Individual and Institutional

Investor Behavior

by
Noah S. Stoffman

A dissertation submitted in partial fulfillment

of the requirements for the degree of
Doctor of Philosophy
(Business Administration)
in the University of Michigan
2008

Doctoral Committee:
Associate Professor Tyler G. Shumway, Chair
Professor Miles S. Kimball
Associate Professor Uday Rajan
Assistant Professor Feng Li
Assistant Professor Kathy Z. Yuan

c Noah S. Stoffman

2008
For my parents

ii
 Acknowledgements

The work presented here would not have been possible without the help of my

teachers, colleagues, family, and friends.

In particular, many helpful discussions with Sreedhar Bharath, Bob Dittmar,

Ryan Israelsen, Amiyatosh Purnanandam, Amit Seru, Sophie Shive, and Clemens

Sialm are gratefully acknowledged.

I thank each of the members of my dissertation committee for their time and

many helpful suggestions. I have particularly benefited from my numerous inter-

actions with Uday Rajan and Kathy Yuan.

I am especially indebted to Tyler Shumway, who not only guided me through

the dissertation process, but has also become a good friend. The first chapter of

this dissertation was written jointly with him and Amit Seru, and I am grateful to

both for their collaboration.

I also thank Jussi Keppo, who helped obtain the data set used throughout this

dissertation, and explained some of the idiosyncracies of the Finish stock market.

Finally, I am eternally grateful to Jennifer Richler, without whose love and

support I would never have been able to complete this dissertation.

iii
 Table of Contents

Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

Chapter

I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

II. Learning By Trading . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 Hypotheses and Methods . . . . . . . . . . . . . . . . . . . . 14

2.1.1 Measuring performance . . . . . . . . . . . . . . . 15
2.1.2 Measuring disposition . . . . . . . . . . . . . . . . 19
2.1.3 Survivorship and heterogeneity . . . . . . . . . . 24
2.1.4 Predictions about heterogeneity and survival . . . 27
2.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3.1 Performance tests and results . . . . . . . . . . . . 32
2.3.2 Disposition tests and results . . . . . . . . . . . . . 34
2.3.3 Heterogeneity in learning . . . . . . . . . . . . . . 40
2.3.4 Survivorship effects . . . . . . . . . . . . . . . . . 43
2.3.5 Other Tests . . . . . . . . . . . . . . . . . . . . . . . 47
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

III. Who Trades with Whom? . . . . . . . . . . . . . . . . . . . . . . . . 63

3.1 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2 Data and methods . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.1 Data description . . . . . . . . . . . . . . . . . . . 72

iv
 3.2.2 Complications . . . . . . . . . . . . . . . . . . . . . 73
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3.1 Investor interaction . . . . . . . . . . . . . . . . . . 80
3.3.2 Daily returns . . . . . . . . . . . . . . . . . . . . . 84
3.3.3 Vector autoregressions . . . . . . . . . . . . . . . . 87
3.3.4 Alternate horizons . . . . . . . . . . . . . . . . . . 92
3.3.5 Returns following trade . . . . . . . . . . . . . . . 97
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

IV. When Are Individual Investors Informed? . . . . . . . . . . . . . 114

4.1 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . 118

4.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.3 Methods and Results . . . . . . . . . . . . . . . . . . . . . . 120
4.3.1 Unconditional post-transaction returns . . . . . . 121
4.3.2 Substitution trades . . . . . . . . . . . . . . . . . . 123
4.3.3 Returns to substitution trades . . . . . . . . . . . . 125
4.3.4 Returns and Aggregate Buy Ratios . . . . . . . . . 131
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

v
 List of Tables

Table

2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

2.2 Disposition Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 52

2.3 Simple Learning Model: Estimates at Individual Level . . . . . . . 53

2.4 Simple Learning Model: Disposition Estimates at Aggregate Level 54

2.5 Heterogeneity in Learning . . . . . . . . . . . . . . . . . . . . . . . 55

2.6 Learning with Individual Fixed Effects . . . . . . . . . . . . . . . . 56

2.7 Learning with Survival Controls . . . . . . . . . . . . . . . . . . . 57

2.8 Risk Taking and Experience . . . . . . . . . . . . . . . . . . . . . . 58

3.1 Stock Returns Following Trading by Institutions and Households 102

3.2 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

3.3 Trader Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

3.4 Trader Interaction—Cross-Sectional Statistics . . . . . . . . . . . . 105

3.5 Returns and Group Interaction—Daily . . . . . . . . . . . . . . . . 106

3.6 Returns and Group Interaction—VAR Results . . . . . . . . . . . . 107

3.7 Returns and Group Interaction—Weekly and Monthly . . . . . . . 108

3.8 Returns and Group Interaction—Intraday Price Evidence . . . . . 109

3.9 Returns Following Trade . . . . . . . . . . . . . . . . . . . . . . . . 110

vi
3.10 Institutions’ Response to Individual Trading . . . . . . . . . . . . 111

4.1 Post-Transaction Returns in U.S. Data—Summary Statistics . . . . 136

4.2 Post-Transaction Returns in U.S. Data—Regression Tests . . . . . 137

4.3 Post-Transaction Returns in U.S. Data—Cross-sectional Means . . 138

4.4 Trade Classifications . . . . . . . . . . . . . . . . . . . . . . . . . . 139

4.5 Post-Transaction Returns in U.S. Data . . . . . . . . . . . . . . . . 140

4.6 Post-Transaction Returns in Finland Data . . . . . . . . . . . . . . 142

4.7 Returns and Aggregate Buy Ratios . . . . . . . . . . . . . . . . . . 143

4.8 Returns and Aggregate Buy Ratios . . . . . . . . . . . . . . . . . . 144

vii
 List of Figures

Figure

2.1 Participation By Year . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2.2 Returns Persistence . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2.3 The Disposition Effect in Aggregate . . . . . . . . . . . . . . . . . 60

2.4 Returns by Disposition Quintile . . . . . . . . . . . . . . . . . . . . 61

2.5 Proportion of Accounts Who Exit . . . . . . . . . . . . . . . . . . . 61

2.6 Trading Intensity and Experience . . . . . . . . . . . . . . . . . . . 62

3.1 Stylized Timeline of Price Path Around Trade . . . . . . . . . . . . 112

3.2 Cumulative Price Impact Functions . . . . . . . . . . . . . . . . . . 113

viii
 Chapter I

Introduction

Academic research has recently documented a wide range of behaviors among

individuals that is difficult to reconcile with the classical view of a rational homo

economicus. In contrast to the “smart money” controlled by professional money

managers, it seems plausible that individual investors may consistently suffer

from costly behavioral biases. These biases may affect not only the performance

of individual investors, but also aggregate to cause price distortions. Given the

critical role for prices in capital allocation, such distortions could be economically

important.

