SSRN Id307440

Institutional Herding
Richard W. Sias*
February 24, 2002
*
Department of Finance, Insurance, and Real Estate, College of Business and Economics, Washington State
University, Pullman, Washington, 99164-4746, (509) 335-2347, [email protected]. I thank James Bennett, Jimmy
Senteza, Laura Starks, Harry Turtle, and David Whidbee for helpful comments and suggestions. I am grateful to the
Gary Brinson Research Fellows Program for providing financial support.
Abstract
Institutional investors’ demand for a security this quarter is positively correlated with
their demand for the security last quarter. These results are attributed to institutional
investors following each other into and out of the same securities (“herding”) and
institutional investors following their own lag trades. Consistent with previous work, we
find institutional investors are “momentum” traders. Little of their herding, however,
results from momentum trading. Moreover, institutional demand is more strongly related
to lag institutional demand than lag returns. Institutional herding declines over time and
differs across capitalizations and investor types. Our results are most consistent with the
hypothesis that institutional investors herd as a result of inferring information from each
other’s trades.
Over the past 30 years the fraction of market capitalization held by institutional investors has nearly
doubled.1 The growing institutional presence has led to a common perception that institutional herding
(i.e., institutional investors following each other into and out of the same securities) impacts security
prices and leads to excess volatility and market fragility. In a recent episode of Wall $treet Week, for
example, Louis Rukeyser claimed, “Who really did do the panicking at the bottom? We found what we
had long suspected. The real patsies were the large institutional traders who’s congenitally shaky nerves
get so much sensationalized media attention.”2 Consistent with these assertions, recent studies (Grinblatt
and Titman, 1989, 1993; Grinblatt, Titman, and Wermers, 1995; Wermers, 1999, 2000; Nofsinger and
Sias, 1999; Jones, Lee, and Weis, 1999; Cai, Kaul, and Zheng, 2000; Sias, Starks, and Titman, 2001)
document a strong positive relation between changes in institutional ownership and returns measured over
the same period. Moreover, Chakravarty (2000), Dennis and Weston (2000), and Sias, Starks, and
Titman (2001) conclude that the relation between changes in institutional ownership and returns measured
over the same period results primarily from the price impact of institutional trading. Although these
studies are consistent with the hypothesis that institutions trading in the same direction impact security
prices, they do not allow one to conclude institutional investors are herding because some securities will
experience net changes in institutional ownership simply by chance.
The primary goal of this paper is to examine this important outstanding question: Do institutional
investors herd? Specifically, we examine the cross-sectional correlation between institutional demand for
a security last quarter and institutional demand for the security this quarter. We demonstrate that the
fraction of institutions buying this quarter will covary (across securities) with the fraction buying last
1
Institutional investors accounted for 28 percent of total U.S. equity ownership in 1970 versus 50 percent in 1999.
2
Quoted by Richard Strozinsky (1998). Other examples include: (1) a recent Associated Press wire story (Cunniff,
2000) noting, “Institutions, of which mutual funds are numerically the largest segment, account for most of each
day’s trading and much of the volatility. . . And so, when you hear of references to investors and what they did in the
marketplace, think not of some prototypical individual, but visualize the herd instead,” (2) a recent Detroit Free
Press article (Yue, 1998) noting, “Pull back the stock market curtain and you’ll find institutional investors feverishly
working the strings,” and (3) a recent New York Times article (Barboza, 1998) claiming, “The selling was furious
early Tuesday … This surge in activity raises a question: Who exactly is moving the market this summer? The
answer: institutional investors…”
1
quarter if institutional investors follow each other into and out of the same securities (herd) or follow
themselves into and out of the same securities. Moreover, the cross-sectional correlation between the
fraction of institutions buying over adjacent quarters can be directly decomposed into the portion that
results from individual institutional investors following their own trades and the portion that results from
institutional investors following other institutional investors’ trades.
Herding is defined as a group of investors following each other into (or out of) the same securities
over some period of time.3 Thus, tests for herding implicitly or explicitly examine the temporal patterns
in the cross-section of institutional traders, i.e., whether they follow each other’s trades. The Lakonishok,
Shleifer, and Vishny (1992) herding measure, for example, tests for cross-sectional temporal dependence
by recognizing that if institutional investors follow each other into (out of) the same stocks over some
period of time, institutional investors will be primarily buyers (sellers) of the security over that time
period. We take a more direct approach by measuring the cross-sectional temporal dependence directly,
i.e., the extent that traders follow each other over adjacent quarters.
The rich theoretical foundation for institutional herding can be divided into five categories −
informational cascades, investigative herding, reputational herding, fads, and characteristic herding.4
Informational cascades result from institutional investors ignoring their own noisy information and
trading with the herd because they infer information from each other’s trades (Banerjee, 1992;
Bikhchandani, Hirshleifer, and Welch, 1992). Investigative herding occurs when institutional investors’
information is positively cross-sectionally correlated, possibly because they follow the same signals
(Froot, Scharfstein, and Stein, 1992; Hirshleifer, Subrahmanyam, and Titman, 1994). Reputational
herding is a consequence of institutional investors facing a reputational cost from acting different from
3
Occasionally, herding is loosely referred to as a group of investors buying or selling the same securities, “at the
same time.” Because trades occur sequentially, however, investors cannot buy or sell at the same time. Thus,
herding is accurately defined as a group of traders following each other into (or out of) the same securities over
some period. Note, however, this does not preclude the possibility that a group of investors accumulate positions
over the same period. For example, investor A may purchase a security on Monday followed by investor B
purchasing the security on Tuesday followed by investor A purchasing more of the security on Wednesday.
4
See Graham (1999), Wermers (1999), and Nofsinger and Sias (1999) for further discussions of these
classifications.
2
the herd (Scharfstein and Stein, 1990; Trueman, 1994). Institutional investors may also herd as a result of
fads (Freidman, 1984; Dreman, 1979; Barberis and Shleifer, 2001). Last, institutional investors may herd
because they are attracted to securities with specific characteristics (Falkenstein, 1996; Del Guercio,
1996; Gompers and Metrick, 2001; Bennett, Sias, and Starks, 2001).5 Thus, institutional herding has
important implications for how information, agency problems, fads, and security characteristics can affect
security selection decisions and asset prices.
Despite these perceptions and strong theoretical foundations, empirical evidence of institutional
herding is nearly non-existent. Lakonishok, Shleifer, and Vishny (1992), for example, conclude, “The
emerging image is that institutions follow a broad range of styles and strategies and that their trades offset
each other…” Similarly, Grinblatt, Titman, and Wermers (1995) and Wermers (1999) find little evidence
of systematic institutional herding. These studies, however, use the same test for institutional herding –
the Lakonishok, Shleifer, and Vishny “herding measure.”6
Our results, in contrast, reveal compelling evidence of institutional herding. In fact, while
previous studies (Grinblatt and Titman, 1989, 1993; Grinblatt, Titman, and Wermers, 1995; Wermers,
1999, 2000; Nofsinger and Sias, 1999; Jones, Lee, and Weis, 1999; Cai, Kaul, and Zheng, 2000; Sias,
Starks, and Titman, 2001) document strong evidence that institutions momentum trade, we demonstrate
that institutional demand is more strongly related to lag institutional demand than lag returns.
Specifically, we document a strong positive relation between the fraction of institutions buying over
adjacent quarters, consistent with both institutional herding and institutions following their own lag
trades. The decomposition reveals both factors play an important role in explaining the correlation
between institutional demand this quarter and last quarter.
5
Momentum trading serves as an example of characteristic herding, i.e., institutional investors are attracted to
(repelled by) securities with high (low) past returns.
6
In an interesting working paper, Pirinsky (2002), takes a different approach to testing for institutional herding by
examining time-series correlation (for individual securities) in changes in the fraction of shares held by institutional
investors. Consistent with our results, Pirinsky documents a positive relation between this quarter’s change in
fractional ownership and last quarter’s change in fractional ownership.
3
The balance of the paper focuses on better understanding institutional herding. We begin by
examining a special case of characteristic herding – “habit investing.” If institutional investors are
attracted to securities with the same characteristics, managers will hold many of the same securities. If
net flows exhibit positive time-series correlation, net flows are cross-sectionally correlated, and these
managers simply proportionally adjust their current portfolio when facing net flows, then these managers
will follow each other into and out of the same securities over adjacent quarters, i.e., “herd.”
To determine if habit investing explains institutional herding, we focus on the cross-sectional
correlation between the fraction of institutions increasing portfolio weights (i.e., the fraction of their
assets held in security k) over adjacent quarters. The results reveal little evidence that institutional
herding is driven by cross-sectional and time-series correlation in net flows to institutional investors.
We examine returns, momentum trading, and herding by capitalization to further test why
institutional investors herd. We document evidence of institutional momentum trading consistent with
characteristic herding (i.e., institutional investors are attracted to securities with large lag returns).
Momentum trading, however, accounts for little of the herding. Moreover, consistent with recent studies
(e.g., Grinblatt and Titman, 1989, 1993; Grinblatt, Titman, and Wermers, 1995; Wermers, 1999, 2000;
Nofsinger and Sias, 1999; Jones, Lee, and Weis, 1999; Sias, Starks, and Titman, 2001; Parrino, Sias, and
Starks, 2001), we document that institutional demand is positively correlated with contemporaneous
returns and weakly positively correlated with returns over the following year. The absence of negative
correlation between institutional demand and future returns suggests that herding is primarily related to
the way information is incorporated into security prices consistent with informational cascades and
investigative herding.
To differentiate informational cascades from investigative herding, we examine herding by
capitalization quintile. We use Wermers’ (1999) argument that informational cascades are more likely to
occur in small capitalization securities (where signals are noisier). In addition, we posit that investigative
herding may be more likely to occur in large capitalization securities (where signals are less noisy).
Although herding occurs in securities of all capitalizations, herding is strongest in small capitalization
4
securities. Our results are most consistent with models that suggest herding results from institutions
inferring information from each other’s trades (informational cascades).
We conclude the study by examining herding over time and by investor type. Specifically, we
investigate how institutional investors following their own and others’ trades changes over time as
institutional ownership grows and market liquidity improves. We find evidence consistent with the
hypothesis that institutional investors are less likely to follow their own lag trades as market liquidity
improves. Theoretically, it is not clear how increased institutionalization should impact herding.
