Chinese institutional investors’ sentiment

Int. Fin. Markets, Inst. and Money 18 (2008) 374–387

Available online at www.sciencedirect.com

Chinese institutional investors’ sentiment

Gerhard Kling a,∗, Lei Gao b

a Bristol Business School, Coldharbour Lane, Bristol BS16 1QY, UKb School of Business, Shantou University, Shantou 515063, PR China

Received 10 October 2005; accepted 4 April 2007Available online 10 April 2007

Abstract

We use daily survey data on Chinese institutional investors’ forecasts to measure investors’ sentiment.Our empirical model uncovers that share prices and investor sentiment do not have a long-run relation;however, in the short-run, the mood of investors follows a positive-feedback process. Hence, institutionalinvestors are optimistic when previous market returns were positive. Contrarily, negative returns trigger adecline in sentiment, which reacts more sensitively to negative than positive returns. Investor sentiment doesnot predict future market movements—but a drop in confidence increases market volatility and destabilizesexchanges. EGARCH models reveal asymmetric responses in the volatility of investor sentiment; however,Granger causality tests reject volatility-spillovers between returns and sentiment.© 2007 Elsevier B.V. All rights reserved.

JEL classification: G14; C22

Keywords: Shanghai stock exchange; Institutional investor; Investors’ sentiment

1. Introduction

Research on investors’ sentiment has focused on small shareholders for two main reasons: (1)data on the sentiment of large investors are not available; (2) small shareholders are supposed toact like noise traders; thus, the impact of noise trading can be analyzed.1 In contrast, we measurethe sentiment of institutional investors in China by collecting daily survey data on investors’ fore-casts. Hence, we know how many institutional investors expect a positive or negative change inthe Shanghai stock exchange (SSE) index on the next day. The interrelation between institutional

∗ Corresponding author. Tel.: +44 11732 83418.E-mail addresses: [email protected] (G. Kling), [email protected] (L. Gao).

1 Kelly (1997) showed that the likelihood of being a noise trader declines with household income; thus, the analysis ofnoise trading is limited to small investors. We define noise trading in the sense of DeLong et al. (1990).

1042-4431/$ – see front matter © 2007 Elsevier B.V. All rights reserved.doi:10.1016/j.intfin.2007.04.002

mailto:[email protected]

mailto:[email protected]

dx.doi.org/10.1016/j.intfin.2007.04.002

G. Kling, L. Gao / Int. Fin. Markets, Inst. and Money 18 (2008) 374–387 375

investors’ sentiment and stock markets might differ from former findings, as institutional investorsare more professional in analyzing public information, conduct large-scale transactions, andpossess informational advantages.

In China, several authorized closed-end investment funds exist, and open-end funds wererecently introduced. Yet other forms of institutional investment like pension funds or managedportfolios of insurance companies are negligible. Based on World Bank data for 2001, Chineseinstitutional investors’ assets reached 19% of GDP; this figure was lower than the correspondingfigures for Hungary (26%), the Czech Republic (32%), and further developed Asian markets, e.g.,South Korea (82%). As Chinese institutional traders are less important than their US and Europeancounterparts, one can regard our analysis as a lower estimate of the importance of institutionalinvestors’ sentiment in other markets.

Three questions are at the core of our interest: (1) what determines the forecasts of institutionalinvestors; (2) can the sentiment index predict future stock market movements; (3) does a dropin sentiment or a higher volatility of investors’ sentiment destabilize stock markets by triggeringhigher market volatility? Besides these core issues, we analyze asymmetric responses of investorsentiment based on EGARCH and TARCH models. In addition, bivariate GARCH models allowus to explore the time-varying correlation between sentiment and stock returns.

The literature on investors’ sentiment covers our research questions; however, only the sen-timent of small investors has been studied so far. Our first question, the search for factors thataffect investors’ sentiment, is still under discussion—albeit the formal model of investor senti-ment developed by Barberis et al. (1998) has provided a baseline for economists. The literaturefocused mainly on our second question: Neal and Wheatley (1998), for instance, quantified thepredictability of stock returns using different measures of investors’ sentiment. Lee et al. (2002)discussed our third question for the US in analyzing the influence of investors’ sentiment onvolatility and excess returns.

The remainder of our paper is organized as follows: Section 2 reviews the recent literatureon theoretical considerations concerning sentiment and market activity, institutional trading,sentiment of small investors, survey methods, and indirect measures of sentiment. Section 3describes our data sources, followed by our empirical analysis that starts with a univariate-timeseries approach followed by unit-root tests and multivariate methods, namely VAR and bivariateGARCH models. Based on our empirical findings, we discuss consequences for mature financialmarkets and peculiarities of the Chinese stock exchange.