The three papers presented in this dissertation explore different aspects of this

issue. Chapter I examines whether individual investors learn from their trading

experience to avoid a particular behavior bias and improve their returns. If indi-

vidual investors suffer from behavioral biases but quickly learn to avoid them as

they become more experienced, then the average degree of bias in the population

will diminish without a continuous influx of new, bias-prone investors. Previous

estimates of the speed of learning suggest that biases would quickly die out in the

population, but the results presented here suggest that in fact the speed at which

1
individuals learn is much slower. The relatively slow speed at which investors

learn suggests that we cannot rule out the possibility that individual investors

suffering from behavioral biases could distort asset prices.

Given these results, a natural question to ask is whether individual investors

actually affect stock prices. This is explored in Chapter II. In particular, I exam-

ine in which direction prices move when trading occurs between two individu-

als, between individuals and institutions, or between two institutions. I find that

prices move in the direction of institutional trading. That is, when individuals

buy shares from institutions, prices fall; and when individuals sell shares to indi-

viduals, prices rise. These results are consistent with the notion that individual

investors supply liquidity to institutions, and not that individuals actively move

prices.

In Chapter III, I implement a strategy for identifying those trades placed by

individual investors that are particularly likely to be based on information. I then

examine whether these trades earn higher ex post returns, and I find that they do.

This indicates that while many individual investors may generate poor invest-

ment performance, at least some individual investors are able to trade on infor-

mation. The technique presented here could also be used to estimate the level of

informed trading that occurs among individuals at any particular time.

2
1.1 Data

In each of the studies presented here, I use a data set that includes the com-

plete trading records of all individual and institutional investors in Finland over a

nine-year period. The remarkable richness of these data allow a uniquely detailed

examination of the behavior of individual and institutional investors in a financial

market. The data come from the central register of shareholdings in Finnish stocks

maintained by Nordic Central Securities Depository (NCSD), which is responsible

for the clearing and settlement of trades in Finland. Finland has a direct holding

system, in which individual investors’ shares are held directly with the CSD. Since

our data come from the CSD, they reflect the official record of holdings and are

therefore of extremely high quality. In particular, shares owned by individuals

but held in street name by a brokerage firm are identified as belonging to the indi-

vidual, and shares for each individual are aggregated across brokerage accounts,

regardless of whether they are held in street name. This allows a clean identifica-

tion of which type of investor owns shares. Grinblatt and Keloharju (2000, 2001a,

2001b) use a subset of the same data, comprising the first two years of our sam-

ple period.1 The data include the transactions of nearly 1.3 million individuals

and firms, beginning in January, 1995 and ending in December, 2003. Summary

statistics relevant to the analysis in each chapter are presented separately.

While our dataset includes exchange-traded options and certain irregular eq-

uity securities, we focus on trading in ordinary shares. Trading in Finland is con-

1 These references provide a detailed discussion of the data.

3
ducted on the Helsinki Stock Exchange, which is owned by OMX, an operator of

stock exchanges in Nordic and Baltic countries. Trading on the Helsinki exchange

begins with an opening call from 9:45–10:00 a.m., and ends with a closing call from

6:20–6:30 p.m. Continuous trading during regular hours is conducted through a

limit order book.

The transaction data include the number of shares bought or sold, correspond-

ing transaction prices, and the trade and settlement dates, although trades are not

time-stamped. Additional demographic data, such as the account-holder’s age,

zip code, and language are also included, as are initial account holdings at the

beginning of our sample period.

Only the direct holdings and transactions of individuals are available. This

means that for an individual who directly trades shares of Nokia and holds a

Finnish mutual fund that owns shares of Nokia, we will observe only trades in the

former. The trades of the mutual fund are included in the dataset, but are identi-

fied as holdings of the mutual fund company, and cannot be tied to the individual.

However, our wealth calculations allow us to compare the importance of the in-

dividual investors as a group to that of other market participants. On average,

individuals hold 12.6 percent of all equity held by Finnish investors, including fi-

nancial institutions, government funds, nonprofit organizations and nonfinancial

corporations. This is more than financial institutions, which hold an average of

9.6 percent during our sample period. The majority of equity is held by the gov-

ernment (34.7%) and nonfinancial firms (33.4%), although these investors trade

4
relatively less and may do so for strategic reasons that are not directly linked to

profit-maximization.

5
 Chapter II

Learning By Trading

Academics have recently shown interest in the investment behavior and per-

formance of individuals, a field that has been called ‘household finance’ by Camp-

bell (2006). Over the past decade, several researchers have documented a number

of behavioral biases among individual investors. More recently, researchers have

found evidence that some individual investors are more informed or skilled than

others.1 Considering these findings, it is natural to ask how skilled or informed

investors acquire their advantage. For example, do investors learn by trading? If

so, to what extent do investors improve their ability and to what extent do they

learn about their inherent ability? And how quickly do investors learn? In this

chapter, we exploit trading records to study both average investor performance

and the strength of the behavioral bias known as the disposition effect.2 We corre-
1 Coval, Hirshleifer, and Shumway (2005) document significant performance persistence among

individuals. Ivković and Weisbenner (2005) find that individuals place more informed trades in
stocks of companies located close to their homes, and Ivković, Sialm, and Weisbenner (2006) show
that individuals with more concentrated portfolios tend to outperform those who are more diver-
sified. Linnainmaa (2007) finds that individuals who trade with limit orders suffer particularly
poor performance.
2 The disposition effect is the propensity of investors to sell assets on which they have ex-

perienced gains and to hold assets on which they have experienced losses. The effect was

6
late performance and disposition with investor experience and investor survival

rates to determine whether and how investors learn by trading.

Motivated by the existing economics literature on learning, we consider two

specific ways in which investors can learn. First, in the spirit of classical learning-

by-doing models (Arrow 1962, Grossman, Kihlstrom, and Mirman 1977), investors

might improve their ability as they trade. Second, as investors trade they might

realize that their inherent level of ability is low and decide to stop actively trading.

