Академический Документы
Профессиональный Документы
Культура Документы
Alok Kumar
Cornell University
Department of Economics
May 2003
(First Version: September 2002)
Comments Welcomed
Please address all correspondence to Alok Kumar, Department of Economics, Uris Hall, Cornell University, Ithaca, NY 14853, email: ak272@cornell.edu OR Charles M.C. Lee, Johnson Graduate School of
Management, Cornell University, Ithaca, NY 14853, email: CL86@cornell.edu. We would like to thank Xiang Cai, Richard Frankel, William Goetzmann, Dong Hong, Zoran Ivkovic, Robert Masson, Victor McGee,
Ted ODonoghue, Vicente Pons, Mark Seasholes, Kent Womack, and seminar participants at MIT Sloan
School for several helpful discussions and valuable comments. We thank Itamar Simonson for making the
investor data available to us and Terrance Odean for answering numerous questions about the investor
database. Finally, we would like to thank Mark Hulbert and Andrew Metrick for providing the data on
investment newsletter recommendations. All remaining errors and omissions are our own.
Introduction
The traditional view of returns comovement relies on the present value model of security
valuation where the current price of a stock reects the present discounted value of a stream
of future earnings. According to this view, any comovement in stock prices (or returns)
reects either the common movements in earnings (i.e., stock fundamentals) or changes in the
discount rates which in turn depend upon various macro-economic variables. Demand shocks
or shifts in investor sentiment have no role to play in this traditional view of comovement
where such shocks can be easily oset by the actions of arbitrageurs.
An alternative theory of returns comovement posits that stock prices are established
through a dynamic interplay between noise traders and rational arbitrageurs (e.g., Shiller
(1984), Shleifer and Summers (1999)). According to this view, in addition to innovations
in fundamentals and macro-economic variables, other factors such as the correlated trading
activities of noise traders, can induce comovement in stock returns. Arbitrage forces may
not be able to fully absorb these correlated demand shocks due to a variety of factors,1 but
most notably due to a non-fundamental risk introduced by the correlated actions of these
investors themselves (Black 1986, DeLong, Shleifer, Summers, and Waldmann 1990).
In this study, we test a particular form of the noise trader model in which individual (or
retail) investor sentiment can aect stock returns. For example, DeLong, Shleifer, Summers,
and Waldmann (1990) (hereafter DSSW) conjecture that because the shares of closed-end
funds are held primarily by individual investors, the discounts on these funds capture the
dierential sentiment of these investors. Consistent with this view, Lee, Shleifer, and Thaler
(1991) (hereafter LST) nd that the returns of stocks with lower institutional ownership
and lower market capitalization are positively correlated with changes in closed-end fund
discounts. In the same spirit, Lakonishok and Maberly (1990) and Abraham and Ikenberry
(1994) nd that the buy-sell imbalance in odd-lot trades exhibit weekly uctuations consistent with the Day-of-the-week eect.
See Shleifer and Vishny (1997) for a theoretical exposition of this argument and Rashes (2001) for an
example where arbitrage fails to discipline the market even in a very simple setting. In the same spirit,
Lamont and Thaler (2002) present the case of equity carve-outs where high transaction costs associated with
short sales significantly diminish the power of arbitrage.
1
A key point of contention with these studies is whether closed-end fund discounts, or
odd-lot trades, are appropriate proxies for the sentiment of individual investors. The LST
ndings, in particular, proved controversial and has spawned a number of follow-up studies
(e.g., DeLong and Shleifer (1991), Chen, Kan, and Miller (1993), Chopra, Lee, Shleifer, and
Thaler (1993), Elton, Gruber, and Busse (1998)). Although closed-end funds are mainly held
by individuals, questions remain as to what movements in the discounts represent. Similar
inference issues apply to the use of odd-lots trades. In short, so long as the proxy used
to measure small investor sentiment remains controversial, questions as to whether small
investor sentiment aects asset prices will remain.
In this study, we use a large set of investor trading data from a major discount brokerage
house to construct a direct measure of individual investor sentiment. Specically, we compute
the buy-sell imbalance (BSI) of trades initiated by individual investors across dierent stock
portfolios. We then compute a residual BSI by regressing each portfolio BSI on the overall
market returns. Using this residual measure of individual investor trading activity, we explore
the relation between individual investor sentiment and stock returns.
Our investigation is focused on three issues central to the noise trader framework. First,
we evaluate the extent to which the buy-sell activities of individual investors are correlated
across dierent stock portfolios. The noise trader model asserts that the aggregate behavior
of noise traders is aected by waves of common sentiment. In other words, the BSI of noise
traders should be correlated across non-overlapping stock portfolios. In the absence of this
type of systematic behavior, it is unlikely that noise trader sentiment can aect stock returns.
Our tests document a strong common component in trading activities of individual investors that is orthogonal to the overall market movement. Using the residual BSI measure,
we nd that monthly pair-wise correlations for this variable between non-overlapping portfolios average around 0.44. This evidence shows the buy-sell activities of individual investors
are correlated across stocks, even after controlling for overall market movements when they
buy (or sell) stocks in one portfolio, they also tend to do so in other portfolios.
Second, we examine the determinants of individual investor sentiment, i.e., what factors
seem to aect the proclivity of individual investors to buy or sell stocks? Noise trader models
3
are generally silent with respect to the source of investor sentiment. Our tests are aimed at
providing additional insights on this issue. In addition to factors suggested by the empirical
literature, i.e., returns to mimicking portfolios based on rm size (SMB), book-to-market
(HML), price momentum (UMD), and macro-economic variables (unexpected ination (UI),
monthly growth in industrial production (MP), value spread (VS), and term spread (TS)), we
also consider the eect of monthly changes in the expert advice obtained from investment
newsletters2 (NLBSI).
We nd that changes in the aggregate sentiment of investment newsletters (residual
NLBSI) has a strong inuence on the residual BSI measure these two measures have a
contemporaneous correlation of 0.50. Residual BSI is positively but weakly correlated with
the SMB and the HML factors (the correlations are 0.001 and 0.054 respectively) and it
is negatively correlated with the UMD factor (the correlation is 0.389). Examining the
correlations between residual BSI and the macro-economic variables, we nd that BSI is
negatively correlated with UI and MP (the correlations are 0.365 and 0.047 respectively)
but positively correlated with VS and TS (the correlations are 0.345 and 0.152 respectively).
In sum, these results suggest that aggregate individual investor sentiment is weakly related to
the standard risk factors, moderately related to the macro-economic variables, and strongly
related to the aggregate sentiment of investment newsletters.
Third, we examine the extent to which individual investor sentiment has incremental
explanatory power for cross-sectional as well as seasonal patterns in stock returns. Noise
trader models suggest that the eect of noise trading will be greatest in those stocks that they
dominate, and in stocks with relatively high arbitrage costs. To examine these hypotheses,
we form stock portfolios by rm size, stock prices, B/M ratios, and institutional ownership,
and examine the incremental eect of residual BSI on the returns of these portfolios, after
controlling for market excess returns (RMRF), SMB, HML, and UMD.
We nd that residual BSI has incremental explanatory power for small stocks, value
2 Hulbert
Financial Digest (HFD) tracks the stock recommendations of a large number of investment
ommendations where a newsletter either explicitly recommends a portfolio of stocks (a model portfolio) or
provides a ranked list of stocks that can be used to construct a portfolio.
more than one model portfolio.
stocks, stocks with low institutional ownership, and stocks with lower prices. The direction
of the relation indicates that when individual investors are relatively bullish (i.e., when BSI
is more positive), the stocks in these portfolios enjoy higher excess returns. The magnitude
of their inuence is aected by factors associated with arbitrage costs, i.e., rm size, stock
price, and the percentage of institutional ownership. These ndings are largely consistent
with the predictions of a noise trader model in which individual (retail) investor sentiment
aects returns, particularly among stocks in which these investors are concentrated.
To further explore the role of individual investor sentiment in the return generating
process, we consider the relation between residual BSI and two robust seasonal patterns in
stock returns the January eect and the Day-of-the-Week eect. We nd strong evidence
that individual investor sentiment partially explains the January eect.3 However, contrary
to Lakonishok and Maberly (1990) and Abraham and Ikenberry (1994), we nd no signicant
relation between our measure of individual investor sentiment and weekly patterns in stock
returns.
In sum, our study provides a number of empirical ndings consistent with the predictions
of noise trader models. First, we show that the buy-sell activities of individual investors are
correlated across non-overlapping stock portfolios. We nd a strong degree of correlation even
after removing the eect of overall market movements. Second, we show that these buy-sell
activities are weakly related to macro-economic variables and excess returns to small rms
and value rms. Furthermore, investors trading activities are strongly inuenced by changes
in the expert advice conveyed in investment newsletters. Finally, we show that individual
investor sentiment has incremental eect in explaining returns for small rms, low priced
rms, value rms, and rms with low institutional ownership. We also nd some support for
the view that the January eect is related to individual investors propensity to buy small
stocks (sell large stocks) in January.