Empirically, institutional herding is larger in the 1980s than the 1990s, and is primarily driven by a
decline in herding in the largest capitalization securities.
Previous studies of herding focus on pension funds (Lakonishok, Shleifer, and Vishny, 1992) or
mutual funds (Grinblatt, Titman, and Wermers, 1995; Wermers, 1999). Previous work (e.g., Del Guercio,
1996; Bennett, Sias, and Starks, 2001) also reveals that different types of institutional investors operate in
different environments (e.g., prudence regulations, holding periods, competition). Differing environments
may influence the likelihood that these investors herd. Moreover, as discussed in detail later, several
proposed theoretical motives for herding suggest specific differences in herding across investor
classifications. Although institutional investors in every classification exhibit statistically significant
evidence of herding, we document substantial differences across classifications. Moreover, we find some
evidence that institutional investors are more likely to follow similarly classified institutional investors
than differently classified institutional investors.
The remainder of the paper is organized as follows. We discuss the data in the next section. We
present tests for institutional herding in Section 2 and examine the possible explanations for institutional
herding in Section 3. In Sections 4 and 5 we examine changes in herding over time and differences across
investor types, respectively. A summary is presented in the final section.
5
1. Data
The data for this study come from two sources. Returns, shares outstanding, and firm capitalizations are
taken from the Center for Research in Security Prices (CRSP) monthly tapes for all New York Stock
Exchange, American Stock Exchange, and NASDAQ stocks. Ownership by each institutional investor
for each security comes from CDA-Spectrum and arises from institutional investors’ 13F filings.7 CDA-
Spectrum classifies each institutional investor as one of five types – bank trust departments, insurance
companies, mutual funds, independent investment advisors, and unclassified institutions.8 Institutional
ownership data are quarterly from March 1983 through December 1997, for a total of 60 quarters.
We begin by calculating, at the beginning and end of each quarter, each institutional investor’s
position in each security as a fraction of the security’s shares outstanding. For each security and quarter,
an institutional investor is defined as a buyer if their ownership in the stock increases and a seller if their
ownership decreases (e.g., an investor moving from holding 0.01 percent of IBM’s shares to 0.02 percent
of IBM’s shares is defined as a buyer).9 For each stock-quarter, we then calculate the fraction of
institutional investors trading the security that are buyers.10 We denote this ratio as the “raw fraction of
institutions buying” security k during quarter t:
# Institutions buying k ,t
Raw∆ k ,t = (1)
# Institutions buying k ,t + # Institutions selling k ,t
7
Institutional investors with $100 million or more under management in exchange-traded or NASDAQ-quoted
equity securities are required to file 13F reports within 45 days of the end of the calendar quarter. Institutions are
required to report all equity positions greater than either 10,000 shares or $200,000 in market value. Because 13F
reporting is aggregated across different units within an institution, the number of institutions reflects the number of
unrelated institutions buying or selling the security.
8
Unclassified institutions primarily consists of foundations, endowments, ESOPs, and internally managed pension
funds. Managers are defined by CDA-Spectrum based on their majority assets. If, for example, a mutual fund
manager also acts as an independent investment advisor, the manager will be classified as mutual fund manger if at
least 50 percent of their assets are in mutual funds.
9
Occasionally, due to small changes in the number of shares outstanding, fractional ownership changes slightly
while the numbers of shares held by the investor does not. Therefore, if the fraction of shares held by a given
investor is the same at beginning and end of the quarter or if the number of shares held by a given investor is the
same at the beginning and end of the quarter, the investor is classified as neither a buyer nor a seller.
10
Because we can only observe institutional ownership at the quarter-end, a manager that buys and sells the same
number of shares within a quarter will not be counted as a trader.
6
We run several filters over the data prior to computing the raw fraction of institutions buying. To
be included in the sample, a manager must hold at least one security at both the beginning and end of the
quarter. Because substantial changes in shares outstanding often accompany changes in CUSIPs (e.g.,
mergers), each security is required to have the same beginning- and end-of-quarter CUSIP to be included
in the sample. In addition, a security must have at least one institutional investor trading the security
during the quarter to be included in the analysis (so the denominator is greater than zero).
The second through sixth columns of Table 1 (Panel A) report the total number of institutional
investors (overall and by investor type) filing 13F reports at five points in our sample period
(approximately every 15 months). The first column reports the time-series average across all 59
quarters.11 Panel B reports the number of securities with at least one, five, 10, or 20 institutional traders,
as well as the number of securities with at least five traders of each type. Panel C reports the average
number of securities held by each investor across all institutions and by each type of institution. On
average, there are 894 managers and 5,505 securities with at least one institutional trader each quarter.
The average manager holds 273 securities in their portfolio. Consistent with Gompers and Metrick
(2001), the results in Table 1 reveal substantial growth in the number of institutions required to file 13F
reports, primarily driven by the growth in independent investment advisors.
[Insert Table 1 about here]
2. Tests for Institutional Herding
If institutional investors follow each other into and out of the same securities (herd), or if individual
institutional investors follow their own last quarter trades, then the fraction of institutions buying in the
current quarter will be positively correlated with the fraction of institutions buying in the previous quarter.
To allow for aggregation over time and to directly compare coefficients in our investigation of
momentum trading (equation (8)), across capitalizations, different measures of institutional demand, and
11
Sixty quarters of data yield 59 quarters that we can measure the raw fraction of institutions buying, i.e., we require
beginning- and end-of-quarter values.
7
investor types, we standardized both the dependent and independent variable(s) such that both have zero
mean and unit variance.12 Specifically, we define the standardized fraction of institutions buying security
k in quarter t (denoted ∆k,t) as:
Raw∆ k ,t − Raw∆ t
∆ k ,t = (2)
σ( Raw∆ k ,t )
where Raw∆ t is the cross-sectional average (across K securities) raw fraction of institutions buying in
quarter t and σ( Raw∆ k ,t ) is the cross-sectional standard deviation (across K securities) of the raw fraction
of institutions buying in quarter t.13 Because standardization is simply a linear rescaling, correlations
between the variables and the R2s associated with the regressions are not affected.
2.1 Does institutional demand predict institutional demand?
We begin the analysis by estimating, each quarter, a cross-sectional (across K securities) regression of the
standardized fraction of institutions buying security k (∆k,t) in the current quarter on the standardized
fraction of institutions buying security k the previous quarter (∆k,t-1):14
∆k,t = βt∆k,t-1 + εk,t (3)
The average coefficient from the 58 regressions and associated t-statistic (computed from the
time-series standard error) are reported in the first row of the first column of Table 2 Panel A.15 The
12
Aggregation or direct comparison of raw data regression coefficients across time, independent variables, different
measures of institutional demand, capitalization quintiles, or investor types is inappropriate because coefficients
depend upon the scale of the data.
13
From a practical standpoint, however, such standardization has little effect on the slope coefficient given in
equation (3). For example, using raw data, the average coefficient associated with lag raw fraction of institutions
buying (see equation (1)) is 0.1193 (t-statistic=12.79) which is partitioned (see equation (4)) into a following their
own trades component of 0.0614 (t-statistic=7.31) and a herding component of 0.0579 (t-statistic=10.24). This is
not surprising because both the dependent and independent variables are divided by the cross-sectional standard
deviation of the raw fraction of institutions buying (at time t and t-1, respectively). Because the cross-sectional
standard deviations of the raw fraction of institutions buying are similar at time t and t-1, standardization will not
have a large impact on the slope coefficient. Subtracting the mean in equation (2) does not affect the slope
coefficients (i.e., it simply scales the intercept to zero).
14
Because both the dependent and independent variables are rescaled to zero mean, the intercept is zero.
15
Because the regression is of the fraction of institutions buying on the lag fraction buying, we estimate 58 cross-
sectional regressions from 60 quarters of data.
8
results reveal strong evidence that institutional investors follow other institutional investors and/or
themselves into and out of the same securities. The coefficient associated with the lag standardized
fraction of institutions buying averages 0.1194 and differs significantly from zero at the 1 percent level.
Because the data are standardized and there is a single independent variable, the coefficients in the
regressions of institutional demand on lag institutional demand have a straightforward interpretation – the
coefficients are the correlations.16 That is, the cross-sectional correlation between institutional demand
this quarter and last quarter averages 11.94 percent.
2.2 Following their own or others’ trades?
The correlation between institutional demand this quarter and institutional demand last quarter has two
components. First, the positive correlation could result from institutional investors following each other
into and out of the same securities (i.e., herding) over adjacent quarters. Second, the positive correlation
could result from individual institutional investors following themselves into and out of the same
securities over adjacent quarters.17
More formally, because the fraction of institutions buying can be written as a sum across
institutional investors, the correlation between the current quarter’s and lag quarter’s fraction of
institutions buying computed from N institutional investors across K securities can be decomposed into
the portion that results from an institutional investor following themselves into and out of the same
securities over adjacent quarters and the portion that results from institutional investors following other
16
See appendix for proof.
17
Assume, for example, there is only one institutional investor and that investor reduces their IBM position in
quarters one and two and increases their Microsoft position in the same quarters. For this lone investor and two
securities, the fraction of institutions buying in the first quarter will be positively correlated (across securities) with
the fraction of institutions buying in the second quarter. For this extreme example, the cross-sectional (population)
covariance of the raw fraction of institutions buying is 0.25, i.e., the fraction of institutions buying IBM is 0 percent
in quarters one and two and 100 percent for Microsoft in quarters one and two. Thus, the covariance is: ((0-0.5)(0-
0.5)+(1-0.5)(1-0.5))/2 = 0.25. The cross-sectional standard deviation (at both time t and t-1) of the raw fraction of
institutions buying is ((0-0.5)2+(1-0.5)2)1/2=√0.25. Thus, the correlation between the raw or standardized fraction of
institutions buying is one.
9
institutional investors into and out of the same securities. Specifically, the fraction of traders that are
buyers can be written as the sum of a series of dummy variables for each trader (that equal one if the
trader is a buyer and zero if the trader is a seller) divided by the number of traders. As a result, the slope
coefficient in equation (3) can be written as (proof in appendix):
 1 ( )(
 K  N k ,t Dn ,k ,t − Raw∆ t Dn ,k ,t −1 − Raw∆ t −1 )
β = ρ(∆ k ,t , ∆ k ,t −1 ) =  ∑∑ 
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  n =1 N k ,t N k ,t −1 
 1 ( )( )
 K  N k ,t N k ,t −1 Dn ,k ,t − Raw∆ t Dm,k ,t −1 − Raw∆ t −1 
(4)
+ ∑∑ ∑
  
