Investor Herds in the Taiwanese Stock Market

Investor Herds in the Taiwanese Stock Market

Rıza Demirer †* , Chun-Da Chen** and Ali M. Kutan*** *

Southern Illinois University Edwardsville ** Tennessee State University *** Southern Illinois University Edwardsville, The William Davidson Institute, University of Michigan Business School, and The Emerging Markets Group, Cass Business School, London.

January 2008

†

Corresponding author. E-mail: [email protected]; Tel: 618−650−2939; Fax: 1−618−650-3047. Please do not quote without permission.

Investor Herds in the Taiwanese Stock Market Abstract This paper compares four different testing methodologies designed to test the existence of investor herds. We use firm level data on 689 firms traded in the Taiwan Stock Exchange, classified into 18 different sectors. We find that the cross sectional standard deviation (CSSD) based testing methodology, which imposes a linear relation between return dispersions and market return, fail to correctly test investor herds and yields no significant evidence of herding among Taiwanese investors. However, the non-linear model proposed by Chang et al. (2000) and the state space based models of Hwang and Salmon (2004) lead to consistent results indicating strong evidence of herd formation in all sectors with Electronics displaying the greatest impact. The fact that Electronics happen to be the most volatile sector among all and the most heavily invested one by foreign institutional investors is consistent with the literature suggesting that the trades of large institutional investors destabilize market prices and increase the volatility of the market due to positive feedback and herding among investors. We also find that the herding effect is more prominent during down movements of the market, indicating that it is the prospect of a loss which triggers herd behavior. JEL Classification Code: G14, G15 Keywords: Herd behavior, Equity return dispersion, Taiwan Stock Exchange.

2

1. Introduction A major function, among others, of security markets is the process of price discovery. Constant research and profit maximizing behavior of buyers and sellers in these markets make sure that every piece of information is reflected in the values of traded securities, hence affecting the values of corporations that issue them. Informational efficiency of a security market therefore is not only important from an academic point of view, as many theories are based on the concept of efficiency, but also crucial for investors and policy makers who rely on the information that these markets signal about the health of corporations as well as the overall economy. Clearly, the acquisition and transparency of information as well as how this information is processed by investors is a major concern for academics as financial theories that we have built depend on the assumption of investor rationality and rational processing of information. Formation of investor herds has been proposed as an alternative explanation of how investors process information and make investment choices. Herd behavior is simply defined as a strategy based on mimicking other investors’ actions (Bikhchandani and Sharma, 2000). Prior studies differ in their explanation to what might trigger such behavior. Devenow and Welch (1996) use the arguments of investor psychology where investors feel a sense of security in following the crowd. Another view suggests that the actions of more informed traders may reveal useful information which may not be accessible to individual investors (Chari and Kehoe, 1999, Calvo and Mendoza, 1998, and Avery and Zemsky, 1998). Investors perhaps cannot identify true or false information rapidly especially if they are in an unfamiliar market. In such a situation, investors might follow other investors’ trades blindly suppressing their own beliefs. Finally, a third approach focuses on the principal-agent relationship where fund managers might want to imitate others as a result of the incentives provided by the compensation scheme or in order to maintain their reputation (Scharfstein and Stein, 1990, Rajan, 1994, and Maug and Naik, 1996).

3

This study has two main contributions. First, we compare four different methodologies proposed in the literature that are designed to test the existence of investor herds. The first two models we estimate are based on return dispersions among individual firms; more specifically, cross sectional standard deviations (CSSD) and cross sectional absolute deviations (CSAD) across a particular sector. These methodologies have been applied to different markets in prior studies by Christie and Huang (1995), Chang, Chen, and Khorana (2000) and Gleason, Lee and Mathur (2003), Gleason, Mathur and Peterson (2004) and Demirer and Kutan (2006). The last two models are based on a state space model specification proposed by Hwang and Salmon (2004). These two sets of models differ in the sense that the first two focus on the cross-sectional variability of returns whereas the last two focus on the cross-sectional variability of factor sensitivities. Our study, therefore, is the first attempt to compare different testing methodologies on a large scale data. The second contribution of this study is to extend herding tests to the Taiwanese stock market and provide insight to this emerging market. We apply the four testing methodologies to firm level data using a large number of stocks traded in the Taiwan Stock Exchange. As shown in Table 1, domestic investors, mostly individual, account for almost 80% of the investment amount in this market, however there has been increasing interest by foreign investors over the past six years. Unlike investment trusts, foreign investors, and security dealers, most individual investors tend to have less professional knowledge and cannot access information accurately and easily. This information asymmetry may lead individuals to follow the actions of more informed institutional and foreign investors which may lead to herding formation in this market. For this purpose, it is especially interesting to examine whether herd formation exists in this emerging Taiwanese market with mostly domestic individual investors. Our findings indicate that the testing methodology based on CSSD as a measure of herding fail to identify existence of herds. This methodology yields no significant evidence of herding among Taiwanese investors. However, the non-linear model proposed by Chang et al. (2000) and the state 4

