Investors' Herd Behavior: Rational or Irrational?

Authors


  • Acknowledgements: The authors thank the editors and the two anonymous referees for their constructive comments. The authors also acknowledge the financial support from National Science Council of Taiwan (Grant no. NSC 99-2410-H-032-030-), as well as the Taiwan Stock Exchange for data support.

Abstract

This study examines the relationships between the herding of various investor groups and trading noise in the Taiwan stock market to determine whether any of the investor groups tend to herd rationally. The study uses a unique and comprehensive data set on intraday transactions and limit order books of the Taiwan Stock Exchange (TWSE). We calculate the high-frequency herding measures and trading noise in a call auction market. We find that institutional investors are likely to be informed traders and herd rationally based on superior information. Institutional investors' herding has a negative impact on trading noise. Their buy (sell) herding predicts positive (negative) future market returns. By contrast, the herding of individual investors tends to contain limited information, as it increases trading noise; the buy (sell) herding of individuals is negatively correlated with future market returns. These findings are more significant for stocks with higher turnover.

1. Introduction

Herding is a common phenomenon in behavioral finance. Nofsinger and Sias (1999) define herding as a group of investors trading in the same direction over a period of time. Recently, investor herd behavior has attracted increasingly more attention, both within and outside of academia.1

Previous empirical work has mainly focused on apparent herding or its impact. However, it might be equally important to focus on the causes driving investors to cluster their trades, because diverse causes of herding may result in distinct effects on financial markets. This paper examines whether investor herd behavior is rational or irrational, and distinguishes information-based from non-information-based herding.

Welch (1996) reports that proponents of the view that herd behavior is irrational believe that investors follow one another blindly and forgo rational analysis. Devenow and Welch (1996) classify herding into rational herding and irrational herding. Rational herding is information-based; rational investors with similar stock preferences adopt the same response to similar information about company characteristics and fundamentals. When the herding of investors is rational in response to new information, herding moves prices toward the fundamental value of assets; price movement is not likely to reverse. By contrast, irrational herding occurs when investors with insufficient information and inadequate risk evaluation disregard their prior beliefs and blindly follow other investors' actions. Non-information-based herding might lead to market inefficiencies, drive asset prices away from fundamental values, and cause asset mispricing (Froot et al., 1992; Hirshleifer et al., 1994; Hwang and Salmon, 2004; Hung et al., 2010).

It is not easy to precisely distinguish rational herding from irrational herding. Prior research suggests that trading noise in financial markets mainly stems from investor irrationality and information asymmetries (DeLong et al., 1990; Admati, 1991). Stoll (2000) suggests that trading noise is friction that accompanies investor trading. Friction can be higher among investors who are not equipped with sufficient information. Hu (2006) extends the notion of friction to propose a valid measure of trading noise in a call auction market. The current paper argues that trading noise in the market increases if many investors trade without information, in the case of herding caused by many uninformed investors blindly following one another. This paper adopts the measure of trading noise used in Hu (2006) to examine the relationship between herding and trading noise on the Taiwan Stock Exchange (TWSE). We determine the rationality of herding in a particular investor group by assessing the level of trading noise in the market after herding by the investor group.

There is growing concern in the literature about the herding behavior in emerging stock markets, especially the Taiwan stock market, which is a market dominated by individual investors. Unlike the U.S. stock market, in which institutional investors hold a large proportion of shares, individual investors held 69% of total security trading value on the TWSE as of the end of 2010. However, there is relatively little research that explores the herding behavior of individuals in the TWSE.2

To the best of our knowledge, there is only one study that investigates herding behavior of individuals in the Taiwanese stock market: Chang et al. (2012) use individual herding indicators and institutional herding indicators to test for the presence of strategy returns. To fill this gap, this paper investigates not only individual herding but also institutional herding. In addition, we further divide institutions into foreign institutions and domestic institutions (proprietary dealers and investment trusts).

The current paper distinguishes from previous studies in the following three aspects. First, this paper adopts an intraday measure of herding intensity, using the comprehensive data set of intraday orders and trades. Gleason et al. (2004) point out that the use of low-frequency data may underestimate investor herding. Using the data set of high-frequency order book, we divide each trading day into nine intervals of 30 minutes and examine the herding of four investor groups within each 30-minute interval over the sample period. The order books allow direct observation of the direction and quantity of orders submitted by investors, enabling us to construct sequential order flows of investor groups and to better understand the trading behavior of each investor group. The herding measure in Lakonishok et al. (1992; hereafter referred to as LSV) is computed based on the proportion of buying or selling volume. Although recent work has commonly used the LSV measure to gauge investor herding, it cannot be applied to high-frequency data to capture short-term dynamic order flows (Christoffersen and Tang, 2009). Hence, we adopt the herding measure of bootstrap order runs developed by Patterson and Sharma (2005) to determine the intensity of buying and selling orders and to capture dynamic herd behavior based on sequences of intraday order flows.

Second, we investigate the relationships between the herding of four investor groups and the subsequent trading noise. With a unique and remarkably comprehensive data set of orders and trades, we categorize filled orders according to four investor groups: individuals, foreign institutions, proprietary dealers, and investment trusts. We compute the 30-minute herding measure for each investor group and explore the rationality of investor herding in terms of the impact of herding on subsequent trading noise and returns. The literature generally views institutional investors as better informed and more rational because of their advantages in acquiring and analyzing information (Busse, 1999; Li and Wang, 2010). By contrast, individuals are typically less informed and are often regarded as noise traders because they are inferior at obtaining information (Brennan, 1995; Kaniel et al., 2008). Information disadvantages may cause individuals to imitate other investors' trading behavior. As information-based herding predicts future returns and reduces trading noise in subsequent periods, uninformed traders tend to cluster their trades in the opposite direction of future market movements and increase trading noise in subsequent periods.