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  n =1 m =1,m≠ n N k ,t N k ,t −1 
where Nk,t is the number of institutional investors trading stock k in quarter t and Dn,k,t is a dummy
variable that equals one (zero) if trader n is a buyer (seller) of security k in quarter t. Similarly, Nk,t-1 is the
number of institutional investors trading stock k in quarter t-1, Dn,k,t-1 is a dummy variable that equals one
(zero) if trader n is a buyer (seller) of security k in quarter t-1, and Dm,k,t-1 is a dummy variable that equals
one (zero) if trader m (m≠n) is a buyer (seller) of security k in quarter t-1.
The first term on the right-hand side of equation (4) is the portion of correlation (or equivalently,
the portion of the slope coefficient in equation (3)) that results from institutional investors following
themselves into and out of the same securities. If institutional investors tend to follow their own trades
over adjacent quarters, the first term on the right-hand side of equation (4) will be positive, i.e., if investor
n buys security k in quarters t and t-1 or sells security k in both quarters, the first term will be positive.
Alternatively, if individual institutional investors’ transactions this quarter are independent of their own
transactions last quarter, the first term will be zero. The first term will be negative if institutional
investors tend to reverse their last quarter’s transactions.
The second term on the right hand side of equation (4) is the portion of the correlation that results
from institutional investors following other institutional investors into and out of the same securities.
This term will be positive if institutional investors tend to follow each other into and out of the same
securities, i.e., if investor m buys (sells) security k in quarter t-1 and investor n buys (sells) security k in
quarter t, the second term will be positive. If institutional investors tend to abandon (accumulate)
10
securities that other institutional investors purchased (sold) the previous quarter, this term will be
negative. If institutional investors’ transactions this quarter are independent of other institutional
investors’ transactions last quarter, this term will be zero.
Time-series averages of each term in equation (4) and associated t-statistics (computed from
time-series standard errors) are reported in the second and third columns of the first row in Panel A of
Table 2. The results reveal that, on average, about half of the correlation (i.e., 0.0617/0.1194) between
the fraction of institutions buying this quarter and the fraction of institutions buying last quarter results
from individual institutional investors continuing to buy (sell) the securities they bought (sold) the
previous quarter (statistically significant at the 1 percent level). Similarly, herding accounts for about half
the correlation (i.e., 0.0576/0.1194) between the fraction of institutions buying this quarter and last
quarter (statistically significant at the 1 percent level).18 Thus, the results reveal that institutional
investors follow themselves and each other (herd) into and out of the same securities.
2.3 Additional tests – Herding and the number of institutional traders
Because the regressions are estimated from the cross-section of securities with at least one institutional
trader, it is possible that our results are driven by securities with relatively few institutional traders. That
is, the fraction of institutions buying a given security this quarter may be strongly related to the fraction
buying the security last quarter because the only institution trading the security last quarter was a buyer
and the only two traders this quarter were buyers. Such a security will contribute more to the cross-
sectional correlation than a security that had 90 percent of 100 traders buying last quarter and 90 percent
of 200 traders buying this quarter.19 To consider this possibility, we repeat the analysis reported in the
first row of Table 2 Panel A but restrict the sample to those securities with at least five, 10, or 20
institutional traders.
18
The relative importance of each component is further investigated in Section 3.4.
19
That is, the product of ∆k,t and ∆k,t-1 in equation (A3) in the appendix would be larger for the first security than the
second.
11
Contrary to the example provided in the previous paragraph, restricting the sample to securities
with at least five institutional traders (first row of Panel B in Table 2) yields an even stronger average
correlation between the fraction of institutions buying this quarter and the fraction of institutions buying
last quarter. Moreover, both components of the correlation – institutions following their own trades and
institutions following other institutions’ trades – are greater in Panel B than Panel A. Restricting the
sample to securities with an even greater number of traders (i.e., Panels C and D), however, does not yield
substantial change in the coefficients.
2.4 Comparison with previous work
Contrary to previous studies, we document strong evidence of institutional herding. It is possible,
however, that differences between earlier results and our own arise from different samples rather than
different methodologies. To disentangle differences in methodology from differences in samples, we
compute the Lakonishok, Shleifer, and Vishny (1992) herding measure for our sample.
As noted in the introduction, Lakonishok, Shleifer, and Vishny (1992) test for cross-sectional
temporal dependence in institutional demand by recognizing that if institutional investors follow each
other into (out of) the same securities over some period of time, then institutional investors will primarily
be buyers (sellers) of those securities over that time period. Specifically, the Lakonishok, Shleifer, and
Vishny herding measure for security k in quarter t is defined as:
H k ,t = Raw∆ k ,t − Raw∆ t − AFk ,t (5)
where the raw fraction of institutions buying is as defined in equation (1) and Raw∆ t is the cross-
sectional average (across K securities) raw fraction of institutions buying in quarter t as in equation (2).
To demonstrate how the Lakonishok, Shleifer, and Vishny (1992) measure captures cross-sectional
temporal dependence, assume that, within a given quarter, institutional investors are just as likely to be
buyers as sellers (i.e., Raw∆ t = 0.5). If in each security institutional investors do not follow each other’s
trades within the quarter (i.e., cross-sectional temporal independence), the number of institutional buyers
12
will, on average, equal the number of institutional sellers and Raw∆ k ,t − Raw∆ t will tend toward zero.
Alternatively, if within each security institutional investors tend to follow each other into (or out of) the
security within the quarter (i.e., cross-sectional temporal dependence), Raw∆ k ,t − Raw∆ t will be non-zero
(i.e., positive if institutions follow other institutions into a given security and negative if institutions
follow other institutions out of a given security) and the absolute value will be greater than zero for most
securities.
The last term in equation (5), AFk,t (adjustment factor for security k in quarter t), accounts for the
fact that even if institutional trades exhibit cross-sectional temporal independence, the expected value of
the absolute value of Raw∆ k ,t − Raw∆ t is greater than zero. That is, simply by chance (or an odd number
of traders), there will often be more or less buyers than sellers. AFk,t is computed for each security-
quarter in which there is at least one institutional trader, by assuming the number of institutional investors
buying security k in quarter t follows a binomial distribution with probability Raw∆ t .20
The average Lakonishok, Shleifer, and Vishny (1992) herding measure across our pooled cross-
sectional time-series sample of 324,770 security-quarters that experienced at least one institutional buyer
or seller is 0.0178 − slightly less than the 0.0270 average reported by Lakonishok, Shleifer, and Vishny.
When we restrict the measure to those stocks with at least five, 10, or 20 institutional investors trading,
the average increases, but remains similar to the averages reported by Lakonishok, Shleifer, and Vishny.21
We conclude that differences between previous studies and our own do not result from differences in
samples.
The key difference is the Lakonishok, Shleifer, and Vishny (1992) measure indirectly tests for
cross-sectional temporal dependence by recognizing that within a period, later institutional traders
20
See Lakonishok, Shleifer, and Vishny (1992) for a detailed discussion of their herding measure and computation
of the adjustment factor.
21
For our data, the average Lakonishok, Shleifer, and Vishny (1992) herding measure for the samples limited to
five, 10, or 20 traders are 0.0246, 0.0283, and 0.0329, respectively. All averages differ significantly from zero at the
1 percent level.
13
following earlier institutional investors’ trades will result in most institutional traders on one side of the
trade within that period. Alternatively, we directly test whether institutional investors follow each other’s
trades by examining the cross-sectional correlation between institutional investors’ trades in one period
and other institutional investors’ trades the next period.
3. Why Do Institutional Investors Herd?
We note five possible motives for institutional herding in the introduction − informational cascades,
investigative herding, reputational herding, characteristic herding (including habit investing), and fads.
These motives, of course, are not mutually exclusive − institutional investors may herd for a number of
reasons. Nonetheless, in this section we evaluate habit investing, momentum trading, returns, and
differences across capitalizations in an attempt to discriminate between these models.
3.1 Habit investing
A special case of characteristic herding may drive the results – herding could result from cross-sectional
and time-series correlation in the net flow of funds to groups of institutional investors.22,23 Specifically, if
some investors experience positive time-series correlation in their net flows and simply invest (divest)
flows into (from) their existing portfolios, then these investors will follow themselves into (when flows
over consecutive quarters are positive) and out of (when flows over consecutive quarters are negative) the
same securities over adjacent quarters. Moreover, if (1) subsets of institutional investors favor securities
with specific characteristics (and therefore hold many of the same securities), (2) flows to these investors
22
Habit investing is a special case of characteristic herding because institutional investors follow each other into and
out of the same stocks (herd) as a result of their attraction to securities with the same characteristics that results in
these institutions holding similar portfolios. That is, if there were no relation between security characteristics and
institutional ownership (i.e., each institution’s security preferences were independent of every other institution’s
preferences), time-series and cross-sectional correlation in the net flows to investors would not result in herding.
23
The possibility that habit investing explains herding is not unique to our tests. The Lakonishok, Shleifer, and
Vishny (1992) herding measure, for example, should capture herding resulting from habit investing regardless of
inter-quarter time-series correlation in net flows. That is, if net flows to subsets of institutional investors exhibit
positive cross-sectional correlation and these investors favor specific stocks (e.g., the securities they already hold),
the Lakonishok, Shleifer, and Vishny measure should be positive.
14
exhibit positive time-series and cross-sectional correlation, and (3) these investors simply invest (divest)
flows into (from) their existing portfolios, then these investors will tend to follow each other into (when
flows over consecutive quarters are positive) and out of (when flows over consecutive quarters are
negative) the same securities over adjacent quarters.24
To test whether habit investing explains herding and following their own lag trades, we examine
the correlation between the fraction of institutions increasing their portfolio weights this quarter and last.
If institutional investors follow themselves and each other into and out of the same securities as a result of
habit investing, then portfolio weights should be independent over adjacent quarters. For example, if
money were flowing into technology funds, technology funds would be buyers of Microsoft. However, if
these mutual funds buy additional shares of Microsoft in proportion to their current holdings, the funds’
Microsoft portfolio weights would not change. Alternatively, if institutional investors follow themselves
and each other into the same securities for reasons other than time-series and cross-sectional correlation in
net flows, then the fraction of institutions increasing portfolio weights will be positively correlated over
adjacent quarters.
We focus on changes in “return-adjusted” portfolio weights rather than changes in raw portfolio
weights to purge return-induced noise from the measure of the fraction of institutions increasing portfolio
weights.25 The return-adjusted portfolio weight is defined as what the end-of-quarter portfolio weight
24
These patterns could occur across similar managers within a classification. For example, technology-oriented
mutual funds may experience a net inflow of funds over consecutive quarters, while a set of mutual funds that focus
on utility stocks experience a net outflow of funds over the same quarters. Alternatively, these flows may occur
across classifications. Specifically, consistent with Gompers and Metrick (2001), Table 1 demonstrates that markets
have experienced substantial differences in the growth rates of different institutional investor classes. Moreover,
while recent work (Falkenstein, 1996; Del Guercio, 1996; Gompers and Metrick, 2001; Bennett, Sias, and Starks,
2001) demonstrates that institutional investors are attracted to securities with specific characteristics, these
preferences differ across classifications (Del Guercio, 1996; Bennett, Sias, and Starks, 2001). For example, mutual
funds may experience a net inflow of funds over consecutive quarters, while bank trust departments experience a net
outflow of funds over the same quarters. In this case, institutional investors (as a group) would herd to those
securities favored by mutual funds and out of securities favored by bank trust departments.
25
We focus on return-adjusted portfolio weights because the fraction of institutions increasing raw portfolio weights
is highly correlated with same period returns (the average cross-sectional correlation between the fraction of
institutions increasing their raw portfolio weight and same quarter returns is 61 percent, on average). For example,
if the investor begins with an equal-weighted portfolio, a stock that experiences a return greater than the average
return of other stocks in the investor’s portfolio will exhibit a larger portfolio weight at the end of the quarter than
the beginning if the investor does not rebalance the portfolio.
15
would be if the investor did not rebalance their portfolio.26 We begin by defining Vn,k,t as the value of
investor n’s position in security k at the end of quarter t (i.e., price at the end of quarter t times the number
of shares held by investor n at the end of quarter t). Investor n is classified as increasing their return-
adjusted portfolio weight (i.e., a buyer) if their end-of-quarter portfolio weight is greater than their return-
adjusted beginning-of-quarter portfolio weight:
V n , k ,t Vn ,k ,t −1 (1 + Rk ,t )
K
> K
(6)
∑V
k =1
n , k ,t ∑V
k =1
n , k ,t −1 (1 + Rk ,t )
where Rk,t is the return for security k over quarter t. Investor n is classified as a seller if the sign is
reversed in equation (6). If an investor buys or sells shares in direct proportion to their portfolio weights
at any time during the quarter, the left- and right-hand sides of equation (6) will be equal and the investor
will not be classified as a buyer or a seller.
We then compute the fraction of institutional investors that increase their return-adjusted portfolio
weight as a fraction of the institutional investors that increase or decrease their return-adjusted portfolio
weight for the security. Thus, the raw fraction of institutions increasing their security k return-adjusted
portfolio weights in quarter t is defined as:
# Inst. inc. return - adjusted weight k,t