space based models of Hwang and Salmon (2004) lead to consistent results indicating strong evidence of herd formation. Therefore, regarding the comparison of testing methodologies, we find that CSSD based methodology, which assumes a linear relation between return dispersions and market return, fail to correctly test investor herds. Regarding test results using Taiwanese market data, we find strong evidence to herd formation in all sectors. Further analysis of test results across sectors indicates that herding has a bigger impact particularly on Electronics sector. Considering the fact that Electronics happen to be the most volatile sector among all and the most heavily invested one by foreign institutional investors, it is not surprising to find stronger evidence of investor herds among investors of securities in this sector. This finding is also consistent with the literature suggesting that the trades of large institutional investors destabilize market prices and increase the volatility of the market due to positive feedback and herding among investors. Individual investors in highly volatile markets are more likely to follow more informed investors’ trades such as foreign investors, and institutional investors. We also find that the herding effect is more prominent during down movements of the market, indicating that it is the prospect of a loss which triggers herd behavior. An outline of the remainder of the paper is as follows. Section 2 briefly summarizes the previous studies on tests of investor herds. Section 3 provides the details of different testing methodologies employed and data description. Section 4 presents empirical results and a comparison of the findings from the return dispersion based models and state space models. Finally, Section 5 concludes the paper and discusses further research. 2. Previous Studies on Investor Herds Different methodologies have been suggested in the literature to test the existence of investor herds. Testing methodologies based on return dispersions among a group of securities focus on 5

cross sectional standard (or absolute) deviations of returns. Prior studies include Christie and Huang (1995) on U.S. equities, Chang, Cheng and Khorana (2000) on international equities, Gleason, Lee and Mathur (2003) on commodity futures traded on European exchanges, Gleason, Mathur and Peterson (2004) on Exchange Traded Funds, and Demirer and Kutan (2006) on Chinese stocks. In general, these studies provide results in favor of the rational asset pricing theories and conclude that herding is not an important factor in determining security returns during periods of market stress. A different testing methodology based on cross sectional variability of factor sensitivities, instead of returns, is suggested by Hwang and Salmon (2004). Their analysis of daily stock returns in the South Korean market provides support for herd formation in this market. Finally, Weiner and Green (2004) employed both parametric and non-parametric methodologies and found little evidence of herding in heating oil and crude-oil futures. 3. Data and Methodology 3.1 Data The data set used in this study contains daily returns for 689 Taiwanese stocks traded on the Taiwan Stock Exchange over the January 1995−December 2006 period. Data are obtained from the Taiwan Stock Exchange Corporation (TSEC). Herding tests in the literature are based on the suggestion that a group is more likely to herd if it is sufficiently homogeneous, i.e. each member faces a similar decision problem, and each member can observe the trades of other members in the group (Bikhchandani and Sharma, 2000). Prior studies have, therefore, applied the tests on groups of stocks categorized on the basis of industry classification (e.g. Christie and Huang, 1995), exchange or country assignment (e.g. Gleason, Mathur and Peterson, 2004 and Chang, Chen, and Khorana, 2000), of course, after the impact of fundamentals has been factored out. Therefore, we assign each of the 689 firms to one of eighteen sector groups including Cement, Food, Plastics, 6

Textile, Electrical Appliances, Wire & Cable, Chemicals, Glass & Ceramics, Pulp & Paper, Steel, Rubber, Automobile, Electronics, Construction, Transportation, Tourism, Banking & Securities, and Retailing. We then calculate portfolio returns based on an equally weighted portfolio of all firms in each sector classification. 3.2 Methodology We employ four different testing methodologies; two return dispersion based models and two state space models. This section summarizes the methodology for each testing methodology. 3.2.1 Return Dispersion Models The first two testing methodologies employed are based on cross sectional standard deviations (CSSD) and cross-sectional absolute standard deviations (CSAD) among individual firm returns within a particular group of securities. Christie and Huang (1995) use CSSD as a measure of the average proximity of individual asset returns to the realized market average in order to test herd behavior. Chang et al. (2000) use CSAD in a non-linear regression specification in order to examine the relation between the level of equity return dispersions and the overall market return. The first methodology employed in this study is based on return dispersions as measured by CSSD. This methodology has been used by Christie and Huang (1995), Chang, Cheng, and Khorana (2000), Gleason, Lee and Mathur (2003), Gleason, Mathur and Peterson (2004), and Demirer and Kutan (2006) as explained in Section 2. Cross-sectional standard deviations (CSSD), used as a measure of return dispersion, is formulated as follows: N

CSSDt =

∑ (r

i, t

− rp , t )

i =1

N −1

2

(1)

where n is the number of firms in the aggregate market portfolio, ri ,t is the observed stock return on firm i for day t and rp ,t is the cross-sectional average of the n returns in the market

7

portfolio for day t. This measure can be regarded as a proxy to individual security return dispersion around the market average. The main idea in this methodology is based on the argument that the presence of herd behavior would lead security returns not to deviate far from the overall market return. The rationale behind this argument is the assumption that individuals suppress their own beliefs and make investment decisions based solely on the collective actions of the market. On the other hand, rational asset pricing models offer a conflicting prediction suggesting that dispersions will increase with the absolute value of market return since each asset differs in its sensitivity to the market return. This methodology also suggests that the presence of herd behavior is most likely to occur during periods of extreme market movements, as they would most likely tend to go with the market consensus during such periods. Hence, we examine the behavior of the dispersion measure in (1) during periods of market stress and estimate the following linear regression model:

CSSDt = α + β D DtL + β U DtU + ε t

(2)

where DtL = 1, if the return on the aggregate market portfolio on day t lies in the lower tail of the return distribution; 0 otherwise, and DtU = 1, if the return on the aggregate market portfolio on day

t lies in the upper tail of the return distribution; 0 otherwise. Although somewhat arbitrary, in the literature, an extreme market return is defined as one that lies in the one (and five) percent lower or upper tail of the return distribution. The dummies in equation (2) aim to capture differences in return dispersions during periods of extreme market movements. As herd formation indicates conformity with market consensus, the presence of negative and statistically significant β D (for down markets) and β U (for up markets) coefficients would indicate herd formation by market participants. The second methodology employed in this paper is suggested by Chang et al. (2000) and uses the 8

cross-sectional absolute deviation of returns (CSAD) as a measure of return dispersion. CSAD is expressed as

CSAD t =

1 N

N

∑| r

i, t

− rm , t |

(3)

i =1

Chang et al. (2000) challenge the CAPM assumption that return dispersions are an increasing function of the market return and that this relation is linear. They suggest that during periods of market stress, one would expect the relation between return dispersion and market return to be non-linearly increasing or even decreasing. Therefore, they propose a testing methodology based on a general quadratic relationship between CSADt and rm,t of the form: CSADt = α + γ 1 rm.t + γ 2 rm2.t + ε t

(4)

According to this methodology, herding would be evidenced by a lower or less than proportional increase in the cross-sectional absolute deviation (CSAD) during periods of extreme market movements. As a result, if herding is present, then the non-linear coefficient, γ2 will be negative and statistically significant; otherwise a statistically positive γ2 would indicate no evidence of herding. 3.2.2 State Space Models

The next two testing methodologies we employ in this study are suggested by Hwang and Salmon (2004). Rather than returns, Hwang and Salmon (2004) focus on the cross-sectional variability of factor sensitivities. Considering a one factor model with the factor being the market return, they formulate a herding measure based on the relative dispersion of the betas for all assets in the market. Next, we briefly explain these two methodologies. Consider the following CAPM in equilibrium, E r (rit ) = β imt Et (rmt )

(5)

where rit and rmt are the excess returns on asset i and the market at time t , respectively, 9

β imt is the systematic risk measure, and Et (⋅) is conditional expectation at time t . When herding behavior is present, investors disregard the equilibrium relationship of Equation (5) and trade in such a way that matches individual asset returns with that of the market. When that happens, the β and expected rate of return presents a bias in a way that would reflect this matching of individual asset returns with that of the market. So, considering CAPM again, when herding behavior is present, real β coefficient obeys the following relation which replaces equation (5): Etb (rit ) b = β imt = β imt − hmt ( β imt − 1) , Et (rmt )

(6)

b are the market’s biased short run conditional expectation on the excess where Etb (rit ) and β imt

returns of asset i and its beta at time t , and hmt is a latent herding parameter that changes over time, hmt ≤ 1 , and conditional on market fundamentals. In general, when (0 < hmt < 1) , one could argue that some degree of herding exists in the market determined by the magnitude of

hmt . Since the form of herding we discuss represents market-wide behavior and Equation (6) is assumed to hold for all assets in the market, we should calculate the level of herding using all assets in the market rather than a single asset, thereby removing the effects of idiosyncratic b b movements in any individual β imt . The standard deviation of β imt can be formulated as

b Std c ( β imt ) = E c (( β imt − hmt ( β imt − 1) − 1) 2 )

= E c (( β imt − 1) 2 ) (1 − hmt )

= Std c ( β imt )(1 − hmt ) ,

(7)

where EC (⋅) represents the cross-sectional expectation. Taking the logarithm of Equation (7) b and assuming that Std c ( β imt ) can be time-varying, we rewrite Equation (7)as

10

b log[Std c ( β imt )] = μ m + υ mt

where μ m = E[log[Std c ( β imt )]] and υ mt ~ iid (0, σ m2 υ ) . State Space Model 1 then can be

estimated as b log[Std c ( β imt )] = μ m + H mt + υ mt ,

(8)

H mt = φ m H mt −1 + η mt ,

(9)

where H mt = log(1 − hmt ) and η mt ~ iid (0, σ m2 η ) . Equations (8) and (9) are the standard state-space model with Kalman filter estimation. In this methodology, we only focus on the dynamic pattern of movements in the latent state variable, H mt , the state equation. When

σ m2 η = 0 , there is no herding, i.e., H mt = 0 for all t . We allow herding, H mt , to evolve over time and follow a dynamic process. A significant value of σ m2 η can be interpreted as the existence of herding and a significant φ supports this particular autoregressive structure. An alternative model can be formulated when we add to Equation (8) market volatility, log σ mt , and return , rmt , as independent variables. This leads to State Space Model 2 formulated as b log[Std c ( β imt )] = μ m + H mt + c m1 log σ mt + c m 2 rmt + υ mt .