This decision to either continue or stop trading is a feature of the recent learning

model of Mahani and Bernhardt (2007). In our analysis we take either of these two

scenarios as evidence of learning by trading. Although these types of learning are

different, they are not mutually exclusive.

To clarify the model of learning we have in mind, consider the case of an in-

dividual who decides to begin trading. The investor must decide which of the

myriad sources of market information and investment advice available to her to

take seriously. She could consult standard news sources, internet sites, invest-

ment newsletters, and neighbors or friends. She might also consider the advice of

brokers, news analysts, authors of books and magazines, and finance professors.

To the extent that these sources fail to agree completely, individuals must deter-

mine how much decision weight to assign to each source. Moreover, the quality
first proposed by Shefrin and Statman (1985), and was subsequently documented in a sample
of trading records from a U.S. discount brokerage firm by Odean (1998). The effect has been
found in other contexts, including in Finland (Grinblatt and Keloharju 2001a), China (Feng and
Seasholes 2005, Shumway and Wu 2006), and Israel (Shapira and Venezia 2001); among profes-
sional market makers (Coval and Shumway 2005), mutual fund managers (Frazzini 2006), and
home sellers (Genesove and Mayer 2001); and in experimental settings (Weber and Camerer 1998).
We focus on the disposition effect because it is a robust empirical finding and it is relatively easy
to measure.

7
of these sources is likely to differ across individuals: some investors may know

executives at a firm, while others will not. As investors begin trading, they can

learn to which of the various sources they should pay more attention. This can

be thought of as improving ability through learning. Investors who only have

access to poor sources of information cannot improve by focusing more on par-

ticular sources. Instead, they will learn that they have no useful information and

will stop trading actively, choosing instead to invest in a passive investment such

as an index fund.3 This is also a type of learning, but rather than improving her

ability, the investor learns about her ability.

This example helps to elucidate the two types of learning we examine in this

chapter. First, investors may learn in such a way that their ability improves, per-

haps by learning which of their sources of information are particularly useful.

We measure this type of learning by correlating investors’ performance and sus-

ceptibility to the disposition effect with their experience. Second, investors may

learn about their ability, perhaps by learning whether they have access to useful

information sources. Because investors who learn that their ability is poor will

cease to trade, we identify this type of learning by examining attrition in our data.

Once we control for unobserved individual heterogeneity and attrition, any learn-

ing that remains will be of the first type, a direct improvement in ability. Since we

track individuals over time—both their survival rates as well as their performance
3 This example is very similar in spirit to well-known ‘bandit’ problems. For example, Bolton

and Harris (1999) study the strategic interaction of agents in an experimentation game. In our set-
ting, we can think of investors as being randomly assigned slot machines with different unobserv-
able expected payoffs, and experimenting to learn what the payoffs are. The random assignment
is analogous to the assignment of inherent ability.

8
and their level of the disposition effect—we are able to differentiate between these

two types of learning. Ours is the first paper to identify and measure both types

of learning.

Differentiating between types of learning allows us to determine how quickly

investors learn in a meaningful way. Estimating learning without controlling for

heterogeneity and attrition results in inflated improvement estimates, which do

not correspond to the experience of any particular type of investor. Measuring the

speed of learning in a meaningful way is important for a number of reasons. If

investors learn quickly and there is low turnover in the population of investors,

behavioral biases are unlikely to significantly affect asset prices. Moreover, if they

learn relatively quickly then the ‘excessive’ trading documented by Odean (1999)

and Barber and Odean (2001) may be justified, because investors may optimally

choose to trade more actively if they know they will improve with experience.

Finally, policy makers should consider the speed at which individuals learn when

determining the costs and benefits of different trading mechanisms.

The type of learning we observe, besides being important from a theory per-

spective, also has significant implications for policy. For instance, if investors can

significantly improve their ability from experience then investor education ini-

tiatives may be worthwhile and perhaps individual investors should be encour-

aged to trade. However, if the ability to invest successfully is relatively fixed,

then screening mechanisms, or tests that measure and reveal inherent ability, have

more value than education. Finally, thinking about both the speed and the type

of learning, has implications for market efficiency. For example, if many inexperi-

9
enced investors begin trading around the same time, and they learn slowly, their

trading could lead to time-varying market efficiency.

We test our hypotheses with a remarkable dataset that includes the complete

trading records of investors in Finland from 1995 to 2003, including more than 22

million observations of trades placed by households. We use these data to esti-

mate disposition and calculate performance at the account level. Our disposition

estimates indicate that a median individual in our sample is 2.8 times more likely

to sell a stock when its price has risen since purchase than when its price has

fallen. We exploit the panel structure of our data to examine whether individual

investors learn to avoid the disposition effect and improve their performance as

they trade. In particular, we estimate the mean return and the disposition effect

for each account and year in our sample and relate these estimates to experience,

past returns, and various demographic controls.

We measure investing experience with both the number of years that an in-

vestor has been trading and with the cumulative number of trades that an in-

vestor has placed. Of course, investors may gain experience by actively trading

securities and observing the results of each trade. If this is the primary way in

which investors learn, then cumulative trades will predict future investment per-

formance and the disposition effect. However, investors may also learn by observ-

ing market quantities and considering the outcomes of hypothetical trades based

on, for example, a particular information source. If this is the primary way that

10
investors learn, then years of experience will be a better predictor of investment

performance and the disposition effect than cumulative trades.4

Our tests provide robust evidence of learning by trading at the individual

level. We show that in a simple model of learning, performance improves and the

disposition effect declines as investors become more experienced. An extra year

of experience is associated with an improvement in average returns of approxi-

mately 40 basis points (bp) over a 30-day horizon, and a reduction in the dispo-

sition effect of about 5 percent. However, we argue that individual heterogeneity

and survivorship effects are likely to significantly affect these simple estimates,

making them difficult to interpret.