Raw RAPW∆ k ,t = (7)
# Inst. inc. return - adjusted weight k,t + # Inst. dec. return - adjusted weight k,t
As before, we standardize the data (i.e., analogous to equation (2)), estimate cross-sectional regressions of
the standardized fraction of institutions that increase their return-adjusted portfolio weights on lag
standardized fraction of institutions that increase their return-adjusted portfolio weights (i.e., analogous to
equation (3)), and decompose the slope coefficient (i.e., analogous to equation (4)).
26
Intuitively, one may expect institutional investors to sell securities that have increased in value to rebalance their
portfolios toward initial weights. Extant work, however, documents strong positive correlation between institutional
demand and same quarter returns, inconsistent with the expected negative correlation associated with rebalancing. It
is possible, of course, that the positive correlation is driven by new traders rather than existing shareholders. To
consider this possibility, we compute the quarterly cross-sectional correlation between the fraction of existing
institutional shareholders that are trading who are buyers and returns the same quarter. We find positive correlation
between existing shareholder demand and same quarter returns, inconsistent with the expected negative correlation
associated with rebalancing.
16
The time-series average correlation (or, equivalently, slope coefficient), its components, and
associated t-statistics (computed from time-series standard errors) are reported as regression 2 in Table 2
Panel A. The results reveal that the fraction of institutions increasing their return-adjusted portfolio
weights is strongly correlated over adjacent quarters. Similar to the first regression, the correlation
between the fraction of institutional investors increasing their return-adjusted portfolio weights and the
lag fraction is about equally attributed to individual institutional investors following their own return-
adjusted portfolio weight changes (0.0615/0.1195) and following other institutional investors’ return-
adjusted portfolio weight changes (0.0580/0.1195). The results are inconsistent with the hypothesis that
institutional herding is primarily driven by institutional investors’ desire to hold their existing portfolios
and cross-sectional and time-series correlation in institutional investors’ net flows (i.e., habit investing).
Moreover, the results are inconsistent with the hypothesis that institutional investors follow their own lag
trades as a result of time-series correlation in their net flows and investing flows into their existing
portfolios.27,28
As with the first regression, we restrict the sample to those securities with at least five, 10, or 20
institutions changing their return-adjusted portfolio weights. The time-series average correlations,
correlation components, and associated t-statistics (computed from time-series standard errors) are
reported as regression 2 in Panels B through D of Table 2. As before, the results reveal that evidence
27
We also examine the time-series of net flows to individual institutional investors and find little evidence of
systematic positive time-series correlation. Specifically, we limit the sample to managers with at least 20 quarters of
data and define the quarterly net flow to each manager as the difference between the value of their end-of-quarter
total equity position and their return-adjusted beginning-of-quarter portfolio value (i.e., what their portfolio would
be worth at the end of the quarter if they did not trade and experienced zero net flows). We then estimate time-series
regressions of quarterly net flows on lag quarterly net flows for each manager. Of the 1,097 managers with
sufficient data, 52 percent exhibit positively correlated net flows and 48 percent exhibit negatively correlated net
flows. The average time-series autocorrelation in quarterly net flows is 0.0224. Ninety percent (88 percent) of the
negative (positive) autocorrelations do not differ significantly from zero at the 5 percent level or better.
28
As a robustness test of the importance of time-series and cross-sectional correlation in net flows, we compute a
more restrictive definition of buyers and sellers. Specifically, to be defined as a buyer, an investor must purchase
additional shares (i.e., satisfy the first definition) and the purchase must be large enough that the security accounts
for more of the investor’s portfolio at the end of the quarter than the beginning of the quarter after adjusting for the
security’s return (i.e., satisfy the second definition). Similarly, to be defined as a seller, an investor must decrease
their position in the security and decrease their return-adjusted portfolio weight. The results remain qualitatively
identical to those reported in Table 2.
17
institutional investors follow their own and others’ lag trades is stronger when the sample excludes
securities with little institutional trading. Specifically, the correlations (and their components) for the
fraction of institutions increasing return-adjusted portfolio weights based on five or more traders (Panel
B) are greater than the analogous correlations and components reported for the sample based on at least
one trader (Panel A). Unlike the analysis of changes in positions (the first regression), the correlation
tends to increase for the return-adjusted portfolio weight measure as the sample is restricted to securities
with an even greater number of traders (Panels C and D).
In sum, the analysis of changing portfolio weights is inconsistent with the habit investing
explanation. Institutional investors appear to herd for reasons other than time-series and cross-sectional
correlation in their net flows.
3.2 Momentum trading
Recent work (e.g., Grinblatt and Titman, 1989, 1993; Grinblatt, Titman, and Wermers, 1995; Wermers,
1999, 2000; Nofsinger and Sias, 1999; Jones, Lee, and Weis, 1999; Sias, Starks, and Titman, 2001)
suggests institutional investors are momentum traders. Institutional momentum trading is a case of
characteristic herding, i.e., institutional investors herd to (away from) stocks with high (low) past returns.
Moreover, inter-quarter herding may result from institutional investors’ momentum trading because the
fraction of institutions buying is positively correlated with same quarter returns. That is, institutional
investors may follow each other into and out of the same stocks because institutional investors are
momentum traders and the lag fraction of institutions buying proxies for lag return. On the other hand,
extant evidence of institutional momentum trading may result from institutional herding, i.e., institutional
investors may herd to (away from) stocks with large (small) lag returns because institutional demand is
correlated with lag institutional demand and lag returns proxy for lag institutional demand.
To evaluate momentum trading and its role in explaining the relation between the fraction of
institutions buying and the lag fraction of institutions buying, we add lag return as a standardized
independent variable to equation (3). Specifically, each quarter, we cross-sectionally regress (across K
18
securities) the quarterly standardized fraction of institutions buying on the lag quarterly standardized
fraction of institutions buying and lag quarterly standardized return:29
∆k,t = β1,t∆k,t-1 + β2,tRk,t-1 + εk,t (8)
Average coefficients from the 58 regressions and associated t-statistics (computed from time-
series standard errors) are reported in Table 3. In addition, we repeat the analysis for the standardized
fraction of institutions increasing return-adjusted portfolio weights. Results for the sample that includes
all securities with at least one, five, 10, or 20 institutional trader(s) are reported in Panels A through D,
respectively.
Consistent with earlier work, institutional investors, as a group, engage in momentum trading.
For example, in the regression of the standardized fraction of institutions buying for the sample including
all securities with at least one trader (the first regression in Panel A), the coefficient associated with lag
standardized returns averages 0.0691 (statistically significant at the 1 percent level). Because the data are
standardized, this suggests that a one standard deviation increase in last quarter’s return is associated with
a 6.91 percent standard deviation greater fraction of institutions buying this quarter. Institutional
momentum trading, however, accounts for little of the institutional herding. That is, adding standardized
lag return to the regression has little impact on the average coefficient associated with last quarter’s
fraction of institutions buying, e.g., the average coefficient reported in the first row of Panel A moves
from 0.1194 in Table 2 to 0.1133 in Table 3. Because all variables are standardized, the coefficients
associated with the lag fraction of institutions buying and lag return are directly comparable − the average
coefficient associated with the lag fraction of institutions buying is approximately 64 percent larger than
the average coefficient associated with lag return. Thus, on average, a one standard deviation change in
lag institutional demand predicts a 64 percent greater change in next quarter’s institutional demand than a
29
Standardized return is calculated by replacing the fraction of institutions buying with return in equation (2).
19
one standard deviation change in lag return. A paired t-test of the hypothesis the coefficients are equal is
rejected at the 1 percent level (t-statistic=3.61).
Repeating the analysis for the standardized fraction of institutions increasing return-adjusted
portfolio weights (the second regression) yields similar results. In addition, we find similar results for the
analysis restricted to securities with at least five, 10, or 20 institutional traders (Panels B through D). In
sum, the fraction of institutions increasing their position or return-adjusted portfolio weights is positively
related to lag returns as well as their own lag values.
There are two important results to emphasize in Table 3. First, the relation between institutional
demand this quarter and last quarter changes little after accounting for momentum trading (i.e., the
coefficients associated with lag institutional demand reported in Tables 2 and 3 are similar). Thus,
momentum trading does not appear to be the primary source of institutional herding. Second, although it
appears to be widely accepted that institutional investors engage in momentum trading (e.g., Grinblatt,
Titman, and Wermers, 1995; Wermers, 1999, 2000; Nofsinger and Sias, 1999; Jones, Lee, and Weis,
1999; Cai, Kaul, and Zheng, 2000; Sias, Starks, and Titman, 2001), institutions’ demand is much more
strongly related to their own lag demand than lag returns.30
3.3 Institutional demand and subsequent returns
Recent work (e.g., Grinblatt and Titman, 1989, 1993; Grinblatt, Titman, and Wermers, 1995; Wermers,
1999, 2000; Nofsinger and Sias, 1999; Jones, Lee, and Weis, 1999; Sias, Starks, and Titman, 2001;
Parrino, Sias, and Starks, 2001) documents a positive relation between measures of institutional demand
and returns the same and previous quarters. Moreover, these studies suggest that institutional demand is
weakly positively correlated with future returns. Assuming institutional herding impacts security returns,
the fads, reputational herding, and characteristic herding models suggest that subsequent returns should be
30
We can reject the hypothesis (at the 5 percent level or better) that the average coefficient associated with lag
standardized institutional demand equals the average coefficient associated with lag standardized return for each of
the eight regressions summarized in Table 3.
20
inversely related to institutional demand.31 That is, if price movements result from something other than
information, we expect to observe subsequent return reversals. Thus, existing evidence is inconsistent
with the hypothesis that institutional herding is primarily motivated by agency problems, attraction to
security characteristics, or fads.
Because our data and measures differ from the data and measures used in previous studies, we
also evaluate the relation between institutional demand and returns. Specifically, each quarter, we
measure the cross-sectional (across K securities) correlation between the fraction of institutions buying
(equation (1)) and returns measured over the previous quarter, same quarter, following quarter, following
six months, and following year.32 The time-series average of these correlations and associated t-statistics
(computed from time-series standard errors) are reported in Table 4.
Consistent with previous studies, we find institutional demand is positively correlated with both
current and prior quarter returns. Specifically, the correlation between the fraction of institutions buying
in the current quarter and returns the same (previous) quarter is 7.84 percent (7.44 percent). Restricting
the sample to securities with at least five, 10, or 20 institutional traders (the second through fourth rows in
Table 4) reveals even stronger relations between the fraction of institutions buying and returns the same
and previous quarters. Most important, consistent with previous work, we document a weak positive
correlation between the fraction of institutions buying and returns in the following quarter or six months.
For the sample including all firms, the evidence suggests the fraction of institutions buying and returns
over the following year are nearly independent. However, when restricting the sample to firms with at
least five, 10, or 20 institutional traders, we document stronger positive relations between the fraction of
institutions buying and returns over the following quarter, six months, or year (all are statistically
significant at the 1 percent level).
31
See Chakravarty (2000), Dennis and Weston (2000), and Sias, Starks, and Titman (2001) for evidence that
institutional demand impacts security prices.
32
Because standardization is a linear re-scaling, these are also the correlations between the standardized fraction of
institutions buying and returns.
21
The results reported in Table 4 are inconsistent with the joint hypothesis that institutional demand
impacts security prices (Chakravarty, 2000; Dennis and Weston, 2000; Sias, Starks, and Titman, 2001),
institutional herding is driven by fads, reputational herding, or characteristic herding, and that the impact
of such non-informational herding is reversed within the following year. The results, however, are
consistent with models of information-based herding – informational cascades and investigative herding.
3.4 Herding by firm size
The lack of subsequent return reversals is consistent with the hypothesis that institutional herding reflects
the manner in which information is impounded into security prices. Thus, the results are most consistent
with informational cascades (i.e., institutional investors infer information from the trades of other
institutions) and investigative herding (i.e., signals are positively correlated across institutional investors).
Wermers (1999) posits that informational cascades are more likely in small capitalization securities
because institutional investors would put a relatively larger weight on what the herd is doing and less
weight on their own noisy private information. Similarly, we hypothesize that the cross-sectional
correlation between signals (possibly resulting from following the same indicators) is likely to be stronger
in larger stocks with less noisy signals. Thus, if institutional herding primarily arises from inferring
information from each other’s trades (informational cascades), herding should be strongest in small
capitalization securities. Alternatively, if institutional herding primarily arises from correlated signals
(investigative herding), herding should be strongest in large capitalization securities.
The analysis by firm size is complicated by the fact that there are typically many more institutions
trading in a large capitalization security than there are trading in a small capitalization security. This
pattern will affect both the decomposition of the correlation coefficient (i.e., the relative importance of the
two terms in equation (4)) and the cross-sectional correlation between the fraction of institutional
investors buying this quarter and the fraction buying last quarter. It affects the decomposition because the
number of “herding” terms increases much faster than the number of “following their own trades” terms
22
as the number of investors increases.33 It affects the correlation because as the number of traders
increases, the cross-sectional standard deviation of the fraction of institutions buying tends to fall.34
Despite these limitations, we begin by examining the cross-sectional correlation between the
fraction of institutions buying this quarter and the fraction buying last quarter for firms within each
capitalization quintile to ensure there is evidence of herding and following their own trades within each
capitalization quintile. Because our focus is on understanding what drives herding, we limit the sample to
securities with at least five institutional traders.35 Specifically, at the beginning of each quarter we sort all
securities with at least five institutional traders into five groups based on beginning-of-quarter
capitalization. To allow direct comparison of regression coefficients across capitalization quintiles, we
standardize (within each capitalization quintile, each quarter) both the dependent and independent
variables to have zero mean and unit variance. Next, we estimate the regression given in equation (3) and
the decomposition given in equation (4), each quarter, limiting the sample to firms within a given
capitalization quintile. The time-series average of the standardized regression coefficients (or
equivalently, the correlations), their components, and associated t-statistics (computed from time-series
standard errors) are reported in Panel A of Table 5 for firms within each capitalization quintile.
33
Specifically, for each firm there are approximately Nk,t “following their own trades” terms (i.e., the first term in
equation (4)), Nk,t*(Nk,t-1-1) herding terms (i.e., the second term in equation (4)), and Nk,t*Nk,t-1 total terms. This
holds exactly if every trader who trades in the first period (for firm k) trades in the second period. If an investor
trades in only one period, then that trader has no following their own trades term but still participates in the
calculation of the herding term (see appendix for further discussion). Thus, as the number of traders (Nk,t) increases,
the fraction of total terms accounted for by following their own trades terms declines and the fraction of total terms
accounted for by herding terms increases.
34
Because the distribution of the fraction of institutions buying is smaller for larger firms, a given level of herding
or following their own trades will be able to explain a larger portion of next quarter’s fraction of institutions buying.
In other words, the denominator in the calculation of the correlation is inversely related to capitalization. If small
firms often have one or two traders, these firms will often have 100 percent of traders buying or 100 percent of the