In addition, we explore the rationality of investor herding during crisis and non-crisis periods. Our sample period involves the recent global financial crisis of 2008. Economou et al. (2011) point out that market prices may deviate from fundamental values due to liquidity constraints and information asymmetries during a financial crisis. We are interested in whether the extreme market conditions will affect the rationality of investor herding. We divide our sample period into the crisis period and the non-crisis period and examine the impact of herding on subsequent trading noise and returns.

Our main findings suggest rational herding by institutional investors and irrational herding by individuals. Herding of foreign institutions reduces trading noise in the subsequent periods during both the crisis period and the non-crisis period, while individual herding results in persistently high trading noise. Although domestic institutions present informational herding, they cannot acquire information as well as foreign institutions. Their herding increases subsequent trading noise during the non-crisis period. Furthermore, institutional investors' buy (sell) herding predicts future upward (downward) price movement, while individual investors' buy (sell) herding negatively correlates with future returns. The findings generally support the view that institutional investors are informed traders while individuals are uninformed.

The results are robust in subgroups categorized by turnover; each subgroup, the high-turnover group and the low-turnover group, exhibits similar patterns of informed institutional herding and uninformed individual herding. The impact of investor herding is greater for stocks with high turnover for all investor groups. The results suggest distinct preferences of institutional investors and information disadvantages of individual investors. Our findings are also robust after we merge split orders submitted by the same trader to diminish a potential bias in herding measures due to order splitting.

The remainder of this paper is organized as follows. Section 'Data and Institutional Background' introduces the institutional background of the TWSE; Section 'Methodology' describes the data and empirical methodology; the empirical results are discussed in Section 'Empirical Results and Discussion'; and Section 'Robustness Checks' reports on robustness checks.

2. Data and Institutional Background

For this paper, we obtained intraday order book data, transaction data, and quote data from 2005 and 2010 from the TWSE, which is one of the most important financial markets in the pan-Pacific region, with high trading volume and turnover. The Monthly Bulletin of Statistics of the Republic of China (October 2012) reports 758 firms listed on the TWSE at the end of 2010, with a total market value of approximately US$0.8 billion and an average daily trading value of US$3840 million.

The TWSE is an order-driven market, in which there is no designated dealer or specialist serving as a market maker. Orders submitted by investors accumulate beginning at 8:30 AM, and transactions occur from 9:00 AM to 1:30 PM. The market-clearing price matches most shares. All buy orders quoted above the market-clearing price and sell orders quoted below the market-clearing price are executed at the market-clearing price. After a match, the best bid price is the highest price of unexecuted buy orders and the best ask price is the lowest price of unexecuted sell orders. Unfulfilled orders remain in the order book and are eliminated at the end of the trading day.

The order book records the date; the time in hours, minutes and seconds; the stock code; and the submitted price and quantity. As the order book also contains flags identifying each investor either as a proprietary dealer, an investment trust, a foreign institution, or an individual, we are able to classify investors into four investor groups. The samples adopted in this study include 495 stocks, excluding stocks with irregularities and unusual exchange sanctions. The TWSE adjusted the tick size of prices beginning on March 1, 2005. To avoid an adjustment effect on this study, the sample period runs from March 1, 2005 to December 31, 2010, for a total of 1459 trading days. We further divide our sample into two periods, the crisis period and the non-crisis period. The crisis period is from September 15, 2008 to March 31, 2009, following the definition in Aït-Sahalia et al. (2012); the remaining trading days are defined as the non-crisis period. Each trading day is divided into daily sessions between 9:00 AM and 1:30 PM, comprising nine 30-minute intervals.

3. Methodology

3.1. Noise Estimation

The traditional estimation of trading noise relies on the trade indicator, which determines whether a trade is initiated by the buyer or the seller (buyer- or seller-initiated), because the trade initiators are likely to contain private information (Glosten and Harris, 1987; Huang and Stoll, 1996). However, in a call auction market such as the TWSE, there is no individual who initiates a trade. Orders are accumulated for a period and executed at a price that maximizes trading volume. As such, implementing the sense of trade initiation to estimate noise in a call auction market is problematic. Therefore, we adopt an estimation of noise proposed by Hu (2006). Hu (2006) decomposes transaction prices into permanent and noise components, and suggests that one can estimate noise using the expectation of price reversal and some instruments that are correlated with noise. He suggests that lagged order imbalances and lagged trade-by-trade returns can capture noise. When buy orders exceed sell orders temporarily due to noise, the price will go up, but it will drop afterwards. Similarly, the price will reverse if lagged returns are attributed to noise. Therefore, lagged order imbalances and lagged returns can be used as instruments for estimating the noise. Moreover, because the relationship between noise and the length of the trading interval is ambiguous in the literature, the coefficients of instruments in estimating noise are allowed to differ across various lengths of the trading intervals. The estimation equation for noise is as follows:

display math(1)

where Pt is the logarithmic transaction price at time t and Pt is the price observed thirty minutes after t; nt is the noise component of the transaction price. Lagged returns (r) and lagged order imbalances (OIB) are instruments. rt is defined as Pt − Pt−1; OIBt equals 1 if the depth at the best bid is greater than the depth at the best ask after immediate transaction at time t, and OIBt equals −1 if the depth at the best bid is less than the depth at the best ask. We also allow coefficients on return and order imbalance to differ across various lengths of the trading interval through dummy variables, Di,t−k. Figure 1 shows the frequency distribution of the length of the trading interval between matches in our sample period. Approximately 81.8% of our samples are matched within 60 seconds and 90.8% are matched within 120 seconds. Based on Figure 1, we set dummy variables D1,t, D2,t, and D3,t, to be 1 when the length of the trading interval between t−1 and t is less than 60 seconds, between 60 and 120 seconds, and more than 120 seconds, respectively.