We adjust for survivorship and heterogeneity with a modified Heckman selec-

tion model that allows for individual fixed effects. The Heckman selection model

is a two-stage instrumental variables model that adjusts for the possibility that

the composition of the sample is endogenous. As instruments, we construct two

variables that satisfy the necessary exogeneity conditions—that is, they are likely

to affect the probability of an investor remaining in the sample, but are unlikely to

affect changes in the investor’s performance or disposition effect except through

their effect on survival. The first variable is an indicator for whether the indi-

vidual inherited shares in the previous year due to the death of a relative. We

conjecture that an individual who inherits shares is more likely to trade in the

future, perhaps because their wealth has increased, or perhaps because the new
4 It
is also possible, of course, that investors learn by considering the returns to hypothetical
trades before they ever start trading. If this is the only way in which investors learn then we will
find no evidence of learning by trading.

11
shares cause them to pay more attention to the stock market. This satisfies the

exogeneity condition since inheritance of shares from a relative is unlikely to di-

rectly affect changes in the performance or the disposition effect of an individual.

The second variable is the variation in the returns across all positions taken by

an account in the previous year. This variation is a measure of the consistency of

an investor’s performance, and we conjecture that an investor with more variable

performance is more likely to stop trading. While an investor’s previous average

performance is likely to be related to her future performance, there is no reason

to believe that the consistency of an investor’s past performance should directly

affect changes in her future performance or disposition effect.

In our first-stage estimates we confirm that investors with poor performance

are those who are more likely to cease trading. Moreover, more successful in-

vestors continue to trade actively. We also find that adjusting for survivorship

and heterogeneity in our learning reduces our learning coefficients by about one-

half to three-quarters. This suggests that while individuals do learn to improve

their trading ability with experience, they primarily learn about their own inher-

ent trading ability and they cease to trade if their ability is low.

Some evidence about learning by trading exists. Feng and Seasholes (2005)

give evidence that investors, in aggregate, display significantly less disposition

over time, estimating that for sophisticated investors the disposition effect is es-

sentially attenuated after about 16 trades. These estimates do not appear to be

consistent with the findings of Frazzini (2006), which shows that mutual fund

managers, who trade substantially more than most individuals, display a signif-

12
icant disposition effect. Furthermore, since Feng and Seasholes (2005) do not ad-

just for heterogeneity and attrition, it is impossible to tell whether their estimates

imply that a particular investor who trades 16 times will no longer exhibit the dis-

position effect or whether they imply that most investors who exhibit the disposi-

tion effect will cease trading before they place 16 trades. Nicolosi, Peng, and Zhu

(2004) show that the trading performance of individuals appears to improve with

trading experience, estimating that individuals can improve their risk-adjusted

portfolio return by about two percent per year (or about 0.8 bp per day) over a

3-year period. This seems quite large. Again, it is not possible to tell whether

this estimate is driven by the most successful investors surviving or by the least

successful investors improving their ability.5

While our tests have some features in common with existing papers, they differ

from the literature in a number of important respects. First, unlike other papers,

our tests use measures of performance and estimates of the disposition effect that

are specific to individuals, allowing us to track particular individuals over time.

This allows us to control for investor heterogeneity and survivorship effects and

allows us to separate the two types of learning as well as to estimate the speed

of learning. It also ensures that each observation in almost all of our tests is an

average or regression coefficient for one individual in one particular year. Thus,

we can be sure that our results hold for the average active investor. Put another

way, investors who trade disproportionately often do not receive disproportionate

5 Otherpapers that find some learning in various settings include List (2003), Barber, Odean,
and Strahilevitz (2004), Linnainmaa (2006), and Choi, Laibson, Madrian, and Metrick (2007).

13
weight in our estimates. Finally, using only one observation per person per year

reduces the likelihood that our standard error estimates are incorrect because of

correlation among our regression residuals. Given the unique features of our data

and our test methods, the results of our hypothesis tests add significantly to the

literature on financial learning.

The rest of the chapter is organized as follows. Section 2.1 describes the hy-

potheses we test and our statistical methods, while Section 2.2 provides detail on

our data. Section 2.3 discusses our results and Section 2.4 concludes.

2.1 Hypotheses and Methods

We test three hypotheses in this chapter. Our first hypothesis (H1) is that in-

dividual investors learn by trading. More importantly, we are interested in un-

derstanding how investors learn by trading. Our second and third hypotheses

clarify the type of learning in H1. Our second hypothesis (H2) is that investors

learn about their inherent ability by trading, and our third hypothesis (H3) is

that investors learn to improve their ability over time. Of course, if H1 is false,

then H2 and H3 will also be false. We test H2 by examining the importance of

individual heterogeneity and attrition in our sample. The notion in H2 is that

individuals stop trading if they realize their ability is low. Once we control for

unobserved individual heterogeneity and attrition, any learning that remains will

support H3—that is, it suggests that individuals learn to directly improve their

ability. To examine our hypotheses we test a number of related predictions. Our

14
prior belief for each of these predictions is that investors both learn about their

inherent ability and learn to improve their ability. This section motivates and de-

scribes our tests in more detail. It also describes some of the statistical methods of

our tests.

2.1.1 Measuring performance

Investor performance is the primary variable that we correlate with experi-

ence to test our hypotheses. Measuring the performance of individual investors

is a significant challenge. Our data, like others that are available, do not include

all non-equity securities that may be held by an investor, so it is impossible to

measure the return for the investor’s entire portfolio. This is made more diffi-

cult by the fact that the amount of money an individual has invested in equities

often fluctuates significantly over time. Since we cannot accurately measure port-

folio returns, we measure performance by examining the average return of stocks

purchased. However, this generates a new problem—comparing the returns on

holding periods of different lengths. For example, it is particularly difficult to

compare the performance of one investor who holds a stock for one week and

earns a holding period return of 3 percent to that of another investor who holds a

stock for one year and earns a holding period return of 15 percent.

Given the challenge of calculating performance, we take a straightforward ap-

proach that is nevertheless likely to capture much of the relevant information in

the individual’s returns. In particular, we calculate the returns earned by the pur-

chased stock in the 30 trading days following each of an investor’s purchases. We

15
choose to examine 30-day returns because the median holding period in our data

is 39 trading days, but all of our findings remain unchanged if we use a 10-, 45-, or

60-day holding period. Importantly, we truncate this calculation window at the

length of the actual holding if it is shorter than 30 days. That is, the 30-day return

for investor i holding stock j is,

Pj (t + min(s, 30))
Ri,j (t) = − 1,
Pj (t)

where Pj (·) denotes the stock’s closing price adjusted for splits and dividends, t

denotes the purchase date, and s denotes the actual holding period.