Although our paper is not the rst to examine the role of individual investor sentiment
3 The
January effect refers to the unusually high stock returns in January, especially for smaller firms, and
the Day-of-the-Week effect refers to the generally lower returns observed on Mondays. Ritter (1988) suggests
a link between the January effect and individual investors, Lakonishok and Maberly (1990) and Abraham and
Ikenberry (1994) suggest the negative sentiment of small investors might account for the Day-of-the-Week
effect.
in explaining stock returns, we believe our results present the most convincing evidence so
far of small investor sentiment-induced comovement in stock returns. By using a direct,
trading-based measure of individual investor sentiment, we avoid the diculties that arise
from establishing the appropriateness of sentiment proxies such as changes in closed-end
fund discounts or odd-lot trades.
The rest of the paper is organized as follows: in the following section, we review the
literature on sentiment-induced stock price movements. A brief description of the datasets
used in the study is provided in Section III. In Section IV, we dene an investor sentiment
measure and establish that there exists a systematic component in the trading activities of
individual investors. In Section V, we provide insights into the determinants of individual
investor sentiment. Estimation of multi-factor time-series models that include the investor
sentiment measure are carried out in Section VI. In Section VII, we examine the role of
investor sentiment in explaining the seasonal variations in stock returns. Finally, in Section
VIII we summarize our ndings and discuss future research.
II
Related Research
In sum, various studies have attempted to derive indirect measures of individual investor
sentiment without direct data on investor trading activities. Those studies that use direct
trading data have focused on documenting individual psychological biases and do not address
the asset-pricing implications of individual investors systematic trading activities, which is
the main focus of our study.
III
Several datasets are used in this study. A brief description of each one follows.
To gauge how representative our individual investor sample is of the overall population
of individual investors in the US, we compare the stock holdings of investors in our sample
with those reported by the Census Bureau4 (Survey of Income and Program Participation
(SIPP), 1995) and the Federal Reserve5 (Survey of Consumer Finances (SCF), 1992, 1995).
According to the 1992 SCF, a typical household held $8,700 in stocks (median was $16,900).
The stock ownership declined marginally in the 1995 SCF where a typical US household
held $8,000 in stocks (median was $15,300). In another independent survey conducted by
the Census Bureau, the real median value of stocks and mutual funds held by households
increased from $7,331 in 1993 to $9,000 in 1995. The median portfolio size of an investor in
our sample ($13,869) matches quite well with the numbers reported in SCF 1992, SCF 1995,
and SIPP 1995. Overall, these comparisons suggest that our sample probably captures a
reasonable cross-section of U.S. households that invest in stocks.
http://www.census.gov/hhes/www/wealth/1995/wealth95.html.
5 The report is available at http://www.federalreserve.gov/pubs/oss/oss2/95/scf95home.html.
Also see Kennickell, Starr-McCluer, and Sunden (1997).
6A
detailed description of the institutional ownership data can be found in Gompers and Metrick (2001).
average, 19% of the stocks have zero institutional ownership and approximately 30% have low
(i.e., less than 5%) institutional holding. As expected, the level of institutional ownership
is even lower for small stocks. Panel B reports results for stocks in the smallest decile
using beginning-of-year NYSE size thresholds. These stocks are important because (i) they
represent a disproportionately large percentage of the stocks held by individual investors,
and (ii) individual investors are likely to play a larger role in the pricing of these stocks.
Panel B results show that approximately 25% of the decile 1 stocks have zero institutional
ownership, and approximately 42% have low (less than 5%) institutional holding. Overall,
these results show that individual investors are likely to be more highly concentrated among
the decile 1 stocks.
thank Andrew Metrick for providing a clean version of the newsletter stock recommendations dataset.
10
newsletter analysts in our sample and during the 1991-96 period, these analysts made 267, 278
recommendations on 4, 548 stocks. This includes recommendations on a considerable number
of stocks in all size and B/M based stock categories (see Panels B and C). The set contains
132, 872 negative and 134, 406 positive recommendations. In any given year, the group of
newsletters recommend approximately 2, 000 stocks and for each stock, as a group, they
make an average of 19 recommendations per year. A typical newsletter analyst made 994
recommendations during the 1991-96 sample period (median is 147). A quarter of them
made less than 38 recommendations, 5% of them made more than 3, 600 recommendations,
and the most active newsletter analyst made 35, 539 recommendations.8
For each stock in our sample, we obtain the daily and the monthly security prices and returns data from CRSP as well as the market capitalization and book value of common equity
data from COMPUSTAT. The monthly time-series of several macro-economic variables (unexpected ination, monthly growth in industrial production, value spread, and term spread)
are obtained from the economic data library of the Federal Reserve Bank of St. Louis.9
Finally, we obtain several datasets from Ken Frenchs data library.10 Specically, we use the
monthly time-series of the 3 Fama-French factors and the momentum factor. We also obtain
the NYSE size and B/M break-points for each month, which are used to construct size and
B/M portfolios.
IV
detailed description of the newsletter dataset is available in Metrick (1999). This study investigates
the stock-selection ability of newsletter analysts and finds that, at an aggregate level, newsletters exhibit
very weak stock-picking and market-timing abilities.
9 The
http://research.stlouisfed.org/fred2/
11
VBijt
Dt
j =1 VBijt +
Dt
j =1
Dt
j =1
Dt
j =1
VSijt
VSijt
(1)
VBijt
VSijt
(2)
where Np is the number of stocks in portfolio p. The portfolio BSI in any given month is
computed using only the stocks within the portfolio in which the investors in our sample
trade during that month.11
Most of the analyses in this paper have been carried out using monthly aggregated measures of investor trading activities. Given the relatively sparse trading activities of individual
investors in our sample (see Table II), a daily buy-sell imbalance (BSI) measure for a particular stock is likely to be unreliable. On any given day, on average, less than 10% of the
stocks in a portfolio are likely to be traded by the investors in our sample. Therefore, to
ensure sucient trading, we compute a BSI time-series for portfolios of at least 250 stocks.
This ensures, on average, approximately 20-25 stocks in each portfolio will be traded daily.
Using this approach, and monthly aggregation, we nd that approximately 50-60% of the
stocks in a typical portfolio are traded each month by the investors in our sample.
measure of portfolio BSI gives the same weight to each stock in computing the portfolio sentiment
measure. An alternative approach is to first compute the aggregate dollar volume in-flow (AVB) and aggregate dollar volume out-flow (AVS) for all the stocks in a portfolio, and define BSI as
AV BAV S .
AV B +AV S
However,
under this alternative approach, a single stock can strongly influence the portfolio BSI in a particular month,
especially around information events such as earnings announcements and stock recommendation changes
when the stock trading volume is unusually high. Our measure avoids such a bias.
12
form non-overlapping portfolios and construct a monthly buy-sell imbalance time-series for
each portfolio. We then compute correlations between pairs of the buy-sell imbalance (BSI)
indices of these non-overlapping portfolios and generate an empirical distribution of BSI
correlation. Specically, we form non-overlapping portfolio pairs where stocks are chosen
randomly from the set of stocks in our sample. 1000 pairs of non-overlapping 250-stock
portfolios are formed and a BSI time-series for each of the random portfolio is obtained.
To remove the common dependence of BSI on the market factor, we perform the following
regression:
BSIpt = b0 + b1RMRFt + pt.
(3)
Here, BSIpt is the buy-sell imbalance index for portfolio p in month t, RMRFt is the market
return in excess of the riskfree rate in month t, and pt is the residual BSI for portfolio p in
month t. We compute the correlations among the residual BSI time-series over 71 months
(January 1991 to November 1996) for each of the portfolio pairs.
Figure 1 shows an empirical distribution of these pairwise correlations. The average
residual BSI correlation is 0.44 (median is 0.45). The average residual BSI correlation decreases monotonically with portfolio size. For instance, for 100-stock portfolios, the average
residual BSI correlation is 0.34 while for 50-stock portfolios, this measure is 0.23. These results provide clear evidence of a systematic component in investor trading activities, which
is uncorrelated with the movements of the market index. In additional (untabulated) tests,
we nd that these correlations are slightly higher when stocks are chosen from similar size
groups. For example, when stocks are chosen from size quintiles 1 and 5, the average residual
BSI correlations are 0.49 and 0.48 respectively.
We carry out Monte Carlo simulations to obtain an estimate of average residual BSI
correlation in the absence of a systematic component in the trading activities of our investors
(i.e., a benchmark average residual BSI correlation). Specically, we generate a BSI matrix
where, for each stock, we keep the frequency and timing of trades xed but we randomly
assign a BSI in each month where BSI (1, 1). Using this simulated BSI matrix and
following the procedure described earlier, we generate an empirical distribution of residual
BSI correlations by computing the correlations between 1000 pairs of non-overlapping 25013
stock portfolios. This entire process is repeated 500 times and a distribution of the average
residual BSI correlation is obtained.12
Figure 2 shows an example of the residual BSI correlation distribution obtained in one of
our randomization tests. Clearly, the average residual BSI correlation of 0.01 is signicantly
lower than the observed average residual BSI correlation of 0.44. In fact, we nd that
the average BSI correlation is lower than 0.44 in each of the 500 repetitions. The evidence
indicates that the observed strong BSI correlations is unlikely to have occured by pure chance
(p-value < 0.002).