2. Literature review

Classical finance theory neglects the role of investor sentiment (see Gomes et al., 2003), asinvestors are supposed to be rational. Even if some investors are not rational, arbitrageurs canexploit their irrational behavior, thus causing prices to reflect future discounted cash flows. Byintroducing demand shifts driven by irrational speculation and binding arbitrage constraints,sentiment can influence stock returns. Hence, a mispricing triggered by uninformed demandshocks could occur. This is in line with the idea that sentiment can be regarded as the propensityto speculate and thus reflects the optimism or pessimism of investors. Especially if arbitrage is riskybecause of subjective valuations,2 high volatility, short selling constraints,3 and thin trading, the

2 A broad range of valuations increases noise trader risk (see DeLong et al., 1990; Shleifer and Vishny, 1997).3 Arbitrage is limited by short selling constraints (see e.g., Jones and Lamont, 2002). In China, the short selling of stock

is hardly possible and derivate instruments are not developed.

376 G. Kling, L. Gao / Int. Fin. Markets, Inst. and Money 18 (2008) 374–387

impact of sentiment on returns should be more pronounced. In particular, the situation in Chinais characterized by the lack of an earnings history, apparently unlimited growth potential, andunsophisticated investors; hence, the role of investor sentiment should be tremendous.

As institutional investors account for more than 50% of trading in the US (see Cai and Zheng,2004), their influence on stock markets is of considerable interest. Previous studies have mainlyfocused on the impact of institutional trading on stock returns, and assessed the consequencesof large-scale transactions, positive-feedback trading, and the herding behavior of institutionalinvestors (see Lakonishok et al., 1992). Nevertheless, Bohl and Brzeszczynski (2006) studiedthe influence of institutional ownership on volatilities. They found that institutional investorsstabilize exchanges, as volatility drops after institutional investors become principal shareholders.These studies, however, have in common that they analyze the effect of institutional trading andownership, whereas our paper tries to measure the relevance of institutional investors’ sentimentfor stock returns and volatilities. This requires data on institutional investors’ sentiment, which areusually not available as surveys or alternative measurement methods focus on small shareholders’sentiment.

Results on whether the sentiment of small investors is a good indicator of future marketmovements are somehow ambiguous. Brown and Cliff (2001) provided only weak evidence ofshort-run predictability, but found long-run predictability of returns for a period of about 2–3 years.Kenneth and Statman (2000) detected a significantly negative correlation between US investors’mood and returns of the S&P500. Apart from the discussion about whether investor sentiment canpredict future market movements, several studies identified an interrelation with market volatility.By using GARCH models, Lee et al. (2002) analyzed the impact of investors’ sentiment on marketvolatility. They uncovered that a decline in investors’ sentiment triggers higher volatility.

Several direct and indirect measures of sentiment have been used in the recent literature. ForUS studies, the sentiment index provided by Investors’ Intelligence of New Rochelle based ona weekly survey of 135 independent advisory services, is commonly applied.4 Survey methodshave provoked the skepticism of many economists. Yet Blinder (1991) argued that a well-designedsurvey can provide stylized facts and valuable clues that go beyond standard models and are notavailable to econometricians. The so-called indirect approach serves as an alternative to surveydata. For US studies, the fluctuations of closed-end fund shares are used as a proxy of sentimentbecause predominantly private investors hold these shares.5 Besides observing closed-end funds,Baker and Stein (2004) argued that market liquidity could be regarded as a proxy of investors’sentiment. They used bid-ask spreads and turnovers as a measure of liquidity. This measure refersto the whole market; thereby, it is impossible to distinguish between different types of investors.Consequently, we favor a survey approach in order to isolate the opinion of institutional investorsand their market impact from all other market participants.

3. Data

Since 20 April 2001, the Chinese Central Television Station has been surveying 75 leadinginstitutional investors. On a daily basis, they are asked about their forecast regarding the Shanghaistock exchange index’s movements on the following day. The survey is conducted at 4 p.m. each

4 Many studies like Siegel (1992) used this index and showed that it can predict market movements.5 Lee et al. (1991) proposed this method, and Neal and Wheatley (1998) showed that fluctuations of closed-end funds

can predict the returns of small caps.


Fig. 1. Predicted values for the optimism of Chinese institutional investors. Using the results of our univariate analysisdescribed in model (2), we obtain predicted values for the change of investor sentiment. Our extended ARMA modelcaptures the systematic behavior of the sentiment index. The residuals follow a white-noise process.

day, and the results are published online.6 As the market closes before 4 p.m., institutional investorscould use today’s market development as information for their forecast. The surveyed institutionalinvestors have three choices: bullish, bearish, and tie. The web page contains the numbers ofinvestors opting for each type of forecast—but individual opinions remain confidential. Based onthese figures, we calculate for each day the percentage of optimistic institutional investors and setup the daily index of optimism. This index is constructed in a similar way as the index providedby Investors’ Intelligence of New Rochelle, which is used for US studies. The sentiment indexreaches 0 if everyone expects a decline in the stock market, and 100 if everyone is optimistic.Henceforth, the sentiment index is bounded between 0 and 100; however, the variable is normallydistributed, as extreme cases of 0% or 100% optimistic responses are rare.7

Three major aspects make the Chinese daily survey unique: first, the survey is based on theopinion of institutional investors, which allows the measurement of their sentiment. Second, itprovides daily information on investors’ mood and thus has the highest possible frequency ofinformation. Sentiment indices for the US are usually on a weekly basis. Third, the forecastinghorizon is only 1 day, as investors are asked about their forecast regarding tomorrow’s marketmovement. Generally, there is no hint that the survey we use contains any incentives to give falseinformation, as individual forecasts remain confidential.