traders selling. Alternatively, if a large firm has 50 institutional traders, it is unlikely 100 percent of those traders
will be on one side of the trade. Empirically, we document a monotonic inverse relation between the cross-sectional
standard deviation of the fraction of institutions buying and capitalization quintile.
35
Evidence of herding is strongest when securities with few institutional traders are excluded from the analysis (see
Table 2). We find similar results, however, for the analysis that includes all securities with at least one institutional
trader.
23
The first column reveals that the fraction of institutions buying this quarter is positively correlated
with the fraction of institutions buying last quarter for every capitalization quintile (statistically
significant at the 1 percent level for each capitalization quintile). The second and third columns partition
the correlation (or, equivalently, the slope coefficient) for firms within each capitalization quintile into the
portion that results from institutional investors following their own trades and the portion that results from
institutional investors following other institutional investors’ trades. The results reveal statistically
significant (at the 1 percent level) evidence of both herding and following their own trades in every
capitalization quintile. Not surprisingly, given the above discussion, the relative contribution of herding
(following their own trades) is larger (smaller) in large capitalization securities.
As discussed above, both the magnitude and the decomposition of the correlation are influenced
by the number of institutional traders. Thus, although the results in Panel A of Table 5 reveal institutional
demand can better explain subsequent institutional demand (i.e., the correlation is stronger) for stocks in
the top capitalization quintile than stocks in the middle three capitalization quintiles, the results do not
necessarily suggest institutions exhibit greater herding in large capitalization stocks. Therefore, we next
examine measures of following their own trades and following each other’s trades that are independent of
the number of institutions trading. Specifically, we compute the average “following their own trades
contribution” for each security-quarter as the numerator of the first term on the right-hand side of
equation (4) restricted to security k divided by the number of terms used in the first term on the right-hand
side of equation (4) for security k in quarter t:
N k* , t
(D )(
− Raw∆ t Dn , k ,t −1 − Raw∆ t −1 )
Average following their own trades contribution k,t = ∑n =1
n , k ,t
N k*,t
(9)
where N k*,t is the number of managers trading security k in both quarter t-1 and quarter t. Similarly, we
compute the average “herding contribution” for each security-quarter as the numerator of the second term
on the right-hand side of equation (4) restricted to security k divided by the total number of terms used in
the second term on the right-hand side of equation (4) for security k in quarter t:
24
N k ,t N k* ,t −1
(D )(
− Raw∆ t Dm ,k ,t −1 − Raw∆ t −1 )
Average herding contribution k,t = ∑ ∑
n =1 m =1, m ≠ n
n , k ,t
N k ,t N k*,t −1
(10)
where Nk,t is the number of managers trading security k in quarter t and N k*,t −1 is the number of different
managers trading security k in quarter t-1. Equations (9) and (10) measure the average contribution to the
correlation of following their own trades and following each other’s trades, respectively, for each firm.
The number of traders or the cross-sectional standard deviation of the fraction of institutions buying,
therefore, do not affect these measures.
Equations (9) and (10) are estimated for each security-quarter with at least five institutional
traders. We then compute the cross-sectional average across securities in each capitalization quintile each
quarter. Table 5 (Panel B) reports the time-series average of the 58 cross-sectional averages and
associated t-statistics (computed from time-series standard errors) for securities within each capitalization
quintile. In addition, the last row in Panel B reports F-statistics associated with the null hypothesis that
the estimates are equal across capitalization quintiles.36
Consistent with Panel A, the results in the first column of Panel B reveal strong evidence of
following their own trades for every capitalization quintile. Moreover, we document a monotonic inverse
relation between the average following their own trades contribution and capitalization. The F-statistic
reveals that we can reject the hypothesis (at the 1 percent level) that the average following their own
trades contribution is equal across capitalization quintiles. The second column in Panel B reveals the
average herding contribution is positive and statistically differs from zero (at the 1 percent level or better)
for every capitalization quintile. In addition, we document a monotonic inverse relation between the
average herding contribution and capitalization. Institutional investors’ trades this quarter are more likely
to follow other institutions’ trades last quarter in smaller capitalization securities. The F-statistic reveals
that we can reject the hypothesis (at the 1 percent level) that the average herding contribution is equal
across capitalization quintiles.
36
These F-statistics are computed from the 290 cross-sectional averages (5 capitalization quintiles * 58 quarters).
25
The difference between the average contribution from following their own trades and the average
contribution from following others’ trades suggests that a given institutional investor is much more likely
to follow their own lag trade than the lag trade of another institutional investor. The results suggest that
institutional investors accumulate and dispose of positions over time through multiple orders. Although
the results are not well-explained by the empirical evidence on the execution of a given order (Keim and
Madhavan, 1995, 1997; Chan and Lakonishok, 1995), the pattern across firm size is consistent with the
hypothesis that the tendency to follow their own trades is related to trading costs.37,38 These trading costs
include both a liquidity component (e.g., Grossman and Miller, 1988) and an informed trading component
(e.g., Kyle, 1985). Specifically, institutional investors may build a position over time because they fear
that the liquidity premium associated with a single large order or even a series of orders within a quarter
would be too great. Similarly, institutions may build or dispose of positions over time because they fear
that a single order or a series of orders within a quarter would impact the price by the (real or perceived)
informational content of the order flow imbalance. Moreover, both of these costs are likely to be higher
for less liquid smaller capitalization securities.
Taking longer to accumulate a position will reduce the liquidity premium paid by the investor.
The effect on the informed trading cost, however, is less clear. Although building the position over more
than one quarter will force the investor to reveal their interest (through the 13F filing), this may be a
weaker signal than that given by the order flow imbalance if the investor tried to establish or liquidate a
large position within a given quarter. That is, market participants likely infer information from both the
37
Keim and Madhavan (1995, 1997) and Chan and Lakonishok (1995) find evidence that many institutional investor
orders are executed over multiple days. Given the evidence that individual orders are usually fully executed within a
day or two, however, it seems unlikely that executing a given order over several days would explain why
institutional investors tend to follow their own last quarter transactions. That is, the results are unlikely to reflect the
gradual execution of single orders, but more likely reflect institutional investors issuing multiple orders for the same
security over time.
38
It is also possible that institutional investors follow their own trades as a result of habit investing, i.e., time-series
correlation in their net flows and investing (divesting) those flows in (from) their existing portfolio. This
interpretation, however, is inconsistent with the analysis of changes in portfolio weights. Moreover, the habit
investing explanation suggest the pattern should not be related to capitalization.
26
change in position (revealed by the 13F filing) and the demand for immediacy associated with that change
(revealed by the order flow imbalance).39
In sum, the results reveal evidence of herding and following their own trades in every
capitalization quintile. In addition, the results reveal that institutional investors are more likely to follow
their own prior quarter trades in smaller securities consistent with the hypothesis that institutions
following their own lag trades is related to trading costs. Most important, institutional investors are more
likely to herd in smaller capitalization securities. Assuming informational cascades are more likely in
small capitalization securities and investigative herding is more likely in large capitalization securities,
the results suggest institutional herding primarily results from institutions inferring information from each
other’s trades. Given the lack of return reversals, however, our analysis suggests these informational
cascades are rational.
4. Changes Over Time
The number of institutional investors filing 13F reports increases from 556 in December 1983 to 1,274 in
December 1997. It is not clear, however, how increased institutionalization should affect herding. If the
number of institutions doubles by duplicating each of the existing institutional investors and therefore the
fraction of institutions buying does not change, the correlation between the fraction of institutional buyers
this quarter and last quarter does not change. Nonetheless, several factors lead one to suspect herding
may change over time. First, the environment faced by institutional investors is dynamic (Bennett, Sias,
and Starks, 2001), suggesting the incentives to herd may vary over time. Second, the growth in
institutional ownership is not uniform across investor types. Thus, the analysis over time reflects a
decline in the relative importance of bank trust departments and an increase in the relative importance of
independent investment advisors (see Table 1). Third, market characteristics change considerably over
39
For example, a recent Wall Street Journal article (Lucchetti and Angwin, 2002) noted, “Since last March, Janus
has sold about 25% of its 245 million-share position in AOL, and last week, the firm disclosed in regulatory filings
that it had accelerated (emphasis added) its selling during the fourth quarter, disposing of an additional 32 million
shares.”
27
time. For example, markets experienced substantial increases in share turnover and firm-specific risk
over the past two decades (Chordia, Roll and Subrahmanyam, 2001; Campbell, Lettau, Malkiel, and Xu,
2001). Fourth, the number of securities traded by institutional investors increases over time.
We begin the analysis of changes over time by dividing the sample into two 29 quarter periods
(September 1983 to September 1990 and December 1990 to December 1997) and comparing the
distributions of standardized coefficients (and the decomposed coefficients) computed from the cross-
sectional regressions for the entire sample of firms as well as for subsamples of firms within each
capitalization quintile. As before, we limit the sample to securities with at least five institutional traders.
Panel A in Table 6 reports the mean correlations over each sub-period and a t-statistic associated with the
null hypothesis that the time-series mean in the first period equals the time-series mean in the second
period. The results reveal that we can reject the hypothesis that the correlation is equal in both periods for
the sample as a whole and for the middle three capitalization quintiles. The decomposition reveals that
this primarily results from a decline in the portion attributed to institutional investors following their own
lag trades.
Similar to the analysis by firm size, however, comparison of the average coefficient across
periods does not allow us to infer changes in the propensity to herd or follow their own lag trades over
time.40 Thus, to allow direct comparison across time we compute the average following their own trades
contribution and average herding contribution (equations (9) and (10), respectively) for each stock-
quarter. We then compute the cross-sectional average of these contributions each quarter and report (in
Panel B of Table 6) the time-series average of these cross-sectional averages over each sub-period. The
third row reports a t-statistic (from a difference in means test) associated with the null hypothesis that the
time-series mean in the first period equals the time-series mean in the second period.
40
Specifically, comparison of correlations across time is clouded because the number of traders is not constant over
time. Thus, both the correlation and the decomposition will be affected (see footnote 33 for additional detail).
28
The results reveal that institutional investors have become less likely to follow their own and
other institutional investors’ trades over time (i.e., the differences in means across periods for both
contribution measures are statistically significant at the 1 percent level). Evaluation across capitalization
quintiles reveals that the decline in the average following their own trades contribution is statistically
significant in all but the largest capitalization quintile. The decline in the propensity to follow their own
trades is consistent with the hypothesis that increases in market liquidity over time (Chordia, Roll and
Subrahmanyam, 2001) have increased the speed with which institutional investors are willing to enter and
exit positions. Thus, this interpretation is also consistent with the hypothesis that institutions’ tendency to
follow their own trades is related to trading costs (see Section 3.4). Alternatively, the decline in herding
is primarily driven by the largest stocks. We find no evidence of meaningful declines in the average
herding contribution for the bottom three capitalization quintiles. Assuming institutional herding is
primarily driven by informational cascades (see Section 3.4), the decline in the average herding
contribution suggests that institutional investors’ tendency to infer information from each other’s trades
has declined in large capitalization stocks over time. This may have resulted from increased efficiency in
large capitalization stocks. That is, if information is impounded faster in large capitalization stocks in the
more recent period, institutional investors will infer less information from each other’s previous quarter’s
trades.
5. Herding by Investor Type
Previous studies of herding behavior focus on pension funds (Lakonishok, Shleifer, and Vishny, 1992) or
mutual funds (Grinblatt, Titman, and Wermers, 1995; Wermers, 1999). Previous work (Del Guercio,
1996; Bennett, Sias, and Starks, 2001) also documents, however, important differences in the
environments faced by different types of institutional investors (e.g., regulatory requirements, holding
periods, competition) that may, in turn, influence the likelihood these investors herd and whether herding
is only within classifications or if some types of institutions lead other types of institutions. Our data
provide a unique opportunity to examine these issues.
29
Examination of herding by investor type also provides further tests of the reputational and
characteristic herding explanations. If institutional herding is driven by institutions favoring stocks with
the same characteristics and those preferences differ across institutional classes (Del Guercio, 1996;
Bennett, Sias, and Starks, 2001), then institutions should be more likely to follow similar type institutions
than different types of institutions. Moreover, if institutional herding is driven by reputational concerns,
then institutional investors should also be more likely to follow similar type institutions than different
types of institutions, e.g., independent investment advisors should be more likely to follow other
independent investment advisors with whom they directly compete than bank trust departments, insurance
companies, mutual funds, or unclassified investors. In addition, we hypothesize that mutual funds and
independent investment advisors are most likely to experience changes in their net flows resulting from
changes in their reputations.41 Thus, the reputational herding explanation suggests mutual funds and
independent investment advisors should exhibit the strongest tendency to herd.
We begin the analysis by computing the standardized fraction of bank trust departments trading
each security that are buyers, each quarter. We then regress the standardized fraction of bank trust
departments buying this quarter (denoted ∆bk , t ) on the standardized fraction of institutional investors (all
classifications) buying last quarter and limit the sample to securities that have at least five bank trust
departments trading this quarter and at least five institutional investors trading last quarter:
∆bk , t = βt∆k,t-1 + εk,t (11)
We then repeat the analysis replacing the dependent variable with each of the other types of institutional
investors. The time-series average of these 58 standardized regression coefficients (or equivalently,
correlation coefficients) are reported in the first column of Panel A in Table 7 (t-statistics, reported in
41
Unclassified institutional investors (which primarily consist of foundations, endowments, ESOPs and internally
managed pension funds) and insurance companies are less likely to experience changes in net flows resulting from
changes in their reputations. One could argue that bank trust departments could also be subject to such patterns
(e.g., bank trust departments with strong recent performance might experience net inflows while those with poor
recent performance might experience net outflows). Nonetheless, we maintain that given the more conservative
nature of investors utilizing bank trust departments (Del Guercio, 1996; Bennett, Sias, and Starks, 2001), bank trust
departments are less likely to experience these patterns than mutual funds or independent investment advisors.
30
parentheses, are based on time-series standard errors). The results reveal a statistically significant (at the
1 percent level) positive relation between the fraction of traders buying in each classification and the lag
aggregate fraction of traders buying.
As before, the fraction of buyers for a given institutional investor class may be related to the lag
fraction of aggregate institutional buyers for one of two reasons – individual institutional investors
following their own lag trades or institutional investors following the lag trades of other institutional
investors. Moreover, the second term can be further decomposed into two components – following other
traders of the same institutional classification or following traders that are members of a different
institutional class. In the case of bank trust departments, for example, the correlation can be decomposed
as (proof in appendix):
 1 ( )(
 K  Bk ,t Db ,k ,t − Raw∆bt Db,k ,t −1 − Raw∆ t −1 )
b
β = ρ(∆bk ,t , ∆ k ,t −1 )
= ∑∑
 
