Content uploaded by Ning Zhu
Author content
All content in this area was uploaded by Ning Zhu on Feb 07, 2018
Content may be subject to copyright.
Do Individual Investors Learn from Their Trading Experience?∗
Gina Nicolosi
College of Business, University of Cincinnati
Liang Peng∗∗
College of Business, University of Cincinnati
Ning Zhu
Graduate School of Management, University of California Davis
Current Draft: March 2004
Abstract
This paper investigates whether individual investors adjust their stock trading according
to their stock selection abilities, which can be inferred from their trading history. Fixed-
effect panel regressions provide strong evidence that the ability to forecast future stock
returns significantly affects investors’ trading activity: investors purchase more actively if
they are more likely to have stock selection ability. Furthermore, trading experience –
measured by the number of purchases, the number of different stocks purchased, and the
variance of purchase dollar amounts – significantly helps improve investors’ portfolio
performance. In addition, we find that learning behavior varies across investors, which
corroborates the heterogeneity of individual investors.
JEL classification: D19, G14
Key words: individual investors, learning, rationality, trading
∗ We thank Brad Barber, Jim Clayton, Michael Ferguson, Brain Hatch, Yong Kim, Brian Kluger, Steve
Slezak, Steve Wyatt and seminar participants at University of Cincinnati for constructive comments. All
errors are ours alone.
∗∗ Corresponding author, PO Box 210195, Cincinnati, OH 45221-0195, Phone: (513)556-6829, Fax:
(513)556-0979, Email: liang.peng@uc.edu.
2
Do Individual Investors Learn from Their Trading Experience?
Rationality of economic agents is a classic assumption in modern economics and finance. This
assumption simplifies the decision-making processes associated with constrained optimization
problems, so that economic phenomena can be analyzed with mathematical models. An
important justification for this assumption is that agents are not likely to make systematic
mistakes. For instance, the assumption of rational expectations “does not deny that people often
make forecasting errors, but it does suggest that errors will not persistently occur on one side or
the other” (Sargent, 1993). The argument is certainly appealing, but not necessarily
substantiated. This paper empirically tests whether a special group of agents - individual
investors - learn about their stock selection ability from their own trading experience and adjust
their trading behavior accordingly. This study provides direct evidence regarding the
fundamental argument of agent rationality: rational investors learn from their mistakes and thus
mistakes should not be repeated systematically.
The study of individual learning behavior has important economic implications. First, the
growing literature of behavioral economics and finance provides strong evidence that agents are
not always fully rational at the aggregate level (see, e.g., Shleifer, 2000; Barberis and Thaler,
2002, for useful surveys). However, there is much less evidence on whether individual
economic agents learn to reduce their mistakes over time. Our study helps fill this gap. Second,
our study provides empirical evidence regarding the appropriateness of assumed rationality.
Specifically, if individuals do not learn, the rationality assumption, as well as the numerous
ensuing economic theories, would be challenged. Furthermore, our study is important because it
helps facilitate future research concerning the behavior of economic agents by disentangling two
possible sources of limited rationality at the aggregate level. One possibility is that individuals
are not fully rational and do not learn. Another possibility is that while individuals do learn to
become more rational over time, the representative agent remains limitedly rational because
amateur agents continually join the economy. In this case, population composition plays a
significant economic role.1
1 See Bakshi and Chen (1994) and Ang and Maddaloni (2003) for examples about relations between economy and
demography.
3
We test two hypotheses regarding individual investors’ learning behavior. The first hypothesis
concerns whether trading history affects future trading activity. We assume individual investors
are able to infer their stock selection ability from their trading history. If investors are rational,
their inferred stock selection ability should affect their future trading patterns. Specifically,
when they think they are able (unable) to select winning stocks, they should trade more (less)
actively. The second hypothesis concerns whether trading experience helps improve future
portfolio performance. Traders with more experience have more data (i.e., past trading activity)
to infer their ability. Therefore, their inferences should be more accurate, which should lead to
better trading decisions and hence better portfolio performance.
We use the nonparametric statistical model developed by Henriksson and Merton (1981),
modified by Cumby and Modest (1987), and applied by Hartzmark (1991) and others to infer
two types of stock selection ability: the ability to forecast the signs of future excess stock returns
as well as the ability to forecast both the signs and magnitudes of future excess stock returns.
We construct a time-series of inferred ability for each investor, and then use a fixed-effect panel
data approach to investigate the effect of the inferred ability on investor trading, controlling for
other possible trading-driving variables.
Our empirical results provide evidence that both types of abilities to forecast over the coming
month (20 trading days) significantly affect investors’ equity purchase decisions, while the
ability to make short term (5 trading days) forecasts do not affect purchases. We also find that
trading experience helps improve portfolio performance. Particularly, the more experienced an
individual investor is (in terms of the number of purchases, the number of stocks purchased, and
the variance of purchase dollar amounts), the higher is the future risk-adjusted return of her
portfolio. On the other hand, our results also demonstrate behavior variations across different
categories of investors. For example, active traders adjust their trading according to their
inferred ability more dramatically than inactive traders. Overall, our empirical findings suggest
that (i) individual investors rationally learn from their own experience and adjust their stock
purchases accordingly, and (ii) learning behavior is heterogeneous across individual investors.
4
We study individual investors because they are normally considered to be the most uninformed
and unskilled economic agents. They seem to be the real-world counterparts to the noise traders
described by behavioral models (see, for example, De Long et al., 1990; Shleifer and Vishny,
1997, among others). Individual investors have been modeled or documented to behave in many
naïve ways. For example, they may under-react or over-react to news; they may insufficiently or
naïvely diversify their portfolios; they may hold too many local or domestic equities; they may
make investment decisions based on familiarity instead of utility maximization; and confusion
concerning stock tickers may even lead them to respond to news incorrectly. 2 In short,
individual investors seem to be making a variety of mistakes that have significant economic
consequences. Therefore, they are suitable subjects for investigating potential learning behavior.