Figure 1.

Frequency distribution of the length of the interval between trades

To obtain the estimation of noise, we run regressions with Eq. (1) for each stock and each month separately following the method in Hu (2006), and then use the fitted value as the estimation of noise in each transaction price. After estimating the noise nt, we compute the variance of returns [Var(rt)], variance of noise [Var(nt − nt−1)], and the ratio of the two (math formula) in each 30-minute interval. The ratio is used as a measure of trading noise3 and is expressed as:

display math(2)

Table 1 reports the summary statistics of trading noise. During the overall sample period, an average of approximately 36% of the variance of trade-by-trade returns is due to trading noise, and trading noise ranges between 24% and 52%. Table 1 also reveals that, compared with the non-crisis period, trading noise is markedly higher during the financial crisis period. The minimum trading noise is 42% and the average increase is 48%, indicating that nearly half of the return variation can be attributed to trading noise during the financial crisis, whereas approximately one-third of the return variation is due to trading noise during the non-crisis period.

Table 1. Summary statistics of intraday trading noise
 MeanMedianS.D.MinMaxObs
  1. This table reports descriptive statistics for intraday trading noise during the overall sample period and two subperiods, the crisis period and the non-crisis period. Trading noise is defined as the ratio of variance of noise to variance of returns within each 30-minute interval, which is expressed in Eq. (2). The crisis period is defined as September 15, 2008 to March 31, 2009. *denotes the significant difference from the non-crisis period at 1%.

Overall period0.36290.31840.08510.24680.52464 128 194
Crisis period0.4804*0.48840.02370.41680.5157102 994
Non-crisis period0.35110.31580.08000.24680.52464 025 200

3.2. Herding Measurement

Because the LSV measure does not apply to high-frequency data, we use the bootstrap run test developed by Patterson and Sharma (2005) to measure herding intensity. Patterson and Sharma (2005) argue that a long sequence of buying or selling orders can be observed if investors herd. The runs of orders in a fixed interval are relatively smaller in the presence of investor herding. Consequently, the bootstrap test is based on the number of buying- or selling- order runs in a fixed interval with the discontinuity adjustment as in Mood (1940). We define a random variable x(w, p, j) as a type of run (buy when = 1 and sell when = 2) of stock j in interval p, which is expressed as:

display math(3)

where Rw is defined as the actual number of runs and N is the total number of orders of stock j during interval p; math formula is the discontinuity adjustment parameter; and probw denotes the probability of buy or sell runs. We also eliminate the observations with fewer than five orders in an interval to prevent the inaccuracy of herding intensity measurements. The random variable x(w, p, j) follows asymptotic normal distribution with zero mean and variance as follows:

display math(4)

Therefore, the herding measure is the random variable divided by the standard deviation, and is expressed as follows:

display math(5)

Investor herd behavior tends to lower runs more than expected. Hence, the value of HM(w, p, j) diminishes as herding intensity increases. Under the assumptions of identical independent distribution and continuous distribution, HM(w, p, j) follows an asymptotic normal distribution of N (0,1). However, due to the transaction prices generated by the discrete process, HM(w, p, j) is not a normal distribution. Therefore, we follow Patterson and Sharma (2005) to construct critical values using the bootstrap method. For each group of investors, we calculate the herding measure for every 30-minute interval and bootstrap their 10% critical values separately for each stock. Using the critical values, we can construct the significant herding dummy H, which equals 1 if the herding measure is smaller than the 10% critical value and equals 0 otherwise.

We also compute the directional herding dummy DH to indicate buy and sell herding. DH equals 1 if a particular group of investors' herding is significant and the number of buy orders submitted by that group of investors exceeds the number of sell orders; i.e., DH equals 1 in the case of significant buy herding. DH equals −1 if investor herding is significant and the number of sell orders exceeds the number of buy orders; i.e., DH equals −1 in the case of significant sell herding. Otherwise, DH equals 0.

The summary statistics for herding measurements of investor groups are reported in Table 2. Among investor groups, foreign institutions exhibit the strongest herding, followed by individuals, during the overall period. On the TWSE, as foreign institutions and individuals are the most active investors, they also have the greatest tendency to cluster their trades. By contrast, proprietary dealers tend to use a wide variety of trading strategies to make profits. They do not adjust their portfolios very often in intraday periods. Thus, the herding intensity of proprietary dealers is relatively lower. Moreover, we also find that all four investor groups show stronger herding intensity during the crisis period. Christie and Huang (1995) and Demirer et al. (2010) use the dispersion of stock returns to detect herding and suggest that, during periods of extreme market movements or market losses, investors tend to disregard their own private information and to mimic the trading decisions of others. Using the order data, we also find that investors are more likely to cluster their trades during the crisis period, which is an extreme market condition.

Table 2. Summary statistics of herding measurements by investor group
 MeanMedianS.D.MinMaxObs
  1. The table reports descriptive statistics herding measurements for the four groups of investors during overall sample period, and two subperiods of the crisis period and the non-crisis period. The herding measurements of investor groups are computed by Eq. (5), using the filled orders submitted by each investor group within each 30-minute interval. The crisis period is defined as September 15, 2008 to March 31, 2009. ** and **denote the significant differences from the non-crisis period at 1% and 5%, respectively.