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b=1
b
Bk ,t N k ,t −1 

 1  K  Bk ,t Bk ,t −1
( )(
Db ,k ,t − Raw∆bt Dm,k ,t −1 − Raw∆ t −1 ) 
+  ∑∑ ∑   (12)
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b =1 m=1,m≠b,m∈B 
b
Bk ,t N k ,t −1

 1 ( )(
 K  Bk ,t N k ,t −1 − Bk ,t −1 Db,k ,t − Raw∆bt Dm,k ,t −1 − Raw∆ t −1 )
+  ∑∑ ∑
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b =1 m=1,m∉B
b
Bk ,t N k ,t −1 

where Db,k,t is a dummy variable that equals one (zero) if bank trust department b is a buyer (seller) of
security k in quarter t and Bk,t is the number of bank trust departments trading stock k in quarter t. Bk,t-1 is
similarly defined for quarter t-1. Nk,t-1 – Bk,t-1 is the number of non-bank trust department institutional
investors trading stock k in quarter t-1. The first term in equation (12) is the portion of the correlation
attributed to bank trust departments following their own lag trades. The second term is the portion of the
correlation attributed to bank trust departments following other bank trust departments. The last term is
the portion of the correlation attributed to bank trust departments following other non-bank trust
department institutions.
31
The second through fifth columns in Panel A of Table 7 report the time series average
decomposition of each of the four components for each of the five types of institutional investors (t-
statistics are computed from time-series standard errors). The results reveal statistically significant
evidence (at the 1 percent level) of following their own trades (second column) and following other
institutional investors trades (third column) for every institutional classification. Moreover, both
components of the second term – following the same type traders (fourth column) and following different
type traders (fifth column) – differ significantly from zero (at the 5 percent level or better) for every
institutional classification.
Comparison of the average coefficient across investor types reveals that lag changes in
institutional ownership has the greatest ability to explain cross-sectional variation in subsequent changes
in ownership by bank trust departments and independent advisors. Again, however, the comparison of
correlations across investor types does not allow us to infer the propensity to herd due to differences in
the number of traders within each classification.42 Thus, to allow direct comparison across investor types,
we again compute the average following their own trades contribution and average herding contribution
for each stock-quarter.43 In addition, we compute the average herding contribution from bank trust
departments following other bank trust departments for each stock-quarter as:
Bk ,t B*k ,t −1
(D )(
− Raw∆bt Dm,k ,t −1 − Raw∆ t −1 )
∑ ∑
b ,k ,t
Ave. same type herding contributionkb,t = (13)
b =1 m =1, m ≠ b , m∈B Bk ,t Bk*,t −1
where Bk*,t −1 is the number of different bank trust departments trading security k in quarter t-1. Similarly,
we compute the average herding contribution from bank trust departments following non-bank trust
departments as:
42
As with the comparison across firm size, comparison of correlation across investor types is clouded because the
number of traders is not constant across investor types (see footnote 33 for additional detail).
43
The average following their own trades contribution is equivalent to equation (9), but limited to bank trust
departments (or whichever group is examined). The average herding contribution is equivalent to equation (10), but
the first summation is from b=1 to the number of bank trust departments trading security k in quarter t. The second
summation still includes all traders (regardless of type) trading security k in period t-1.
32
Bk ,t N k ,t −1 − Bk ,t −1
(D )(
− Raw∆bt Dm,k ,t −1 − Raw∆ t −1 )
∑ ∑
b , k ,t
Ave. dif . type herding contributionkb,t = (14)
b =1 m =1,m∉B Bk ,t (N k ,t −1 − Bk ,t −1 )
where N k ,t −1 − Bk ,t −1 is the number of non-bank trust department institutions trading security k in quarter t-
1. Equations (13) and (14) measure the average contribution to the correlation of following other bank
trust departments and following other types of institutional investors, respectively. Similar to equations
(9) and (10), these measures are not affected by the number of traders nor the cross-sectional standard
deviation of institutions buying.
Limiting the sample to securities with at least five bank trust departments trading this quarter and
five institutional investors trading last quarter, we compute the quarterly cross-sectional average
following their own trades contribution (i.e., analogous to equation (9)), herding contribution (i.e.,
analogous to equation (10)), following same type herding contribution (equation (13)), and following
different type herding contribution (equation (14)). The time-series averages (and associated t-statistic
computed from the time-series standard error) of these cross-sectional averages are reported in the first
row of Panel B in Table 7. The analysis is then repeated for each of the other types of institutional
investors. The last row in Table 7 reports F-statistics associated with the null hypothesis that the
estimates are equal across investor types.44 The last column in Panel B reports the time-series average
difference between the cross-sectional average contribution from following the same type of traders (third
column) and the cross-sectional average contribution from following different types of traders (fourth
column). The t-statistic associated with the value in the last column is based on a paired t-test for
difference in means.
The results in the last row of Table 7 reveal that we can reject the hypothesis (at the 1 percent
level) of equality across investor types for all four “average contribution” measures. Mutual funds,
independent advisors, and unclassified investors more often follow their own lag trades than bank trust
departments or insurance companies. Bank trust departments, unclassified investors, and independent
44
These F-statistics are computed from the 290 cross-sectional averages (5 investor types * 58 quarters).
33
advisors more often herd than mutual funds or insurance companies. The results are consistent with the
hypothesis that the tendency to herd is influenced by the differing environments faced by these investors.
The last column reveals mixed support for the reputational and characteristic herding hypotheses.
Specifically, bank trust departments and insurance companies are more likely to follow similar type
institutions than different types of institutions. That is, the average contribution from following same
type traders is greater than the average contribution from following different type traders for these two
investor types (statistically significant at the 1 percent level). Mutual funds, independent investment
advisors, and unclassified institutional investors, however, exhibit no such pattern inconsistent with
reputational and characteristic herding models. In fact, unclassified institutional investors are more likely
to follow differently classified traders than similarly classified traders (statistically significant at the 5
percent level). The two trader types that should be most concerned about their reputations (mutual funds
and independent advisors) exhibit little evidence of focusing on their most direct competitors’ trades. In
addition, mutual funds exhibit the weakest tendency to herd, inconsistent with the reputational herding
hypothesis.
In sum, the results in Table 7 reveal evidence of following their own and others’ trades for each
institutional classification. Nonetheless, we document substantial variation in herding across investor
types – bank trust departments exhibit the strongest propensity to herd and mutual funds and insurance
companies exhibit the lowest propensity to herd.45 In addition, two of the five trader types are more likely
to follow similarly classified traders than differently classified traders, consistent with reputational and
characteristic herding models. Three of the five trader types, however, fail to exhibit this pattern.
6. Summary
Despite strong theoretical foundations for institutional herding and the seemingly common perception that
professional investors herd, existing studies find little evidence that institutional investors engage in
45
As discussed in Section 2, mutual fund data are at the “fund family” level. Analysis of individual mutual funds
may lead to different conclusions.
34
herding behaviors. This study takes a new approach to testing for institutional herding by directly
examining the cross-sectional temporal dependence in institutional demand over adjacent quarters. We
demonstrate that the fraction of institutions buying this quarter will positively covary with the fraction of
institutions buying last quarter if institutional investors follow their own lag trades or if institutional
investors follow each other into and out of the same securities (herd).
Our results reveal that the fraction of institutions increasing their position in a security this
quarter is strongly related to the fraction increasing their position last quarter. This correlation is equally
attributed to institutions following their own last quarter’s trades and institutions following each other’s
last quarter’s trades. Analysis of changes in portfolio weights reveals institutional investors’ tendency to
follow their own lag trades does not result from time-series correlation in their net flows and investing net
flows in their existing portfolios. Analyses across capitalizations and time suggest that institutional
investors’ tendency to follow their own lag trades is related to trading costs. Specifically, the results are
consistent with the hypothesis that institutional investors accumulate and liquidate positions over time to
minimize the price impact of their trading.
The analysis of changes in portfolio weights reveals that herding does not primarily result from
time-series and cross-sectional correlation in the flows to investors (“habit investing”). Although we find
evidence of institutional momentum trading, such trading accounts for little of the herding. Moreover,
our results suggest that institutional demand is more strongly related to lag institutional demand than lag
returns. In addition, we find no evidence that institutional herding drives prices from fundamental values.
To the contrary, we find that institutional demand is weakly, but positively, related to returns over the
following year suggesting institutional herding reflects that manner in which information is impounded
into securities prices. We also find evidence that herding declines over time (in large stocks) and differs
across investor types. Evaluation across capitalizations suggests the results are most consistent with the
hypothesis that institutional herding primarily results from institutions inferring information from each
other’s trades.
35
References
Banerjee, A., 1992, “A Simple Model of Herd Behavior,” American Economic Review, 88, 724-748.
Barberis, N. and A. Shleifer, 2001, “Style Investing,” working paper, University of Chicago and Harvard
University.
Barboza, D., “Marketplace Amid Market Turmoil, Small Investor is Steadfast,” New York Times (August
13, 1998).
Bennett, J., R. Sias, and L. Starks, 2001, “Greener Pastures and the Impact of Dynamic Institutional
Preferences,” working paper, University of Massachusetts – Boston, Washington State University,
and The University of Texas.
Bikhchandani, S., D. Hirshleifer, and I. Welch, 1992, “A Theory of Fads, Fashion, Custom, and Cultural
Change as Informational Cascades,” Journal of Political Economy, 100, 992-1026.
Cai, F., G. Kaul, and L. Zheng, 2000, “Institutional Trading and Stock Returns,” working paper, University
of Michigan.
Campbell, J., M. Lettau, B. Malkiel, and Y. Xu, 2001, “Have Individual Stocks Become More Volatile?
An Empirical Exploration of Idiosyncratic Risk,” Journal of Finance, 56, 1-43.
Chakravarty, S., 2000, “Stealth Trading: Which Traders’ Trades Move Prices?” forthcoming, Journal of
Financial Economics.
Chan, L. K. C. and J. Lakonishok, 1995, “The Behavior of Stock Prices Around Institutional Trades,”
Journal of Finance, 50, 1147-1174.
Chordia, T., R. Roll, and A. Subrahmanyam, 2001, “Market Liquidity and Trading Activity,” Journal of
Finance, 56, 501-530.
Cunniff, J., “Don’t Blame the Public,” Associated Press wire story (February 15, 2000).
Del Guercio, D., 1996, “The Distorting Effect of the Prudent-man Laws on Institutional Equity
Investment,” Journal of Financial Economics, 40, 31-62.
Dennis, P. and J. Weston, 2000, “Who’s Informed? An Analysis of Stock Ownership and Informed
Trading,” working paper, University of Virginia and Rice University.
Dreman, D., 1979, Contrarian Investment Strategy: The Psychology of Stock Market Success, Random
House, N.Y.
Falkenstein, E. G., 1996, “Preferences for Stock Characteristics as Revealed by Mutual Fund Portfolio
Holdings,” Journal of Finance, 51, 111-135.
Friedman, B. M., 1984, “A Comment: Stock Prices and Social Dynamics,” Brookings Papers on Economic
Activity, 2, 504-508.
Froot, K. A., D. S. Scharfstein, and J. C. Stein, 1992, “Herd on the Street: Informational Inefficiencies in a
Market with Short-term Speculation,” Journal of Finance, 47, 1461-1484.
Gompers, P. and A. Metrick, 2001, “Institutional Investors and Equity Prices,” Quarterly Journal of
Economics, 116, 229-260.
Graham, J. R., 1999, “Herding Among Investment Newsletters: Theory and Evidence,” Journal of Finance,
54, 237-268.
Grinblatt, M. and S. Titman, 1989, “Portfolio Performance Evaluation: Old Issues and New Insights,” Review
of Financial Studies, 2, 393-422.
Grinblatt, M. and S. Titman, 1993, “Performance Measurement without Benchmarks: An Examination of
Mutual Fund Returns,” Journal of Business, 66, 47−68
Grinblatt, M., S. Titman, and R. Wermers, 1995, “Momentum Investment Strategies, Portfolio Performance,
and Herding: A Study of Mutual Fund Behavior,” American Economic Review, 85, 1088-1105.
Grossman, S. and M. Miller, 1988, “Liquidity and Market Structure,” Journal of Finance, 73, 617-633.
Hirshleifer, D., A. Subrahmanyam, and S. Titman, 1994, “Security Analysis and Trading Patterns when
Some Investors Receive Information Before Others,” Journal of Finance, 49, 1665-1698.
Jones, S., D. Lee, and E. Weis, 1999, “Herding and Feedback Trading by Different Types of Institutions and
the Effects on Stock Prices,” working paper, Indiana University – Indianapolis Campus, Kennesaw State
University, and Merrill Lynch.
Keim, D. B. and A. Madhavan, 1995, “Anatomy of the Trading Process: Empirical Evidence on the Behavior
of Institutional Traders,” Journal of Financial Economics, 37, 371-398.
Keim, D. B. and A. Madhavan, 1997, “Transactions Costs and Investment Styles: An Inter-exchange
Analysis of Institutional Equity Trades,” Journal of Financial Economics, 46, 265-292.
Kyle, A. S., 1985, “Continuous Auctions and Insider Trading,” Econometrica, 53, 1315-1335.
Lakonishok, J., A. Shleifer, and R. W. Vishny, 1992, “The Impact of Institutional Trading on Stock Prices,”
Journal of Financial Economics, 32, 23-43.
Lucchetti, A. and J. Angwin, “Janus Hastens to Unload Shares in AOL,” Wall Street Journal (February
20, 2002).
Nofsinger, J. and R. W. Sias, 1999, “Herding and Feedback Trading by Institutional and Individual
Investors,” Journal of Finance, 54, 2263-2295.
Parrino, R., L. Starks, and R. Sias, 2001, “Voting with their Feet: Ownership Changes around the time of
forced CEO Turnover,” working paper, University of Texas and Washington State University.
Pirinsky, C., 2002, “Herding and Contrarian Trading of Institutional Investors,” working paper, Texas
A&M University.
Scharfstein, D. S. and J. C. Stein, 1990, “Herd Behavior and Investment,” American Economic Review, 80,
465-479.
Sias, R., L. Starks, and S. Titman, 2001, “The Price Impact of Institutional Trading,” working paper,
Washington State University and The University of Texas.
Strozinsky, R., 1998, “Inside the Stock Market” web site (www.standford.edu/~stroz/).
Trueman, B., 1994, “Analyst Forecasts and Herding Behavior,” The Review of Financial Studies, 7, 97-
124.
Wermers, R., 1999, “Mutual Fund Trading and the Impact on Stock Prices,” Journal of Finance, 54, 581-
622.
Wermers, R., 2000, “Mutual Fund Performance: An Empirical Decomposition into Stock-Picking Talent,
Style, Transactions Costs, and Expenses,” Journal of Finance, 55, 1655-1695.
Yue, L., “Large Investors Control Market,” Detroit Free Press (August 21, 1998).
Table 1
Descriptive statistics
Average June 1983 Dec. 1986 Sept. 1990 June 1994 Dec. 1997
all periods
Panel A: Number of institutional investors
# Institutions 894 556 703 908 1074 1274
# Banks 189 198 190 200 189 164
# Ins. companies 66 61 63 68 70 68
# Mutual funds 58 46 50 56 61 81
# Indep. advisors 502 173 325 491 681 892
# Unclassified 79 78 75 93 73 69
Panel B: Average number of securities with:
≥ 1 trader 5,505 3,863 4,697 5,040 6,512 7,437
≥ 5 traders 3,783 2,246 3,055 3,262 4,710 5,745
≥ 10 traders 2,851 1,654 2,209 2,432 3,658 4,464
≥ 20 traders 1,966 1,143 1,430 1,637 2,519 3,308
≥ 5 bank traders 4,623 1,563 1,979 1,812 2,867 3,267