If they are able to learn rationally, it seems plausible to argue that other more sophisticated
economic agents should be able to learn as well.
We study learning behavior pertaining to trading because, unlike other economic processes in
which different types of behavior could all be rational, it is easier to identify trading mistakes.
Since trading is costly, it is reasonable to argue that individual investors should not trade unless
they have legitimate reasons. Legitimate reasons include having (i) security analysis ability –
the ability to select individual stocks that will outperform the market; (ii) market timing ability as
defined by Merton (1981) – the ability to forecast stock market performance market relative to
fixed-income securities; (iii) a need to rebalance their portfolios due to changes in personal
preference or changes in the covariance matrix of asset returns; (iv) a desire to exploit tax
benefits; or (v) a need to meet liquidity demands.
Our analysis is based on a sizeable panel data set containing the trading activities of individuals
belonging to a large national discount brokerage firm (see Odean 1998 and Barber and Odean
2000 for more descriptions of the data). This data set has several features that dramatically
motivate our research. First, the data allow us to observe individuals’ detailed trading history
(including prices and shares of stock bought and sold on a given date), dramatically facilitating
the inference of investor security analysis ability. Second, the data allow us to follow individuals
2 See for example Alpert and Raiffa (1982), Barber and Odean (2000), Benartzi (2001), Benartzi and Thaler (2001),
Coval and Moskowitz (1999), Daniel et al. (1998), Fischoff et al. (1977), Goetzmann and Kumar (2003), Hong and
5
over a period of six years, which is beneficial because the learning process is naturally a
temporal process. Finally, the data set contains a large number of individual investors, therefore
we are able to attain more inference power by exploiting the variation across individuals and thus
mitigate possible sample selection problems.
It is worth noting that rational learning behavior (i.e., whereby investors adjust their trading
according to their inferred ability) relates to but also differs from irrational performance-
feedback trading (i.e., whereby investors become overconfident after good performance and
consequently trade more actively). First, rational learning captures the relation between ability
and trading activity, while performance-feedback trading captures the relation between
performance and trading activity. Good performance does not necessarily suggest skill or ability
since performance can be due entirely to luck. Ability is estimated using methodology
established in Henriksson and Merton (1981) and modified by Cumby and Modest (1987), which
is based on significant statistical relations that are unlikely to be outcomes of chance. Second,
performance-feedback trading may suggest asymmetric reinforcement. Good performance may
lead to overconfidence and thus more trading, but poor performance does not necessarily lead to
less trading. On the other hand, rational learning suggests that investors should adjust their
trading activity according to their ability, regardless of their past performance (i.e., good or bad).
Finally, rational learning behavior suggests that trading experience should help improve future
investor performance, while performance-feedback trading lacks this implication.
This paper relates to literature regarding individual investor trading, which has documented
several important, and often not fully rational, behaviors. 3 For example, investors trade
excessively: investors who trade the most earn the lowest average returns after transaction costs.
Interestingly, on average, men trade more actively yet perform worse than women. Investors
who switch from phone-based trading to online trading also trade more and earn lower returns.
Additionally, individual investors are often reluctant to realize their losses. Finally, the trading
activity of individual investors is affected by past returns and historical price patterns. Our
empirical results provide evidence that, despite these irrational behaviors, individual investors
Stein (1999), Huberman (2001), Ivkovich and Weisbenner (2003), Rashes (2001), Zhu (2003), among others.
6
rationally learn from their trading experience and adjust their trading according to their stock
selection ability, thus indicating the complexity of human behavior.
This paper also relates to recent finding concerning individual investors’ ability to beat the
market. Coval et al. (2003) show that a portion of individual investors persistently earn average
excess returns, suggesting that at least some individual investors are able to select stocks. Our
empirical results suggest that individual investors can become better traders over time: if they
have (do not have) stock selection ability, they will trade more (less) actively, and trading
experience helps them achieve better portfolio performance.
Our empirical evidence corroborates recent findings pertaining to individual learning behavior in
marketplaces. Using experiments, List (2003) finds that market experience plays a significant
role in eliminating the endowment effect - individual behavior converges to neoclassical
predictions as market experience increases. Dhar and Zhu (2003) find that trading experience
helps alleviate individual investors’ tendency to sell winning stocks too soon and hold losing
stocks too long. We find that individual investors, despite their various investment mistakes, are
able to achieve better portfolio performance by adjusting their stock purchases according to their
trading experience. This finding emphasizes the importance of heterogeneity among economic
agents in the economy. The composition of investors, e.g. experienced vs. inexperienced, could
potentially influence how the market functions.
The paper proceeds as follows. Section I describes the data. Section II discusses our research
design. Section III reports our estimation results. Section IV concludes.
I. Data
The data used in this study come from a large discount brokerage firm, and cover the
investments of 78,000 households from January 1991 to December 1996.4 The data have three
important components: (1) Position data record the sample households’ end-of-month portfolio
3 See Barber and Odean (2000, 2001, 2002a, 2002b), Odean (1998, 1999), and Grinblatt and Keloharju (2001)
among others.
4 The end-of-month portfolio position data is available from January 1991 to January 1997. The trade data is
available from January 1991 to November 1996.
7
positions. (2) Trade data record all trades made by sample investors. Both the trading and
position files include common stocks, mutual funds, and other securities (i.e., American Deposit
Receipt, fixed income securities, and options). (3) An investor characteristics file includes
investor characteristics such as investors’ income levels, occupation categories, ages, and when
they opened their accounts.