(10)

4. Empirical results 4.1 Descriptive statistics

Table 2 provides summary statistics for average daily log returns, return dispersions, and the average number of firms used to compute these statistics for each sector. Since the number of stocks in a sector does not stay constant over time, we report the average number of firms over the sample period in the second column of Table 2. Examining Table 2, we observe that the average daily returns for all sectors are positive while electronics and automobile sectors have the highest average daily returns. In terms of volatility, 11

construction and electronics sectors are ranked as the most volatile sectors. Over the past few years, the electronics sector has shown rapid growth and stocks in this sector have been among the most actively traded stocks in Taiwan. According to the annual report of Taiwan Stock Exchange Corporation, the trading volume and value of securities in this sector account for 60% and 70% of the total volume and value of the Taiwan stock market respectively. Heavy trading volume and investor interest in this sector may be a reason for the high return volatility that we observe in the electronics sector. Considering the high volatility, one might expect herding behavior to be more likely in this sector relative to other sectors. Due to heavy trading volume and high uncertainty in the behavior of stocks, investors in this sector may find it rational to follow institutional investors’ trading strategies as well as each other. Furthermore, foreign shareholding percentage of electronic sector is about 35% which may further affect investor behavior in terms of their trades, making it more likely to observe herd behavior. The foreign shareholding percentage of stocks in the banking and securities sector is about 25%, ranked second place in the stock market; however, we observe that the average and the standard deviation of daily stock returns in this sector are not as high as other sectors. Moreover, during the sample period, the government of Taiwan has actively executed the so-called “Bank Merger Policy”. Having this in mind, one might argue that investors would not be inclined to follow others’ strategies promptly during the peak of a bank merger. Due to these two factors, one may expect herd behavior in the banking and securities sector to be less likely. Compared to electronics (Hi-Tech) sector, the remaining more traditional sectors tend to have a lower foreign shareholding percentage. However, despite being a traditional sector, we observe a higher average return and standard deviation for construction sector. One reason for this observation may be due to the recession in the housing market during the sample period. In order to stimulate growth in the housing market, the government of Taiwan has executed different policies frequently, governmental intervention policy might have led to the higher returns and 12

corresponding higher volatility in this sector. Panel B in Table 2 reports summary statistics for daily cross sectional standard deviations within each sector. Consistent with the findings from Panel A, we observe the highest cross sectional volatility in Electronics followed by Construction. 4.2 Results of return dispersion models

Table 3 presents estimation results for the CSSD based model in Equation 2. Given the significant variation in dispersions and strong correlation, all estimations are done using the Newey-West heteroskedasticity and autocorrelation consistent standard errors 1 . We use the Taiwan Stock Exchange Composite Index to represent the market and use the upper and lower one and five percentiles of the market return to represent periods of market stress. For a majority of the sectors analyzed, we do not find any evidence in favor of herd formation during periods of large market swings. The regressions yield statistically significant and positive β estimates indicating that equity return dispersions increase during periods of large price changes as predicted by CAPM. The only exception to this is Electronics where we observe significantly lower return dispersions when market is in the upper or lower one percentile, indicating herd formation during extreme moves of the market index. Later in this section, we will provide explanations as to why it is more likely to observe herd behavior in Electronics. However, for the moment, keep in mind that Electronics happen to be the most volatile sector among all and the most heavily invested one by foreign investors. Table 4 presents estimations results for the non-linear CSAD based model in Equation 4. Following Chang et al. (2000), we run three separate regressions for each sector: one using the whole sample, and two restricting the data to up (or down) movements of the market index. Running separate models in this manner allows us to examine whether there is any asymmetric effect of herd behavior. Our findings with the non-linear model lead to completely different results 1

We also estimated the models using GARCH models; the results were qualitatively the same. 13

than the first methodology. Examining Table 4, we find evidence to herd formation in all sectors except for Tourism and Automobile. The regressions yield statistically significant and negative γ2 estimates indicating a non-linear and decreasing relation between equity return dispersions and the market return. However, when we examine regression results run with data restricted to up and down markets separately, we observe that herding effect is mostly prominent during market losses. It is more likely to observe herd formation during periods of market losses. Furthermore, when we compare the absolute values of γ2 estimates, we see that return dispersions during downside movements of the market are much lower than those for upside movements. Therefore, one can conclude that it is the prospect of losing which pushes investors to follow others’ trades and display herd behavior. However, such behavior is not the case during up markets indicating investor confidence when investors are confident in their decisions. In the next subsection, we summarize our findings from the state space models estimated. 4.3 Results of state space models

Table 5 presents estimation results for State Space Model 1. The results indicate strong evidence of herding through H mt . We observe that H mt is highly persistent with large and significant values of φˆm . More importantly, the estimates for σ mη are highly significant providing support for herd behavior. In particular, Electronics, Textile, and Banking & Securities display higher H mt estimates implying that investors may be following institutional investors’ trading behavior.

The findings make more sense considering the fact that Electronic and Banking & Securities sectors are two of the most volatile sectors and the two favorite trading targets for foreign institutional investors. In a separate but related study Gabaix et al. (2005) present a model in which the trades of large institutional investors destabilize market prices and increase the volatility of the market due to positive feedback and herding among investors. Unlike investment trusts, foreign investors, and