≥ 5 ins. co. traders 3,080 500 569 787 1,261 1,898
≥ 5 mut. fund traders 2,863 366 323 641 1,150 2,481
≥ 5 indep. adv. traders 5,033 1,522 2,215 2,616 3,795 4,742
≥ 5 unclassified traders 2,305 320 534 667 882 1,029
Panel C: Average number of securities held by each manager
All institutions 273 224 249 259 300 332
Banks 440 293 391 395 521 625
Ins. companies 329 202 211 295 422 567
Mutual funds 313 182 174 241 368 659
Indep. advisors 211 213 201 203 220 227
Unclassified 245 112 183 245 294 383
Each quarter between March 1983 and December 1997 we calculate the number of institutional investors (overall
and by type), the number of securities traded by at least one, five, 10, or 20 institutional investors, the number of
securities traded by at least five institutional investors of each type of institution, and the cross-sectional average
number of securities held by each institutional investor (overall and by manager type). Panels A, B, and C report the
time-series averages of these figures as well as their values (at approximately) every 15 quarters.
Table 2
Tests for herding
∆k,t = βt∆k,t-1 + εk,t
Average coefficient (β) Partitioned slope coefficient Average R2
Institutions following Institutions following
their own trades others’ trades
Panel A: Securities with at least one institutional trader
Regression 1 – Buyer if increased position
0.1194 0.0617 0.0576 1.93%
(12.67)** (7.33)** (10.12)**
Regression 2 – Buyer if increased return-adjusted portfolio weight
0.1195 0.0615 0.0580 2.38%
(9.26)** (4.57)** (8.61)**
Panel B: Securities with ≥ 5 institutional traders
0.1755 0.0674 0.1081 3.05%
(25.54)** (17.72)** (18.90)**
0.2048 0.0675 0.1373 4.63%
(23.38)** (9.25)** (15.21)**
Panel C: Securities with ≥ 10 institutional traders
0.1727 0.0554 0.1173 3.16%
(30.78)** (23.32)** (23.62)**
0.2476 0.0565 0.1910 6.44%
(33.67)** (14.52)** (26.40)**
Panel D: Securities with ≥ 20 institutional traders
0.1602 0.0414 0.1188 2.73%
(29.86)** (25.42)** (23.96)**
0.2959 0.0446 0.2513 9.01%
(43.64)** (22.23)** (38.59)**
For each security and quarter between March 1983 and December 1997 we calculate the fraction of institutional
traders that increase their position in the security and the fraction that increase their return-adjusted portfolio weight
in the security. An investor is defined as increasing their position if they hold a greater fraction of the firm’s shares
at the end of the quarter than they held at the beginning. An institution is defined as increasing their return-adjusted
portfolio weight if the security’s weight in the investor’s portfolio at the end of the quarter is larger than what it
would be if the manager did not rebalance their beginning-of-quarter portfolio. All data are standardized (i.e.,
rescaled to zero mean, unit variance) each quarter. We then estimate 58 quarterly cross-sectional regressions of each
measure on their lag value. Because there is a single independent variable in each regression and the data are
standardized, these regression coefficients are also the cross-sectional correlations between each measure and its lag
value. The first column reports the time-series average of these 58 correlation coefficients and associated t-statistic
(in parentheses, computed from time-series standard errors). The second and third columns report the portion of the
correlation (see equation (4)) that results from institutional investors following their own lag trades and the portion
that results from institutions following the previous trades of other institutions (herding). Panels B, C, and D report
the averages when limiting the sample to securities with at least five, 10, or 20 institutional traders, respectively. **
indicates statistical significance at the 1 percent level; * indicates statistical significance at the 5 percent level.
Table 3
Standardized regression of institutional demand on lag institutional demand and lag return
∆k,t = β1,t∆k,t-1 + β2,tRk,t-1 + εk,t
Average coefficient associated Average coefficient associated Average R2
with lag institutional demand (β1) with lag return (β2)
Panel A: Securities with at least one institutional trader
0.1133 0.0691 2.59%
(11.84)** (12.10)**
0.1154 0.0545 2.86%
(9.00)** (9.42)**
Panel B: Securities with ≥ 5 institutional traders
0.1624 0.0789 4.19%
(22.25)** (11.90)**
0.1953 0.0774 5.50%
(22.32)** (10.89)**
Panel C: Securities with ≥ 10 institutional traders
0.1590 0.0671 3.90%
(26.26)** (8.98)**
0.2340 0.0801 7.34%
(31.97)** (11.16)**
Panel D: Securities with ≥ 20 institutional traders
0.1490 0.0476 3.23%
(26.90)** (6.54)**
0.2802 0.0658 9.79%
(42.75)** (8.04)**
traders that increase their position in the security and the fraction that increase their return-adjusted portfolio weight
in the security. An investor is defined as increasing their position if they hold a greater fraction of the firm’s shares
at the end of the quarter than they held at the beginning. An institution is defined as increasing their return-adjusted
portfolio weight if the security’s weight in the investor’s portfolio at the end of the quarter is larger than what it
would be if the manager did not rebalance their beginning-of-quarter portfolio. We then estimate 58 quarterly cross-
sectional regressions of each measure on their lag value and lag quarterly return. To allow direct comparison
between the coefficients associated with the independent variables, we standardized (i.e., rescale to zero mean, unit
variance) all data each quarter. The average (standardized) coefficients from the 58 cross-sectional regressions and
associated t-statistics (computed from time-series standard errors) are reported in Panel A. Panels B, C, and D
report the averages when limiting the sample to securities with at least five, 10, or 20 institutional traders,
respectively. ** indicates statistical significance at the 1 percent level; * indicates statistical significance at the 5
percent level.
Table 4
Correlation between institutional demand and returns
Sample Previous Same quarter Following Following 6 Following 12
quarter return return quarter return month return month return
≥ 1 Institutional 0.0744 0.0784 0.0077 0.0060 -0.0009
trader (14.45)** (16.94)** (1.96) (1.33) (-0.25)
≥ 5 Institutional 0.1116 0.1466 0.0221 0.0241 0.0139
traders (17.10)** (22.31)** (4.36)** (4.71)** (3.52)**
≥ 10 Institutional 0.1101 0.1773 0.0291 0.0326 0.0219
traders (14.12)** (25.68)** (5.17)** (5.79)** (4.25)**
≥ 20 Institutional 0.0998 0.2079 0.0257 0.0300 0.0205
traders (12.17)** (28.47)** (3.47)** (4.33)** (3.43)**
traders that increase their position in the security. An investor is defined as increasing their position if they hold a
greater fraction of the firm’s shares at the end of the quarter than they held at the beginning. Each quarter, we then
calculate the cross-sectional correlation between the fraction of institutions buying and returns over the previous
quarter, the same quarter, the following quarter, the following six months, and the following year. The time-series
averages of these cross-sectional correlations and associated t-statistics (computed from time-series standard errors)
are reported for the sample that includes all securities with at least one institutional trader and samples limited to
securities with at least five, 10, or 20 institutional traders. ** indicates statistical significance at the 1 percent level;
* indicates statistical significance at the 5 percent level.
Table 5
Herding and following their own trades by capitalization quintile (≥ 5 institutional traders)
Panel A: Coefficient and slope decomposition
Capitalization Average coefficient Partitioned slope coefficient Average R2
quintile (β) Institutions following Institutions following
Small firms 0.1801 0.0892 0.0909 3.65%
(21.17)** (15.05)** (11.13)**
Quintile 2 0.1469 0.0743 0.0726 2.53%
(18.30)** (16.23)** (10.06)**
Quintile 3 0.1307 0.0618 0.0690 1.95%
(20.22)** (17.43)** (13.31)**
Quintile 4 0.1368 0.0409 0.0959 2.20%
(17.94)** (15.36)** (13.58)**
Large firms 0.1644 0.0269 0.1375 3.07%
(20.60)** (20.17)** (17.98)**
Panel B: Average contributions
Average contribution from following Average contribution from
their own trades following others’ trades
Small firms 0.0439 0.0036
(16.91)** (10.40)**
Quintile 2 0.0372 0.0020
(18.86)** (9.94)**
Quintile 3 0.0340 0.0014
(19.79)** (12.36)**
Quintile 4 0.0285 0.0013
(22.61)** (12.77)**
Large firms 0.0255 0.0001
(27.44)** (17.15)**
F-statistic 16.33 28.97
(p-value) (0.01) (0.01)
traders that increase their position in the security. An investor is defined as increasing their position if they hold a
greater fraction of the firm’s shares at the end of the quarter than they held at the beginning. Firms (with at least
five institutional traders) are then sorted, each quarter, into quintiles based on beginning of quarter capitalization.
Within each capitalization quintile, all data are standardized (i.e., rescaled to zero mean, unit variance) each quarter.
We then estimate 58 quarterly cross-sectional regressions of the fraction of institutional traders buying on the lag
fraction of institutional traders buying for firms within each capitalization quintile. Because there is a single
independent variable in each regression and the data are standardized, these regression coefficients are also the
cross-sectional correlations between the fraction of institutional traders buying this quarter and the fraction of
institutional traders buying last quarter. The first column in Panel A reports the time-series average of these 58
correlation coefficients and associated t-statistic (in parentheses, computed from time-series standard errors). The
second and third columns in Panel A report the portion of the correlation (see equation (4)) that results from
institutional investors following their own lag trades and the portion that results from institutions following the
previous trades of other institutions (herding). We then compute the cross-sectional average contribution from
following their own trades (equation (9)) and following others’ trades (equation (10)) for firms within each
capitalization quintile, each quarter. Time-series averages and associated t-statistics (computed from time-series
standard errors) of these 58 cross-sectional averages are reported in Panel B. The last row in Panel B reports F-
statistics associated with the null hypothesis that the values are equal across capitalization quintiles. ** indicates
statistical significance at the 1 percent level; * indicates statistical significance at the 5 percent level.
Table 6
Analysis over time (≥ 5 institutional traders)
Sample Panel A: Coefficient and slope decomposition
Average coefficient (βt) Partitioned slope coefficient
Institutions following their Institutions following
own trades others’ trades
All firms (8309 to 9009) 0.1982 0.0823 0.1158
All firms (9012 to 9712) 0.1527 0.0524 0.1003
t-statistic 3.65** 4.57** 1.37
Small firms (8309 to 9009) 0.1933 0.1058 0.0857
Small firms (9012 to 9712) 0.1668 0.0708 0.0960
t-statistic 1.58 3.38** -0.63
Quintile 2 (8309 to 9009) 0.1683 0.0932 0.0751
Quintile 2 (9012 to 9712) 0.1255 0.0553 0.0702
t-statistic 2.83** 4.91 0.34
Quintile 3 (8309 to 9009) 0.1456 0.0777 0.0679
Quintile 3 (9012 to 9712) 0.1159 0.0458 0.0700
t-statistic 2.39* 5.55** -0.20
Quintile 4 (8309 to 9009) 0.1682 0.0508 0.1175
Quintile 4 (9012 to 9712) 0.1055 0.0301 0.0744
t-statistic 4.86** 4.20** 3.30**
Large firms (8309 to 9009) 0.1707 0.0269 0.1438
Large firms (9012 to 9712) 0.1581 0.0270 0.1312
t-statistic 0.79 -0.02 0.82
Average contribution from following Average contribution from following
All firms (8309 to 9009) 0.0401 0.0028
All firms (9012 to 9712) 0.0284 0.0021
t-statistic 4.60** 2.26**
Small firms (8309 to 9009) 0.0523 0.0035
Small firms (9012 to 9712) 0.0355 0.0037
t-statistic 3.57** -0.30
Quintile 2 (8309 to 9009) 0.0456 0.0022
Quintile 2 (9012 to 9712) 0.0287 0.0018
t-statistic 5.16** 0.90
Quintile 3 (8309 to 9009) 0.0421 0.0015
Quintile 3 (9012 to 9712) 0.0259 0.0013
t-statistic 5.95** 0.80
Quintile 4 (8309 to 9009) 0.0321 0.0017
Quintile 4 (9012 to 9712) 0.0249 0.0010
t-statistic 3.08** 4.12**
Large firms (8309 to 9009) 0.0242 0.0011
Large firms (9012 to 9712) 0.0269 0.0009
t-statistic -1.43 2.31*
For each security and quarter between March 1983 and December 1997 we calculate the fraction of institutional traders that
increase their position in the security. Firms (with at least five institutional traders) are then sorted, each quarter, into quintiles
based on beginning of quarter capitalization. Within each capitalization quintile, all data are standardized (i.e., rescaled to zero
mean, unit variance) each quarter. We then estimate 58 quarterly cross-sectional regressions of the fraction of institutional
traders buying on the lag fraction of institutional traders buying for firms within each capitalization quintile. We also repeat the
analysis for the sample including all firms with at least five institutional traders. Because there is a single independent variable in
each regression and the data are standardized, these regression coefficients are also the cross-sectional correlations between the
fraction of institutional traders buying this quarter and the fraction of institutional traders buying last quarter. Panel A reports
the time-series average of these correlation coefficients and the portion of the correlation (see equation (4)) that results from
institutional investors following their own lag trades and the portion that results from institutions following the previous trades of
other institutions (herding) over the first 29 quarters in the sample period (8309-9009) and the second 29 quarters in the sample
period (9012-9712). The third row reports t-statistics from a difference in means test of the hypothesis that the mean value in the
first period equals the mean value in the second period. Panel B repeats the analysis for the average contribution from following
their own trades (equation (9)) and following others’ trades (equation (10)) for firms within each capitalization quintile and the
sample including all firms (with at least five institutional traders). ** indicates statistical significance at the 1 percent level; *
indicates statistical significance at the 5 percent level.
Table 7
Analysis by investor type (≥ 5 institutional traders)
Trader type Panel A: Coefficient and slope decomposition
Average Partitioned slope coefficient Average R2
coefficient (β) Following Institutions following others’ trades
their own All Same type Different type
trades
Banks 0.1293 0.0207 0.1086 0.0499 0.0587 1.92%
(19.55)** (8.15)** (18.94)** (15.25)** (15.61)**
Ins. cos. 0.0495 0.0164 0.0331 0.0047 0.0284 0.45%
(8.29)** (5.75)** (4.84)** (4.23)** (4.42)**
Mutual funds 0.0521 0.0134 0.0387 0.0029 0.0358 0.57%
(7.14)** (8.69)** (5.56)** (2.35)* (5.54)**
Ind. advisors 0.1338 0.0587 0.0750 0.0344 0.0407 1.98%
(23.01)** (26.03)** (13.24)** (11.55)** (10.86)**
Unclassified 0.0469 0.0081 0.0388 0.0068 0.0320 0.66%
(5.32)** (4.85)** (4.36)** (3.42)** (3.90)**
Average Average Average Average Average same
contribution from contribution from contribution from contribution from cont. less average
following their following others’ following same following different cont.
own trades trades type traders different type
traders
Banks 0.0200 0.0023 0.0033 0.0018 0.0015
(11.97)** (18.44)** (14.06)** (14.37)** (6.66)**
Ins. cos. 0.0248 0.0006 0.0012 0.0006 0.0006
(13.18)** (5.99)** (5.46)** (5.33)** (3.06)**
Mutual funds 0.0347 0.0006 0.0009 0.0006 0.0002
(22.05)** (5.11)** (2.83)** (4.98)** (0.81)
Ind. advisors 0.0367 0.0016 0.0016 0.0017 -0.0001
(28.13)** (12.51)** (11.40)** (10.43)** (-0.38)
Unclassified 0.0410 0.0017 0.0015 0.0019 -0.0005
(25.10)** (8.99)** (7.48)** (8.57)** (-2.35)*
F-statistic 25.71 34.96 17.49 22.05
(p-value) (0.01) (0.01) (0.01) (0.01)
For each security and quarter between March 1983 and December 1997 we calculate the fraction of institutional traders overall
and by each investor type (bank trust departments, insurance companies, mutual funds, independent investment advisors, and
unclassified institutional investors) that increase their position in the security. We then restrict the sample to securities with at
least five bank trust departments trading this quarter and five institutional investors (of any classification) trading last quarter and
estimate 58 quarterly cross-sectional regressions of the fraction of bank trust department traders buying on the lag fraction of all
institutional traders buying. Because there is a single independent variable in each regression and the data are standardized, these
regression coefficients are also the cross-sectional correlations between the fraction of bank trust department traders buying this
quarter and the fraction of all institutional traders buying last quarter. The first column in Panel A reports the time-series
average of these 58 correlation coefficients and associated t-statistic (in parentheses, computed from time-series standard errors).
The second and third columns report the portion of the correlation (see equation (4)) that results from bank trust departments
following their own lag trades and the portion that results from bank trust departments following the previous trades of other
institutions (herding). The fourth and fifth columns report the portion of the coefficient that results from bank trust departments
following other bank trust department trades and the portion that results from bank trust departments following non-bank trust
department trades, respectively (see equation (12)). The analysis is then repeated for each of the other four institutional
classifications. Panel B repeats the analysis for the average contribution from following their own trades (equation (9)),
following others’ trades (equation (10)), following similarly classified traders (equation (13)), and following differently classified
traders (equation (14)). The last row in Panel B reports F-statistics associated with the null hypothesis that values are equal
across investor classifications. ** indicates statistical significance at the 1 percent level; * indicates statistical significance at the
5 percent level.
Appendix: Proof of equations (4) and (12)
Equation (2) defines the standardized fraction of institutions buying security k in quarter t as the raw
fraction less the cross-sectional average raw fraction in quarter t divided by the cross-sectional standard
deviation of the raw fraction of institutions buying in quarter t:
Raw∆ k ,t − Raw∆ t
∆ k ,t = (A1)
σ( Raw∆ k ,t )
where Raw∆k,t is the fraction of institutional traders that buy security k in quarter t. Because Raw∆k,t is the
sum of the number of institutional investors buying security k this quarter divided by the number of
institutional investors trading the security this quarter, the standardized fraction of institutions buying
security k in quarter t can be written as:
 N k ,t Dn ,k ,t 
 ∑
 n =1 N k ,t 
 − Raw∆ t
∆ k ,t = (A2)
σ( Raw∆ k ,t )
where Nk,t is then number of institutional investors trading security k in quarter t and Dn,k,t is a dummy
variable that equals one (zero) if institutional investor n buys (sells) security k in quarter t. The slope
coefficient in equation (3) is given as:
1 K
cov(∆ k ,t , ∆ k ,t −1 ) ∑
K − 1 k =1
∆ k ,t ∆ k ,t −1
1 K
β=
σ 2 (∆ k ,t −1 )
=
σ 2 (∆ k ,t −1 )
= ∑
K − 1 k =1
∆ k ,t ∆ k ,t −1 (A3)
where K is the number of securities included in the regression. The numerator in equation (A3) is the
cross-sectional covariance between the standardized fraction of institutions buying this quarter and last
quarter. Because the data are standardized, the denominator (σ2(∆k,t-1)) in equation (A3) is one, i.e., the
slope coefficient is the covariance between the standardized data. Moreover, because the cross-sectional
standard deviation of the current quarter’s fraction of institutions buying is also one (i.e., σ(∆k,t) = σ(∆k,t-1) =
1), the covariance is also equal to the correlation. Thus, the slope coefficient from the standardized
regression is equal to the correlation. Furthermore, because standardization is simply a linear re-scaling of
the original data, the correlation between the standardized variables also equals the correlation between the
current and lag quarter’s raw fraction of institutions buying:
β = ρ(∆ k ,t , ∆ k ,t −1 ) = ρ( Raw∆ k ,t , Raw∆ k ,t −1 ) (A4)
Substituting (A2) into the right-hand side of equation (A3) an equivalent expression of this
correlation (or slope coefficient) is:
 1  K  N k ,t Dn , k , t − Raw∆ t  N k ,t −1 Dn , k , t −1 − Raw∆ t −1 
β=
( K − 1) σ ( Raw∆ ) σ ( Raw∆ ) ∑ ∑
 