In this study, we focus on investors’ stock selection ability so we exclude 11,535 investors who
do not hold common stocks in any month of our sample period. Sample households can open
multiple accounts at the discount brokerage firm. An average sample household has two
accounts. The most common reason for two accounts is the tax-preferred status of retirement
accounts. Common stocks make up roughly 60 percent of the account values and slightly more
than 60 percent of all trades.
The dataset is filtered as follows: First, if a household has multiple trading accounts, we treat
these accounts as one large account, aggregating monthly trades across all accounts associated
with a particular household. Then, only households that opened their first account in 1990 or
1991 are examined because their entire timeline of trading activity beginning with their very first
trade can be observed. Finally, to be in the sample at time t, household i must have non-
missing data for all variables utilized in this analysis. These requirements result in a sample of
65,118 household-months for 2,973 households, spanning the calendar time March, 1991 to
November, 1996.
Table 1 summarizes the trading activities and shows that individual investors have very different
trading behavior. For example, the average number of trades per month varies dramatically from
a minimum of 0 to a maximum of 113.0; the average number of buys per month varies from 0 to
49.67; and the average number of sells varies from 0 to 63.33.
II. Research Design
We test two null hypotheses regarding investors’ learning behavior: (i) investors’ stock
purchases are not affected by their security analysis ability, which they estimate using past
purchases and post-purchase stock returns; (ii) trading experience does not help investors
8
improve the performance of their portfolios. To test the first hypothesis, we construct monthly
series of security analysis ability proxies for each investor, then run fixed effect regressions of
the number of stock purchases on the ability proxies, controlling for other variables that may
affect stock purchases. Two types of security analysis abilities are studied: the ability to forecast
the signs of future risk-adjusted excess stock returns and the ability to forecast both the signs and
the magnitudes of future risk-adjusted excess stock returns, where the future is defined as the 5
and 20 trading days following the purchase, respectively. To test the second hypothesis, we
construct time series of 3 different measures of trading experience for each household, and run
fixed effect regressions of investors’ portfolio risk-adjusted excess return on the trading
experience measures.
We follow Coval et al. (2003) and use trades that initiate or expand existing positions in
companies in order to infer investors’ abilities. Essentially, we consider buys as predictions of
future price increases, but do not consider sales as predictions of future price decreases. The
rationale is, as Coval et al. (2003) argue, that sales are often not strongly driven by private
information or specific analysis of the sold stock. Instead, investors may sell to satisfy liquidity
needs or to move into other firms expected to outperform the market, etc. We also ignore short-
sales since there are very few occurrences.
Does Ability Affect Future Purchases?
The evaluation of investors’ forecasting ability is a well-studied problem in finance (see Becker
et al., 1999, Henriksson and Merton, 1981, Cumby and Modest, 1987, Hartzmark, 1987 and 1991,
Jagannathan and Korajczyk, 1986, and Merton, 1981, among many others). It is conventional
(e.g., Merton, 1981) to partition forecasting skills into “microforecasting” and
“macroforecasting,” defined as forecasting individual stocks’ price movements relative to the
market and stock market price movements relative to fixed-income securities, respectively. The
former is frequently referred to “security analysis” and the latter is termed “market timing.” This
paper focuses on investors’ security analysis abilities.
The first type of security analysis ability we study is the ability to forecast the signs of future
risk-adjusted excess stock returns. A classic approach to evaluate this type of ability is
9
developed by Henriksson and Merton (1981) and later modified by Cumby and Modest (1987)
and concerns whether the conditional probability of correctly forecasting the signs of future price
changes significantly differs from 0.5. This approach and its extensions are widely used to study
the performance of mutual funds and futures market traders, as well as the forecasting ability of
newsletters.5 Since we study only purchases, inferring security analysis ability is dramatically
simplified under the assumption that investors are equally capable of forecasting future price
increases and decreases.
The inference of investors’ abilities to forecast the signs of future risk-adjusted excess returns
consists of the following steps. First, we estimate the risk-adjusted excess return of purchased
stocks in the week (5 trading days) and month (20 trading days) after the purchases, respectively.
The different horizons are chosen to investigate investors’ ability to forecast short-term and
longer-term excess returns. To do this, for each purchased stock, we first run a time series
regression of its daily returns net the Treasury-bill rates on the Fama-French factors using a time
window spanning 100 trading days before to 100 days after the purchase day. The stock’s excess
return over the week or month following the purchase is the sum of its estimated regression
intercept and the error terms during the period in question; or equivalently, the realized return
minus the sum of the estimated factor loadings times the realized value of each of the factors.
We start with the day following the purchase to mitigate possible price impacts.
For investor i at the beginning of month t, the information with which the investor can infer her
ability includes ,it
N and ,it
G. ,it
N is the number of purchases made at least 5 or 20 trading days
before t (depending on the forecasting time horizon). ,it
G is the number of good purchases,
which are purchases with nonnegative risk-adjusted excess returns over the 5 or 20 subsequent
trading days (depending on the forecasting time horizon). Assuming the sign of the excess return
is generated from a binomial process, the null hypothesis is that the probability that the risk-
adjusted excess return conditional upon purchase will be positive is 0.5. A two-sided test of the
hypothesis is straightforward, and we follow Hartzmark (1991) and define a proxy for the
security selection ability as
5 See Bollen and Busse (2001), Chance and Hemler (2001), Graham and Harvey (1996), and Womack (1996),
among others.