14

security dealers (informed traders), most individuals have less professional knowledge and cannot access information accurately as easily. This information asymmetry would further lead individuals to perform momentum strategies which might lead to herding effects. Similarly, Luo (2003) find that mutual fund investors create excess volatility as fund flows lead to higher subsequent market volatility. In another related study, Dennis and Strickland (2002) find that stocks that move the most during large moves of the market index happen to be those that have relatively larger institutional holdings. They suggest that this finding is due to herd formation among mutual funds and pension plan sponsors. Other sectors we analyze also have significant and highly persistent H mt estimates and exhibit herding behavior; however, H mt values are lower than those for Electronics and Banking & Securities sectors. Table 6 presents estimation results for State Space Model 2. Our findings are similar to those we observe with the first model results. Once again, taking into account the level of market volatility and return this time, we find strong evidence of herding through H mt as the standard deviation of η mt is significantly different from zero and H mt is highly persistent with the φˆm estimate being significant. Please note the difference between the two state space models. The second model contains two market variables in the measurement equation allowing us to analyze the degree of herding given the state of the market. Furthermore, we focus on the coefficient of market volatility and market returns. The findings for Plastics, Electrical App., Chemicals, Pulp, b Steel, Automobile, Tourism, and Retailing sectors indicate that Std c ( β imt ) values decrease with

market volatility, since log σ mt values have significant and negative coefficients. These results are consistent with previous studies which suggest that herding is more likely to occur during periods of market stress.

15

b However, the results show that Std c ( β imt ) values increase as market volatility rises for Cement,

Food, Textile, Wire & Cable, Glass, Electronics, and Banking & Securities since log σ mt have significant and positive coefficients. It is interesting to note that these sectors with positive log σ mt coefficient estimates have higher daily trading activities and higher foreign shareholding percentage. Thus, we believe that investors in these sectors usually exhibit herding behavior regardless of market situation. Furthermore, comparing the estimates for φˆm across sectors we notice that the effect of herd behavior in Electronics is greater compared to financial and traditional sectors. 5. Conclusions

Herd formation has been proposed as an alternative explanation of how investors make investment choices. The existence of investor herds is a major concern for academics as financial theories that we have built depend on the assumption of investor rationality as well as rational processing of information. Such behavior also presents a concern to policymakers as such behavior might aggravate volatility of returns and hence destabilize financial markets, especially in crisis conditions. In this paper, we compare four different methodologies to test herd formation in Taiwan Stock Exchange. We use firm level data on 689 firms and classify them into 18 different sectors. We then estimate four different models to test the existence of investor herds in these markets. The first two models we estimate are based on return dispersions among individual firms; more specifically, cross sectional standard deviations (CSSD) and cross sectional absolute deviations (CSAD) across a particular sector. The last two models are based on a state space model specification. These two sets of models differ in the sense that the first two focus on the cross-sectional variability of returns whereas the last two focus on the cross-sectional variability of factor sensitivities. Our study is the first attempt to compare different testing methodologies that 16

are designed to test the existence of investor herds. We apply these tests to firm level data using a large number of stocks traded in the Taiwan Stock Exchange. Our findings indicate that Christie and Huang (1995) methodology based on CSSD as a measure of herding fail to identify existence of herds. This methodology yields no significant evidence of herding among Taiwanese investors. However, the non-linear model proposed by Chang et al. (2000) and the state space based models of Hwang and Salmon (2004) lead to consistent results indicating strong evidence of herd formation. Therefore, regarding the comparison of testing methodologies, we find that CSSD based methodology, which assumes a linear relation between return dispersions and market return, fail to correctly test investor herds. One of the implications of the results is that future studies should capture such non-linear relationships in the data. Regarding test results using Taiwanese market data, we find strong evidence to herd formation in all sectors. Comparison of test results across sectors, we see that herding has a bigger impact particularly on Electronics sector. Considering the fact that Electronics happen to be the most volatile sector among all and the most heavily invested one by foreign institutional investors, it is not surprising to find stronger evidence of investor herds among investors of securities in this sector. This finding is also consistent with the literature suggesting that the trades of large institutional investors destabilize market prices and increase the volatility of the market due to positive feedback and herding among investors. Individual investors in highly volatile markets are more likely to follow more informed investors’ trades such as foreign investors, and institutional investors. We also find that the herding effect is more prominent during down movements of the market, indicating that it is the prospect of a loss which triggers herd behavior. Our study has several limitations. We rely on only a stock market to provide inferences about the robustness of different testing methodologies on herding behavior. In addition, we use data on an emerging market. The results could be sensitive to using data from advanced markets. Hence, our 17

results are preliminary but could be used as a yardstick for future studies. Further evidence from several –both emerging and advanced – markets is necessary to generalize the results of the paper.

18

References

Avery, C., Zemsky, P., 1998. Multidimensional Uncertainty and Herd Behavior in Financial Markets. American Economic Review 88, 724-748. Bikhchandani, S., Sharma, S., 2000. Herd Behavior in Financial Markets: A Review. IMF Working Paper WP/00/48 (Washington: International Monetary Fund). Calvo, G., Mendoza, E., 1998. Rational Herd Behavior and Globalization of Securities Markets. Mimeo, University of Maryland. Chang, E. C., Cheng, J. W., Khorana, A., 2000. An Examination of Herd Behavior in Equity Markets: An International Perspective. Journal of Banking and Finance 24, No. 10, 1651-1699. Chari, V. V., Kehoe, P., 1999. Financial Crises as Herds. Mimeo, Federal Reserve Bank of Minneapolis. Christie, W.G., Huang, R. D., 1995. Following the Pied Piper: Do individual Returns Herd around the Market? Financial Analyst Journal, July-August 1995, 31-37. Demirer, R. and Kutan, A. M., 2006, Does herding behavior exist in Chinese stock market? Journal of International Financial Markets, Institutions and Money 16, 123-142. Dennis, P. and D. Strickland, 2002. Who blinks in volatile markets, individuals or institutions?, Journal of Finance 51, 111-135. Devenow, A., Welch, I., 1996. Rational Herding in Financial Economics, European Economic Review 40, 603-615. Gabaix, X., P. Gopikerishman, V. Plerou and H. E. Stanley, 2005, Institutional investors and stock market volatility, Quarterly Journal of Economics, forthcoming. Gleason, K. C., Lee, C. I., Mathur, I., 2003. Herding Behavior in European Futures Markets. Finance Letters 1, 5-8 Gleason, K. C., Mathur, I., Peterson, M. A., 2004. Analysis of Intraday Herding Behavior among the Sector ETFs. Journal of Empirical Finance 11, 681–694 Hirshleifer, D., Teoh, S. H., 2001. Herd behavior and Cascading in Capital Markets: A Review and Synthesis. Working Paper. Hwang, S., and Salmon, M., 2004, Market stress and herding, Journal of Empirical Finance 11, 585-616 19