Our approach is an attempt to deal with the problem of comparing returns

over similar holding periods while ensuring that the actual selling decisions of

investors affect their performance. By measuring returns this way, we hope to

capture the value of short-term signals that the investor may have received. Look-

ing over longer horizons would introduce considerable noise into our return esti-

mates.

To adjust for risk, some of our results use risk-adjusted returns (or alphas). We

use a four-factor model to adjust returns for known risk factors. In addition to

a value-weighted market return, we construct three factor-mimicking portfolios

(SMB, HML, and UMD; see Fama and French (1993) and Carhart (1997)). To con-

struct the HML and SMB factors, we augment our data with information from

Thomson Financial’s DataStream database. We take quarterly data on shares out-

16
standing and book value of equity using definitions as similar as possible to those

of Fama and French.

Each quarter, we sort firms into deciles independently along three dimensions:

market capitalization (price per share times shares outstanding), market-to-book

(price per share divided by book value of equity per share), and past returns (over

months t − 12 to t − 2). The small-minus-big (SMB) portfolio consists of a long

position in the top half of stocks by market capitalization, and a short position in

the bottom half. The high-minus-low (HML) portfolio consists of a long position

in those stocks in the top 30 percent of market-to-book, and short position in stocks

in the lowest 30 percent. The up-minus-down (UMD) portfolio consists of a long

position in stocks in the top decile of past return performance, and a short position

in stocks in the bottom decile. We use the overnight interbank lending rate in

Finland as a proxy for the risk-free rate. Beginning in 1999, this is equivalent to

the Euribor rate.

To measure a firm’s risk in a particular year, we regress daily returns in excess

of the risk-free rate on the daily returns of the four factor-mimicking portfolios

and a constant. To ensure that our estimates are not contaminated by well-known

problems associated with nonsynchronous trading, we follow the two-stage least

squares approach of Scholes and Williams (1977, pp. 316–319).

We use the estimated factor betas from these regressions to calculate risk-

adjusted daily returns for each stock, and we use these adjusted returns in the

same performance regressions we previously estimated with raw returns. The

17
risk adjusted return is defined as

 
αi,t = Ri,t − β iM RMRFt + βSi SMBt + β iH HMLt + βU
i UMDt , (2.1)

where Ri,t denotes the raw return of stock i on date t, β iM is stock i’s market beta,

βSi is the beta on the size factor (SMB), β iH is the beta on the value/growth factor

(HML), and βU
i is the beta on the momentum factor (UMD). The returns RMRF,

SMB, HML, and UMD are the returns on the market portfolio, and the three

factor-mimicking portfolios, respectively. Given the daily time-series of factor-

mimicking portfolios constructed as described above, we have a daily series of

risk-adjusted returns, which we can then use instead of raw returns in several

of our tests. As with the raw returns, we calculate holding period returns over a

fixed interval of 30 days, but truncate the holding period at the length of the actual

holding period.

Given our measure of performance, we perform a few tests to be sure that it has

similar properties to the measures used previously, such as poor investor perfor-

mance on average and persistence in performance. However, our main prediction

about performance relates directly to the central hypothesis of our paper:

Prediction 1. Individuals with more investment experience have better investment per-

formance.

We test this conjecture at this point by regressing our performance measure on

our experience measures. Performance is measured as an investor’s average re-

turn over the 30 trading days following each purchase. Our primary experience

18
variables include the number of years that an investor has been in our data and

the cumulative number of trades the investor has placed. We include a quadratic

term for each experience variable to allow investors to learn more slowly over

time. Notably, evidence for this prediction is consistent with investors learning

by trading as predicted by H1, but it does not allow us to differentiate between

H2 and H3. Later in the chapter we adjust for individual heterogeneity and po-

tential survivorship bias. Results of the tests related to performance are presented

in Section 2.3.1.

2.1.2 Measuring disposition

Another outcome variable that we track to evaluate whether investors learn

with experience is the disposition effect. Previous researchers have measured the

disposition effect in a number of ways. Odean (1998) compares the proportion of

losses realized to the proportion of gains realized by a large sample of investors at

a discount brokerage firm. Grinblatt and Keloharju (2001a) model the decision to

sell or hold each stock in an investor’s portfolio by estimating a logit model that

includes one observation for each position on each day that an account sells any

security. Days in which an account does not trade are dropped from their analysis.

As Feng and Seasholes (2005) point out, a potential problem with these and

similar approaches is that they may give incorrect inferences in cases in which

capital gains or losses vary over time. Hazard models, which have been exten-

sively applied in a number of fields including labor economics and epidemiol-

ogy, are ideally suited to our setting. Feng and Seasholes (2005) implement a

19
parametric hazard model. However, they pool their data and estimate the haz-

ard regression only once for all investors together, ignoring most individual-level

heterogeneity and any survivorship bias. Since our focus is on estimating dispo-

sition at an individual level, we estimate the hazard regression for each investor

and year. Implementation of the hazard model uses all data about the investor’s

trading and the stock price path, rather than just data on days when a purchase

or sale is made. That is, it implicitly considers the hold-or-sell decision each day.

This improves our disposition estimates, and gives us more power with which to

investigate learning. More important, individual level estimates enable us to test

H2 and H3 by explicitly accounting for investor attrition and heterogeneity.

We use a Cox proportional hazard model with time-varying covariates to mo-

del the probability that an investor will sell shares that he currently holds. We

count every purchase of a stock as the beginning of a new position, and a position

ends on the date the investor first sells part or all of his holdings. Alternative

definitions of a holding period, such as first purchase to last sale, or requiring

a complete liquidation of a position, do not substantively alter our results. Our

time-varying covariates include daily observations of some market-wide variables

and daily observations of whether each position corresponds to a capital gain or

loss.

Proportional hazard models make the assumption that the hazard rate is

λ(t) = φ(t) exp( x (t)0 β). (2.2)

20
Here, the hazard rate, λ(t), is the probability of selling at time t conditional on

holding a stock until time t, and φ(t) is referred to as the baseline hazard. The

term exp( x (t)0 β) allows the expected holding time to depend on covariates that

vary over time. In our specification, each of the covariates changes daily. Since

we estimate the hazard model for each investor-year, the baseline hazard rate de-

scribes the typical holding period of just one investor in one particular year. The

Cox proportional hazard model does not impose any structure on the baseline

hazard, and Cox’s (1972) partial likelihood approach allows us to estimate the

β coefficients without estimating φ(t). The method also allows for censored ob-

servations, which is important in our setting because investors have not always

closed a position by the end of a given year. Details about estimating the propor-

tional hazard model can be found in Cox and Oakes (1984).

