters is large and all parameters are in a matrix form

    so the likelihood function may be non-differentiable or even differentiable derivatives are

    hard to be calculated. Another numerical procedure, the Nelder-Mead algorithm5, as a

    modification of BHHH, works reasonably well for non-differentiable and complicated

    functions by approximating the derivatives using function values. To search for the optimal

    value efficiently, a proper initial value is essential. The estimation can be divided into two

    steps. The first step is to find out all the significant univariate models. The second step is to

    extend the univariate set to a multivariate model through the Nelder-Mead algorithm.

    Therefore, the initial parameter matrices we use are all diagonal matrices which come

    from univariate models.

    In the larger order BEKK model, the extended form of the matrix equation, which

    means the transformation from BEKK to 𝑣𝑒𝑐ℎ form, is hard to specify. Thus the

    interpretations of coefficients are not easy to understand. In this case, we apply the

    4 BHHH, introduced by Berndt, B. Hall, R. Hall and J. Hausman in “Estimation and Inference in Nonlinear Structural Models”

    (1974), is a non-linear optimization algorithm similar to the Gauss-Newton algorithm. 5 The Nelder-Mead method was first proposed by John Nelder & Roger Mead (1965) in “A simplex method for function

    minimization”. It is a nonlinear technique for minimizing an objective function in a many-dimensional space.

  • 13

    conditional covariance to measure the linkage between volatilities.

    4. Empirical Results

    In this section, we use the GARCH and BEKK models to model our stock returns data.

    The approaches by the GARCH method to Period I&II are not efficient in this case.

    Significant models and reasonable diagnostics cannot be given under the present study.

    Therefore we concentrate on the longest Period III (Depression Period) to continue

    searching for the existence of time-varying returns and volatilities in each series as well as

    market linkages. Then we analyze the fitness of models through comparison between the

    moving averages volatility and estimated ones. We try different approaches to the model,

    step by step, from univariate to multivariate. In Section 4.1, we report the ARCH effects in

    the stock return series and estimate 4 univariate GARCH models to do the preliminary

    fitting. Then we improve the univariate model set to a combined multivariate BEKK model

    in Section 4.2. In the next Section 4.3, we perform some diagnostic checking on the BEKK

    model. Then in Section 4.4, we report the market linkages demonstrated by time-varying

    conditional correlations. In Section 4.5, we introduce the VIRF function to distinguish the

    strength of linkages.

    Table 2. Estimated coefficients for univariate models

    SHCOMP SIASA FSTE100 SP500

    𝜇 0.870 *

    (0.368)

    1.043 **

    (0.340)

    𝜔 5.782*

    (2.747)

    10.078 *

    (4.926)

    3.048 **

    (1.437)

    2.538 **

    (0.876)

    𝛼 0.050 ***

    (0.013)

    0.048 ***

    (0.012)

    0.101 ***

    (0.022)

    0.123 ***

    (0.019)

    𝛽 0.922 ***

    (0.021)

    0.919 ***

    (0.024)

    0.882 ***

    (0.024)

    0.866 ***

    (0.017)

    AIC 8.236 8.569 7.858 7.924

    ARCH-LM Stat 11.340

    (p=0.000)

    5.276

    (p=0.021)

    9.314

    (p=0.002)

    76.342

    (p=0.000)

    Note: Values in parentheses are standard errors. ***, **and * represent significant levels at 0.1%, 1%

    and 5%.

  • 14

    Figure 3. Comparison between 7-days moving average and model estimates

    4.1. The evidence of ARCH effects

    As often done in previous studies, our first approach is to fit our series by

    AR(1)-GARCH(1,1) and get some estimations. The estimates of the autoregressive

    coefficients are not significantly different from zero and are therefore removed from the

    mean equation. To guarantee the significance of the model, we remove the constant terms

    in the mean equation of the Chinese series. Table 2 shows the results of the estimation

    without autoregressive effects in the mean equation including diagnostic statistics. From

    the table, except for two autoregressive constants in SHCOMP and FSTE100, all other

    coefficients are significant, implying a positive model fitting. The ARCH-LM tests for

    examining the existence of ARCH effects also prove the reasonable ARCH coefficients.