10
()
,
,,
,
1 probability significance level sign of 0.5
it
it it
it
G
FS N
=− × −
(1)
This proxy incorporates information on the direction and significance of the ability. For example,
if ,,it it
GN equals 0.8 and the probability significance level is 0.15, then
()
,1 0.15 1 0.85
it
FS =− ×= . Therefore, the range
(
)
,
11
it
FS
−
≤≤
encompasses the universe of
investors: those with statistically significant inferior ability, those with no ability, and those with
statistically significant superior ability. To differentiate the ability estimated using the risk-
adjusted excess return in the week after the purchase from that estimated using the excess return
over the following month, we denote the former by 5
,it
F
S and the latter by 20
,it
F
S.
Hartzmark (1991) defines the big hit ability as the capability to predict both the magnitude and
direction of future price changes. Investors with superior big hit ability will establish larger long
positions when higher excess returns are anticipated. At time t, denote by ,ij
R
the subsequent
risk-adjusted excess return of the
j
th stock purchased by investor i, and by ,ij
D the dollar
amount of the purchase (the number of shares multiplied by the transaction price). At the
beginning of month t, we run a times-series regression investor i,
,,,,ij it ij ij
RD
α
βε
=
++, (2)
using all purchases made 5 or 20 trading days before t, respectively. To incorporate information
on the sign of the estimator of
β
, standard errors, and degree of freedom into one aggregate
measure, we follow Hartzmark (1991) and define the big hit ability for investor i at t, ,it
FB , as
()
(
)
,,,
1 probability significance level sign of
it it it
FB
β
=− × . (3)
Apparently, the range of ,it
FB is still
(
)
1,1−. To differentiate ,it
FB estimated using the excess
return in the week after the purchase from that estimated using the excess return in the following
month, we denote the former by 5
,it
F
B and the latter by 20
,it
F
B.
We study the relation between investors’ purchases and security analysis abilities using the
following fixed effect regression.
11
(
)
titttmtm
tmtmtititmtitiiti
JANDECVGR
SDSDPRRTFcT
,1091,81,7
1,,61,51,41,31,2,1,
3
ερρρρ
ρ
ρ
ρ
ρ
ρ
ρ
+++++
−
+
+
+
+
+
+=
−−
−−−−− (4)
In equation (4), ,it
T is the number of purchases made by investor i in month t; ,it
F is investor
i’s security analysis ability and takes the value of 5
,it
FS , 20
,it
FS , 5
,it
FB and 20
,it
FB , respectively;
,1mt
R− is the lagged S&P500 index return; ,1it
R
−
is the lagged personal portfolio return of the
investor; 1, −ti
Pis the market value of investor i’s portfolio (consisting of both mutual funds and
stocks) at the end of month 1t−; ,mt
SD is the cross-sectional standard deviation of all stock
returns in the market, and ,,1mt mt
SD SD
−
− captures the change in the cross-stock return standard
deviation; 1,
3−tm
GR is the compounded return of S&P500 over the past three years; 1, −tm
V is the
dollar value of NYSE/AMEX/NASDAQ stocks traded in the prior month; t
DEC is a dummy
variable for December; and t
JAN is a dummy variable for January.
Equation (4) includes ,it
F as an explanatory variable to study whether security analysis ability
affects trading. If an investor is rational, when ,it
F is positive, she should capitalize on her
ability by purchasing stocks more actively, and when ,it
F is negative, she should reduces her
trading activity. Therefore, a positive 1
ρ
is consistent with the existence of rational investor
learning.
Besides ,it
F, we include other explanatory variables in equation (4) to control for other trading
needs. First, we include lagged purchases to control for possible serial correlation in trading
activity. Second, we use the change in cross-sectional standard deviation of all stocks returns in
the market as a proxy for the change in the covariance of stock returns to control for rational
portfolio rebalancing. In addition, we include lagged market and personal portfolio returns. The
lagged personal portfolio return helps control for performance-feedback trading, and the lagged
long-run historical market performance helps capture the possible evolution of trading behavior,
such as that caused by changing investor sentiment correlated with market trends. In addition,
we include the lagged market value of investors’ portfolios to control for possible relations
between trading and portfolio size. For instance, some investors may cease trading simply
12
because they have consumed almost all of their wealth and thus have no money with which to
trade. We also include lagged market trading volume in dollar terms to control for fads and/or
herding in the investment market. Furthermore, we use December and January dummy variables
to capture possible trading seasonality (see e.g., Kumar and Lee, 2002). Finally, we include an
individual time-invariant intercept to capture trading associated with time-invariant components
of latent variables such as the investor’s risk attitude, gender, personal income, education, etc.
Other trading needs are captured by the error term.
Does Experience Help Improve Performance?
If investors rationally learn from past trading, as an investor’s trading experiences increases (e.g.,
more purchase transactions and/or more stocks purchased) so does the information set with
which the investor estimates her stock selection ability. This larger sample size should improve
the precision of the ability estimates, and subsequent trading decisions should improve
performance more significantly.
We first construct monthly time series of risk-adjusted portfolio returns, ,it
R
ER , for each investor
by adding the estimated intercept term to the estimation residuals of the regression of monthly
investor portfolio returns net the Treasury-bill rates on the Fama-French factors. These risk-
adjusted portfolio returns are measures of investors’ portfolio performance. We use three
measures of the trading experience, denoted by ,it
E. The first measure is simply the number of
all purchases made before t: 1
,
1
t
is
sT
−
=
∑. The second measure is the number of different stocks an
investor has ever purchased prior to time t. The third measure is the variance of the dollar
amounts of purchases prior to t. It is not implausible to argue that the more purchases an
investor has made and the more different stocks an investor has purchased, the better equipped
she is to infer her security analysis ability. Furthermore, the larger is the variance of the
explanatory variable, which is the dollar amounts of purchases in regression (2), the more
accurately the big hit ability can be estimated.
We investigate the relation between trading experience and portfolio performance using the
following fixed effect regression.