 k =1  n =1 N k ,t

∑
 n =1 N


(A5)
 k ,t k , t −1   k , t −1 
where Nk,t-1 is the number of institutional investors trading stock k in quarter t-1 and Dn,k,t-1 is a dummy
variable that equals one (zero) if trader n is a buyer (seller) in quarter t-1. Using the distributive property of
the product of sums, equation (A5) can be written as:
 1 ( )(
 K  N k ,t Dn , k , t − Raw∆ t Dn , k , t −1 − Raw∆ t −1 )
β = ρ(∆ k , t , ∆ k , t −1 ) =  ∑∑ 
 ( K − 1)σ( Raw∆ k , t )σ( Raw∆ k , t −1 )  k =1  n =1 N k , t N k , t −1


(A6)
 1
N N
( )(
 K  k ,t k ,t −1 Dn , k , t − Raw∆ t Dm, k ,t −1 − Raw∆ t −1  )
+ ∑∑ ∑
 
 ( K − 1)σ( Raw∆ k , t )σ( Raw∆ k , t −1 )  k =1  n =1 m =1, m ≠ n N k , t N k , t −1


The first term on the right-hand side of equation (A6) is the portion of correlation that results from
institutional investors following themselves into and out of the same securities, e.g., if investor n buys
security k in quarters t and t-1 or sells security k in both quarters, the first term will be positive.46 The
second term on the right hand side of equation (A6) is the portion of the correlation that results from
institutional investors following each other into and out of the same securities, e.g., if investor m buys
(sells) security k in quarter t-1 and investor n buys (sells) security k in quarter t, the second term will be
positive.
46
Although the sum is from n=1 to Nk,t, there are not necessarily Nk,t “terms.” If an institutional investor trades only in
quarter t or only in quarter t-1, the investor will not be included in the “following their own trades” term. The trader
will be included in the cross-correlation (herding) term. Assume, for example, there are three institutional investors
that trade security k in both quarters. In this case, there will be nine total terms (3*3) – three “following their own
trades” terms and six herding terms. Alternatively, assume that three investors trade in quarter t-1. One of these
investors does not trade in quarter t. A new trader, however, trades in quarter t, such that there is still a total of three
traders in both quarters. In this case there will be two following their own trades terms and seven herding terms
(again for nine total terms).
Next, define the standardized fraction of bank trust departments buying security k in quarter t as the
raw fraction of bank trust departments less the cross-sectional average raw fraction of bank trust
departments buying in quarter t divided by the cross-sectional standard deviation of the raw fraction of
bank trust departments buying in quarter t:
Raw∆bk ,t − Raw∆bt
∆bk ,t = (A7)
σ( Raw∆bk ,t )
where Raw∆bk ,t is the fraction of bank trust department traders that buy security k in quarter t. Because
Raw∆bk ,t is the sum of the number of bank trust departments buying security k this quarter divided by the
number of bank trust departments trading the security this quarter, the standardized fraction of bank trust
departments buying security k in quarter t can be written as:
 Bk ,t Db ,k ,t 
 ∑ B
 b =1 k ,t 
 − Raw∆bt
∆bk ,t = (A8)
σ( Raw∆bk ,t )
where Bk,t is then number of bank trust departments trading security k in quarter t and Db,k,t is a dummy
variable that equals one (zero) if bank trust department b buys (sells) security k in quarter t. The slope
coefficient from the regression of the standardized fraction of bank trust departments buying this quarter on
the standardized fraction of institutional investors buying last quarter is given as:
cov(∆bk ,t , ∆ k ,t -1 )
1
∑
K - 1 k =1
∆bk ,t ∆ k ,t -1 K
b
β =
σ 2 (∆ k ,t -1 )
=
σ 2 (∆ k ,t -1 )
=
1
∑
K - 1 k =1
∆bk ,t ∆ k ,t -1 (A9)
where K is the number of securities included in the regression. The numerator in equation (A9) is the
cross-sectional covariance between the standardized fraction of bank trust departments buying this quarter
and the standardized fraction of institutional investors buying last quarter. Because the data are
standardized, the denominator (σ2(∆k,t-1)) in equation (A3) is one, i.e., as before, the slope coefficient is the
covariance between the standardized data. Moreover, because the cross-sectional standard deviation of the
current quarter’s fraction of bank trust departments buying is also one (i.e., σ(∆ bk ,t ) = σ(∆k,t-1) = 1), the
covariance is also equal to the correlation between the two variables.
Substituting (A2) and (A8) into the right-hand side of equation (A9) an equivalent expression of
this correlation (or slope coefficient) is:
 1  K  Bk ,t Db ,k ,t − Raw∆bt  N k ,t −1 D


n , k ,t −1 − Raw∆ t −1  
b
β =
( K − 1) σ ( Raw ∆b
) σ ( Raw ∆ )
 ∑ ∑
 k =1  n =1 Bk ,t  ∑ N k ,t −1 
(A10)
 k ,t k ,t −1   n =1 
Using the distributive property of the product of sums, equation (A10) can be written as:

β b = ρ(∆bk ,t , ∆ k ,t −1 ) = 
1
∑∑
 
( )(
 K  Bk ,t Db , k ,t − Raw∆bt Db ,k ,t −1 − Raw∆ t −1 

)
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b =1 
b
Bk ,t N k ,t −1
 (A11)
+
 1
  ∑∑ ∑
N
( b
)(
 K  k ,t k ,t −1 Db ,k ,t − Raw∆ t Dm ,k ,t −1 − Raw∆ t −1 
B

)
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b =1 m =1, m ≠b 
b
Bk ,t N k ,t −1

The first term on the right-hand side of equation (A11) is the portion of correlation that results from bank
trust departments following themselves into and out of the same securities. The second term on the right
hand side of equation (A11) is the portion of the correlation that results from bank trust departments
following other institutional investors (both other bank trust departments and other non-bank trust
department institutional investors) into and out of the same securities.
The second term in equation (A11) can be further partitioned by investor type. Specifically, the
term can be divided into two components – bank trust departments following other bank trust departments
and bank trust departments following other types of institutional investors. Again using the distributive
property of the product of sums, equation (A11) can be written as:

β b = ρ(∆bk ,t , ∆ k ,t −1 ) = 
1
∑∑
 
( )(
 K  Bk ,t Db,k ,t − Raw∆bt Db,k ,t −1 − Raw∆ t −1 )
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b =1 
b
Bk ,t N k ,t −1


+
1  K  Bk ,t
 ∑∑ ∑ 
Bk ,t −1
( )(
Db,k ,t − Raw∆bt Dm,k ,t −1 − Raw∆ t −1 ) 
 (A12)
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b =1 m=1,m ≠b ,m∈B 
b
Bk ,t N k ,t −1


+
1
 ∑∑ ∑
( )(
 K  Bk ,t N k ,t −1 − Bk ,t −1 Db,k ,t − Raw∆bt Dm,k ,t −1 − Raw∆ t −1 )
 ( K − 1)σ( Raw∆ k ,t )σ( Raw∆ k ,t −1 )  k =1  b =1 m =1,m∉B 
b
Bk ,t N k ,t −1

The first term on the right-hand side of equation (A12) is as define in equation (A11). The second term in
equation (A12) is the portion of correlation that results from bank trust departments following other bank
trust departments into and out of the same securities, e.g., if bank trust department b buys (sells) security k
in quarter t and investor m is a different bank trust department (i.e., m∈B, m≠b) that buys (sells) security k
in quarter t-1, the second term will be positive. Similarly, the third term in equation (A12) is the portion of
correlation that results from bank trust departments following other types of institutional investors into and
out of the same securities, e.g., if bank trust department b buys (sells) security k in quarter t and
institutional investor m is not a bank trust department (i.e., m∉B) that buys (sells) security k in quarter t-1,
the third term will be positive.

SSRN Id307440

Uploaded by

Copyright:

Available Formats

SSRN Id307440

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id307440

Uploaded by

Copyright:

Available Formats

Institutional Herding

February 24, 2002

experience net changes in institutional ownership simply by chance.

institutional investors following other institutional investors’ trades.

security selection decisions and asset prices.

the Lakonishok, Shleifer, and Vishny “herding measure.”6

between institutional demand this quarter and last quarter.

To determine if habit investing explains institutional herding, we focus on the cross-sectional

To differentiate informational cascades from investigative herding, we examine herding by

inferring information from each other’s trades (informational cascades).

decline in herding in the largest capitalization securities.

classifications. Although institutional investors in every classification exhibit statistically significant

than differently classified institutional investors.

investor types, respectively. A summary is presented in the final section.

institutions buying” security k during quarter t:

reports, primarily driven by the growth in independent investment advisors.

[Insert Table 1 about here]

2. Tests for Institutional Herding

k in quarter t (denoted ∆k,t) as:

2.1 Does institutional demand predict institutional demand?

fraction of institutions buying security k the previous quarter (∆k,t-1):14

∆k,t = βt∆k,t-1 + εk,t (3)

this quarter and last quarter averages 11.94 percent.

[Insert Table 2 about here]

2.2 Following their own or others’ trades?

securities over adjacent quarters.17

coefficient in equation (3) can be written as (proof in appendix):

one (zero) if trader m (m≠n) is a buyer (seller) of security k in quarter t-1.

investors tend to reverse their last quarter’s transactions.

investors’ transactions last quarter, this term will be zero.

2.3 Additional tests – Herding and the number of institutional traders

substantial change in the coefficients.

2.4 Comparison with previous work

Contrary to previous studies, we document strong evidence of institutional herding. It is possible,

different methodologies. To disentangle differences in methodology from differences in samples, we

Vishny herding measure for security k in quarter t is defined as:

H k ,t = Raw∆ k ,t − Raw∆ t − AFk ,t (5)

and other institutional investors’ trades the next period.

3. Why Do Institutional Investors Herd?

differences across capitalizations in an attempt to discriminate between these models.

3.1 Habit investing

negative) the same securities over adjacent quarters.24

adjusted beginning-of-quarter portfolio weight:

will not be classified as a buyer or a seller.

portfolio weights in quarter t is defined as:

# Inst. inc. return - adjusted weight k,t

with an even greater number of traders (Panels C and D).

correlation in their net flows.

3.2 Momentum trading

fraction of institutions buying and lag quarterly standardized return:29

∆k,t = β1,t∆k,t-1 + β2,tRk,t-1 + εk,t (8)

[Insert Table 3 about here]

rejected at the 1 percent level (t-statistic=3.61).

related to lag returns as well as their own lag values.

strongly related to their own lag demand than lag returns.30

3.3 Institutional demand and subsequent returns

security characteristics, or fads.

(computed from time-series standard errors) are reported in Table 4.

[Insert Table 4 about here]