    Actually in a long-term financial returns case, ARCH effects on volatility exist widely.

    Another point in Table 2 is that we could find that returns in a similar economic

    environment (SHCOMP&SIASA, FSTE100&SP500) have similar scale of coefficients in the

  • 15

    variance equation. The larger 𝛽 and smaller 𝛼 in the Chinese cases means the changing

    volatility is more possibly caused by cumulative effects and keeps more stable in the

    long-term financial development.

    Figure 3 shows the comparison of volatilities between moving averages and model

    estimates. I choose the lag as 7 in the moving average to describe weekly volatilities, which

    could somehow express the characters of conditional volatilities. Through each row in

    Figure 3, we find that most of the sharpest peaks from the moving average are removed in

    the model estimates while the overall scale of volatilities remains the same. As we

    mentioned in Data Section, the volatilities in theChinese markets vary more dramatically

    than in the western markets. Almost in the whole period, the Chinese volatilities keep a

    heavy fluctuation, especially within 2007 to 2009. Compared to the Chinese markets, the

    British and American markets only have obvious changes in the second half year of 2008.

    The reason is known as the world-wide financial crisis. The more fluctuation in Chinese

    markets may be a result of the domestic multi-macroeconomic regulation and control.

    4.2 The evidence of market linkages

    The VAR-BEKK models are estimated by the maximum log-likelihood method and the

    results are reported in Table 3 and Table 4.

    Table 3 reflects models associated with SHCOMP. The elements of the three matrices,

    γij, aij and gij, represent the impact from series i to series j of returns and volatilities.

    Since we have removed the autoregressive terms, the first two rows of coefficients, 𝛾12

    and 𝛾21, capture the instant cross-market effects in the mean equation. γ12 captures the

    instant impact from SHCOMP to FTSE100/S&P500 and γ21 captures the instant impact

    from FTSE100/S&P500 to SHCOMP. The mutual interaction exists due to the significant

    r12 and r21. The positive γ12 and γ21 show the positive synergies between SHCOMP and

    western exchange. That is an increase in western indices/SHCOMP also promotes an

    increase in SHCOMP/western indexes, and a decrease in western indexes/SHCOMP will

    lead a decrease in SHCOMP/western indices. The following 𝑎1𝑖 and 𝑎2𝑖 measure the

    ARCH effect in the variance equation. In model 1 and model 2, the significant a1i means

    SHCOMP has a reasonable impact on the western stock volatilities. On the opposite, not

  • 16

    both British and American indexes have a significant effect on SHCOMP’s volatilities. The

    impact from FTSE100, measured by a211 , is insignificant but the impact from S&P500,

    measured by a212 , is significant. At last, the rest of the 𝐺 matrix, consisting of 𝑔1𝑖 and 𝑔2𝑖,

    are all statistically significant, indicating volatility spillovers between FTSE100/S&P500

    and SHCOMP. From the two models, we could conclude that the American market has

    multiple impacts on SHCOMP.

    Table 3. Estimated coefficients for bivariate VAR-BEKK model focus on SHCOMP

    Model 1: SHCOMP & FTSE100 Model 2: SHCOMP & SP500

    SHCOMP FTSE100 SHCOMP SP500

    𝛾1𝑖 0.055*** 0.0844***

    𝛾2𝑖 0.044*** 0.088***

    𝑎1𝑖 -0.117* 0.188*** 0.054* -0.059***

    𝑎2𝑖 0.025 0.166*** -0.009*** 0.368***

    𝑔1𝑖 0.990*** 0.018*** 0.996*** 0.007***

    𝑔2𝑖 0.013*** 0.954*** 0.011*** 0.926***

    LBQ(10) 8.895 8.150 7.497 10.433

    Probability 0.542 0.614 0.677 0.403

    LBQ(20) 21.137 19.725 21.315 17.408

    Probability 0.389 0.475 0.378 0.626

    LBQs(10) 18.273 52.529 52.566 18.067

    Probability 0.050 0.000 0.000 0.054

    LBQs(20) 28.950 60.077 72.236 25.303

    Probability 0.088 0.000 0.000 0.190

    LLR -10597.49 -10638.09

    Note: In the following context, we use an upper index γij1 and γij

    2 to distinguish γij in Model 1

    and Model 2, as well as aij and gij. LBQ(10) and LBQ(20) represent the Ljung-Box Q-statistic of

    standardized residuals with a lag equals to 10 and 20. LBQs(10) and LBQs(20) represent the

    Ljung-Box Q-statistic of squared standardized residuals with a lag of 10 and 20. These notes are

    also available in Table 4.