Panel A: Proprietary dealers
Overall period−0.7685−0.80130.3665−1.69380.42171 250 508
Crisis period−0.9784***−0.95950.2611−1.6339−0.3866124 467
Non-crisis period−0.7475−0.76620.3690−1.69380.42171 126 041
Panel B: Investment trusts
Overall period−1.1176−1.11260.4567−2.3774−0.0198551 744
Crisis period−1.3319***−1.34120.2430−1.9482−0.733160 677
Non-crisis period−1.0961−1.07510.4675−2.3774−0.0198491 067
Panel C: Foreign institutions
Overall period−1.9872−1.94510.4185−3.7361−0.41632 179 460
Crisis period−2.0701**−2.02960.5540−3.7361−0.4163189 324
Non-crisis period−1.9789−1.93810.4017−3.6747−0.53571 990 136
Panel D: Individuals
Overall period−1.1447−1.13270.2995−2.4766−0.39615 898 274
Crisis period−1.1772**−1.16140.2098−1.7725−0.5294508 192
Non-crisis period−1.1415−1.12590.3069−2.4766−0.39615 390 082

4. Empirical Results and Discussion

Table 3 reports the mean intraday trading noise in each 30-minute interval from market opening to closing during the crisis period and the non-crisis period. The trading noise within a 30-minute interval corresponding to the day of the week (Monday to Friday) and intraday interval (from market opening to closing) is averaged for each stock. Table 3 shows that trading noise exhibits a reverse-J shape. Trading noise is the highest at market opening, becomes increasingly lower in the middle of the day, then increases when approaching market closing. The reverse-J pattern is more obvious during the crisis period. Lee et al. (2001) find that both informed and non-informed investors tend to place more orders both at market opening and closing. As a consequence, trading volume is typically higher during these two intervals. High trading volume accompanies high trading noise at market opening and closing, suggesting that uninformed investors seem to be more influential. The return variation at market opening is not due to information revealed overnight, but rather to trading noise.

Table 3. Intraday trading noise
 9:00–9:30 9:30–10:00 10:00–10:30 10:30–11:00 11:00–11:30 11:30–12:00 12:00–12:30 12:30–13:00 13:00–13:30
  1. This table reports the mean intraday trading noise within the 30-minute interval corresponding the day of the week (Monday to Friday) and intraday interval (from market opening to closing) during the crisis period and the non-crisis period. Trading noise is defined as the ratio of variance of noise to variance of returns within each 30-minute interval, which is expressed in Eq. (2). The crisis period is defined as September 15, 2008 to March 31, 2009.

Panel A: Crisis period
Mon0.49470.47930.47700.47170.47620.47670.48100.48190.4822
Tue0.49820.48420.47740.48110.47620.47420.48170.48010.4880
Wed0.49650.47990.48050.48010.47810.47250.47780.48090.4816
Thu0.49090.48500.48460.48010.47950.47920.47590.47700.4793
Fri0.49720.48310.48410.47560.47180.47610.47040.47580.4762
All0.49590.48280.48050.47780.47650.47610.47800.47930.4807
Panel B: Non-crisis period
Mon0.35400.35050.35120.35070.35080.35100.35140.35190.3531
Tue0.35440.35060.35170.35160.35170.35070.35150.35200.3530
Wed0.35420.35220.35230.35070.35290.35170.35370.35420.3526
Thu0.35320.34970.34990.35020.35050.35040.35140.35190.3530
Fri0.35180.34860.34920.34940.34830.34920.35020.35130.3517
All0.35330.35020.35070.35040.35070.35040.35150.35210.3528

We next investigate the relationship between herding intensity and subsequent trading noise to determine whether the trading noise will increase or decrease when herding intensity is more pronounced. We classify herding measurements of investor groups for every 30-minute interval into three herding intensity groups, separately for the crisis period and the non-crisis period. For each stock, herding measurements of investor groups are sorted from the lowest to the highest and are divided equally into three groups. The lowest (highest) herding intensity group comprises the highest (lowest) one-third herding measurements of investor groups. We then average the trading noise in the next interval corresponding to the herding intensity group for each stock. Table 4 reports the mean subsequent trading noise (that is, trading noise in the next 30-minute interval) corresponding to the three groups classified by the herding intensity of each investor group. Panel A shows that during the crisis period the herding intensity of individuals has a positive relationship to subsequent trading noise. This suggests that the greater herding intensity for individuals is not driven by information, and therefore reduces market efficiency and increases subsequent trading noise. By contrast, the herding intensity of institutions (including foreign institutions and domestic institutions) and subsequent noise exhibits a negative relationship. This shows that the higher herding intensity of institutions may be information-based, resulting in reduced subsequent trading noise.

Table 4. Relationship between herding intensity and subsequent trading noise
 Proprietary dealersInvestment trustsForeign institutionsIndividuals
  1. This table reports the mean subsequent trading noise (that is, trading noise in the next 30-minute interval) corresponding to three herding intensity groups during the crisis period and the non-crisis period. The lowest (highest) herding intensity group comprises the highest (lowest) one-third herding measurements of investor groups. Trading noise is defined as the ratio of variance of noise to variance of returns within each 30-minute interval, which is expressed in Eq. (2). The crisis period is defined as September 15, 2008 to March 31, 2009.