4. Empirical models

4.1. Univariate analysis of investor sentiment

Daily institutional investors’ sentiment fluctuates remarkably over time (see Fig. 1). To modelthe variance pattern, we use an ARCH (p) model based on the seminal paper of Engle (1982). The

6 See http://www.eestart.com/zxzx/20040220/558105.html.7 Shapiro-Wilk and Shapiro-Francia W tests indicate, with p-values of 0.084 and 0.122, that the null hypothesis of the

sentiment index being normally distributed cannot be rejected at the 95% level of confidence.

http://www.eestart.com/zxzx/20040220/558105.html


log of the sentiment index is labeled yt, and its conditional variance is �2t . Hereafter, we always

refer to the logarithmically transformed sentiment index except when stated otherwise.

yt = α0 + εt, E(εt|Ωt) = 0, var(εt|Ωt) = σ2t = β0 +

p∑j=1

βjε2t−j (1)

To determine the appropriate lag structure of the ARCH (p), we conduct LM tests by regressingsquared residuals of Eq. (1) on their lagged values.8 Justified by these outcomes, we specify anARCH (2) process.

Besides dealing with a nonlinear structure in conditional variances, one should capture theautoregressive behavior of the sentiment index. Accordingly, the next step is to specify an ARMA(p, q) model and insert this structure into the mean Eq. (1).9 By looking at the autocorrelationfunction and the partial autocorrelation function, we can justify an ARMA (3, 4) specification.A reduction of the lags does not lower information criteria like the Akaike criterion; thus, weincorporate an ARMA (3, 4) term into our mean equation.

yt = α0 +3∑

j=1

γjyt−j +4∑

j=1

δjεt−j + εt, �2t = β0 +

2∑j=1

βjε2t−j (2)

Fig. 1 depicts the actual and predicted values of the sentiment index. To ensure that our modelcaptures all nonlinearities in the time series, we confirm through white noise tests that the residualsexhibit erratic behavior.10 The Barlett’s periodogram-based test procedure fails to reject the nullthat the series is white noise, and the Portmanteau Box-Pierce test points in the same direction.11

How can one interpret these findings? The negative autocorrelation suggests that a positive forecastis likely to be followed by a negative one. Yet the importance of former assessments dies outquickly. The following section focuses on the interrelation between sentiment, stock returns, andvolatility.

4.2. The interrelation between sentiment and stock returns: a VAR analysis

Before using more elaborate techniques, a simple cross-correlogram can help to detect whetherreturns or forecasts move first. Fig. 2 plots the cross-correlation for different lags and leads. Apositive correlation of about 0.4106 between lagged market returns and forecasts reveals aninteresting interdependency. Previous market returns influence the current mood of institutionalinvestors positively. This is a highly interesting finding, as current market returns do not affect

8 Multiplying the number of observations and the R2 achieved for different lag structures yields a test statistic that isChi-squared distributed with p degrees of freedom. It is not possible to reject the null hypothesis that ARCH terms arejointly insignificant for an ARCH (3) specification and higher orders—but with a p-value of 0.001 for the χ2-test, anARCH (2) specification has significant coefficients.

9 To reliably estimate an ARMA specification for the sentiment index, the time series has to be stationary. By lookingat Fig. 1, one obtains the impression that this requirement is fulfilled. Also, ADF tests indicate that the time series doesnot possess a unit root. A subsequent section discusses this issue more thoroughly.10 Before performing white noise tests, one has to fill two gaps in the time series of the residuals. Because the problem

of missing values is rather limited, we propose a very simple solution for this problem. According to this, the two missingvalues in the time series are replaced by the previous values of the residuals.11 The Barlett’s test statistic reaches 0.950 (p-value: 0.327), and the Portmanteau Q statistic is equal to 39.622 (p-value:

0.487).


Fig. 2. Cross-correlogram between market returns and forecasts. The cross-correlogram shows that market returns of theprevious day are positively correlated with the current value of the forecasting index.

forecasts despite the fact that the survey is conducted after the market is closed. Institutionalinvestors could use current returns to improve their forecasts, yet they rely on past informa-tion. To study this time structure and correlation further, our strategy is as follows. First, byapplying unit-root tests, we test whether both time series are stationary. Second, we modelthe short-term interrelation by a VAR approach and select the lag structure based on informa-tion criteria. Thereafter, Granger causality tests show whether forecasts affect market returns orvice versa.