13
,,,it i it it
RER c E
ρ
ε
=
++ (5)
In (5), ,it
R
ER is the estimated risk-adjusted excess return of investor i’s portfolio at time t; ,it
E
is investor i’s trading experience at time t, and can assume three measures respectively; i
c is an
individual time-invariant intercept that captures unobserved individual specific factors; ,it
ε
is a
zero mean error term. A positive
ρ
is consistent with the existence of rational investor learning.
Sub-sample Analysis
We conduct a variety of sub-sample studies. First, we re-estimate equations (4) and (5) using
data in odd and even transaction months. Second, we re-estimate the equations for two
categories of investors: active traders (complete at least 25 trades in the sample period) and
inactive traders. Finally, we re-estimate the equations after categorizing investors according to
their average excess portfolio performance as winners (top one third), average (middle one third),
and losers (bottom one third).
The results of these sub-sample studies may have different interpretations. Under the assumption
of homogenous investors, these studies are robustness checks. However, allowing for the
heterogeneity of investors, these studies may reveal systematic differences between different
types of individual investors.
III. Empirical Results
Table 2 reports the results of regression (4) based on all investors in our sample. First, we find
that the coefficients of both 20
,it
FS and 20
,it
FB are significantly positive at the 1% level. Therefore,
we reject the null hypothesis that investors’ stock purchases are not affected by their abilities to
forecast the signs and magnitudes of excess stock returns over the time horizon of 20 business
days. Furthermore, the positive signs of the coefficients are consistent with the rational learning
hypothesis: when investors believe they are able to forecast excess returns, they purchase more
actively; when they do not believe in their ability, they reduce their stock purchases. Note that
the past performance of investors and the stock market have been controlled, so our results are
not caused by performance-feedback trading. At the same time, we find that 5
,it
F
S significantly
14
reduces stock purchases and 5
,it
FB is insignificant, which suggests that investors do not adjust
their purchases according to their ability to forecast excess returns over the time horizon of 5
business days. Consequently, we can not reject the null hypothesis that investors are not learning
from their trading experience. However, it is worth noting that the negative or insignificant
coefficients do not necessarily support the null hypothesis. There are a variety of possible
reasons why rational learning investors may not adjust their stock purchases according to their
ability to forecast short term excess returns. For example, individual investors may simply
intend to profit in a longer time horizon.
Tables 3 to 9 report the results of regression (4) using different sub-samples. Tables 3 and 4
report the results based on observations in each investor’s odd and even transaction months.
Tables 5 and 6 examine active and inactive investors. Tables 7, 8, and 9 are for households
whose arithmetic average monthly portfolio return in excess of the market return belonged in the
top, middle, and bottom one third of the distribution. These tables suggest that, on one hand, the
learning behavior we find in table 2 is observed in most sub sample studies. First, the coefficient
of 20
,it
F
S is significantly positive except for winners and even transaction months, for which the
coefficient is insignificant. It is worth noting that an insignificant coefficient does not
necessarily suggest that investors do not learn. It is possible that the insignificant coefficient of
winners is an artifact. For example, the number of purchases made by winners may be relatively
stable even if they actively learn and trade accordingly, because investors’ purchasing activity
can be limited by their wealth levels. Second, the coefficient of 20
,it
FB is significantly positive
except for inactive traders, average performance and losers, for which the coefficient is
insignificant. On the other hand, tables 3 to 9 also suggest the learning behavior differs across
investors. For instance, losers respond more dramatically to their ability to forecast the direction
of future excess returns, while winners do not seem to respond. However, the differences may
not be surprising since the heterogeneity of economic agents is an expected and established fact.
Table 10 reports the results of regression (5) using the full sample along with different sub-
samples. All three proxies are significantly positive, which rejects the null hypothesis that
trading experience does not help improve portfolio performance and is consistent with rational
15
investor learning behavior. The sub-sample studies, on the other hand, suggest heterogeneity
across investors. For example, inactive traders benefit much more significantly from the number
of prior purchases and the number of different purchased stocks than active traders. In addition,
investors with average performance benefit more from their trading experience than winners and
losers.
IV. Conclusions
This paper investigates whether a special group of economic agents - individual investors - learn
about their stock selection ability from their own trading experience and then accordingly adjust
their stock trading activity. We find that stock selection ability – particularly the ability to
forecast the signs and both the signs and magnitudes of excess stock returns in coming month –
significantly affects stock purchases. However, the ability to make short term (5 trading days)
forecasts does not affect purchasing activity. We also find that trading experience helps improve
portfolio performance. Particularly, as an investor completes more purchase transactions and
purchases more unique stocks, her portfolio’s subsequent risk-adjusted monthly return is higher.
Overall, our empirical findings are consistent with the hypothesis that individual investors
(despite making numerous documented mistakes) learn from their own trading experience, adjust
their stock purchases accordingly, and achieve higher portfolio performance. Our empirical
results also highlight the importance of individual investor heterogeneity. Specifically, learning
behavior varies across different categories of investors. For example, active traders adjust their
stock purchases according to their abilities more dramatically than inactive traders.
16
References
Alpert M., and H. Raiffa, 1982. A Progress Report on the Training of Probability Assessors,” in
D. Kahneman, P. Slovic and A. Tversky, eds., Judgment under Uncertainty: Heuristics and
Biases, Cambridge: Cambridge University Press.
Ang, Andrew and Angela Maddaloni, 2003. Do Demographic Changes Affect Risk Premiums?
Evidence from International Data. Journal of Business, forthcoming.
Bakshi, Gurdip and Zhiwu Chen, 1994. Baby Boom, Population Aging, and Capital Markets.
Journal of Business, 67 (2), 165-202.
Barber, Brad M. and Terrance Odean, 2000. Trading Is Hazardous to Your Wealth: The
Common Stock Investment Performance of Individual Investors. Journal of Finance, 55, 773-
806.