Overall, our results show that correlations between BSI indices are strongly positive over
non-overlapping stock portfolios. Our randomization tests suggest these strong correlations
could not have occured by chance alone. Taken together, the evidence points to the existence
of a systematic (or market-wide) component in the trading activities of individual investors.
Specically, when individual investors buy (sell) one basket of stocks, they are likely to
simultaneously buy (sell) other stock baskets.
What factors might induce systematic trading among individual investors? The noise trader
literature oers little guidance on this issue. For example, Fischer Black contrasts noise with
information, and denes noise trading as trading on noise as if it were information (Black
1986, pp. 529). Shiller (1984) suggests that common sentiments arise when investors trade
on pseudo-signals, such as price and volume patterns, popular models, or the forecasts of
Wall Street gurus.13 However, most past studies have attempted to document the existence
of investor sentiments rather than explore their origins.14 Our analysis begins with a credible
12 For robustness, we also carried out a similar test where instead of choosing BSI randomly from (1, 1),
we resample BSI from the observed BSI matrix. In addition, we repeated our tests using 50, 100, and 500
stock portfolios. The results are qualitatively similar in all these cases.
13 Pseudo-signals are signals that are non-informative in estimating a firms fundamental value, but that
may be nevertheless persuasive in their own right.
14 There are a few notable exceptions. In a recent study, Barber, Odean, and Zhu (2003) examine whether
psychological biases lead to correlated trading among individual investors. They find that the correlated
buying decisions are driven by attention (Barber and Odean 2001) and extrapolation of past trends (Kahneman and Tversky 1973) while their correlated selling activities result from investors reluctance to realize
losses (Shefrin and Statman 1985). In another recent study, using a Chinese dataset, Feng and Seasholes
14
measure of sentiment, and provides some insights on how this sentiment is correlated with
several other commonly used variables in empirical nance.
As a starting point, we investigate the relations between our measure of investor sentiment
and several empirically-inspired risk factors that appear in the literature. Specically, Fama
and French (1992) nd that rms market capitalization (size) and book-to-market ratio
(B/M) explain a signicant portion of the cross-sectional variation in stock returns. In a
related study, Fama and French (1993) form mimicking portfolios based on size (SMB) and
B/M (HML), and show that when the standard market model is augmented by these two
variables, a number of pricing anomalies disappear. In this study, we examine the relation
between BSI, SMB, and HML. We also include a price momentum factor (UMD), as suggested
by Jegadeesh and Titman (1993) and Carhart (1997).
In addition to these empirically inspired risk factors, we also examine whether individual
investor sentiment is inuenced by innovations in macro-economic variables (Chen, Roll, and
Ross 1986). Under the present value model of security valuation, the current price of a stock
reects the present discounted value of a stream of expected future cash ows. If investors
adopt a present value model (or some variant) to formulate their trading decisions, variables that inuence expected future cash ows or the discount rate may inuence individual
investor sentiment.
Following Chen, Roll, and Ross (1986) and Ferson and Schadt (1996), we consider the
following four macro-economic variables as potential determinants of investor sentiment: (i)
UI: unexpected ination where the average of 12 most recent ination realizations is used to
estimate the expected level of ination, (ii) MP: monthly growth in industrial production,
(iii) TS: the term-spread (a measure of the term structure) which is the dierence between the
yield of a constant-maturity 10-year Treasury bond and the yield of a 3-month Treasury bill,
and (iv) VS: the value-spread (a measure of risk premium) which represents the dierence
between the yields of Moodys BAA-rated corporate bond and AAA-rated corporate bond
(or long term government bond).
Finally, we examine the relation between investor sentiment and the sentiment of a group
(2002) find that the trading activities of investors that live within a certain geographic region are strongly
correlated.
15
of investment newsletter writers (experts). It is possible that individual investors seek expert advice from professional information providers such as investment newsletter writers.
It is also possible that individual investors and newsletter writers are inuenced by common
waves of investor sentiment. In either case, we would nd a positive correlation between BSI
and the opinion of these newsletter writers.
To examine the relation between investor sentiment and the sentiment of investment
newsletters, we dene a BSI-like measure for capturing the sentiment (or the degree of
bullishness) of newsletter analysts. Specically, we code an increase (decrease) in portfolio
weight for a stock, or an addition (deletion) of a stock into a portfolio, as a positive (negative)
recommendation. We compute a newsletter sentiment measure for each stock i in month
t (NLBSIit ) as:
NLBSIit =
POSijt
Dt
j =1 POSijt +
Dt
j =1
Dt
j =1
Dt
j =1
NEGijt
NEGijt
(4)
where Dt is the number of days in month t, POSijt is the number of new positive recommendations for stock i on day j in month t, and NEGijt is the number of new negative
recommendations for stock i on day j in month t. We then dene the newsletters degree of
bullishness for portfolio p in month t (NLBSIpt ) as:
100 p
NLBSIit
Np i=1
N
NLBSIpt =
(5)
where Np is the number of stocks in portfolio p for which new recommendations are available
in month t.
Examining the correlations between residual BSI and the macro-economic variables (see
Table IV, Panel B), we nd that BSI is negatively correlated with UI and MP (the contemporaneous correlations are 0.365 and 0.047 respectively) but positively correlated with VS
and TS (the contemporaneous correlations are 0.345 and 0.152 respectively). These moderate correlations suggest that macro-economic variables may have an inuence on aggregate
investor sentiment. Specically, individual investors are more bullish when the term spread
or the risk premium are higher.
The residual BSI is most strongly correlated with the residual NLBSI (see Table IV,
Panel C) the contemporaneous as well as the lagged correlations are strongly positive. The
contemporaneous correlation between residual BSI and residual NLBSI is 0.499 while the
two lagged correlations are 0.539 and 0.457 respectively. To better understand the nature
of this relation, we estimated the following time-series regression model, in which we control
for the eect of the macro-economic variables:
t = 1, 2, . . . , T.
(6)
Here, BSIt is the residual aggregate investor sentiment in month t, NLBSIt is the residual
aggregate newsletter sentiment in month t, UIt is the unexpected ination in month t, MPt
is the monthly growth in industrial production in month t, TSt is the term-spread in month
t, VSt is the value-spread in month t, and t is the error term.
Table V reports the regression results. It is clear that the contemporaneous investornewsletter sentiment relation is strong and positive (1 = 0.204, t-value = 2.666) even after
controlling for the eects of macro-economic variables on BSI. The unexpected ination
variable has a negative and signicant loading, suggesting that investors respond to changes
in the rate of ination. However, the loadings on other three macro-variables are statistically
insignicant.
Overall, these results indicate that aggregate individual investor sentiment is weakly
related to the standard risk factors, moderately related to the macro-economic variables,
and strongly related to the sentiment of investment newsletters.
17
V.B
To better characterize the lead-lag relation between the aggregate investor and newsletter
sentiments (residual BSI and residual NLBSI respectively), we also estimated the following
bivariate vector auto-regression (VAR) model:
BSIt
NLBSIt
= b10 + b11
b20
b21
b12
b22
BSIt1
NLBSIt1
+ 1
t
2t
(7)
Table VI presents the VAR estimates and Granger causality probabilities. Both sentiments
exhibit strong persistence (b11 = 0.33 with a t-value of 2.18 and b22 = 0.53 with a t-value of
4.62). More interestingly, we nd that the lagged aggregate newsletter sentiment (NLBSI)
predicts the current aggregate investor sentiment (BSI), but lagged BSI has no power to
predict the current NLBSI. This evidence of individual investor sentiment predictability
suggests that newsletter recommendations inuence individual investor trading behavior.
The p-values from Granger causality tests (see Table VI, Panel B) summarize and reinforce
these observations.15
VAR estimates and the Granger causality tests are sensitive to the choice of the lag length, we
carried out the VAR estimation using lag length of 1, 2, 4, and 6 months.
lagged aggregate newsletter sentiment (NLBSI) predicts the current aggregate investor sentiment (BSI) but
lagged BSI has no power to predict current NLBSI. For robustness, we also re-estimated equation (7) in the
presence of four chosen macro-economic variables.
p-values
from the Granger causality tests for this conditional VAR model are virtually unchanged from those reported
in Table IV. These results are available from us upon request.
18
activities of investors among small stocks is likely to be more positively correlated with the
sentiment of investment newsletter analysts.
Figure 3 shows the residual BSI correlation between the newsletter and investor BSI
indices for each of the ten size-decile portfolios. Among the lower size-decile portfolios,
the BSI correlation is strong and positive for the rst three size deciles, the residual BSI
correlations are 0.464, 0.358, and 0.460 respectively. Moving from lower to higher size-decile
portfolios, we nd that the BSI correlation decreases almost monotonically. In fact, for
both deciles 8 and 9, the BSI correlations are negative (-0.041 and -0.125 respectively). The
correlation increases again for the largest size decile where it is positive but weak (0.033).