We test for stationarity by calculating ADF based tests.12 Due to the ‘traditionally’ weakexplanatory power of ADF tests, we also report the results of KPSS tests that use a different nullhypothesis.13 If both test procedures point in the same direction, the result can be regarded asinformative. Both test procedures point in the same direction and thus confirm that both timeseries are stationary.14 As investor sentiment is I(0) and stock prices are I(1), both time seriescannot be cointegrated and a long-run relationship does not exist. However, we analyze theshort-run dynamics by a VAR model and determine the number of lags as suggested by infor-mation criteria, namely Schwarz Bayesian (SBIC), Akaike (AIC), and Hannan-Quin (HQIC).As expected,15 the SBIC favors the inclusion of two lags, whereas AIC and HQIC request oneadditional lag. Hence, we estimate a VAR with two and three lags, respectively, and carry outGranger causality tests. Table 1 presents the outcomes. The interpretation is twofold. Only themarket return rt−1 of the previous day affects the current forecasts yt. Moreover, Granger causalitytests reveal that investor sentiment does not influence returns and hence is not a valuable infor-mation for the market. It is noteworthy that a positive market return triggers a positive forecast of

12 Elliott et al. (1996) modified ADF tests by a GLS approach that accounts for serial correlation.13 Kwiatkowski et al. (1992) constructed this test procedure.14 The ADF test statistic reaches −4.511 for the sentiment index and −17.451 for market returns. Hence, the null

hypotheses that the series have a unit root can be rejected on the 99% level of confidence. KPSS test statistics are 0.117for the sentiment index and 0.044 for market returns; hence, the null hypothesis cannot be rejected. The Schwert criteriondetermines the lag structure.15 The SBIC usually favors less complex models.


Table 1Results of the VAR and Granger causality tests for three and four lags

VAR with two lags VAR with three lags

Dependent variable: forecasting index yt

Constant 2.3828 (0.000) 1.9938 (0.000)yt−1 0.1836 (0.000) 0.1348 (0.001)yt−2 0.2163 (0.000) 0.1984 (0.000)yt−3 – 0.1651 (0.000)rt−1 12.0890 (0.000) 12.2988 (0.000)rt−2 −1.3894 (0.215) −0.4192 (0.710)rt−3 – −0.2720 (0.806)

Adjusted R2 0.2633 0.2883

Dependent variable: market returns rt

Constant 0.0133 (0.087) 0.0116 (0.181)yt−1 −0.0019 (0.258) −0.0021 (0.214)yt−2 −0.0016 (0.272) −0.0012 (0.456)yt−3 – 0.0003 (0.846)rt−1 0.0150 (0.720) 0.0253 (0.542)rt−2 −0.0147 (0.748) −0.0206 (0.656)rt−3 – −0.0177 (0.697)

Adjusted R2 0.0073 0.0088

Observations 578 570

Granger causality tests:rt−j Granger cause yt 142.2300 (0.000) 148.5700 (0.000)yt−j Granger cause rt 3.2292 (0.199) 2.6615 (0.447)

Inserting market returns rt and the forecasting index yt as endogenous series into a VAR approach can account for bothdirections of causality. Granger causality tests indicate that market returns affect forecasts—but not vice versa. Due to thefact that the three information criteria do not favor the same lag structure of the VAR, a specification with three and fourlags is used. In the case of Granger causality tests, Wald test statistics appear in columns two and three and are Chi-squareddistributed with three or four degrees of freedom. P-values are in parentheses.

the institutional investors. Accordingly, institutional investors’ sentiment can be interpreted as apositive-feedback process with regard to previous market returns. Variance decompositions high-light that an impulse in returns determines about 20% of the subsequent innovations in investorsentiment.

4.3. Asymmetric responses of investor sentiment to changes in past market returns

A widely discussed question in the literature is whether market participants react more severelyif bad news is released.16 Putting this in other words, one should test the hypothesis that neg-ative market returns influence sentiment more than positive returns. To implement this test intoour ARCH framework, we modify model (2) by inserting an indicator variable labeled I and an

16 Sias (1997) found that the expectations of small investors (compared to institutional investors) react more sensitivelyif the market environment changes.


interaction term of the indicator variable with lagged market returns Irt−1.17 The indicator vari-able I takes the value one when the previous market return rt−1 is negative, and the value zerootherwise. Accordingly, we allow an asymmetric response of investor sentiment to past returns inthe mean-equation. To explore whether asymmetric responses also exist in the variance-equation,we run three different specifications, namely the standard ARCH(2), EGARCH and TARCHmodels.

yt = α0 + α1rt−1 + α2I + α3I rt−1 +3∑

j=1

γjyt−j +4∑

j=1

δjεt−j + εt,

ARCH(2) : �2t = β0 +

2∑j=1

βjε2t−j,

EGARCH : εt = vt

√ht ; ln(ht) = β0 +

2∑j=1

βj

εt−j

h0.5t−j

+2∑

j=1

λj

∣∣∣∣∣ εt−j

h0.5t−j

∣∣∣∣∣ ,

TARCH : εt = vt

√ht ; ht = β0 +

2∑j=1

βjε2t−j +

2∑j=1

λjDt−jε2t−j (3)

Note that vt is a white noise process with variance σ2v = 1, and Dt−j = 1 if εt−j < 0. Table 2 shows

the outcomes of model (3) with an ARCH(2) specification. To discuss asymmetric responses ofinvestor sentiment in the conditional variance, Table 2 also reports TARCH and EGARCH modelresults.18 The coefficient of the indicator variable I is not significant; however, the coefficient ofthe interaction term Irt−1 is highly significant (p-value = 0.000) and has a positive sign regardlessof whether we use an ARCH, EGARCH or TARCH model. Consequently, if the previous returnrt−1 is negative, investor sentiment exhibits a pronounced decline. In particular, the magnitude ofthe impact of past returns is about three times larger in the case of negative returns compared toincreasing stock prices. Based on the threshold ARCH model (see Chen, 1998), we do not findasymmetric responses in the conditional variance of forecasts triggered by past market movements.However, the EGARCH model confirms that negative shocks (measured by −βj + λj) have astronger impact on conditional variance than positive shocks (measured by βj + λj). This confirmsChen’s (1998) finding that investor sentiment exhibits a higher variance after negative shocks.Based on our estimates for the mean and conditional variance equations, we state that negativereturns have a pronounced negative impact on investor sentiment, and that negative shocks increasethe conditional variance of investor sentiment.