Panel A: Crisis period
H1(lowest)0.49600.49170.50000.4940
H20.48560.48930.48610.4959
H3(highest)0.48100.48420.47700.4976
Panel B: Non-crisis period
H1(lowest)0.31560.30610.32670.3087
H20.31890.31270.32960.3360
H3(highest)0.32370.31880.32750.3260

However, the pattern changes during the non-crisis period. There is no monotonic relationship between herding intensity and subsequent trading noise for foreign institutions and individuals. The greater herding intensity of domestic institutions results in greater subsequent trading noise, which is completely the reverse of the results during the crisis period. There are two possible explanations for the contrasting results for domestic institutions. The first explanation is that domestic institutions have more incentive to pay for acquiring information during extreme market conditions. Equipped with the information they have acquired, their high herding intensity will decrease subsequent trading noise during the crisis period. During the non-crisis period, the relatively insufficient information causes their herding intensity to have a positive relationship with subsequent trading noise. Therefore, domestic institutions exhibit informational herding during the crisis period but uninformed herding during the non-crisis period. Another explanation is that domestic institutions acquire information both during the crisis period and the non-crisis period; however, superior skills are required to refine information from the signals or trends revealed in the market under relatively stable market conditions. Some domestic institutions cannot distinguish information from noise with great accuracy during the non-crisis period, resulting on average in a positive relationship between herding intensity and subsequent trading noise during the non-crisis period.

To obtain further evidence to determine that whether the herding of investors is driven by information, we next consider the effect of investor herding on subsequent trading noise and returns. The consequences of information-based herding and non-information-based herding are different. If investor herding results from informational reasons, such herd behavior drives asset prices toward fundamental values. In such a scenario, the information-based herding reduces subsequent trading noise because the information is impounded into the price of the security. In addition, we would expect the informational buy (sell) herding to predict positive (negative) future market returns.

On the other hand, investors could herd without information. The non-informational buy (sell) herding of investors is negatively correlated with future market returns. However, there are two alternative effects of non-informational herding on subsequent trading noise. If the non-informational herding acts as noise trading, it may drive asset prices away from fundamental values and reduce the efficiency of stock prices. Such non-information-based herding increases trading noise in the subsequent period. By contrast, non-information-based herding may reduce subsequent trading noise. The absence of non-information-based herding, which predicts negative future returns and reduces trading noise in the later period, indicates that investors herd as liquidity providers. The liquidity they provide allows informed traders to incorporate their information into the price of the security. Consequently, we can observe reduced trading noise after liquidity providers herd, even if their herding is not driven by information.

We focus on observations of instances in which investors appear to engage in significant herding, and investigate the effect of herding of different investor groups on subsequent trading noise and returns. To define significant herding, we construct the 10% critical values through bootstrapping the sample for each stock separately and use dummy variables that equal 1 if observations of herding are lower than critical values. We also compute the directional herding dummy to indicate buy and sell herding. Buy herding occurs when the herding of a particular investor group is significant and the number of buy orders submitted by that group of investor exceeds the number of sell orders. Similarly, sell herding occurs when investor herding is significant and the number of sell orders exceeds the number of buy orders. After constructing the dummy variables for investor herding, we explore the impact of investor herding on subsequent trading noise and subsequent returns. The regressions are as follows:

display math(6)
display math(7)

where NRj,p+1 denotes trading noise in stock j during interval p + 1; RETj,p+1 denotes the 30-minute returns in stock j during interval p + 1 following the method in Harris (1986). math formula, depending on which investor group is examined in the regression; THj,p denotes the dummy of significant herding of proprietary dealers; MHj,p denotes the dummy of significant herding of investment trusts; FHj,p denotes the dummy of significant herding of foreign institutions; and IHj,p denotes the dummy of significant herding of individuals. FCp is a dummy variable that denotes the crisis period; β1 indicates the effect of investor herding on subsequent trading noise and returns during the non-crisis period, and β2 indicates the incremental effect of investor herding during the crisis period. As the trading noise exhibits the intraday pattern observed in Table 3, we add intraday interval dummy variables into the regressions.

The investor herding variable Hj,p in Eq. (6) may be endogenous because of noise trading, and the directional herding variable DHj,p in Eq. (7) may be endogenous because of positive feedback trading. Trading noise and returns may cause herding; on the other hand, herding might also affect trading noise and returns. Moreover, the endogeneity of Hj,p and DHj,p is likely to entail endogeneity of the interaction variables Hj,p × FCp and DHj,p × FCp. To handle this problem, we estimate regressions by two-stage least squares with fixed effects. In Eq. (6), the lagged endogenous regressors (Hj,p−1) and lagged interaction variables (Hj,p−1 × FCp−1) are used as instrumental variables for Hj,p and Hj,p × FCp. Similarly, DHj,p−1 and DHj,p−1 × FCp−1 are used as instrumental variables in Eq. (7).

Table 5 reports the results. The effect of individual herding increases trading noise in the subsequent interval during both the crisis period and the non-crisis period. Individuals' directional herding also negatively correlates with subsequent returns, indicating that the clustered buy (sell) orders of individual investors are followed by negative (positive) market returns. The relationship between the herding of individuals and subsequent trading noise and returns suggests that individual herding contains limited information and individuals herd as noise traders. Individuals usually suffer from information disadvantages. Their herding increases trading noise in the subsequent periods. Moreover, their buy (sell) herding negatively correlates with future market returns, i.e., stock prices go down (up) after individuals' buy (sell) herding.