Luo, D., 2003. Market volatility and mutual funds cash flows. Working paper, Yale International Center for Finance. Maug, E., Naik, N., 1996. Herding and Delegated Portfolio Management. Mimeo, London Business School. Rajan, R. G., 1994. Why Credit Policies Fluctuate: A Theory and Some Evidence. Quarterly Journal of Economics 436, 399-442. Scharfstein, D., Stein, J., 1990. Herd Behavior and Investment. American Economic Review 80, 465-479. Shiller, R. J., 1981. Do Stock Prices Move too Much to be Justified by Subsequent Changes in Dividends? American Economic Review, 71, 421-436. Summers, L. H., 1986. Does the Stock Market Rationally Reflect Fundamental Values? Journal of Finance 41, 591-601. Weiner, R. J., and Green, M. A., 2004. Do birds of a feather flock together? Speculator herding in dervatives markets. Working Paper George Washington University

20

Table 1. Year 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

Securities Trading Value Percentage by Investors Type (%) Domestic Individual 89.25

Domestic Juridical Person 8.62

Foreign Individual 0.01

Foreign Juridical Person 2.12

90.73 89.73 88.23 86.10 84.41 82.30 77.84 75.94 68.84 70.56

7.55 8.63 9.36 10.27 9.69 10.05 11.51 11.56 13.29 11.04

0.01 0.02 0.01 0.01 0.01 0.97 1.24 1.63 2.41 2.25

1.71 1.62 2.40 3.62 5.89 6.68 9.41 10.87 15.46 16.15

21

Table 2. Summary Statistics: Average Daily Returns and Cross-Sectional Standard Deviations. Industry

# Firms

# Obs.

Mean

Std. Dev.

Min.

Max.

Panel A: Average Daily Returns Cement Food Plastics Textiles Electrical App Wire & Cable Chemicals Glass and Ceramics Pulp & Paper Steel Rubber Automobile Electronics Construction Transportation Tourism Banking & Securities Retailing Others

8 18 20 47 36 14 35 7 7 24 9 5 307 34 18 6 45 10 39

3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154 3154

0.016% 0.026 0.028 0.014 0.040 0.021 0.035 0.003 0.017 0.043 0.044 0.056 0.056 0.051 0.049 0.039 0.019 0.040 0.049

2.457% 2.404 2.735 2.773 2.506 2.670 2.476 3.019 2.707 2.827 2.723 2.212 3.106 3.183 2.641 2.486 2.426 2.433 2.498

-7.000% -7.440 -7.000 -8.090 -7.600 -7.000 -7.000 -19.680 -7.000 -7.000 -7.000 -6.990 -38.240 -7.560 -7.550 -7.000 -7.770 -7.000 -7.200

7.000% 7.000 7.000 14.070 10.880 14.250 9.980 7.000 7.000 14.980 16.720 7.000 277.000 13.680 7.000 7.000 11.470 7.000 11.650

Panel B: Cross-Sectional Standard Deviations Cement Foods Plastics Textile Electrical App Wire and Cable Chemicals Glass and Ceramics Pulp and Paper Steel Rubber Automobile Electronics Construction Transportation Tourism Banking & Securities Retailing Others

1.558% 1.868 1.895 2.119 2.015 1.956 1.923 2.299 1.690 1.917 1.914 1.405 2.433 2.324 1.868 1.655 1.644 1.845 2.073

0.811% 0.705 0.725 0.608 0.648 0.790 0.638 1.084 0.815 0.736 0.843 0.888 0.920 0.730 0.731 0.843 0.639 0.778 0.590

0.106% 0.384 0.104 0.278 0.230 0.102 0.254 0.100 0.068 0.108 0.124 0.000 0.151 0.492 0.100 0.036 0.176 0.108 0.447

5.004% 5.193 4.883 4.462 4.479 4.989 4.619 8.545 5.042 5.358 5.698 6.398 22.296 5.089 4.745 5.795 4.613 5.108 4.938

22

Table 3. Regression Coefficients for CSSDt = α + β D DtL + β U DtU + ε t (t-ratios in parentheses). Return Dispersions Industry