    The Ljung-Box Q statistics for the 10th and 20th orders in the standardized residuals of

    two bivariate models indicate an appropriate specification of the mean equation. However,

    when we examine the Ljung-Box Q statistics for the 10th and 20th orders in squared

    standardized residuals, we cannot always get independent squared residuals. In the

  • 17

    SHCOMP&FTSE100 model, the squared residuals of the conditional variance of FTSE100

    fail to be random. So is SHCOMP in the SHCOMP&SP500 model. Since we strictly take the

    order of variance equation from the univariate model and ensure a right selection for the

    univariate model, we cannot deny the model to be a good approach. And we insist to study

    the linkages from this BEKK model.

    Table 4. Estimated coefficients for bivariate VAR-BEKK model focus on SIASA

    Model 3: SIASA & FTSE100 Model 4: SIASA & SP500

    SIASA FTSE100 SIASA SP500

    𝛾1𝑖 0.206*** 0.351

    𝛾2𝑖 0.035*** 0.000***

    𝑎1𝑖 -0.103* 0.148* 0.003* 0.147***

    𝑎2𝑖 0.007*** 0.171*** -0.033*** 0.251***

    𝑔1𝑖 0.992*** 0.009 0.995*** -0.005***

    𝑔2𝑖 0.022*** 0.959*** 0.029*** 0.941***

    LBQ(10) 11.216 7.811 11.216 7.811

    Probability 0.340 0.647 0.340 0.647

    LBQ(20) 22.560 16.427 22.560 16.427

    Probability 0.310 0.689 0.310 0.689

    LBQs(10) 21.712 59.346 21.713 59.346

    Probability 0.016 0.000 0.016 0.000

    LBQs(20) 59.346 70.012 39.504 70.012

    Probability 0.000 0.000 0.005 0.000

    LLR -10895.70 -10930.87

    Table 4 shows the model estimation associated with SIASA and western markets.

    Similarly as Table 3, most coefficients are statistically significant. One difference is in

    model 4, where γ124 , which means the spillovers from SIASA to S&P500 in the mean

    equation, is insignificant. Another point is, in model 3, where g123 is also insignificant,

    indicating the absence of an unreasonable volatility spillover from SIASA to FTSE100.

    These two insignificant parameters show that SIASA has a weaker external influence than

    SHCOMP, but is influenced more from western markets.

    The Ljung-Box diagnostic on the residuals of the two bivariate models supply a good

    support for the residuals independence. And the Ljung-Box diagnostic on the squared

    residuals shows the same dependent results as above.

  • 18

    4.3 Diagnostic checking on the bivariate VAR-BEKK models

    The stationary condition of the BEKK(1,1) model is that all the eigenvalues of the

    matrix A1 ⊗𝐴1 + 𝐺1 ⊗𝐺1, where ⊗ is the Kronecker product of two matrixes, are

    smaller than one in modulus.6 Table 5 shows the eigenvalues of each model. It is clear that

    the modulus of all the eigenvalues is smaller than 1, guaranteeing stationarity of all the

    models.

    Table 5. Eigenvalues of bivariate BEKK models

    Eigenvalues Model 1 Model 2 Model 3 Model 4

    𝜆1 0.9972 0.9969 0.9992 0.9969

    𝜆2 0.9687 0.9878 0.9721 0.9631

    𝜆3 0.9201 0.9458 0.9324 0.9420

    𝜆4 0.8954 0.9415 0.9072 0.9104

    In the previous section, we reported the results of the bivariate BEKK models and

    analyzed the residuals. Now we carry out the likelihood ratio test to examine if the series

    are appropriately constructed in the models. We begin by checking the variance equation.