Table 5. Regression of subsequent trading noise and returns on current herding
 Proprietary dealersinvestment trustsForeign institutionsIndividuals
  1. Panel A: NRj,p+1 = β0 + β1Hj,p + β2Hj,p × FCp + β3 × FCp+1 + β4 × NRj,p + ∑IntervalDum + εj,p+1,NR

  2. Panel B: RETj,p+1 = β0 + β1DHj,p + β2DHj,p × FCp + β3 × FCp+1 + β4 × RETj,p + ∑IntervalDum + εj,p+1,RET

  3. NRj,p+1 denotes trading noise in interval + 1 for stock j, and RETj,p+1 denotes returns during the 30-minute interval. math formula, depending on which investor group is examined in the regression; THj,p denotes the dummy of significant herding of proprietary dealers, MHk,t denotes the dummy of significant herding of investment trusts, FHj,p denotes the dummy of significant herding of foreign institutions, and IHj,p denotes the dummy of significant herding of individuals. DHj,p equals 1 if investor herding is significant and the number of buy orders exceeds the number of sell orders in interval t for stock k, denoting significant buy herding; DHj,p equals −1 if investor herding is significant and the number of sell orders exceeds the number of buy orders, denoting significant sell herding; otherwise, DHj,p equals 0. FCp is a dummy variable that denotes the crisis period. The results are estimated by two-stage least squares with fixed effects. Hj,p−1 and Hj,p−1 × FCp−1 are used as instrumental variables for Hj,p and Hj,p × FCp in Panel A; DHj,p−1 and DHj,p−1 × FCp−1 are used as instrumental variables for DHj,p and DHj,p × FCp in Panel B. Robust standard errors to control for heteroskedasticity are reported in parentheses. ***, ** and * denote significance at 1%, 5% and 10%, respectively.

Panel A: NRj,p+1
Constant9.528***9.553***9.518***9.676***
(0.0212)(0.0212)(0.0213)(0.0226)
H 1.296***0.895***−0.589***1.388***
(0.0964)(0.0668)(0.123)(0.0844)
H × FC−3.700***−1.995***−1.059**1.332**
(0.590)(0.525)(0.459)(0.646)
FC 5.513***5.461***5.424***5.250***
(0.0542)(0.0508)(0.0543)(0.0900)
Obs3 695 1213 695 1213 695 1213 695 121
Panel B: RETj,p+1
Constant−0.0033***−0.0034***−0.0035***−0.0032***
(0.0010)(0.0010)(0.00010)(0.0009)
DH 0.0314***0.0053*0.0398***−0.438***
(0.00569)(0.0031)(0.00409)(0.00501)
DH × FC0.0316*0.0275**0.0404***−0.0329*
(0.0187)(0.0109)(0.0141)(0.0186)
FC 0.0625***0.0624***0.0631***0.0629***
(0.0011)(0.0011)(0.0011)(0.0011)
Obs6 499 8436 499 8436 499 8436 499 843

On the contrary, foreign institutions herd as informed traders. Herding of foreign institutions reduces subsequent trading noise during not only the crisis period but also the non-crisis period. Furthermore, the directional herding of foreign institutions predicts the positive future returns during both the crisis period and the non-crisis period. These results provide supporting evidence of rational herding of foreign institutions that is driven by information.

The effect of the herding of domestic institutions (proprietary dealers and investment trusts) on subsequent trading noise is consistent with the results in Table 4 that the herding of domestic institutions significantly reduces subsequent trading noise during the crisis period but increases trading noise during the non-crisis period. The effect of directional herding on future returns, however, shows that domestic institutions incur positive returns during both the crisis period and the non-crisis period. Domestic institutions' buy herding predicts positive future market returns and vice-versa. This result is more consistent with the second explanation for increased trading noise after domestic institutions herd during the non-crisis period: domestic institutions acquire information during the crisis period and the non-crisis period; however, some domestic institutions cannot distinguish information from noise with great accuracy during the non-crisis period. On average, therefore, the herding of domestic institutions increases subsequent trading noise during the non-crisis period, while the directional herding is positively correlated with future market returns.

Overall, Table 5 demonstrates that individual herding is likely to be irrational and uninformed. Individuals incur losses from non-informational herding, causing persistently high trading noise in the subsequent periods. By contrast, institutional investors herd rationally. They tend to herd on the basis of information and thereby reduce future trading noise. In addition, their buy (sell) herding predicts positive (negative) future market returns. Our findings are consistent with the general view that institutional investors are informed traders while individual investors are uninformed traders.

5. Robustness Checks

5.1. Turnover

The results reported so far are based on all stocks. In this section, we divide the overall sample into stock groups according to turnover. Stocks are ranked by turnover and are divided equally into three groups, and the lowest and highest turnover groups are selected. The herding rationality of each investor group is analyzed following the same process as in Eq. (6) and (7). Previous studies have found that while institutional investors exhibit greater performance, individual investors have comparative advantages in the low turnover/liquidity stocks (Barber and Odean, 2000; Lin et al., 2007). Thus we expect that the herding of foreign institutions (individuals) more strongly reduces (increases) trading noise for stocks with high turnover, since institutional investors have greater information advantages for those stocks. In addition, high liquidity can ease the degree of information asymmetry. In the high-turnover stock group, domestic institutions can distinguish information from noise more accurately because of the lower degree of information asymmetry. Therefore, smaller increments of subsequent trading noise due to the herding of domestic institutions during the non-crisis period in high-turnover stocks than in low-turnover stocks can be expected. Moreover, we also expect the directional herding of institutional (individual) investors to have a greater positive (negative) impact on future returns for stocks with high turnover.

Table 6 reports the results of subsequent trading noise and returns after investors' herding for high and low turnover stocks, respectively. Consistent with the findings in the previous section, Panel A and B show that the herding of individuals (foreign institutions) increases (reduces) subsequent trading noise for both high- and low-turnover stocks during the crisis period and the non-crisis period. The directional herding of individuals negatively correlates with returns in the subsequent periods for both high- and low-turnover stocks, while the directional herding of foreign institutions has a positive impact on future returns. The effect of domestic institutional herding on subsequent trading noise and returns presents a similar picture to the results in Table 5. Domestic institutional herding reduces subsequent trading noise during the crisis period but increases trading noise during the non-crisis period. Their directional herding is positively correlated with future returns.