Barber, Brad M. and Terrance Odean, 2001. Boys Will Be Boys: Gender, Overconfidence, and
Common Stock Investment. Quarterly Journal of Economics, 141, 261-292.
Barber, Brad M. and Terrance Odean, 2002a. Online Investors: Do the Slow Die First? Review of
Financial Studies, 15, 455-487.
Barber, Brad, and Terrance Odean, 2002b. All that Glitters: The Effect of Attention and News on
the Buying Behavior of Individual and Institutional Investors. Working paper, UC Berkeley.
Barberis, Nicholas and Thaler, Richard, 2002. A Survey of Behavioral Finance. Handbook of the
Economics of Finance, edited by George Constantinides, Milt Harris and Rene Stulz,
forthcoming.
Becker, Connie, David Myers, and Michael Schill, 1999. Conditional Market Timing with
Benchmark Investors. Journal of Financial Economics, 52, 119-148.
Benartzi, Shlomo, 2001. Excessive Extrapolation and the Allocation of 401(k) Accounts to
Compnay Stock. Journal of Finance 56, 1747-1764.
Benartzi, Shlomo, and Richard Thaler, 2001. Naive Diversification Strategies in Defined
Contribution Saving Plans. American Economic Review, 91:79-98.
Bollen, N.P.B. and J.A. Busse, 2001. On the Timing Ability of Mutual Fund Managers. Journal
of Finance, 56 (3): 1075-1094.
Chance, D.M. and M.L. Hemler, 2001. The Performance of Professional Market Timers: Daily
Evidence from Executed Strategies. Journal of Financial Economics, 62 (2): 377-411.
Coval, Joshua, David Hirshleifer, and Tyler Shumway, 2003. Can Individual Investors Beat the
Market? Harvard NOM working paper.
17
Coval, Joshua, and Toby Moskowitz, 1999. Home Bias at Home: Local Equity Preference in
Domestic Portfolios. Journal of Finance, 54, 2045-73
Cumby, Robert E. and David M. Modest, 1987. Testing for Market Timing Ability: A
Framework for Forecast Evaluation. Journal of Financial Economics, 19: 169-190.
Daniel, Kent, David Hirshleifer, and Avanidhar Subrahmanyam, 1998. Investor Psychology and
Security Market under- and overreactions. Journal of Finance, 53, 1839-1886.
Dhar, Ravi and Ning Zhu, 2002. Up Close and Personal: An Individual Level Analysis of the
Disposition Effect. Yale ICF Working Paper.
De Long, J.B., A. Shleifer, L. Summer, and R. Waldmann, 1990. Noise Trader Risk in Financial
Markets. Journal of Political Economy, 98, 703-738.
Erev, Ido and Alvin E. Roth, 1998. Predicting How People May Play Games: Reinforcement
Learning in Experimental Games with Unique, Mixed Strategy Equilibria. American Economic
Review, 88(4), pp. 818-881.
Fischoff B., P. Slovic, and S. Lichtenstein, 1977. Knowing with Certainty: The Appropriateness
of Extreme Confidence. Journal of Experimental Psychology: Human Perception and
Performance, 3, 552-564.
Goetzmann, William, and Alok Kumar, 2003. Diversification Decisions of Individual Investors
and Asset Prices. Yale ICF Working Paper.
Graham, J.R. and Harvey, C.R., 1996, Market Timing Ability and Volatility Implied in
Investment Newsletters' Asset Allocation Recommendations. Journal of Financial Economics 42
(3): 397-421.
Grinblatt, Mark and Matti Keloharju, 2001. What Makes Investors Trade? Journal of Finance,
56(2): 589-616.
Hartzmark, Michael L., 1991. Luck Versus Forecast Ability – Determinants of Trader
Performance in Futures Markets. Journal of Business, 64 (1): 49-74.
Henriksson, Roy D. and Merton, Robert C., 1981. On Market Timing and Investment
Performance. 2. Statistical Procedures for Evaluating Forecasting Skills. Journal of Business,
54(4): 513-533.
Hong, Harrison and Jeremy Stein, 1999. A Unified Theory of Underreaction, Momentum
Trading, and Overreaction in Asset Markets. Journal of Finance, 54, 2143-2184.
Huberman, G., 2001. Familiarity Breeds Investment. Review of Financial Studies, 14, 659-680.
18
Ivkovich, Zoran, and Scott J. Weisbennner, 2003. Local Does as Local Is: Information Content
of the Geography of Individual Investors’ Common Stock Investments. University of Illinois at
Urbana-Champaign Working Paper.
Jagannathan, R., and R. A. Korajczyk, 1986. Assessing the Market Timing Performance of
Managed Portfolios. Journal of Business, 59 (2) 217-235.
Kumar, Alok, and Charles M. C. Lee, 2002, Individual Investor Sentiment and Comovement in
Small Stock Returns, University of Notre Dame Working Paper.
List, John A. 2003. Does Market Experience Eliminate Market Anomalies? Quarterly Journal of
Economics, 118 (1): 41-71.
Merton, Robert C., 1981. On Market Timing and Investment Performance. I. An Equilibrium
Theory of Value for Market Forecasts. Journal of Business, 54 (3): 363-406.
Odean, Terrance, 1998. Are Investors Reluctant to Realize Their Losses? Journal of Finance, 53:
1775-1798.
Odean, Terrance, 1999. Do Investors Trade Too Much? American Economic Review, 89: 1279-
1298
Rashes, 2001. Massively Confused Investors Making Conspicuously Ignorant Choices (MCI-
MCIC). Journal of Finance, 56(5), 1911-1927.