    We introduce the restrictions that all the off-diagonal parameters, the coefficients in the A

    and G matrices are zeros. These restrictions limit the interdependence of each series. The

    bivariate BEKK model with such restrictions, diagonal BEKK model, is equivalent to two

    univariate GARCH models. If we reject the validity of the restrictions, the combination of

    univariate models is not appropriate. The likelihood ratio test 1 statistics reported in Table

    6 are all quite large, rejecting the null hypothesis that the off-diagonal parameters are

    zeros.

    Recall that the coefficient we estimated, the insignificant r12 stands for the linkage of

    returns between Chinese and western markets. To examine whether this linkage exists or

    not, we test the log-likelihood ratio between the present model and the model without this

    linkage. The likelihood ratio test 2 shows that the null hypothesis for the zero linkage is

    6 See proof in section 11.3.1 of “GARCH Models—Structure, Statistical Inference and Financial Applications” by Christian

    Francq and Jean-Michel Zakoian.

  • 19

    rejected statistically. That is, though we get insignificant coefficients, they still make sense

    anyway.

    The likelihood ratio test 3 is to examine the cross-market spillovers in volatility, both

    from lagged volatility and lagged returns. The results prove the conjecture which the

    Chinese stock volatility is significantly affected by British and American markets. The cross

    spillovers in volatility strengthen the global linkage. When the shock happens in a market,

    others will react, too.

    A summary test for the cross spillovers is given by likelihood ratio test 4. All the

    cross-markets effects on Chinese markets are set to be zero in the null hypothesis. And the

    statistics suggest that the cross spillovers should be included in the model, implying a

    widely related global market.

    Table 6. Restriction tests concerning the bivariate BEKK model

    Log-likelihood ratio test

    statistics

    SHCOMP

    &FTSE100

    SHCOMP

    &SP500

    SIASA

    &FTSE100

    SIASA

    &SP500

    1.Diagonal BEKK,

    H0:𝑎𝑖𝑗 = 𝑔𝑖𝑗 = 0 df = 4

    129.56 16.74 39.44 97.44

    2.Instant cross market Impact

    from FTSE100 and S&P500,

    H0: γ12=0 df=1

    58.9

    63.62

    34.68

    47.2

    3.Cross spillovers in volatility

    from FTSE100 and S&P500,

    H0: 𝑎12 = 𝑔12 = 0 df=2

    50.64

    12.06

    6.02

    12.14

    4.Cross spillovers in both

    mean and variance equation,

    H0: γ12 = 𝑎12 = 𝑔120 df=3

    94.88

    74.68

    51.66

    57.26

    4.4 Market linkage demonstration

    Under our study of the depression period, we find that Chinese markets, both in

    Shanghai and Shenzhen, are linked in terms of returns and volatility with western

    developed markets in Britain and U.S. Figure 4 is the illustration of the conditional

    correlation of the series pairs in our models. The figure supports our evidence of market

    linkages nowadays. The fluctuations of dynamic correlations follow a similar trend in the

  • 20

    whole period, showing a co-movement of global stock markets. And from Figure 4, we can

    also observe that the correlation is not so large, mostly between -0.1 to 0.3, except for

    some individual intervals. Thus we know the linkages exist but the influences are limited. It

    is true in the economic environment since the domestic economy is mainly affected by

    domestic factors. The external impacts function on the economy could not threaten the

    basement.

    Figure 4. Estimated conditional correlations during November 2008 and March 2012

    4.5 VIRF for the estimated BEKK models

    In the following, we refer to the VIRF plots given in Figure 5 for the estimation results

    of BEKK models. In order to visualize the function, we produce a shock in the bivariate

    model. The shock in this case is designed as ξ0,1, which means there’s one innovation in

    the second series but the first series keeps the same. In the stationary situation, the shock

    will be consumed by the long-term progress in the model. Thus, the shock in the volatility

  • 21

    will disappear after some periods. We manage this change in the volatility, describing it as

    the cross-series interferences.

    Figure 5. VIRF on Chinese stock after an external shock

    In Figure 5, the red dashed line represents for the shock from FTSE100 to SHCOMP,

    and the red solid line is the shock from S&P500 to SHCOMP. The blue dashed line and blue

    solid line are the shock from FTSE100 to SIASA and the shock from S&P500 to SIASA

    respectively. It is obvious that when we introduce ξ0,1 to the model, the volatility of first

    series will jump to a high level immediately. It may continue increasing for a period or just

    fall down to regain balance.