Overall, our results are consistent with the view that for smaller stocks where publicly
available information is limited, investors are likely to seek expert advice from external
sources such as investment newsletters.16
In sum, our results suggest that the sentiment of individual investors, in particular their
buy-sell activity among small-cap stocks, is strongly inuenced by the sentiment of investment newsletters. Furthermore, individual investor sentiment is only weakly correlated with
the standard risk factors (SMB, HML, and UMD) and macro-economic variables (UI, MP,
TS, and VS). These results suggest that individual investor sentiment may give rise to an
orthogonal source of comovement in stock returns, particularly among small stocks.
VI
In this section, we examine the incremental explanatory power of individual investor sentiment for cross-sectional stock returns. Our investigation follows procedures that have
become standard in recent asset pricing studies. For most of our tests, we employ variants of
a ve-factor time-series model where the rst three factors are from Fama and French (1993),
the fourth factor is the momentum factor (Jegadeesh and Titman 1993, Carhart 1997), and
the fth factor is investors buy-sell imbalance measure for the portfolio. Specically, the
16 Interestingly,
our finding coincide well with Fisher and Statman (2000). The authors use survey data to
show that individual investor sentiment is correlated with the recommendations of newsletter analysts.
19
(8)
t = 1, 2, . . . , T.
Here, Rpt is the rate of return on the portfolio, Rf t is the riskfree rate of return, RMRFt is
the market return in excess of the riskfree rate, SMBt is the dierence between the valueweighted return of a portfolio of small stocks and the value-weighted return of a portfolio
of large stocks, HMLt is the dierence between the value-weighted return of a portfolio of
high B/M stocks and the value-weighted return of a portfolio of low B/M stocks, UMDt is
the dierence between the value-weighted return of a portfolio of stocks with high returns
during months t 12 to t 2 and the value-weighted return of a portfolio of stocks with low
returns during months t 12 to t 2, BSIpt is investors buy-sell imbalance for portfolio p
in month t, and pt is the residual return on the portfolio.
Table VII presents the results of a time-series factor model estimation for each of the ve
size-quintile portfolios. For portfolio 1, the BSI loading is positive (0.069) and statistically
signicant (t-value = 2.467). With a monthly standard deviation of 8.430% for quintile 1
BSI, one standard deviation shift in the sentiment measure corresponds to a 0.582% monthly
shift in the portfolio return, which is economically signicant. For the remaining 4 portfolios,
the BSI loadings are statistically insignicant. These results suggest stocks in the smallest
quintile earn positive (negative) excess returns when individual investor sentiment is more
bullish (bearish). Individual sentiment does not play a signicant role in the comovement of
returns for other size quintiles.
The loadings on the standard risk factors are also revealing. As expected, the loading
on the SMB factor monotonically decreases as we move from portfolio 1 (smallest stocks) to
portfolio 5 (largest stocks). For portfolio 1, the loading on SMB is 1.409 and it decreases
to 0.160 in portfolio 5. However, we also nd a decreasing trend in the loading on the
HML factor. For portfolio 1, the loading on HML is 0.634 and it decreases to 0.039 for
portfolio 5. This suggests that the smaller stocks held by our investors are likely to be value
(high B/M) stocks. However, given each investors relatively smaller portfolio size (see Table
I), this result may also reect individual investors preference for lower priced stocks. The
loading on the UMD factor is negative for all 5 size-quintile portfolios, indicating an overall
preference for stocks that have performed poorly in the recent past.
VI.B Other Uni-dimensional Sorts: B/M, Institutional Ownership and Stock Price
To further examine the incremental power of investor sentiment in explaining returns, we
perform three other uni-dimensional sorts and carry out the ve-factor model estimation for
each set of portfolios. Given that the loadings on the HML factor exhibit a decreasing trend
in the estimates of size-portfolios, it is possible that the B/M ratio rather than rm size is
driving our main results. Previous studies (Barber and Odean 2001, Hong and Kumar 2002)
have documented that investors exhibit value-seeking behavior around dierent types of
news events. Thus, individual investors may be more active among value stocks and their
sentiment may have a greater ability to explain value stock returns.
21
level of institutional ownership at the end of the year, we assign each stock into low (deciles
1-3), medium (deciles 4-7), and high (deciles 8-10) institutional holding categories. We use
this coarser 3 3 grid instead of the more standard 5 5 or 10 10 grid because of the
fewer number of stocks among the higher size deciles in our sample. However, the qualitative
nature of our results do not change with a ner grid.
The factor model estimation results for the 9 size-ownership portfolios are presented in
Panel A of Table IX. For brevity, we only report the BSI loadings. As expected, the BSI
loadings are stronger for stocks with lower institutional ownership. For small stocks with low
institutional ownership, the BSI loading is 0.072 with a t-value of 3.424. The BSI loading is
positive but statistically insignicant for small sizemedium ownership and small sizehigh
ownership portfolios. These results suggest that rm size and institutional ownership are
joint determinants of the strength of sentiment-return relation.
Panel B of Table IX reports results for a nested double-sort along rm size and B/M
dimensions. We nd that among the small stocks category, the BSI loading is positive only
for high B/M portfolio (i.e, small-value stock portfolio). The BSI loading is insignicant
for all other size-B/M portfolios. Apparently small value stocks earn positive excess returns
when individual investors are more bullish.
Finally, we perform a two-way cut on rm size and stock price. Given the relatively
small size of investor portfolios (see Table I), it is likely that our investors prefer lower priced
stocks. If this is the case, individual investor concentration should be higher, and the BSI
loadings would be stronger, among lower priced stocks. To examine this possibility, we
assign stocks into 9 categories using a nested double-sort on rm size and stock price. The
sentiment loading estimates for the 9 size-price portfolios are presented in Panel C of Table
IX. The BSI loading is strongest (0.115 with a t-value of = 2.941) for small, lower-priced
stocks. It is positive and statistically signicant for small sizemedium price portfolio too.
For all other size-price portfolios, the sentiment loading is statistically insignicant.
Overall, the factor model estimates for portfolios obtained from two-dimensional sorts
reveal quite clearly that the sentiment-return relation is inuenced by all four stock characteristics we considered, namely, rm size, institutional ownership, B/M ratio, and stock
23
price. The sentiment-return relation is strongest for small-value stocks with lower level of
institutional ownership and lower prices. In general, these results support the behavioral
view of returns comovement where investor clienteles exist for certain sets of stocks and the
demand shocks generated by these clienteles induce comovement in returns.
VII
Various seasonal patterns in stock returns have been attributed to the trading behavior of
individual investors. For instance, Ritter (1988) suggests that the January eect, i.e., the
unusually high returns earned by rms in January, especially by small rms, may be a result
of systematic patterns in the trading activities of individual investors. Individual investors
are likely to engage in tax-loss selling in December. As this selling pressure subsides at
the end of December, there is a rebound in stock returns in January, especially during the
rst 10-days of the month. In a recent study, Poterba and Weisbenner (2001) compare the
relationship between past stock returns and the turn-of-the-year returns across dierent tax
regimes from the perspective of individual investors. Their results support the view that taxloss selling by individual investors contributes to the unusual patterns in the turn-of-the-year
stock returns.
Other explanations for the January eect are also consistent with the observed patterns
of stock returns at the turn-of-the-year. For instance, Lakonishok, Shleifer, Thaler, and
Vishny (1991) suggest that window-dressing by institutional investors may be an important
determinant of the observed patterns in the turn-of-the-year returns. Managers may have a
strong incentive to sell the losers in their portfolios at end of the year before their portfolios
are evaluated. To avoid revealing that they have held losers during the year, they may
dispose of losing investments in December. However, the window-dressing hypothesis cannot
explain the stronger turn-of-the-year eect among smaller rms where the concentration of
institutional investors is low.
Another widely documented seasonal pattern in stock returns is the Monday eect or the
weekend eect. It has been observed that the expected stock returns vary with the day-of24
the-week (French 1980) and in particular, the average Monday return is negative, especially
among smaller stocks. Prior studies have also attributed Monday eect to the behavior
of individual investors. Lakonishok and Maberly (1990) use odd-lot trades as proxy for
individual investor trading activities and nd that the overall odd-lot trading is higher on
Mondays and more importantly, odd-lot sales are signicantly higher on Mondays compared
with other days of the week. They use this evidence to suggest that the lower returns on
Monday may be due to a relatively higher concentration of individual trades and a higher
propensity of individual investors to be net sellers on Mondays. Abraham and Ikenberry
(1994) also use odd-lot trades as a proxy for individual investor trading activities and conrm
the ndings of Lakonishok and Maberly (1990). They also analyze the conditional return
distributions on Mondays and go on to show that the weekend eect is exacerbated when
negative returns are observed on Fridays.
We use a direct measure of individual investor trading activities and test whether the
sentiment of individual investors is indeed responsible for these two seasonal patterns in
stock returns. Our dataset also allows us to identify the stocks associated with all the
individual trades. This feature of the data allows for more rened tests of our hypotheses.
If the systematic trading activities of individual investors contribute to the January and the
Monday eects, we expect the relation between investor sentiment and the seasonal return
patterns to be stronger among smaller stocks where the concentration of individual investors
is relatively higher.