Sargent, Thomas J., Rational Expectations. The Concise Encyclopedia of Economics. Library of
Economics and Liberty. Retrieved August 24, 2003 from the World Wide Web:
http://www.econlib.org/library/Enc/RationalExpectations.html
Shleifer, A., and R. Vishny, 1997. The Limits of Arbitrage. Journal of Finance, 52, 35-55.
Shor, Mikhael, 2003. Learning to Respond: The Use of Heuristics in Dynamic Games.
Vanderbilt University Working Paper.
Womack, K.L., 1996. Do Brokerage Analysts’ Recommendations Have Investment Value?
Journal of Finance, 51 (1): 137-167.
Zhu, Ning, 2003. The Local Bias of Individual Investors. Yale ICF Working Paper.
19
Table 1
Data Summary: Trading Activities
This table reports summary statistics of the trading activities of households who opened their
first accounts in 1990 and 1991, respectively.
Average number of trades per month
Accounts Min 25% Median 75% Max
1990 0.00 0.17 0.38 0.80 43.16
1991 0.00 0.15 0.34 0.79 113.00
Total 0.00 0.16 0.37 0.79 113.00
Average number of buys per month
Accounts Min 25% Median 75% Max
1990 0.00 0.09 0.21 0.45 21.06
1991 0.00 0.07 0.19 0.45 49.67
Total 0.00 0.08 0.20 0.45 49.67
Average number of sells per month
Accounts Min 25% Median 75% Max
1990 0.00 0.07 0.17 0.37 22.09
1991 0.00 0.05 0.14 0.36 63.33
Total 0.00 0.06 0.15 0.36 63.33
20
Table 2
Security Analysis Ability and Stock Purchases
This table reports the results of the following fixed effect regression.
(
)
titttmtm
tmtmtititmtitiiti
JANDECVGR
SDSDPRRTFcT
,1091,81,7
1,,61,51,41,31,2,1,
3
ερρρρ
ρ
ρ
ρ
ρ
ρ
ρ
+++++
−
+
+
+
+
++=
−−
−−−−−
,it
T is the number of purchases household i made in month t. ,it
F takes on the value of 5
,it
FS , 20
,it
FS ,
5
,it
F
B and 20
,it
F
B, respectively. 5
,it
F
S and 20
,it
F
S measure investors’ abilities to forecast the signs of future
excess returns in the coming week (5 trading days) and month (20 trading days), respectively. 5
,it
FB and
20
,it
FB measure investors’ abilities to forecast both the signs and magnitudes of future excess returns in
the coming week and month, respectively. Additional explanatory variables are lagged purchases ,1it
T
−
,
lagged market return ,1mt
R−, lagged portfolio return ,1it
R
−
, households’ lagged mutual fund and stock
positions 1, −ti
P (coefficients are multiplied by 1,000,000), change of cross-sectional standard deviation of
stock returns in the market ,,1mt mt
SD SD
−
−, lagged three-year S&P500 return 1,
3−tm
GR , dollar value of
NYSE/AMEX/NASDAQ stocks traded in the prior month 1, −tm
V, and December t
DEC and January
t
JAN dummy variables. Coefficients with * are significant at the 5% level, and coefficients with ** are
significant at the 1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
*-0.044 **0.073 0.009 **0.054
2
ρ
**0.262 **0.262 **0.262 **0.262
3
ρ
**0.811 **0.805 **0.811 **0.812
4
ρ
**0.148 **0.147 **0.148 **0.147
5
ρ
**-0.375 **-0.385 **-0.380 **-0.383
6
ρ
**0.421 **0.424 **0.422 **0.421
7
ρ
**0.511 **0.504 **0.509 **0.509
8
ρ
**-0.000 **-0.000 **-0.000 **-0.000
9
ρ
-0.027 -0.027 -0.027 -0.027
10
ρ
*0.045 *0.045 *0.045 *0.045
21
Table 3
Security Analysis Ability and Stock Purchases: Odd Months
This table reproduces Table 2 but only uses data for each household’s odd moths (transaction
time). Coefficients with * are significant at the 5% level, and coefficients with ** are significant
at the 1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
-0.016 **0.108 -0.029 *0.057
2
ρ
**0.291 **0.290 **0.291 **0.290
3
ρ
**0.955 **0.941 **0.955 **0.955
4
ρ
*0.127 *0.127 *0.127 *0.126
5
ρ
-0.077 -0.085 -0.077 -0.084
6
ρ
**0.467 **0.469 **0.465 **0.467
7
ρ
**0.528 **0.521 **0.529 **0.527
8
ρ
**-0.000 **-0.000 **-0.000 **-0.000
9
ρ
-0.015 -0.014 -0.015 -0.015
10
ρ
0.050 0.050 0.050 0.050
22
Table 4
Security Analysis Ability and Purchases: Even Months
This table reproduces Table 2 but only uses data for each household’s even months (transaction
time). Coefficients with * are significant at the 5% level, and coefficients with ** are significant
at the 1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
*-0.071 0.041 0.045 *0.051
2
ρ
**0.231 **0.231 **0.231 **0.231
3
ρ
*0.664 *0.660 *0.661 *0.663
4
ρ
**0.167 **0.167 **0.167 **0.167
5
ρ
**-0.681 **-0.693 **-0.688 **-0.692
6
ρ
*0.379 *0.385 *0.382 *0.383
7
ρ
**0.496 **0.491 **0.492 **0.494
8
ρ
**-0.000 *-0.000 *-0.000 *-0.000
9
ρ
-0.041 -0.041 -0.041 -0.041
10
ρ
0.041 0.042 0.042 0.042
23
Table 5
Security Analysis Ability and Purchases: Active Traders
This table reproduces Table 2 but only uses households that performed at least 25 transactions
throughout the entire sample period. Coefficients with * are significant at the 5% level, and