The particular specification we use for investor i’s hazard to sell position j is

n o
λi,j (t| xi,j (t)) = φi (t) exp βdi I ( Ri,j (t) > 0) + βri R̄ M,t + βsi σM,t + βV V
i M,t . (2.3)

The key variable in this regression is I ( Ri,j (t) > 0), an indicator for whether the

total return on position j from the time of purchase up until time t is positive.

Investors who suffer from the disposition effect are more likely to sell when this

condition is true, so they will have positive values of βdi . In the rest of the chapter,

we refer to βdi as the ‘disposition coefficient.’ The total return variable includes

any dividends or other distributions, and is calculated using closing prices on all

days, including the date of purchase. We use the closing price on the purchase

21
date instead of the actual purchase price to ensure that our results are not con-

taminated by microstructure effects.6 In addition, we include as controls three

5-day moving averages of market-level variables to ensure that we are not cap-

turing selling related to market-wide movements: market returns ( R̄ M,t ), squared

market returns (σM,t ), and market volume (VM,t ). We repeat this estimation via

maximum likelihood each year from 1995–2003 for each individual i who places

at least seven round-trip trades in a year.

To investigate learning, we use our disposition estimates as dependent vari-

ables in cross-sectional regressions. Because these estimates are measured with

noise, we use weighted least-squares (WLS) regressions where the weight given

an observation is proportional to the reciprocal of the estimated variance of the

disposition coefficients. The standard errors used in this calculation are calculated

using the robust ‘sandwich’ estimator, with clustering by security.

We perform a few tests on our disposition coefficients to be sure that our mea-

sure has similar properties to the measures used previously. In particular, for

investors to have an incentive to learn to avoid the disposition effect, it must be a

behavioral bias that is costly to them. We examine whether disposition is a some-

what stable, predictable attribute of a particular investor and whether investors

with more disposition have inferior investment performance. However, our main

prediction about disposition is closely related to our central hypothesis:

6 In addition, our data do not include transaction prices during the first three months of 1995,

but we do have closing prices during this period.

22
Prediction 2. Individuals with more investment experience have a weaker disposition

effect.

We test this prediction initially by regressing disposition coefficients on our expe-

rience measures. As in our performance results, we use the number of years that

an investor has been in our data, and/or the cumulative number of trades to mea-

sure experience. Also similar to our earlier tests on performance, support for this

prediction is consistent with H1, but it does not allow us to differentiate between

H2 and H3. Again, later in the chapter we adjust for individual heterogeneity and

survival to differentiate between the two types of learning.

Our empirical strategy differs from that of Feng and Seasholes (2005), who

estimate one large hazard model with all investors pooled together. Their hazard

model includes an experience variable, and an interaction term of I ( Ri,j (t) > 0)

with their experience variable. We avoid this approach because it gives inordinate

statistical weight to investors who trade frequently. To see the sort of problem that

this approach might generate, suppose that investors trade according to different

styles. Some investors trade like day traders, exploiting short-term supply and

demand imbalances, while others trade like fundamental analysts, looking for

larger and longer-term mispricing. If we aggregate traders by assuming that each

of an individual’s trades is a separate, conditionally independent observation, the

inferences we make about learning will be driven by the types of investors who

trade the most. Accordingly, we could mistakenly attribute the improvements

of a select few individuals to the whole population. While we avoid this sort of

23
problem by making an individual-year the unit of observation, we estimate a few

hazard models with the Feng and Seasholes (2005) method to allow comparison

with their results.

2.1.3 Survivorship and heterogeneity

Our predictions so far have ignored heterogeneity and survivorship. How-

ever, we need to consider heterogeneity and attrition to have persuasive infer-

ences about learning and to test H2 and H3. Through wealth effects, significant

heterogeneity in ability can plausibly make it appear as if there is learning when in

fact there is none. More important, simple models that neglect survivorship may

provide evidence consistent with investors learning, but they will not be able to

determine whether poor investors actually improve their ability or simply drop

out after learning about their inherent inability. To avoid these problems, we care-

fully control for investor heterogeneity and survivorship bias. We use two meth-

ods to disentangle the two effects. First, we control for time-invariant unobserved

individual characteristics by including individual fixed effects in our learning re-

gressions. Any additional improvement over time is then likely attributable to

improvement of ability. Second, we directly examine how much ceasing to trade

(or learning about ability) affects our inferences about learning by using a modi-

fied version of the selection model introduced by Heckman (1976).

The classic Heckman model involves a two-stage procedure. In the first stage,

a selection model is constructed to predict which observations will be observable

in the second stage. In the second stage, the regression of interest is estimated

24
with an adjustment for survivorship bias. We modify this procedure to account

for both survivorship bias and individual heterogeneity, adopting the empirical

strategy of Wooldridge (1995), which modifies the Heckman model to allow for

fixed effects. Intuitively, this approach accounts for survivorship by estimating the

selection model every year and including the inverse Mills ratios (the conditional

probability that an individual would not cease to trade) of each selection equation

in the learning regression model. Individual time-invariant heterogeneity is ac-

counted for in this method by running the learning regression in first differences.

More concretely, the learning regression model we estimate is

∆yi,t+1 = β∆xi,t + ρ2 d2t λi,t + . . . + ρ T dTt λi,t + ei,t , (2.4)

where λi,t are the Inverse Mills ratios from a year t cross-sectional probit model

(the selection model) and d2t , . . . , dTt are year dummies. Including these variables

in the learning regression accounts for the impact of the selection equation. Note

that a joint test of ρt = 0 for t = 2, . . . , T is a test of whether survivorship bias is a

concern.

The first-stage uses cross-sectional probit regressions to predict whether or not

the individual ceases to trade in a given period. The probit regressions include a

constant, linear and quadratic experience terms, the number of different stocks the

investor trades, the individual’s average return in the previous year, and the indi-

vidual’s average daily marked-to-market total portfolio value. As instruments, we

use the following variables: (1) a dummy variable for whether an investor inher-

25
ited shares in the previous calendar year; and (2) the cross-sectional standard de-

viation of the individual’s previous-year 30-day return. As we will explain below,

both these variables are likely to satisfy the necessary exogeneity conditions—that

is, they are likely to affect the probability of remaining in the sample, but are un-

likely to affect changes in an individual’s performance or disposition effect except

through their effect on survival.