These results are broadly consistent with the findings of (Odean 1998) where investors reluctance to
realize losses (disposition effect) decreases during the year, attaining a minimum in December.
17
25
to January, we nd an abrupt and a signicant upward jump in BSI. This pattern is observed
for each of the 5 Januarys in our sample period. In December, the average BSI for portfolio
1 is 19.37% and in January it is 11.53% (see Table VIII), resulting in an average upward
move of 30.90%.
This seasonal pattern is not observed among the larger stock portfolios. Similar to smaller
stocks, we nd a greater tendency by investors to sell throughout the year.18 However, the
increase in year-end selling is not as pronounced as those for the smaller stock portfolios.
Moreover, we nd little evidence of a January rebound in BSI for large stocks.
Figure 5 provides a closer look at the trading behavior of investors at the turn-of-theyear. The average daily variation in the buy-sell imbalance (BSI) for small, medium, and
large stock portfolios during a 40-day period around the rst trading date in January are
plotted. The abrupt change in the average BSI for small stocks from the last trading date in
December to the rst trading date in January is rather striking. These results suggest that
the January bounce back in investors trading activities is strongest for smaller rms.
To investigate more formally the impact of investor sentiment on January returns, we
compare the average January residuals from dierent variants of the multi-factor time-series
model specied in equation (8), some of which include the BSI variable while others dont.
Our (untabulated) results show that, using a single-factor model containing only the market
risk factor, the average January residual is 5.54%. It decreases signicantly to 2.29% when
the three-factor Fama-French model is used. With the introduction of the momentum factor,
we nd a further reduction in the residual to 1.87%. Finally, when the BSI variable is added
to the model, the January residual is 1.49%. The dierence between the residual returns on
the smallest and the largest size portfolio, i.e., the arbitrage return, is also the lowest for the
ve-factor model that includes the investment sentiment factor. The arbitrage returns are
5.58%, 1.85%, 1.53%, and 1.19% for the 1, 3, 4, and 5-factor models respectively.
To summarize, our ndings are strongly suggestive of a link between tax-loss selling
activities of individual investors, and the exceptional performance of stocks (in particular,
18 Note that the negative average BSI for portfolios 4-10 throughout the year does not mean that in each
month investors are net sellers of stocks in each of these portfolios. Since the portfolio BSI is computed by
averaging individual stock BSIs, this simply means that in any given month, investors are likely to be net
sellers for a greater number of stocks in that portfolio.
26
small value stocks) in January. However, with only 6 years of data on investor trading
activities, we are unable to establish this link conclusively.
VII.B
Table XI reports the average daily return and the average daily BSI for each day of the week
and for each of the ten size decile portfolios. Consistent with previously reported results, for
stocks in deciles 1-4, the average return for Monday is negative (and statistically signicant)
during our 1991-96 sample period. For higher size deciles, the average Monday returns
are either negative and statistically insignicant (deciles 5-8) or positive and statistically
insignicant (decile 9) or positive and statistically signicant (decile 10). However, examining
the daily variation in the BSI pattern, we nd that the average Monday BSI is not negative
but rather positive for most (1-7 and 10) size-decile portfolios. In addition, comparing the
BSIs across the days of the week for a particular size decile, we nd that the average Monday
BSI is usually the highest. So, in our sample of individual investors, there is no evidence of
higher selling pressure (or even weaker buying pressure) on Mondays, even among smaller
stocks where the concentration of individual investors is considerably higher.
VIII
The tendency in the nancial press, and among capital market researchers, is to think of the
market as a single unit. This notion permeates our vocabulary and is well entrenched in our
thinking. The market is said to have gone up today, or the market responded negatively to
IBMs latest earnings report. In academic circles, inferences about the impact and usefulness
of a news release have been based on the aggregate market response. Market eciency is
also dened in terms of the speed of the aggregate response to news events, or the extent to
which market prices (an aggregate measure) impound currently available information. In all
these cases, the market is regarded, perhaps too casually, as a monolithic whole.
An alternative view is to regard the market as being composed of various clienteles or
informational subgroups. These subgroups respond to dierent information stimuli and may
27
respond dierently to the same information stimulus. The reactions of these subgroups
are sometimes re-enforcing and sometimes o-setting. To the extent that reactions among
these groups are not perfectly correlated, any theory which attempts to track aggregate
market movements will encounter empirical anomalies caused by the diering reaction of
the subgroups. A richer theory of market behavior would explicitly incorporate the potential
of diering reactions from important market subgroups.
One potentially fruitful partition is between large institutional investors and small individual traders. Extant evidence, primarily from surveys, shows individual investors trade
much less often, spend far less time on investment analysis and typically rely on a very different set of information sources from their professional counterparts.19 To the extent that
individual investors respond to dierent information signals, or have dierent time-varying
liquidity and consumption needs, the buy-sell imbalance in individual trades will dier systematically from overall market movements. If small investors buy-sell patterns do not move
in lock-step with those of larger investors, assets in market segments dominated by small
traders may be characterized by apparent pricing anomalies.
In this study, we have used a large set of data from a major discount brokerage house to
explore the eect of individual investor trading on stock returns. Our results show that the
buy-sell imbalance in individual investors trades contains a systematic component that is
uncorrelated with overall market movements. Using this common component as a measure
of individual investor sentiment, we examined its relation with a number of market pricing
anomalies.
We nd that our measure of individual investor sentiment has incremental explanatory
power for small stocks, value stocks, stocks with low institutional ownership, and stocks
with lower prices. The direction of the relation indicates that when individual investors are
relatively bullish (i.e., when net BSI is more positive), the stocks in these portfolios enjoy
higher excess returns. The magnitude of their inuence is aected by factors associated with
arbitrage costs, i.e., rm size, stock price, and the percentage of institutional ownership.
Evidence on trading frequency can be derived from ownership and volume information provided by NYSE
share ownership surveys as well as the Securities Industry Association Investor Activity reports. Evidence
on decision styles and information sources is presented in Lease, Lewellen, and Schlarbaum (1974), Lewellen,
Lease, and Schlarbaum (1977), Yunker and Krehbiel (1974), and Shiller and Pound (1989).
19
28
Individual investor trading activities also exhibited strong seasonalities that partially explain
the January Eect.
Collectively, these ndings are broadly consistent with the predictions of a noise trader
model in which the systematic activities of individual investors aect the returns of those
stocks in which they are concentrated. Our ndings show that the magnitude of their
inuence is aected by factors associated with arbitrage costs, i.e., rm size, stock price, and
the percentage of institutional ownership.
In sum, our ndings establishes the existence of a systematic component in the trading
activities of individual investors, and show that this sentiment measure has implications for
stock returns. These ndings raise a number of interesting issues. Although our sentiment
measure is correlated with changes in the recommendations of professional newsletters, many
questions remain as to the factors that aect the proclivity of individual investors to buy
or sell stocks. Indeed, these ndings highlight the need to better understand the processes
by which individual investors formulate their trading decisions, including an identication of
the information sources they use in decision making and the nature of their belief updating
process. We hope to address some of these topics in future research.
29
References
Abraham, Abraham, and David L. Ikenberry, 1994, The individual investor and the weekend
eect, Journal of Financial and Quantitative Analysis 29, 263277.
Barber, Brad, and Terrance Odean, 2000, Trading is hazardous to your wealth: The common
stock investment performance of individual investors, Journal of Finance 55, 773806.
, 2001, All that glitters: The eect of attention and news on the buying behavior of
individual and institutional investors, Working Paper, Haas School of Business, University
of California at Berkeley, November 2001.
, and Ning Zhu, 2003, Systematic noise, Working Paper, Haas School of Business,
University of California at Berkeley, April 2003.
Black, Fischer, 1986, Noise, Journal of Finance 41, 529543.
Carhart, Mark M., 1997, On persistence in mutual fund performance, Journal of Finance
52, 5782.
Chen, Nai-Fu, Raymond Kan, and Merton H. Miller, 1993, Are the discounts on closed-end
funds a sentiment index?, Journal of Finance 48, 785800.
Chen, Nai-Fu, Richard Roll, and Stephen A. Ross, 1986, Economic forces and the stock
market, Journal of Business 59, 383403.
Chopra, Navin, Charles M.C. Lee, Andrei Shleifer, and Richard H. Thaler, 1993, Yes, discounts on closed-end funds are a sentiment index, Journal of Finance 48, 801808.
DeLong, J. Bradford, and Andrei Shleifer, 1991, The stock market bubble of 1929: Evidence
from closed-end funds, Journal of Economic History 52, 675700.
, Lawrence H. Summers, and Robert J. Waldmann, 1990, Noise trader risk in nancial
markets, Journal of Political Economy 98, 703738.
30
Elton, Edward J., Martin J. Gruber, and Jerey A. Busse, 1998, Do investors care about
sentiment?, Journal of Business 77, 477501.
Fama, Eugene F., and Kenneth R. French, 1992, The cross-section of expected stock returns,
Journal of Finance 47, 427465.
, 1993, Common risk factors in returns on stocks and bonds, Journal of Financial
Economics 33, 356.