    The first point we find in Figure 5 is that the shock comes from S&P500 is stronger

    than that from FTSE100, i.e. solid line is always beyond dashed line. The peak of the

    reaction of SHCOMP from S&P500 is around 330 but the reaction of SHCOMP from

    FTSE100 is as small as 5.65. It is the same in the case of SIASA. The peak of the reaction of

  • 22

    SIASA from FTSE100 is around 65.51 on the contrast of a peak about 524.54 from S&P500.

    This is an evidence that the U.S. stock market has more influence in Chinese markets,

    compared to the British stock market.

    The second point is that SHCOMP keeps more stability with the western shocks than

    SIASA. It is clear in the figure that the blue solid line is beyond the red solid line and the

    blue dashed line is also beyond the red dashed line. In the meanwhile, tails of red lines are

    on the left of blue lines, implying that it takes a longer time to regain balance in SIASA than

    in SHCOMP. The phenomenon may come from the fact that small companies do not have

    enough anti-risk ability. When the external depression strikes the domestic economics, the

    small companies are more likely to face the problems of losing orders and get trouble with

    capital turnover. However the large companies would have their business distributed to

    multiple operations in case of centralized risk. Even if the shock happens, large companies

    could recover soon due to its leading market position and rich capital reserve. Therefore

    these advantages help large companies to keep their system more stable and the

    disadvantages of small size of the listed companies would undertake more risk in

    economic globalization.

    The last point which cannot be ignored is that the sensitivities of SHCOMP and SIASA

    to the shock from FTSE100 or S&P500 are different even SHCOMP takes more stability to

    resist the shocks. In Figure 5, at 𝑡 = 91 and 𝑡 = 132, the red lines reach to their peaks,

    which means it takes 91 and 132 days to result in the biggest shocks from FTSE100 and

    S&P500 to SCHOMP. Therefore the approximate average shock rates are 5.65/91 = 0.06

    and 330/132 = 2.5 which the latter one is nearly 42 times of the former one. While for

    shocks from FTSE100 and S&P500 to SIASA, it takes 145 and 150 days to reach to the top.

    The average shock rates of SIASA are calculated as 65.51/145 = 0.45 and 524.54/150 =

    3.5, which the rate from S&P500 is only 8 times of that from FTSE100. From the high

    level shock rates of SIASA, it is obvious that the SIASA is more sensitive to the external

    spillovers. On the other hand, spillovers from S&P500 are keeping high level impacts both

    in SIASA and in SHCOMP while spillovers from FTSE100 show different influences in the

    two Chinese markets. In a time of no longer than 150 days, the impact would reach to its

  • 23

    peak. In case of the bankruptcy due to the heaviest strike, all the companies should try to

    guarantee a safe cash flow during this period.

    5. Conclusion

    This study investigates the linkages between Chinese stock exchange markets and

    global developed markets. Under the BEKK approach to daily stock returns from November

    2008 to March 2012, we find the evidence supporting the existence of linkages. The

    spillovers of returns and volatilities, from western markets to Chinese markets, are found

    in the models. On the opposite, the stock spillovers from China to western countries are

    also significant, which indicates that Chinese stock markets are developing from the

    emerging markets and Chinese financial economics is an important part of the world

    financial system.

    We illustrate the dynamic linkages through dynamic conditional correlations. From

    our illustration linkages are observed neither too weak nor too strong. Thus the volatility

    impulse response functions are used to distinguish the different linkages under systematic

    shocks. The VIRF explains the co-movement of global stock interaction. The stock exchange

    in Shanghai, representing the large listed companies in China, is less sensible to the

    western shocks compared to the stock exchange in Shenzhen, which represents mostly the

    medium and small companies. The shocks from U.S. markets are also proved to be stronger

    than those from European markets. Though Europe is the largest trading partner, in the

    financial aspect, U.S. still takes the leading influence on China.

    Our study suggests an anti-risk method for Chinese companies. Through our VIRF

    analysis, since the power of spillovers from U.S and Europe are different, larger companies

    should care more risks from U.S. while small companies should distribute their business

    associated with Europe and U.S. in appropriate proportions to minimize the shock from

    both regions. In the meanwhile, they should prepare for the continuous shocks within a

    specific period.