coefficients with ** are significant at the 1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
*-0.084 **0.084 0.018 **0.075
2
ρ
**0.280 **0.280 **0.280 **0.280
3
ρ
**1.408 **1.391 **1.406 **1.409
4
ρ
**0.207 **0.206 **0.207 **0.206
5
ρ
*-0.414 *-0.429 *-0.422 *-0.426
6
ρ
**0.650 **0.652 **0.651 **0.650
7
ρ
**0.721 **0.716 **0.723 **0.721
8
ρ
-0.000 -0.000 -0.000 -0.000
9
ρ
-0.026 -0.026 -0.026 -0.026
10
ρ
0.070 0.070 0.070 0.071
24
Table 6
Security Analysis Ability and Purchases: Inactive Traders
This table reproduces Table 2 but only uses households who performed less than 25 transactions
throughout the entire sample period. Coefficients with * are significant at the 5% level, and
coefficients with ** are significant at the 1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
**0.051 **0.064 0.006 0.012
2
ρ
0.007 0.007 0.008 0.008
3
ρ
-0.022 -0.023 -0.024 -0.024
4
ρ
0.033 0.033 0.033 0.033
5
ρ
**-0.331 **-0.334 **-0.327 **-0.328
6
ρ
0.025 0.026 0.024 0.024
7
ρ
**0.339 **0.340 **0.343 **0.343
8
ρ
**-0.000 **-0.000 **-0.000 **-0.000
9
ρ
**-0.031 **-0.031 **-0.031 **-0.031
10
ρ
0.009 0.009 0.008 0.008
25
Table 7
Security Analysis Ability and Purchases: Winners
This table reproduces Table 2 but only uses households whose arithmetic average monthly
portfolio return in excess of the market return belonged in the top one third of the sample.
Coefficients with * are significant at the 5% level, and coefficients with ** are significant at the
1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
*-0.114 -0.009 0.026 **0.154
2
ρ
**0.329 **0.329 **0.329 **0.328
3
ρ
**1.292 **1.287 **1.285 **1.278
4
ρ
**0.145 **0.145 **0.145 **0.144
5
ρ
-0.287 -0.299 -0.294 -0.294
6
ρ
*0.471 *0.473 *0.474 *0.471
7
ρ
**0.461 **0.464 **0.464 **0.459
8
ρ
0.000 0.000 0.000 0.000
9
ρ
-0.046 -0.046 -0.046 -0.046
10
ρ
0.067 0.067 0.067 0.067
26
Table 8
Security Analysis Ability and Purchases – Performance Groups (Average)
This table reproduces Table 2 but only uses households whose arithmetic average monthly
portfolio return in excess of the market return belonged in the middle one third of the sample.
Coefficients with * are significant at the 5% level, and coefficients with ** are significant at the
1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
**-0.116 **0.080 0.045 -0.031
2
ρ
**0.177 **0.177 **0.177 **0.177
3
ρ
0.514 0.507 0.516 0.515
4
ρ
**0.286 **0.287 **0.287 **0.288
5
ρ
0.046 0.006 0.026 0.032
6
ρ
**0.404 **0.407 **0.406 **0.406
7
ρ
**0.516 **0.509 **0.513 **0.515
8
ρ
**-0.000 **-0.000 **-0.000 **-0.000
9
ρ
0.003 0.003 0.003 0.003
10
ρ
0.046 0.045 0.045 0.045
27
Table 9
Security Analysis Ability and Purchases – Performance Groups (Losers)
This table reproduces Table 2 but only uses households whose arithmetic average monthly
portfolio return in excess of the market return belonged in the bottom one third of the sample.
Coefficients with * are significant at the 5% level, and coefficients with ** are significant at the
1% level.
5
,it
FS 20
,it
FS 5
,it
FB 20
,it
FB
1
ρ
**0.180 **0.184 *-0.077 0.032
2
ρ
**0.247 **0.246 **0.248 **0.248
3
ρ
0.485 0.477 0.477 0.481
4
ρ
0.060 0.062 0.063 0.064
5
ρ
**-0.915 **-0.915 **-0.902 **-0.904
6
ρ
*0.356 *0.365 *0.358 *0.355
7
ρ
**0.577 **0.578 **0.601 **0.592
8
ρ
**-0.000 **-0.000 **-0.000 **-0.000
9
ρ
*-0.060 *-0.060 *-0.061 *-0.060
10
ρ
0.024 0.024 0.023 0.023
28
Table 10
Trading Experience and Portfolio Performance
This table reports the results of fixed effect regressions that test whether investors’ experience
helps increase the risk adjusted excess returns of their portfolios.
,,,it i it it
RER c E
ρ
ε
=
++
,it
R
ER is the Fama French three factor-adjusted excess portfolio return of household i at month t,
and ,it
E is household i’s trading experience prior to t, which is measured by three different
proxies. The first experience proxy is household i’s number of purchase transactions. The
second proxy is the number of different stocks household i has ever purchased. The third proxy
is the variance of the dollar amounts of household i’s purchases. Coefficients have been
multiplied by 1,000,000. Coefficients with * are significant at the 5% level, and coefficients with
** are significant at the 1% level.
Proxy 1 Proxy 2 Proxy 3
All samples
Coefficient **131.255 **366.745 **0.000
Odd Months
Coefficient 100.198 *340.954 0.000
Even Months
Coefficient **163.004 **391.744 *0.000
Active Traders
Coefficient **116.248 **322.120 *0.000
Inactive Traders
Coefficient **2280.551 **2659.029 0.000
Winners
Coefficient 115.402 316.317 0.000
Average
Coefficient **213.002 **396.693 0.000
Losers
Coefficient 89.677 433.041 0.000