Feng, Lei, and Mark S. Seasholes, 2002, Correlated trading and location, Working Paper,
Haas School of Business, University of California at Berkeley, December 2002.
Ferson, Wayne E., and Rudi W. Schadt, 1996, Measuring fund strategy and performance in
changing economic conditions, Journal of Finance 51, 425461.
Fisher, Kenneth L., and Meir Statman, 2000, Investor sentiment and stock returns, Financial
Analysts Journal 56, 1623.
French, Kenneth R., 1980, Stock returns and the weekend eect, Journal of Financial Economics 8, 5570.
Gemmill, Gordon, and Dylan C. Thomas, 2002, Noise-trading, costly arbitrage, and asset
prices: Evidence from closed end funds, Journal of Finance Forthcoming.
Goetzmann, William N., and Massimo Massa, 2000, Daily momentum and contrarian behavior of index fund investors, Journal of Financial and Quantitative Analysis Forthcoming.
Gompers, Paul A., and Andrew Metrick, 2001, Institutional investors and equity prices,
Quarterly Journal of Economics 116, 229259.
Grinblatt, Mark, and Matti Keloharju, 2000, The investment behavior and performance
of various investor types: A study of Finlands unique data set, Journal of Financial
Economics 55, 4367.
, 2001, What makes investors trade?, Journal of Finance 56, 589616.
31
Hong, Dong, and Alok Kumar, 2002, What induces noise trading around public announcement events?, Working Paper, Department of Economics, Cornell University, February
2002.
Ivkovic, Zoran, and Scott Weisbenner, 2003, Local does as local is: Information content of the
geography of individual investors common stock investments, Working Paper, Department
of Finance, University of Illinois at Urbana-Champaign, March 2003.
Jegadeesh, Narasimhan, and Sheridan Titman, 1993, Returns to buying winners and selling
losers: Implications for market eciency, Journal of Finance 48, 6591.
Kahneman, Daniel, and Amos Tversky, 1973, On the psychology of prediction, Psychological
Review 80, 237251.
Kennickell, Arthur B., Martha Starr-McCluer, and Annika E. Sunden, 1997, Family nances
in the us: Recent evidence from the survey of consumer nances, Federal Reserve Bulletin
83, 124.
Lakonishok, Josef, and Edwin Maberly, 1990, The weekend eect: Trading patterns of individual and institutional investors, Journal of Finance 45, 231243.
Lakonishok, Josef, Andrei Shleifer, Richard Thaler, and Robert Vishny, 1991, Window dressing by pension fund managers, American Economic Review 81, 227231.
Lamont, Owen A., and Richard H. Thaler, 2002, Can the market add and subtract? Mispricing in tech stock carve-outs, Working Paper, Graduate School of Business, University
of Chicago, June 2002.
Lease, Ronald C., Wilbur G. Lewellen, and Gary G. Schlarbaum, 1974, The individual
investor: Attributes and attitudes, Journal of Business 29, 413433.
Lee, Charles M.C., Andrei Shleifer, and Richard H. Thaler, 1991, Investor sentiment and the
closed-end fund puzzle, Journal of Finance 46, 75108.
32
Lewellen, Wilbur G., Ronald C. Lease, and Gary G. Schlarbaum, 1977, Patterns of investment strategy and behavior among individual investors, Journal of Business 50, 296333.
Metrick, Andrew, 1999, Performance evaluation with transactions data: The stock selection
of investment newsletters, Journal of Finance 54, 17431775.
Neal, Robert, and Simon M. Wheatley, 1998, Do measures of investor sentiment predict
return?, Journal of Financial and Quantitative Analysis 33, 17751798.
Odean, Terrance, 1998, Are investors reluctant to realize their losses?, Journal of Finance
53, 17751798.
, 1999, Do investors trade too much?, American Economic Review 89, 12791298.
Pindyck, Robert S., and Julio J. Rotemberg, 1993, The comovement of stock prices, Quarterly Journal of Economics 108, 10731104.
Poterba, James M., and Scott J. Weisbenner, 2001, Capital gains tax rules, tax-loss trading,
and turn-of-the-year returns, Journal of Finance 56, 353368.
Rashes, Michael S., 2001, Massively confused investors making conspicuously ignorant
choices, Journal of Finance 56, 19111927.
Ritter, Jay R., 1988, The buying and selling behavior of individual investors at the turn of
the year, Journal of Finance 43, 701717.
Shapira, Zur, and Itzhak Venezia, 2001, Patterns of behavior of professionally managed and
independent investors, Journal of Banking and Finance 25, 15731587.
Shefrin, Hersh M., and Meir Statman, 1985, The disposition to sell winners too early and
ride losers too long, Journal of Finance 40, 777790.
Shiller, Robert J., 1984, Stock prices and social dynamics, Brookings Papers on Economic
Activity 2, 457510.
33
, and John Pound, 1989, Survey evidence on diusion of interest and information
among investors, Journal of Economic Behavior and Organization 12, 4766.
Shleifer, Andrei, and Lawrence H. Summers, 1999, The noise trader approach to Finance,
Journal of Economic Perspectives 4, 1933.
Shleifer, Andrei, and Robert W. Vishny, 1997, The limits of arbitrage, Journal of Finance
52, 3555.
Swaminathan, Bhaskaran, 1996, Time-varying expected small rm returns and closed-end
fund discounts, Review of Financial Studies 9, 845887.
Yunker, James A., and Timothy L. Krehbiel, 1974, Investment analysis by the individual
investor, Quarterly Review of Economics and Business 28, 90101.
Zhu, Ning, 2002, The local bias of individual investors, Working Paper, International Center
for Finance, Yale School of Management, February 2002.
34
Table I
The primary
dataset consists of trades and monthly portfolio positions of investors at a major discount brokerage house
in the U.S. for the period of 1991-96.
that contain the end-of-month portfolios of all investors, (ii) a trade file that contains all transactions carried
out by the investors in the database, and (iii) a demographics file that contains information such as age,
gender, marital status, income code, occupation code, geographical location (zip code), etc. for a subset of
investors.
Time Period:
Number of households:
Number of accounts:
Number of households with position in equities:
79, 995
158, 031
62, 387
$2.18 billion
$35,629 (Median = $13,869)
41 (Median = 19)
4 (Median = 3)
50 (Median = 48)
Panel C: Securities
Total number of traded common stocks:
10, 486
Number of common stocks for which data is available from CRSP: 9, 893
Panel D: Trades
Total number of trades:
Number of trades in common stocks:
Average Portfolio Turnover:
Average Holding Period:
2, 886, 912
1, 854, 776
7.59% (Median: 2.53%)
187 days (Median = 95)
35
Table II
Year
1991
1992
1993
1994
1995
1996
Num
trades
314915
303863
311569
258220
313551
325737
1991
1992
1993
1994
1995
1996
Panel
35284
38708
44105
41455
46330
45737
36
Num low
insti own
stocks
2036
1904
1849
1810
1733
1308
1425
1472
1453
1426
1372
1073
Table III
Hulbert Financial
Digest. Panel A provides an annual summary of the frequency of recommendations and the number of
This table provides a summary of the investment newsletters database obtained from
stocks covered by the newsletters in our sample. Panels B and C provide details of the newsletter coverage
across size and B/M deciles. Note that the number of stocks covered (row sums) in Panels B and C do not
add up to the total number of stocks covered by newsletters in each of the six years (reported in Panel A)
because of missing size or book-to-market information.
Panel A: Newsletter Coverage and Frequency of Recommendations
Number of
Number of
Number of
Number of
Year
Newsletters Stocks Covered Positive Recos Negative Recos
1991
147
1960
22281
19262
1992
152
1924
17583
15820
1993
156
2187
15653
16674
1994
169
2363
18167
20249
1995
175
2467
25578
23535
1996
175
2687
35144
37332
Year
1991
1992
1993
1994
1995
1996
Panel
D1
386
346
506
548
521
570
B: Newsletter Coverage
D2 D3 D4 D5
311 226 195 153
303 256 186 165
347 261 198 165
363 272 198 197
359 318 230 189
404 321 240 205
Year
1991
1992
1993
1994
1995
1996
37
Table IV
This table reports the correlations among standard risk factors and aggregate sentiment measures during
our 1991-96 sample period. RMRF is the market return in excess of the riskfree rate, SMB is the difference
between the value-weighted return of a portfolio of small stocks and the value-weighted return of a portfolio
of large stocks, HML is the difference between the value-weighted return of a portfolio of high B/M stocks
and the value-weighted return of a portfolio of low B/M stocks, UMD is the difference between the valueweighted return of a portfolio of stocks with high returns during months t 12 to t 2 and the value-weighted
return of a portfolio of stocks with low returns during months
12 to
buy-sell imbalance, and NLBSI is newsletters aggregate buy-sell imbalance. The aggregate BSI in month t
t
t
N
=1
=1
1
BSIit where BSIit =
. D is the number of days in month
is defined as BSIt =
t
t
i
N
=1
=1
D VBijt
Dj VBijt +
j
=1
t,
on day
in month
t,
and
D VSijt
jD
j VSijt
on day
in month
t,
portfolio. Residual BSI is obtained by removing the common dependence of BSI on the market.