Table 6. Regression of subsequent trading noise and returns on current herding, controlling for turnover
 Proprietary dealersInvestment trustsForeign institutionsIndividuals
  1. Panel A: NRj,p+1 = β0 + β1Hj,p + β2Hj,p × FCp + β3 × FCp+1 + β4 × NRj,p + ∑IntervalDum + εj,p+1,NR

  2. Panel B: RETj,p+1 = β0 + β1DHj,p + β2DHj,p × FCp + β3 × FCp+1 + β4 × RETj,p + ∑IntervalDum + εj,p+1,RET

  3. NRj,p+1 denotes trading noise in interval + 1 for stock j; RETj,p+1 denotes returns during the 30-minute interval. math formula, depending on which investor group is examined in the regression; THj,p denotes the dummy of significant herding of proprietary dealers, MHk,t denotes the dummy of significant herding of investment trusts, FHj,p denotes the dummy of significant herding of foreign institutions, and IHj,p denotes the dummy of significant herding of individuals. DHj,p equals 1 if investor herding is significant and the number of buy orders exceeds the number of sell orders in interval t for stock k, denoting significant buy herding; DHj,p equals −1 if investor herding is significant and the number of sell orders exceeds the number of buy orders, denoting significant sell herding; otherwise, DHj,p equals 0. FCp is a dummy variable that denotes the crisis period. The results are estimated by two-stage least squares with fixed effects. Hj,p−1 and Hj,p−1 × FCp−1 are used as instrumental variables for Hj,p and Hj,p × FCp in Panel A; DHj,p−1 and DHj,p−1 × FCp−1 are used as instrumental variables for DHj,p and DHj,p × FCp in Panel B. Robust standard errors to control for heteroskedasticity are reported in parentheses. ***, ** and * denotes significance at 1%, 5% and 10%, respectively.

Panel A: NRj,p+1
High-turnover stocks
Constant6.256***6.245***6.270***6.290***
(0.0355)(0.0354)(0.0356)(0.0375)
H 0.787***0.359*−0.877***0.717***
(0.128)(0.190)(0.114)(0.143)
× FC−4.094***−3.191***−3.080***2.340**
(1.081)(0.787)(0.802)(1.168)
Obs840 236840 236840 236840 236
Low-turnover stocks
Constant12.10***12.11***12.05***12.22***
(0.0609)(0.0609)(0.0612)(0.0635)
H 0.938***0.989***−0.174*0.534***
(0.162)(0.186)(0.105)(0.163)
× FC−3.491***−2.448**0.794−0.501
(0.803)(0.984)(0.937)(1.331)
Obs543 318543 318543 318543 318
High-Low
H

−0.151*

(0.083)

−0.630***

(0.126)

−0.703***

(0.188)

0.183**

(0.081)

H × FC

−0.693***

(0.147)

−0.743*

(0.431)

−2.286***

(0.981)

2.841***

(1.011)

Panel B: RETj,p+1
High-turnover stocks
Constant−0.085***−0.085***−0.085***−0.084***
(0.002)(0.002)(0.002)(0.002)
DH0.383***0.502***0.424***−0.602***
(0.120)(0.104)(0.106)(0.112)
DH × FC0.030−0.0600.096**−0.120**
(0.044)(0.063)(0.044)(0.057)
Obs1 313 0981 313 0981 313 0981 313 098
Low-turnover stocks
Constant0.085***0.085***0.083***0.085***
(0.002)(0.002)(0.002)(0.002)
DH0.031***−0.0670.065−0.298***
(0.009)(0.0761)(0.0077)(0.009)
DH × FC0.069**−0.0060.0350−0.075**
(0.029)(0.051)(0.023)(0.031)
Obs1 313 0071 313 0071 313 0071 313 007
High-Low
DH

0.352***

(0.094)

0.569***

(0.157)

0.359***

(0.097)

−0.304***

(0.112)

DH × FC

−0.039

(0.101)

0.054**

(0.024)

0.061**

(0.024)

−0.005

(0.023)

As we expected, the results in Table 6 also indicate that the magnitude of the coefficients for high-turnover stocks is significantly greater than that for low-turnover stocks. The herding of individuals (foreign institutions) more strongly increases (reduces) trading noise for high-turnover stocks. The increment of subsequent trading noise due to herding of domestic institutions during the non-crisis period is smaller for high-turnover stocks than that for low-turnover stocks. Moreover, the directional herding of individuals (foreign and domestic institutions) has a greater negative (positive) impact on returns in the subsequent periods for high-turnover stocks.

5.2. Stealth Trading

An issue raised by the appearance of a decrease in the number of runs is whether it arises from herding by different traders or splitting of orders by the same trader. It has been found in the literature that institutions and informed traders tend to split their large orders over time to minimize the impact on prices, which is termed “stealth trading” by Barclay and Warner (1993).4

Order splitting by the same informed trader will inflate the herding measure calculated in the paper, as it will decrease the number of runs. Linnainmaa and Saar (2012) exclude orders from the same broker in order to capture trade clustering rather than order splitting by investors. In this section, we address the potential problem of order splitting by investors in order to diminish the bias in herding measures.