Market return in the extreme upper/lower 1% of the return distribution βU βD α

0.245% 0.294%** (1.532) (2.053) *** Food 1.857 0.699 0.442*** (4.812) (4.004) Plastics 1.893 0.168 0.026 (1.019) (0.314) Textile 2.116 0.146 0.149* (1.014) (1.840) * Electrical App. 2.009 0.243 0.304*** (1.646) (3.449) Wire and Cable 1.955 0.089 0.016 (0.512) (0.117) Chemicals 1.919 0.379*** 0.043 (2.649) (0.629) Glass and Ceramics 2.296 0.089 0.186 (0.334) (0.899) Pulp and Paper 1.691 -0.0002 -0.073 (-0.001) (-0.523) Steel 1.914 0.086 0.165 (0.724) (1.488) Rubber 1.916 0.061 -0.219* (0.277) (-1.661) * Automobile 1.395 0.392 0.569*** (1.947) (3.423) *** Electronics 2.330 -0.389 -0.485*** (-3.111) (-5.340) Construction 2.319 0.217 0.274** (1.424) (2.388) ** Transportation 1.862 0.426 0.197* (2.541) (1.945) *** Tourism 1.641 0.838 0.542*** (3.574) (4.162) ** Banking & Securities 1.639 0.320 0.135 (2.559) (1.090) ** Retailing 1.841 0.350 0.100 (2.497) (1.177) Others 2.067 0.348*** 0.210*** (3.019) (2.418) (***, **, and * denote statistical significance at 1%, 5%, and 10%, respectively) Cement

1.553%

Market return in the extreme upper/lower 5% of the return distribution βD βU α 1.527% 1.824 1.857 2.090 1.980 1.928 1.890 2.262 1.670 1.881 1.886 1.367 2.319 2.299 1.828 1.609 1.610 1.811 2.041

0.353%*** (5.057) 0.541*** (8.979) 0.498*** (7.580) 0.339*** (6.507) 0.402*** (6.727) 0.322*** (4.530) 0.476*** (8.739) 0.331*** (3.394) 0.307*** (4.245) 0.404*** (6.770) 0.456*** (5.214) 0.479*** (5.777) 0.069 (1.245) 0.257*** (4.441) 0.514*** (8.813) 0.564*** (6.480) 0.359*** (6.306) 0.448*** (7.214) 0.418*** (9.346)

0.269%*** (4.563) 0.351*** (7.005) 0.260*** (5.704) 0.242*** (6.433) 0.285*** (6.954) 0.238*** (4.001) 0.193*** (4.764) 0.403*** (4.313) 0.099 (1.612) 0.313*** (5.814) 0.113* (1.902) 0.285*** (3.920) -0.012 (-0.260) 0.257*** (5.262) 0.283*** (5.951) 0.344*** (5.384) 0.326*** (7.047) 0.249*** (5.435) 0.221*** (5.410)

23

Table 4. Regression Coefficients for CSADt = α + γ 1 rm.t + γ 2 rm2.t + ε t (t-ratios in parentheses). Absolute Deviation Industry

Whole Sample

α

γ1

Down Market (Rm0)

α

γ1

γ2

1.036

0.095** (2.373) 0.117*** (3.240) 0.221*** (6.373) 0.169*** (5.593) 0.144*** (4.843) 0.122*** (2.993) 0.161*** (5.364) 0.145*** (2.712) 0.146*** (3.954) 0.149*** (3.855) 0.194*** (4.731) 0.002 (0.051) 0.258*** (7.737) 0.145*** (3.847) 0.150*** (4.244) 0.094** (2.339) 0.175*** (4.958) 0.120*** (3.519) 0.147*** (4.957)

-0.003 (-0.263) 0.0004 (0.044) -0.029*** (-3.728) -0.018*** (-2.606) -0.009 (-1.298) -0.013 (-1.195) -0.018*** (-2.604) -0.009 (-0.692) -0.027*** (-3.156) -0.011 (-1.148) -0.033*** (-3.147) 0.020** (1.984) -0.057*** (-7.063) -0.014 (-1.525) -0.012 (-1.439) 0.008 (0.805) -0.017* (-1.774) -0.012* (-1.656) -0.012* (-1.664)

1.221 1.236 1.421 1.345 1.321 1.260 1.544 1.143 1.255 1.266 0.978 1.617 1.610 1.227 1.097 1.050 1.263 1.365

Table 5. State Space Model 1 (t-ratios in parentheses) b log[Std c ( β imt )] = μ m + H mt + υ mt and H mt = φ m H mt −1 + η mt Industry Cement