4.4. The impact of investor sentiment on market volatility

After confirming that previous stock returns affect sentiment but not vice versa, we focus onthe impact of institutional investors’ sentiment on market volatility. To test the impact of laggedsentiment yt−1 on volatility and to capture volatility patterns, we apply a GARCH model of the

17 Note that market returns are weakly exogenous and Granger-cause sentiment; hence, inserting lagged returns does notcause an endogeneity bias.18 We use the EGARCH model developed by Nelson (1991), and the TARCH model based on Glosten et al. (1993).


Table 2Asymmetric responses of investor sentiment in mean and variance equation

Explanatory variables Model (3) withasymmetry in meanequation

EGARCH withasymmetry inmean equation

TARCH withasymmetry in meanequation

Constant in mean equation 4.0302 (0.000) 4.0748 (0.000) 4.0450 (0.000)rt−1 5.7138 (0.003) 5.7534 (0.006) 5.2269 (0.016)I −0.0119 (0.626) −0.0167 (0.511) −0.0185 (0.461)Irt−1 10.2118 (0.000) 10.7557 (0.000) 11.7037 (0.000)yt−1 −0.4453 (0.000) 0.7106 (0.000) −0.1968 (0.003)yt−2 −0.0704 (0.405) 0.8880 (0.000) −0.0453 (0.543)yt−3 0.6782 (0.000) −0.5981 (0.000) 0.7964 (0.000)εt−1 0.6782 (0.000) −0.5131 (0.000) 0.4434 (0.000)εt−2 0.4613 (0.000) −0.7587 (0.000) 0.3472 (0.001)εt−3 −0.2964 (0.005) 0.3856 (0.001) −0.5237 (0.000)εt−4 0.1476 (0.002) −0.0820 (0.183) 0.0369 (0.472)

Constant in variance equation 0.0809 (0.000) −2.2360 (0.000) 0.0834 (0.000)ε2t−1 0.0262 (0.492) – 0.0415 (0.438)

ε2t−2 0.2516 (0.000) – 0.3389 (0.000)

Dt−1ε2t−1 – – −0.0968 (0.187)

Dt−2ε2t−2 – – −0.1860 (0.060)

εt−1/h0.5t−1 – −0.0598 (0.237) –

εt−2/h2t−2 – −0.1448 (0.010) –

|εt−1/h0.5t−1| – 0.0652 (0.468) –

|εt−2/h0.5t−2| – 0.3870 (0.000) –

Observations 593 593 593Portmanteau Q test 46.0209 (0.237) 41.0870 (0.423) 47.0187 (0.207)Barlett’s B statistic 1.0892 (0.186) 0.9785 (0.294) 1.3294 (0.058)

P-values are set in parentheses.

following form:19

rt = α0 + α1yt−1 + ut, E(ut|Ωt) = 0,

var(ut|Ωt) = σ2ut = β0 + β1u

2t−1 + β2σ

2ut−1 + β3yt−1 (4)

We find that the sentiment index has a significantly negative impact on the conditional varianceof market returns. Table 3 reports the results of different specifications of model (4); thus, weestimate a GARCH(1,1) model without any impact of the lagged sentiment index, a specificationwith the lagged sentiment index in the mean and variance equations, and a model that includesthe lagged investor sentiment in the variance equation only. Accordingly, one can confirm that thesentiment has an impact on volatility—but not on returns. If the sentiment is very low, one should

19 Note that sentiment is a stationary variable; hence, taking the first-difference, as done by Lee et al. (2002) for US data,is not required and would be a misspecification. We specify a GARCH (1,1) model after testing for the order of an ARCH(p) model using the LM test as mentioned in a previous section. Because the lag of the ARCH (p) model would exceedthree, one can approximate this complex model by a simpler GARCH (1,1) specification (see Bollerslev, 1986).