For our first instrument, we conjecture that an individual who inherits shares

is more likely to trade in the future, perhaps because their wealth has increased,

or because the new shares cause them to pay more attention to the stock market.

This satisfies the exogeneity condition since inheritance of shares from a relative

is unlikely to directly affect changes in the performance or disposition effect of an

individual. In the data, when an investor dies and shares are transferred to an heir,

it appears as a transaction with the account of the deceased selling shares and the

account of the heir purchasing shares. A special code identifies the transaction as

an inheritance. In our sample death transfers are evenly spread over the sample

(ranging from 62 in 1995 to as high as 443 in 2003). Our second instrument is the

variation in the returns across all positions taken by an account in the previous

year. This variation is a measure of the consistency of an investor’s performance,

and we conjecture that an investor with more variable performance is more likely

to stop trading. Again, there is no reason to believe that the consistency of an in-

vestor’s past performance should directly affect changes in his future performance

or disposition effect.

26
 We construct the sample to be used in the selection model as follows. An ac-

count observation is added to the selection sample if it places one or more trades

in a given year. This differs from our main sample, where we require investors to

have placed at least seven round-trip trades in order to estimate either the average

performance or the disposition coefficient. Once an account is added, it remains

in the selection sample until 2003, which is the end of our data. In some years,

an account will have placed enough round-trip trades to be included in our haz-

ard regressions, so the data will include a performance average and disposition

estimate for this account. However, each year we will also have data on many

accounts for which we do not have performance and disposition estimates. If es-

timates are available, we treat the account as having been selected into our data.

2.1.4 Predictions about heterogeneity and survival

In order to differentiate between H2 and H3, we examine the role of investor

heterogeneity and survival in our data. For these to be the important, the hetero-

geneity for which we want to control must be at least partially observable. This

observation leads to our first prediction about heterogeneity:

Prediction 3. There is substantial predictable heterogeneity in investors’ performance

and behavioral bias. Some heterogeneity is correlated with experience.

We test this conjecture by sorting investors into subsamples based on various char-

acteristics that are ex-ante likely to be related to their financial sophistication. For

example, we sort investors by their wealth (proxied by each investor’s average

27
daily portfolio value), by whether or not they trade options, and by several other

characteristics.7 We then estimate the average disposition and performance of

each subsample, testing whether there is a significant difference between the ex-

ante sophisticated and unsophisticated subsamples. We also use each subsample

to estimate our simple learning regressions, and we look at the experience co-

efficients in these subsample regressions to test whether learning across groups

occurs at the same rate.

Our next prediction examines the effect on learning of individual heterogene-

ity more directly:

Prediction 4. Accounting for individual heterogeneity explains a significant portion of

the learning implied by simple models.

We examine this prediction by estimating our learning regressions once again, this

time controlling for investor heterogeneity by including individual fixed effects.

We also include year fixed effects to adjust for time-series variation in average

market returns. We compare the coefficients of our fixed effects regression to the

simple model that we previously estimated.

We go further, exploring our data for potentially important survivorship ef-

fects. For survivorship effects to be important, there must be significant attrition

in the data, which we confirm by testing our next prediction:

7 Our proxy for wealth includes only the equity holdings of investors. This measure could
be negatively correlated with risk aversion since for any level of total wealth an investor who
allocates more money to equities may be more risk-averse. However, even if this is the case, cross-
sectional variation in risk aversion cannot explain changes in the disposition effect, performance
and survival rates over time.

28
Prediction 5. Many new investors will cease trading within a short period of time. Those

who remain will trade more over time.

We examine this prediction by looking at the rate at which investors who are in

our sample in one year (having placed seven or more round-trip trades that year)

continue to be in the sample in subsequent years. The prediction that surviving

traders will trade with more intensity is a feature of standard learning models

like that of Mahani and Bernhardt (2007). In this model, investors do not initially

know their type. As investors trade, they update their subjective probabilities of

being skilled, learning about their inherent ability. Investors with a sufficiently

low probability of being skilled eventually cease trading while those that survive

increase the intensity of their trades. We examine the intensity part of this pre-

diction by looking at the typical pattern of trading volume for a new account that

continues to trade.

Finally, we examine how investor survivorship affects our learning estimates

in our last prediction:

Prediction 6. Accounting for survivorship in addition to individual heterogeneity ex-

plains a significant portion of the learning implied by simple models.

We carefully control for survivorship bias by estimating our learning regressions

with the modified Heckman (1976) correction. It is important for us to control for

survivorship bias, since it is clear that investors with weaker performance will be

less likely to continue trading long enough for us to estimate their performance

in future periods. Comparing the results of our simple models to those that cor-

29
rect for survivorship allows us to estimate what fraction of learning by trading

is driven by learning about inherent ability (H2) and what fraction is driven by

improved ability (H3).

2.2 Data

The data used here are discussed in Chapter I. In addition to those variables,

we use the data to create proxies for wealth and measures of investor sophistica-

tion. To construct a proxy for wealth, we use opening balances and subsequent

trades to reconstruct the total portfolio holdings of each account on a daily basis.

Using these holdings, we approximate wealth as the average daily marked-to-

market portfolio value for each investor. We also calculate the average value of

trades placed by an investor each year. To measure sophistication, we note that

investors who trade options are likely to be more familiar with financial markets.

This is particularly true in our setting because many of the options in our data are

granted to corporate executives as part of compensation. Therefore, while we do

not include options trades in our estimates of disposition, we use whether an in-

vestor ever trades options as a proxy for sophistication. We also count the number

of distinct securities traded by an investor over the sample period, and use this as

a measure of portfolio diversification.

Table 2.1 provides summary statistics for the new accounts in our dataset. New

accounts are accounts that place their first trade in 1995 or a subsequent year.