  • 24

    References

    1. Ågren, M., 2006. Does oil price uncertainty transmit to stock markets? Working paper in

    Department of Economics of Uppsala University, 2006:23.

    2. Bollerslev, T., 1986. Generalized Autoregressive Conditional Heteroskedasticity. Journal

    of Econometrics, 31: 307-327.

    3. Bollerslev, T., Engle, R., Wooldrige, J., 1988. A capital asset pricing model with

    time-varying covariances. Journal of Political Economy, 96: 116-131.

    4. Brooks, RD., Ragunathan, V., 2003. Returns and volatility on the Chinese stock

    markets. Applied Financial Economics, 13: 747–52.

    5. Chow, R., Lin, J., Wu, C., 1999. Modeling the Taiwan stock market and international

    linkages. Pacific Economic Review, 4: 305-320.

    6. Chow, G., Lawler, C., 2003. A time series analysis of the Shanghai and New York stock

    price indices. Annals of Economics and Finance, 4:17-35.

    7. Diekman, K., 2011. Are there Spillover Effects from Hong Kong and the United States to

    Chinese Stock Markets? Working Papers with number 89 in Institute of Empirical

    Economic Research, 2011-12-12.

    8. Engle, R., 1982. Autoregressive conditional heteroscedasticity with estimates of the

    variance of United Kingdom inflation. Econometrica, 50: 987-1007.

    9. Engle, R., Kroner, K., 1995. Multivariate simultaneous generalized ARCH. Econometric

    Theory, 11: 122-150.

    10. Francq, C., Zakoian, JM., 2010. GARCH Models: Structure, Statistical Inference and

    Financial Applications. Wiley, New York.

    11. Hafner, C., Herwartz, H., 1998. Volatility impulse functions for multivariate GARCH

    models. CORE Discussion Papers from Universite catholique de Louvain, Center for

    Operations Research and Econometrics, 2001-09-01.

    12. Hafner, C., Herwartz, H., 2006. Volatility impulse responses for multivariate GARCH

    models: an exchange rate illustration. Journal of International Money and Finance, 25:

    719-740.

  • 25

    13. Haroutounian, M., Price, S., 2001. Volatility in the transition markets of central Europe.

    Applied Financial Economics, 11: 93-105.

    14. Kim, Y and Shin, J., 2000. Interactions among China-related stocks. Asia-Pacific

    Financial Markets, 7: 97–115.

    15. Li, H., 2007. International linkages of the Chinese stock exchanges: a multivariate

    GARCH analysis. Applied Financial Economics, 17: 285-297.

    16. Li, H., 2011. China’s stock market reforms and its international stock market linkages.

    BMRC-QASS conference on Macro and Financial Economics, 2011-05-24.

    17. Li, H., Majerowska E., 2007. Testing stock market linkages for Poland and Hungary: A

    multivariate GARCH approach. Research in International Business and Finance, 22:

    247-266.

    18. Pen., LY., Sevi, B., 2009. Volatility transmission and volatility impulse response functions

    in European electricity forward markets. Energy Economics, 32: 758-770.

    19. Wang, P., Liu, A., Wang, P., 2004. Returns and risk interactions in Chinese stock

    markets. Journal of International Financial Markets, Institutions and Money, 14: 367–84.

    20. Worthington, A., Higgs, H., 2004. Transmission of equity returns and volatility in Asian

    developed and emerging markets: a multivariate GARCH analysis. International Journal

    of Finance and Economics, 9: 71-80.

    21. Xiang, Z., Weijin, Z., Jie, Z., 2011. Volatility spillovers between the Chinese and world

    equity markets. Pacific-Basin Finance Journal, 20: 247-270.

    22. Yang, J., Hsiao, C., Li, Q., Wang, Z., 2006. The emerging market crisis and stock market

    linkages: future evidence. Journal of Applied Econometrics, 21: 727-744.

    23. Zhou, X., Zhang, W., Zhang, J., 2011. Volatility spillovers between the Chinese and world

    equity markets. Pacific-Basin Finance Journal, 20: 247-270.