Table 3GARCH specifications for stock returns with and without lagged investor sentiment

Explanatory variables GARCH withoutsentiment

Sentiment in variance Sentiment in meanand variance

Constant in mean equation −0.0006 (0.124) −0.0007 (0.115) −0.0020 (0.176)yt−1 – – 0.000 (0.344)Constant in variance equation 0.0000 (0.000) −8.2926 (0.000) −8.1759 (0.000)ε2t−1 0.2514 (0.000) 0.2462 (0.006) 0.2444 (0.005)

ht−1 0.7193 (0.000) 0.7255 (0.000) 0.7258 (0.000)yt−1 – −0.8039 (0.000) −0.8327 (0.000)

Observations 597 594 594Portmanteau Q test 47.6520 (0.190) 47.6520 (0.190) 48.4644 (0.169)Barlett’s B statistic 0.7152 (0.686) 0.7152 (0.686) 0.6193 (0.838)

Note: Based on model (4), we apply a GARCH(1,1) model to stock returns. The Portmanteau Q test cannot reject the nullhypothesis that the residuals follow a white noise process. P-values are set in parentheses.

expect a high volatility on the market.20 Obviously, statistical significance does not necessarilyimply that the impact of sentiment on volatility is an economically important effect. To evaluatethe relevance of institutional investors’ sentiment, we determine the one-step ahead forecasts ofthe conditional variance based on our GARCH(1,1) model with the lagged sentiment. Then, weexpress the contribution of sentiment relative to the predicted conditional variance in percentagepoints. In the case of overwhelming optimism (95 percentile of the sentiment indicator), marketvolatility drops by 12.74% on average. If institutional investors are very pessimistic (5 percentileof the sentiment indicator), volatility rises by 6.60% on average. Consequently, the impact of insti-tutional investors’ sentiment on market volatility has a considerable economic effect. Pessimismof institutional investors destabilizes the stock market by triggering higher market volatility.Having thus confirmed the empirical relevance of our results, do our findings also make sensetheoretically?

Theoretical and empirical studies confirm that stocks with a higher volatility are more sensitiveto shifts in investor sentiment (see Baker and Wurgler, 2004), as they exhibit a broader range ofvaluation, which make these stocks more vulnerable to noise trading (see DeLong et al., 1990;Shleifer and Vishny, 1997). However, we investigate institutional investors’ sentiment and focus onthe opposite causal direction. A theoretical model concerning the linkage between sentiment andmarket volatility does not exist, but there is a broad literature on institutional trading behavior andmarket stability. Institutional investors can be destabilizing because of positive-feedback trading(see Lakonishok et al., 1992), which means that they trade less if past returns were negative,and trade more if past returns were positive. Our study shows that past returns have a positiveimpact on current sentiment; therefore, negative returns trigger a reduction in investor sentiment.If institutional investors are pessimistic about the future development of the stock market, theymight trade less.21 What we observe is a positive-feedback trading or “trend chasing” behavior,which destabilizes exchanges as more liquidity is provided in good times and less in bad times(see Aiken, 1998). As a consequence, market volatility increases (see DeLong et al., 1990; Cutler

20 In contrast to Lee et al. (2002) who used an ARCH-in-mean model, we cannot confirm an impact of the sentimenton stock returns (or excess returns). However, we find an impact of sentiment on volatility, which they found for the USsentiment index.21 Note that we cannot observe the trading behavior of Chinese institutional investors directly.


et al., 1990). Nofsinger and Sias (1999) argued that lagged stock returns act as a common signalfor positive-feedback trading, which confirms our findings that past returns affect sentiment inthe sense of a positive-feedback process.

4.5. Bivariate GARCH models

So far, we have focused on univariate GARCH models, and our VAR model allowed interrela-tions between mean-equations but did not allow volatility-spillovers and correlation between theerror terms of the two-equation system. Accordingly, this section explores multivariate GARCH

Table 4Bivariate GARCH models with and without VAR specification in mean-equation

VAR and BEKK based onVGARCH(1,1)

BEKK based onVGARCH(1,1)

VAR and BEKK based onVGARCH(2,1)

Dependent variable: sentiment yt

Mean equation (VAR)Constant 1.6675 (0.000) 3.9719 (0.000) 1.7634 (0.000)yt−1 0.2098 (0.000) – 0.1988 (0.000)yt−2 0.1878 (0.000) – 0.1907 (0.000)yt−3 0.1822 (0.000) – 0.1672 (0.000)rt−1 11.8689 (0.000) – 11.6644 (0.000)rt−2 −0.7103 (0.607) – −1.0323 (0.435)rt−3 −0.3401 (0.798) – −0.3797 (0.772)

Adjusted R2 0.2690 −0.0060 0.2712

VarianceConstant 0.0231 (0.003) 0.0479 (0.019) 0.0857 (0.000)ε2t−1 0.1235 (0.000) 0.1234 (0.001) 0.0217 (0.551)

ε2t−2 – – 0.2229 (0.000)

ht−1 0.6787 (0.000) 0.5882 (0.000) –

Dependent variable: market returns rt

Mean equation (VAR)Constant 0.0159 (0.016) 0.000 (0.000) 0.0159 (0.016)yt−1 −0.0016 (0.158) – −0.0016 (0.158)yt−2 −0.0023 (0.062) – −0.0023 (0.062)yt−3 −0.0002 (0.840) – −0.0002 (0.840)rt−1 −0.0091 (0.846) – −0.0091 (0.846)rt−2 0.0035 (0.943) – 0.0035 (0.943)rt−3 −0.0027 (0.956) – −0.0027 (0.956)