38
Table V
This table reports the coefficient estimates from the following time-series model:
BSIt = + 1 NLBSIt + 2 UIt + 3 MPt + 4 VSt + 5 TSt + t t = 1, 2, . . . , T .
Here, BSIt is the residual aggregate investor sentiment in month t, NLBSIt is the residual aggregate newsletter
sentiment in month t, UIt is the unexpected inflation in month t, MPt is the monthly growth in industrial
production in month t, TSt is the term-spread in month t, VSt is the value-spread in month t, and t
N
is the error term. The aggregate BSI in month t is defined as BSIt = 100
BSIit where BSIit =
i=1
N
Dt VBijt Dt VSijt
jD=1
. Dt is the number of days in month t, VBijt is the buy volume (measured in dollars) for
Dj=1
t
t
VB
+
ijt
j=1
j=1 VSijt
stock i on day j in month t, VSijt is the sell volume (measured in dollars) for stock i on day j in month t, and N
is the number of stocks in the aggregate investor portfolio. NLBSI is defined as NLBSIt = N1 Ni=1 NLBSIi,t
Dt POSijt Dt NEGijt
jD=1
where NLBSIit = Dj=1
. Here, Dt is the number of days in month t, POSijt is the
t
t
POS
+
ijt
j=1
j=1 NEGijt
number of positive recommendations for stock i on day j in month t, and NEGijt is the number of negative
recommendations for stock i on day j in month t. Residual BSI (NLBSI) is obtained by removing the
common dependence of BSI (NLBSI) on the market. The Newey-West adjusted t-values of the coefficient
estimates are reported in the parentheses.
Intercept
0.017
(0.019)
NLBSI
0.238
(4.674)
-1.575
(-2.735)
-0.237
(-0.420)
0.204
(2.666)
UI
MP
VS
TS
Adj. R2
0.187
-16.544
(-3.758)
1.517
(1.200)
8.981
0.885
(1.979) (1.010)
0.177
-13.664
(-1.907)
1.560
(1.193)
0.276
0.844
(0.048) (1.113)
0.222
39
Table VI
BSIt
NLBSIt
b10
b20
b11
b21
b12
b22
BSIt1
NLBSIt1 +
12tt
BSI is investors aggregate buy-sell imbalance and NLBSI is newsletters aggregate Dbuy-sell imbalance.
The
Dt VSijt
t
VB
ijt
N
j
=1
j
=1
aggregate BSI in month t is defined as BSIt = 100
BSIit where BSIit = Dt VB + Dt VS . Dt
i=1
N
j=1 ijt
j=1 ijt
is the number of days in month t, VBijt is the buy volume (measured in dollars) for stock i on day j in
month t, VSijt is the sell volume (measured in dollars) for stock i on day j in month t, and N is the
number of stocks in the aggregate investor portfolio. NLBSI is defined as NLBSIt = N1 Ni=1 NLBSIi,t
Dt POSijt Dt NEGijt
jD=1
where NLBSIit = Dj=1
. Here, Dt is the number of days in month t, POSijt is the
t
t
POS
+
ijt
j=1
j=1 NEGijt
number of positive recommendations for stock i on day j in month t, and NEGijt is the number of negative
recommendations for stock i on day j in month t. Residual BSI (NLBSI) is obtained by removing the
common dependence of BSI (NLBSI) on the market. The vector auto-regression estimates for the 1991-96
period are reported in Panel A. In Panel B, the p-values from Granger causality tests are shown where a
matrix element represents the impact of column variable on the row variable.
Panel A: Vector Auto-Regressive Model Estimates
Variable
Const
BSIt1 NLBSIt1 Adj R2
BSI t
-0.005
0.325
0.185
(-0.359) (2.182) (3.556)
0.317
NLBSI t
-0.006
(-0.620)
0.192
(1.069)
0.531
(4.616)
40
0.356
Table VII
formed at the end of each year in December using NYSE size break-points and then held fixed throughout
the following year. The following time-series factor model is estimated:
Rpt
Rf t =
1p RMRFt
p SMBt
p HMLt
p UMDt
p BSIpt
pt
= 1, 2,
. . . , T.
Here, Rpt is the rate of return on the size-ownership portfolio, Rf t is the riskfree rate of return, RMRFt is
the market return in excess of the riskfree rate, SMBt is the difference between the value-weighted return of
a portfolio of small stocks and the value-weighted return of a portfolio of large stocks, HMLt is the difference
between the value-weighted return of a portfolio of high B/M stocks and the value-weighted return of a
portfolio of low B/M stocks, UMDt is the difference between the value-weighted return of a portfolio of
stocks with high returns during months
stocks with low returns during months
portfolio in month
as BSIpt =
100
Np
t,
and
pt
12 to
12 to
p
=1 BSIit where BSIit =
adjusted
t-values
Portfolio
Size
Quintile 1
on day
Dt VBijt
Dj=1
t
j=1 VBijt +
N
i
in month
t,
and
Np
Dt VSijt
jD=1
.
t
j=1 VSijt
on day
is defined
in month
t,
0.463
(1.712)
0.165
(0.576)
RMRF
0.867
(9.741)
0.870
(10.137)
SMB
HML
UMD
BSI
1.448
0.694
-0.244
(14.317) (6.338) (-2.551)
1.409
0.634
-0.134
0.069
(14.277) (5.854) (-1.308) (2.467)
-0.050
0.994
Size
(-0.472) (28.475)
Quintile 2 -0.058
0.995
(-0.524) (27.921)
0.924
0.188
(23.305) (4.379)
0.926
0.188
(22.857) (4.347)
-0.166
(-4.407)
-0.161
(-3.951)
-0.148
1.018
Size
(-1.894) (39.611)
Quintile 3 -0.140
1.020
(-1.719) (38.755)
0.732
0.131
(25.053) (4.140)
0.734
0.132
(24.322) (4.130)
-0.018
(-0.643)
-0.015
(-0.535)
0.004
(0.381)
-0.026
1.035
Size
(-0.288) (35.310)
Quintile 4 -0.018
1.036
(-0.161) (33.599)
0.421
0.103
(12.639) (2.845)
0.422
0.103
(12.025) (2.825)
-0.039
(-1.230)
-0.038
(-1.181)
0.001
(0.116)
0.039
-0.068
(1.992) (-3.978)
0.039
-0.070
(1.959) (-4.028)
-0.004
(-0.835)
Size
Quintile 5
0.041
(0.837)
0.003
(0.039)
1.024
(34.167)
1.017
(27.758)
-0.160
(-8.828)
-0.160
(-8.797)
41
Adj. R2
0.814
0.827
0.958
0.004
(0.293)
0.957
0.974
0.974
0.959
0.959
0.985
0.985
Table VIII
The
quintile portfolios are formed at the end of each year in December and then held fixed throughout the
following year. In forming the quintile portfolios, we use NYSE break-points for B/M, size, stock price, and
institutional ownership variables. The following time-series factor model is estimated:
Rpt
Rf t =
1p RMRFt
p SMBt
p HMLt
p UMDt
p BSIpt
pt
= 1, 2,
. . . , T.
Here, Rpt is the rate of return on the size-ownership portfolio, Rf t is the riskfree rate of return, RMRFt is
the market return in excess of the riskfree rate, SMBt is the difference between the value-weighted return of
a portfolio of small stocks and the value-weighted return of a portfolio of large stocks, HMLt is the difference
between the value-weighted return of a portfolio of high B/M stocks and the value-weighted return of a
portfolio of low B/M stocks, UMDt is the difference between the value-weighted return of a portfolio of
stocks with high returns during months
stocks with low returns during months
portfolio in month
as BSIpt =
100
Np
t,
and
pt
p
=1 BSIit where BSIit =
adjusted
t-values
on day
Dt VBijt
Dj=1
t
j=1 VBijt +
N
i
12 to
12 to
in month
t,
and
Np
Dt VSijt
jD=1
.
t
j=1 VSijt
on day
is defined
in month
t,
Sorting Variable
Firm
Size
Q1
0.069
(2.467)
BSI Loadings
Q2
Q3
Q4
0.004
0.004
0.001
(0.293) (0.381) (0.116)
Institutional
Ownership
0.052
(2.970)
0.042
0.024
-0.017
(2.068) (1.074) (-1.518)
-0.010
(-0.752)
Book-ToMarket
0.036
(1.024)
-0.025
(-1.098)
0.008
(0.674)
0.011
(0.486)
0.054
(2.203)
Stock
Price
0.078
(2.485)
0.032
(1.287)
0.029
(1.271)
0.011
(1.004)
-0.001
(-0.616)
42
Q5
-0.004
(-0.835)
Table IX
break-points for B/M, size, stock price, and institutional ownership variables.
Rf t =
= 1, 2,
. . . , T.