μ

φm ***

σ mυ ***

***

-0.685 0.940 0.085 (-20.667) (125.420) (20.057) Food -0.635*** 0.930*** 0.019*** (-26.858) (133.023) ( 3.173) *** *** Plastics -0.580 0.945 0.048*** (-19.496) (153.061) (11.752) *** *** Textile -0.647 0.958 0.040*** (-20.535) (144.139) (10.893) *** *** Electrical App. -0.698 0.913 0.053*** (-28.868) (116.803) (11.126) Wire and Cable -1.456*** 0.927*** 0.015*** (-25.168) (708.846) ( 2.655) Chemicals -0.657*** 0.954*** 0.025*** (-20.784) (143.880) ( 3.276) *** *** Glass and Ceramics -0.577 0.919 0.082*** (-14.992) (100.782) (11.846) *** *** Pulp and Paper -0.809 0.892 0.070*** (-22.822) ( 78.357) ( 7.603) *** *** Steel -0.707 0.949 0.047*** (-21.097) (149.692) ( 6.952) Rubber -0.828*** 0.905*** 0.100*** (-18.120) ( 77.216) (11.559) Automobile -0.457*** 0.920*** 0.022*** (-21.511) (106.424) ( 4.300) *** *** Electronics -1.282 0.997 0.038*** (-21.152) (624.326) (18.075) *** *** Construction -0.622 0.942 0.014 (-29.220) (131.795) ( 1.536) Transportation -0.690*** 0.942*** 0.080*** (-21.305) (156.531) (12.039) Tourism -0.555*** 0.932*** 0.033** (-11.245) ( 76.183) ( 2.383) Banking & Securities -0.757*** 0.956*** 0.039*** (-22.293) (171.375) (11.205) *** *** Retailing -0.606 0.933 0.039*** (-21.166) (121.470) ( 4.920) *** *** Others -1.374 0.916 0.085*** (-10.797) (460.285) (21.045) (***, **, and * denote statistical significance at 1%, 5%, and 10%, respectively)

σ mη 0.099*** (20.354) 0.091*** (35.052) 0.091*** (27.420) 0.068*** (18.509) 0.113*** (28.401) 0.064*** (21.435) 0.081*** (17.644) 0.142*** (21.611) 0.165*** (23.430) 0.103*** (20.184) 0.165*** (20.181) 0.082*** (29.282) 0.054*** (23.420) 0.070*** (17.702) 0.105*** (16.675) 0.113*** (14.307) 0.085*** (28.180) 0.101*** (17.390) 0.091*** (19.421)

25

Table 6. State Space Model 2 (t-ratios in parentheses) b log[Std c ( β imt )] = μ m + H mt + c m1 log σ mt + c m 2 rmt + υ mt and H mt = φ m H mt −1 + η mt Industry Cement

μ

φm ***

σ mη

σ mυ ***

***

0.451 0.941 0.085 (13.803) (131.070) (18.904) Food 0.589*** 0.931*** 0.019*** (24.607) (138.585) ( 3.376) *** *** Plastics -0.848 0.944 0.048*** (-32.457) (137.411) (9.191) *** *** Textile 0.897 0.957 0.040*** (32.465) (172.134) (15.143) *** *** Electrical App. -0.806 0.913 0.053*** (-39.118) (113.037) (9.079) Wire and Cable 1.437*** 0.917*** 0.015*** (22.304) (677.640) (4.315) Chemicals -1.628*** 0.950*** 0.024*** (-56.555) (157.520) (5.994) *** *** Glass and Ceramics 1.330 0.917 0.082*** (36.300) (99.504) (10.205) *** *** Pulp and Paper -2.479 0.892 0.070*** (-70.245) (78.008) (7.548) *** *** Steel -1.470 0.945 0.047*** (-42.510) (150.194) (10.907) Rubber -0.471*** 0.907*** 0.100*** (-10.372) (79.038) (10.540) Automobile -0.541*** 0.919*** 0.021*** (-25.690) (107.362) (4.277) *** Electronics 0.481 0.997 0.038*** (0.527) (604.024) (13.811) *** *** Construction 0.998 0.937 0.015*** (50.873) (150.635) (2.584) *** *** Transportation -0.179 0.941 0.080*** (-5.391) (142.554) (17.725) Tourism -6.787*** 0.926*** 0.032*** (-226.046) (100.682) (4.499) Banking & Securities -0.293*** 0.957*** 0.040*** (-8.487) (169.286) (11.260) *** *** Retailing -2.139 0.926 0.038*** (-90.847) (128.728) (8.629) *** *** Others -0.547 0.954 0.084*** (-14.520) (159.310) (20.375) (***, **, and * denote statistical significance at 1%, 5%, and 10%, respectively)

log σ m ***

0.098 (19.428) 0.091*** (34.885) 0.091*** (18.795) 0.068*** (26.050) 0.113*** (22.353) 0.063*** (33.953) 0.082*** (32.524) 0.142*** (21.199) 0.164*** (23.473) 0.104*** (28.278) 0.165*** (18.077) 0.082*** (29.303) 0.053*** (17.754) 0.069*** (35.092) 0.105*** (21.771) 0.111*** (23.539) 0.085*** (28.132) 0.102*** (31.852) 0.095*** (20.390)

***

0.659 (34.854) 0.682*** (51.060) -0.151*** (-10.268) 0.876*** (56.141) -0.058*** (-4.683) 1.658*** (45.111) -0.545*** (-34.230) 1.092*** (52.221) -0.970*** (-47.373) -0.429*** (-22.200) 0.211*** (7.917) -0.047*** (-4.000) 1.022* (1.933) 0.947*** (82.499) 0.293*** (15.301) -3.535*** (-208.998) 0.264*** (13.559) -0.833*** (-65.823) 0.013 (0.631)

rm -0.005 (-0.040) 0.005 (0.066) 0.046 (0.510) 0.027 (0.374) -0.069 (-0.520) 0.037 (0.820) 0.026 (0.336) -0.068 (-0.415) -0.092 (-0.503) 0.029 (0.266) 0.006 (0.026) 0.017 (0.205) 0.007 (0.118) -0.03 (-0.587) -0.028 (-0.222) -0.228* (-1.823) -0.058 (-0.687) 0.009 (0.074) 0.137 (0.870)

26