Adjusted R2 −0.0092 −0.0051 −0.0092

VarianceConstant 0.0000 (0.000) 0.0000 (0.000) 0.0000 (0.000)ε2t−1 0.2585 (0.000) 0.2454 (0.000) 0.2585 (0.000)

ht−1 0.7215 (0.000) 0.7306 (0.000) 0.7215 (0.000)Observations 590 592 590

Granger causality tests (F-tests with three lags) for variance-spilloversSpillover from sentiment to returns 0.3820 (0.766) 0.3453 (0.793) 0.3824 (0.766)Spillover from returns to sentiment 1.6389 (0.179) 1.6159 (0.185) 1.6383 (0.182)


Fig. 3. Time-varying correlation between investor sentiment and stock returns.

techniques that allow volatility-spillovers and time-varying correlation. We follow the bivari-ate BEKK model by Engle and Kroner (1995) and construct our basic model in the followingmanner:22

VAR :

[rt

yt

]= � +

3∑j=1

�j

[rt−j

yt−j

]+ �t

Vector GARCH(1, 1) : vec(Ht) = ht =

⎡⎢⎣

hr,t

hry,t

hy,t

⎤⎥⎦ = C + A vec(�t−1�′

t−1) + Bht−1

BEKK parameterization : Ht = C′∗C∗ + A′

∗�t−1�′t−1A∗ + B′∗Ht−1B∗

(5)

Based on our univariate GARCH models, we know that the conditional volatility of stock returnscan be modeled by GARCH(1,1), whereas the conditional volatility of investor sentiment fol-lows an ARCH(2) process. Accordingly, we use the standard vector GARCH(1,1) and apply theBEKK parameterization—but we also explore an extension by allowing an ARCH(2) process forinvestor sentiment, GARCH(1,1) for stock returns, and GARCH(1,1) for the covariance term.23

For the sake of comparison, we also run a vector GARCH(1,1) specification without the VARstructure in the conditional mean-equation. Table 4 shows the results of our maximum-likelihoodestimation. To highlight the time-varying correlation coefficients between investor sentiment andstock return, Fig. 3 plots conditional correlation coefficients based on the BEKK parameterizationof a VGARCH(1,1) model.24 Fig. 3 illustrates further that the conditional correlation coefficientfluctuates tremendously over time, and that on average, the correlation coefficient reaches 0.02.

22 There are a huge variety of different bivariate conditional heteroscedasticity models. Bollerslev et al. (1988) extendedthe univariate GARCH model to the vector GARCH– but this model does not guarantee that the variance-covariancematrix is positively semidefinite (see Lien and Luo, 1993). Hence, we rely on the BEKK model by Engle and Kroner(1995), which overcomes the problem of negative variance terms and allows modifications, as it is fairly general.23 We can extend the vector GARCH to a VGARCH(2,1) and impose parameter restrictions.24 Note that the two alternative specifications lead to very similar conditional correlation coefficients.


The 5 percentile is −0.19, and the 95 percentile is 0.21; thus, we can state that the contem-poraneous correlation between the two series is of minor relevance. Nevertheless, Fig. 2, theVAR in Table 1 and the extended VAR in Table 4 point to the fact that lagged returns affectinvestor sentiment (in the mean-equation). Based on the three bivariate GARCH specificationsin Table 4, we carry out Granger causality tests to detect volatility-spillovers. Thus far, Table 3showed that lagged investor sentiment influences the conditional variance of stock returns. Now,we focus on the alleged impact of the lagged conditional variance of investor sentiment on theconditional variance of stock returns and vice versa. However, Granger causality tests reveal thatvolatility-spillovers cannot be confirmed.

5. Conclusions

In contrast to previous research, we analyze institutional investors’ sentiment and revealthat a long-run relation between share prices and sentiment does not exist—but short-termdynamics matter. Granger causality tests show that previous market returns influence investors’sentiment—but not vice versa. Henceforth, institutional investors’ sentiment does not predictfuture stock returns—but stock returns influence sentiment in the sense of a positive-feedbackprocess. Our findings confirm Brown and Cliff (2001) that discovered that the sentiment of smallinvestors is a poor indicator of future stock returns, at least in the short-run. Hence, one can statethat institutional investors do not possess any superior information or above-average analyticalskills to form their expectations about future market developments. An asymmetric response modelemphasizes that institutional investors react by far more sensitively after the market declines. Thisis in line with findings for small shareholders (see Sias, 1997).

Lee et al. (2002) found that the sentiment of small investors and hence noise-trading increasesmarket volatility—but also increases excess returns. This implies that the additional risk of noisetrading is priced. We confirm for institutional investors that their sentiment has a tremendousimpact on volatilities—but we cannot find any influence on stock returns. As investor senti-ment follows a positive-feedback process, we argue that Chinese institutional investors exhibita “trend chasing” behavior, which destabilizes exchanges and hence increases market volatility(see DeLong et al., 1990; Cutler et al., 1990).

As survey data on institutional investors’ sentiment is to date only available for the Chinesemarket, we must carefully assess which implications can be derived for mature financial marketsfrom our study. Institutional investors are relatively unimportant in China; however, their sentimenthas a considerable impact on volatility. Henceforth, one could argue that the discovered impactof institutional investors’ sentiment might be even stronger in mature markets. Nevertheless,Chinese institutional investors have a short history, as the first securities firms entered the marketin September 1992. Due to this short period of learning, Chinese institutional investors lackexperience, which might partly explain our finding that institutional investors’ sentiment hassimilar effects as the sentiment of small shareholders.