Here, Rpt is the rate of return on the size-ownership portfolio, Rf t is the riskfree rate of return, RMRFt is
the market return in excess of the riskfree rate, SMBt is the difference between the value-weighted return of
a portfolio of small stocks and the value-weighted return of a portfolio of large stocks, HMLt is the difference
between the value-weighted return of a portfolio of high B/M stocks and the value-weighted return of a
portfolio of low B/M stocks, UMDt is the difference between the value-weighted return of a portfolio of
stocks with high returns during months
stocks with low returns during months
portfolio in month
as BSIpt =
100
Np
t,
and
pt
12 to
12 to
adjusted
t-values
on day
Dt VBijt
Dj=1
t
j=1 VBijt +
p
=1 BSIit where BSIit =
N
i
in month
t,
and
Np
Dt VSijt
jD=1
.
t
j=1 VSijt
on day
in month
Medium
(D4-D7)
0.001
(0.237)
-0.012
(-0.754)
-0.010
(-0.717)
Large
(D8-D10)
-0.002
(-0.229)
-0.001
(-0.136)
0.001
(0.053)
Firm Size
Small
(D1-D3)
Medium
(D4-D7)
Large
(D8-D10)
is defined
Firm Size
Small
(D1-D3)
Medium
(D4-D7)
Large
(D8-D10)
0.115
(2.941)
0.071
(2.235)
High
0.024
(1.530)
-0.008
(-0.265)
-0.035
(-1.197)
0.001
(0.065)
0.014
(0.647)
0.009
(0.467)
-0.008
(-0.511)
43
Table X
This table reports the average monthly buy-sell imbalance (BSI) and average monthly returns for December,
January and February-November time-periods. The BSI for portfolio p in month t is defined as BSIpt =
1
p
=1 BSIit
N
i
where BSIit =
Dt VBijt
Dj=1
t
j=1 VBijt +
on day
in month
t,
and
Dt VSijt
jD=1
.
t
j=1 VSijt
on day
Np
in month
t,
t,
returns are computed for the 10 size-decile portfolios during the 1991-96 sample period.
Size Decile
D1
D2
D3
D4
D5
D6
D7
D8
D9
D10
Average Monthly
Dec
Jan
-19.365 11.530
-10.126 4.790
-7.464
1.514
-12.434 -3.330
-10.480 -2.035
-11.562 -8.709
-15.698 -11.497
-14.696 -15.563
-18.154 -14.393
-11.981 -10.642
BSI (%)
Feb-Nov
4.784
1.412
-0.765
-2.533
-4.270
-4.969
-10.461
-9.700
-11.056
-6.096
44
Table XI
of days in month
t,
D VBijt D VSijt
Dj VBijt + jD VSijt
j
j
on day
in month
t,
and
Np
on day
in month
t,
VSijt is
portfolio.
Size
Decile
D1
D2
D3
D4
D5
D6
D7
D8
D9
D10
Fri
Mon
4.938
3.520
4.167
4.616
3.020
-0.077
-0.125
0.147
0.259
0.387
(8.456)
(5.613)
(6.286)
(7.181)
(4.234)
(-1.715)
(-3.413)
(4.077)
(7.249)
(11.607)
Mon
Fri
4.899
3.903
4.208
4.539
2.687
-0.112
-0.074
0.122
0.206
0.300
(5.555)
(4.371)
(4.897)
(4.815)
(2.713)
(-3.308)
(-2.351)
(3.874)
(7.412)
(10.840)
4.673
3.692
2.266
3.822
2.925
-0.129
-0.064
0.107
0.203
0.272
(4.353)
(3.337)
(2.041)
(3.299)
(2.401)
(-4.003)
(-2.087)
(3.439)
(7.100)
(9.146)
4.357
3.184
3.607
3.143
0.790
-0.094
-0.049
0.120
0.181
0.239
(3.801)
(2.568)
(3.276)
(2.740)
(0.669)
(-2.575)
(-1.577)
(3.775)
(6.182)
(8.015)
6.550
4.744
3.554
5.433
2.491
-0.056
-0.030
0.127
0.196
0.192
(5.059)
(3.478)
(2.819)
(4.183)
(1.825)
(-1.508)
(-0.909)
(3.792)
(6.377)
(6.080)
5.549
5.364
5.154
5.984
4.522
-0.056
-0.035
0.125
0.174
0.178
(3.825)
(3.536)
(3.432)
(3.867)
(2.757)
(-1.428)
(-1.021)
(3.635)
(5.479)
(5.649)
3.760
4.640
1.791
6.837
4.245
-0.039
-0.026
0.129
0.148
0.139
(2.494)
(2.932)
(1.198)
(3.986)
(2.866)
(-1.030)
(-0.719)
(3.639)
(4.523)
(4.301)
-1.393
2.256
-0.790
3.511
0.943
-0.004
-0.015
0.138
0.141
0.109
(-0.871)
(1.550)
(-0.516)
(2.143)
(0.587)
(-0.109)
(-0.399)
(3.901)
(4.200)
(3.374)
0.459
-2.540
-0.160
0.736
-2.734
0.018
0.009
0.125
0.132
0.074
(0.313)
(-1.638)
(-0.104)
(0.462)
(-1.767)
(0.458)
(0.253)
(3.630)
(3.938)
(2.396)
5.363
0.655
1.680
4.193
0.550
0.121
0.033
0.105
0.079
0.011
(2.622)
(0.298)
(0.814)
(1.836)
(0.241)
(2.982)
(0.883)
(3.206)
(2.203)
(0.284)
45
= 0.44, Mdn = 0.45, = 0.09, 25thP = 0.38, 75thP = 0.51, Skew = 0.36, Kurt = 0.10
0.045
0.04
Proportion of Cases
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0.65
BSI correlation distribution obtained by forming 1000 pairs of non-overlapping 250-stock random-portfolios.
t
t
=1
=1
The BSI for stock i in month t is defined as BSIit =
and the BSI for portfolio
is
t
t
=1
=1
BSIpt =
100
Np
p
=1 BSIit
N
i
where
Dt
D VBijt D VSijt
Dj VBijt + jD VSijt
j
j
i on day j
month
t,
and
Np
in month
t, VSijt
t,
p.
46
i on day j
in
= 0.01, Mdn = 0.01, = 0.12, 25thP = 0.07, 75thP = 0.09, Skew = 0.04, Kurt = 0.39
0.06
0.05
Proportion of Cases
0.04
0.03
0.02
0.01
0.25
0.2
0.15
0.1
0.05
0.05
0.1
0.15
0.2
0.25
Figure 2. An example of the residual buy-sell imbalance (BSI) correlation distribution generated using a randomization procedure. This figure shows the residual BSI correlation distribution
obtained in one of the randomization tests. The BSI time series for each stock is generated by keeping the
frequency of trades fixed but the location as well as the direction (buy or sell) is randomized.
In each of
these tests, 1000 pairs of non-overlapping 250-stock random-portfolios are formed. The BSI for stock
in
t
t
N
=1
=1
and the BSI for portfolio
is BSIpt =
BSIit
month
is defined as BSIit =
i
t
t
N
=1
=1
D VBijt
Dj VBijt +
j
where
day
Dt
in month
t,
D VSijt
jD
j VSijt
t,
p.
p
=1
i on
Np is
100
p
on day
in month
t,
and
on the market.
47
0.5
0.4
0.3
0.2
0.1
0.1
0.2
D1
D2
D3
D4
D5
D6
D7
D8
D9
D10
Size Deciles
Figure 3. Cross-sectional variation in the relation between investor and newsletter sentiments.
This figure shows the residual BSI correlations between the newsletter and investor BSI indices for each of
Dt VBijt Dt VSijt
jD=1
the ten size-decile portfolios. The BSI for stock i in month t is defined as BSIit =
Dj=1
t
t
VB
+
ijt
j=1
j=1 VSijt
N
p
100
and the BSI for portfolio p is BSIpt =
i=1 BSIit where Dt is the number of days in month t, VBijt is
Np
i on day j in month t, VSijt is the sell volume (measured in
Np is the number of stocks in portfolio p. NLBSI is defined
on day
in month
t,
and
as NLBSIpt = 1N00p Ni=1p NLBSIit where Np is the number of stocks for which new recommendations are
Dt POSijt Dt NEGijt
jD=1
available in month t, and NLBSIit = Dj=1
. Here, Dt is the number of days in month t,
t
t
j=1 POSijt + j=1 NEGijt
POSijt is the number of new positive recommendations for stock i on day j in month t, and NEGijt is the
number of new negative recommendations for stock i on day j in month t. Residual BSI (NLBSI) is obtained
by removing the common dependence of BSI (NLBSI) on the market.
48
30
20
10
0
10
20
30
30
20
10
0
10
20
30
Dec 1991
Dec 1992
Dec 1993
Dec 1994
Dec 1995
Dec 1996
t,
and
49
on day
Cumulative BSI
3
40
30
20
10
10
20
30
40
average daily variation in the buy-sell imbalance (BSI) of small (deciles 1-3), medium (deciles 4-7), and large
(deciles 8-10) stock portfolios during 40 days around the first trading day in January. The BSI for portfolio
p
VB VS
=1 BSIit where BSIit = VBitit +VSitit .
N
i
50
on day
t,
and
Np
is