To achieve this robustness check, we use additional intraday order book data, which includes account ID, during the subperiod from September 1, 2007 to August 31, 2008. The account ID enables us to identify each trader directly. An order placed within thirty minutes with the same account ID and direction is regarded as a split order and is merged with the earlier order. We then recompute the herding measure and perform the regressions to explore the impact of herding on subsequent trading noise and returns during this subperiod. Because the subperiod does not involve the global financial crisis, we modify Eq. (6) and (7) as follows:

display math(8)
display math(9)

where the lagged endogenous regressors, Hj,p−1 and DHj,p−1, are used as instrumental variables for Hj,p and DHj,p, respectively.

The results are reported in Table 7. Panel A and Panel B display the result with and without the split orders. The results in Panel A show that during the subperiod, the impact of investor herding on subsequent trading noise and returns is consistent with the impact during the overall sample period. Individual herding has a significantly positive impact on subsequent trading noise, whereas the impact of institutional herding is significantly negative. The directional herding of individuals is still negatively related to future returns, while the directional herding of foreign institutions and investment trusts predicts positive future returns. After we merge the split orders to diminish the potential bias in herding measures, the results of the impact of herding show a similar picture to those in Panel A. Therefore, our findings are robust: institutional herding is information-based while individual herding is uninformed.

Table 7. Regression of subsequent trading noise and returns on current herding, controlling for stealth trading
 Proprietary dealersInvestment trustsForeign institutionsIndividuals
  1. Panel A: NRj,p+1 = β0 + β1Hj,p + β2 × NRj,p + ∑IntervalDum + εj,p+1,NR

  2. Panel B: RETj,p+1 = β0 + β1DHj,p + β2 × RETj,p + ∑IntervalDum + εj,p+1,RET

  3. NRj,p+1, RETj,p+1, Hj,p, and DHj,p are defined the same as in Table 6. Panel A and Panel B display the results with and without split orders from September 1, 2007 to August 31, 2008. Orders within thirty minutes with the same account ID and direction are regard as split orders and are merged to earlier ones. The results are estimated by two-stage least squares with fixed-effects. Hj,p−1 and DHj,p−1 are used as instrumental variables for Hj,p and DHj,p. Robust standard errors to control for heteroskedasticity are reported in parentheses. ***, ** and * denotes significance at 1%, 5% and 10%, respectively.

Panel A: All orders
NR j,p+1
Constant4.467***4.490***4.473***4.430***
(0.0604)(0.0603)(0.0605)(0.0628)
H −0.789***−0.253*−0.300*0.689***
(0.244)(0.154)(0.167)(0.227)
Obs404 715404 715404 715404 715
RET j,p+1
Constant0.0055**0.0048*0.0046*0.0046*
(0.0025)(0.0025)(0.0025)(0.0025)
DH0.008160.0253**0.0300*−0.0717***
(0.0257)(0.0115)(0.0163)(0.0275)
Obs1 095 9271 095 9271 095 9271 095 927
Panel B: Merging split orders
NR j,p+1
Constant8.183***8.239***8.287***8.167***
(0.068)(0.067)(0.072)(0.066)
H −0.811−2.871***−1.245***2.446*
(1.037)(0.356)(0.302)(1.264)
Obs404 715404 715404 715404 715
RET j,p+1
Constant0.005**0.005*0.005*0.004*
(0.003)(0.003)(0.003)(0.002)
DH 0.1170.100*0.033**−1.432***
(0.285)(0.054)(0.014)(0.551)
Obs1 095 9271 095 9271 095 9271 095 927

6. Conclusion

In this paper, we use a comprehensive data set on intraday transactions and limit order books on the TWSE to investigate the impact of herding on trading noise and returns. Contrary to most previous studies, which mainly examine the presence of herding, we focus on the causes that drive investors to cluster their trades. We distinguish information-based herding from non-information-based herding by examining whether investors' herding reduces future trading noise and whether buy (sell) herding predicts future market returns.

The most significant finding is that institutional herding is rational and information-based, but individual herding is not. The herding of foreign institutions reduces trading noise in the subsequent periods during both the crisis period and the non-crisis period, while individual herding results in persistently high trading noise. Although domestic institutions exhibit informational herding, they cannot acquire information as well as foreign institutions. Their herding increases subsequent trading noise during the non-crisis period. Furthermore, institutional investors' buy (sell) herding predicts future upward (downward) price movement, while individual investors' buy (sell) herding is negatively correlated with future returns. The results are robust over different turnover subgroups. The impact of herding is significantly greater for stocks with high turnover for all investor groups. The results suggest a greater advantage (disadvantage) for institutional (individual) investors in trading stocks with high turnover. Our findings are also robust after we merge the split orders to diminish the potential bias in herding measures due to order splitting.

Notes

  1. 1

    See related research, including Bikhchandani et al. (1992), Choe et al. (1998), Jones et al. (1999), Wermers (1999), Bowe and Domuta (2004), Chen and Hong (2006), Uchida and Nakagawa (2007), Lin et al. (2007), and Demirer et al. (2010).

  2. 2

    For instance, Chang et al. (2000) and Demirer et al. (2010) test the existence of herding; Lin and Swanson (2003), Chen et al. (2008), and Lu et al. (2012) investigate the herding of foreign institutions; Lin et al.(2007) examine foreign and domestic institutional herding, and Hung et al. (2010) investigate mutual fund herding.

  3. 3

    Hu (2006) examines the determinants of noise using both the variance of noise and the ratio of the two. However, the variance of noise is decomposed from, and can be largely influenced by, the variance of returns. Therefore, in our paper we use the ratio of the two as trading noise, a standardization of the variance of noise.

  4. 4

    We thank an anonymous referee for bringing this issue to our attention.

Ancillary