References

Aiken, B., 1998. Have institutional investors destabilized emerging markets? Contemporary Economic Policy 16, 173–184.Baker, M., Stein, J.C., 2004. Market liquidity as a sentiment indicator. Journal of Financial Markets 7, 271–299.Baker, M., Wurgler, J., 2004. Investor sentiment and the cross-section of stock returns. NBER Working Paper 10449.Barberis, N., Shleifer, A., Vishny, R., 1998. A model of investor sentiment. Journal of Financial Economics 49, 307–343.Blinder, A.S., 1991. Why are prices sticky? Preliminary results from an interview study. American Economic Review 81,

89–96.


Bohl, M.T., Brzeszczynski, J., 2006. Do institutional investors destabilize stock prices? Evidence from an emergingmarket. Journal of International Financial Markets, Institutions & Money 16, 370–383.

Bollerslev, T., 1986. Generalized autoregressive conditional heteroscedasticity. Journal of Econometrics 31, 307–327.Bollerslev, T., Engle, R.F., Wooldridge, J.M., 1988. A capital asset pricing model with time-varying covariances. Journal

of Political Economy 96, 116–131.Brown, G.W., Cliff, M.T., 2001. Investor sentiment and near term stock market. Working Paper, Kenan-Flagler Business

School, University of North Carolina at Chapel Hill, September 2001.Cai, F., Zheng, L., 2004. Institutional trading and stock returns. Finance Research Letters 1, 178–189.Chen, C.W.S., 1998. A Bayesian analysis of generalized threshold autoregressive models. Statistics and Probability Letters

40, 15–22.Cutler, D., Poterba, J.M., Summers, L.H., 1990. Speculative dynamics and the role of feedback traders. American Economic

Review Papers and Proceedings, 63–68.DeLong, J.B., Shleifer, A., Summers, L.H., Waldmann, R.J., 1990. Noise trader risk in financial markets. Journal of

Political Economy 4, 703–738.Elliott, G., Rothenberg, T.J., Stock, J.H., 1996. Efficient test for an autoregressive unit root. Econometrica 64, 813–836.Engle, R.F., Kroner, K., 1995. Multivariate simultaneous generalized ARCH. Econometric Theory 11, 122–150.Engle, R., 1982. Autoregressive conditional heteroskedasticity with estimates of the variance of UK inflation. Economet-

rica 50, 987–1008.Glosten, L., Jagannathan, R., Runkle, D., 1993. Relationship between the expected value and the volatility of the nominal

excess return on stocks. Journal of Finance 48, 1779–1801.Gomes, J., Kogan, L., Zhang, L., 2003. Equilibrium cross section of returns. Journal of Political Economy 111, 693–732.Jones, C., Lamont, O., 2002. Short sale constraints and stock returns. Journal of Financial Economics 66, 207–239.Kelly, M., 1997. Do noise traders influence stock prices? Journal of Money, Credit, and Banking 29, 351–363.Kenneth, L.F., Statman, M., 2000. Investor sentiment and stock returns. Financial Analysts Journal, 16–23.Kwiatkowski, D., Phillips, P.C.B., Schmidt, P., Shin, Y., 1992. Testing the null hypothesis of stationarity against the

alternative of a unit root. Journal of Econometrics 54, 159–178.Lakonishok, J., Shleifer, A., Vishny, R.W., 1992. The impact of institutional trading on stock prices. Journal of Financial

Economics 32, 23–43.Lee, W.Y., Jiang, C.X., Indro, D.C., 2002. Stock market volatility, excess returns, and the role of investor sentiment.

Journal of Banking and Finance 26, 2277–2299.Lee, C.M.C., Shleifer, A., Thaler, R.D., 1991. Investor sentiment and the closed-end fund puzzle. Journal of Finance 46,

75–109.Lien, D., Luo, X., 1993. Estimating multiperiod hedge ratios in cointegrated markets. Journal of Futures Markets 13,

909–920.Neal, R., Wheatley, S.M., 1998. Do measures of investor sentiment predict return? Journal of Financial and Quantitative

Analysis 33, 1775–1798.Nelson, D.B., 1991. Conditional heteroskedasticity in asset returns: a new approach. Econometrica 59, 347–370.Nofsinger, J.R., Sias, R.W., 1999. Herding and feedback trading by institutional investors. Journal of Finance 54,

2263–2295.Shleifer, A., Vishny, R.W., 1997. The limits of arbitrage. Journal of Finance 52, 35–55.Sias, R.W., 1997. The sensitivity of individual and institutional investors’ expectations to changing market conditions:

evidence from closed-end funds. Review of Quantitative Finance and Accounting 8, 245–269.Siegel, J.J., 1992. Equity risk premia, corporate profit forecasts, and investor sentiment around the stock crash of October

19/87. Journal of Business 65, 557–570.

Documents

Chinese institutional investors’ sentiment