(That is, they have no recorded initial positions.) Panel A includes all new ac-

30
counts that place at least one trade during our sample period (1995–2003), while

Panel B gives results only for those new accounts for which we are able to es-

timate the disposition coefficient at least once. We only attempt to estimate the

disposition coefficient if an individual has placed at least seven round-trip trades

in a given year, although even with this restriction the procedure to maximize the

likelihood function does not always converge. The last two rows of each panel are

indicator variables, taking a value of one if the investor: (a) trades options; or (b)

is female; and zero otherwise.

Comparing Panels A and B, it is apparent that the subset of investors for whom

disposition coefficients are available is somewhat different than the larger popu-

lation. By construction, the accounts in Panel B place more trades, but they also

have larger portfolios, trade larger amounts of money, trade in a wider selection

of securities, and are somewhat younger. As well, investors for whom we can es-

timate disposition are more likely to trade options (17%) than the overall sample

(3%). Since we are only able to estimate disposition for investors who trade with

some frequency, this likely results from the fact that investors who trade options

are simply more likely to trade in general.

Figure 2.1 shows the number of accounts (including both new and existing ac-

counts) that place one or more trades in each year. There is considerable variation

in the number of accounts placing trades over time, from a low of 54,196 accounts

in 1995 to a high of 311,013 accounts in 2000. Additions of new accounts follows

a similar pattern. We discuss entry and exit from the sample in more detail in the

next section.

31
2.3 Results

We present our empirical findings in this section. We begin by presenting the

results of our tests relating to performance in Section 2.3.1 and to the disposition

effect in Section 2.3.2. Section 2.3.3 examines whether there is heterogeneity in

learning and whether our tests that adjust for heterogeneity change our learning

estimates. Section 2.3.4 examines the importance and magnitude of survivorship

effects, and a number of additional tests are presented in Section 2.3.5.

2.3.1 Performance tests and results

We start by performing a few tests on trader performance to be sure that our

performance measure has similar properties to the measures used in the literature.

Previous papers, in particular Odean (1998), have shown that average investor

performance is worse than that of the market portfolio. Poor performance by av-

erage investors is also a prediction of the model in Mahani and Bernhardt (2007).

This motivates our first test—on average, individuals do not earn returns in excess

of the market return. We test this hypothesis by calculating the average return to

a stock purchased by an individual investor net of the market return. Calculating

this average at a 30-day horizon (using our convention of using a shorter holding

period if the individual sells the stock before 30 days) yields an average return

net of the market of −4.9 percent. At a 60-day horizon, the average net return

is −10.1 percent. At both of these horizons, returns net of the market return are

quite statistically significantly negative.

32
 While the average performance of individual investors is likely to be quite

poor, Coval, Hirshleifer, and Shumway (2005) shows that some individuals per-

sistently outperform others. Again, performance persistence among individuals

is an implication of Mahani and Bernhardt (2007). We test whether there is any

persistence in investor performance in three related ways. For the first approach,

we regress each investor’s average 30-day return in year t on the investor’s av-

erage return in year t − 1 and year fixed effects. Using year fixed effects adjusts

for time series variation in average market returns. The estimated coefficient in

this regression is 0.183 (p < 0.0001), very statistically and economically signifi-

cant. Our second approach is to calculate each investor’s average return in two

disjoint time periods, 1995–1999 and 2000–2003. We then calculate the Spearman

rank correlation between the return series from the first period with that from

the second period. This correlation is 0.164 (p < 0.0001), again quite statistically

and economically significant. Our third test method involves sorting investors in

each year into performance quartiles, and then plotting the average performance

of each of those quartiles for the next several years. This plot, which appears in

Figure 2.2, again gives evidence that the most successful investors in the past con-

tinue to outperform the least successful investors for at least a couple of years.

Results calculated with alphas instead of raw returns are qualitatively the same.

These results confirm that there is a degree of persistence in individual returns.

Our main prediction on performance is that more experienced investors have

better investment performance. For now, we test this prediction by simply re-

gressing performance on experience, experience squared (to allow learning to

33
slow down over time) and some control variables, ignoring any individual het-

erogeneity or survivorship bias. Columns 1 and 2 of Table 2.3 report the results

of this regression. When experience is measured either in number of years or cu-

mulative trades, it is positively and significantly related to average returns. An

additional year of experience increases average 30-day post-purchase returns by

41 − 4 = 37 bp, or approximately 3 percent at an annualized rate. An additional

100 trades increases returns at slightly over one-fourth of this rate. Again, results

estimated with alpha instead of raw returns are quite similar (unreported). While

these estimates are encouraging, the speed of learning they imply seems almost

implausibly large. Taking the regression parameters at face value, an investor

with 8 years of experience should outperform a new investor by about 22 percent

per year. While we observe some heterogeneity in investor ability (or some perfor-

mance persistence) it is not nearly large enough to justify these large coefficients.

2.3.2 Disposition tests and results

The disposition effect is quite large in our data. To give an idea of the eco-

nomic significance of the effect, we present some aggregate evidence of the effect

in our data. Figure 2.3 is a plot of the relation between the propensity to sell an

existing position (the hazard ratio) and the position’s holding period return. To

generate this plot, we group all investors and estimate one hazard model each

year. We group the data for this procedure so we can estimate a model with many

covariates, but almost all of the tests that follow are based on individual-level

results. Rather than using only one indicator variable as in Equation (2.3), we

34
use 20 dummy variables corresponding to different 1 percent return ‘bins’ in this

model. In Figure 2.3 we plot the dummy variable coefficients by year. The sum of

these coefficients times their corresponding dummy variables is multiplied by the

baseline hazard rate to give the actual conditional hazard rate. The conditional

hazard ratio is remarkably similar across years. The plot shows an obvious kink

in the hazard ratio near zero: investors are clearly more likely to sell a stock if

it has increased in value since the purchase date. This provides strong support

for the presence of a disposition effect in aggregate, consistent with the extensive

literature cited above.

Turning to our main individual-level disposition regressions, we require that

an investor place at least seven round-trip trades in a year to be included in the

sample, and we run the regression for each investor-year to generate a separate

disposition coefficient whenever possible. While this filter drastically reduces our

sample size, it is necessary to ensure that our coefficients of interest are identified.

Table 2.2 summarizes the distribution of our disposition estimates, which we

use to investigate several of our predictions. Panel A provides information on all

investors for whom we have estimates. There are 18,042 observations in our panel,

and the number of observations each year rises considerably in the first half of the

sample and then declines somewhat in the latter part of our sample. The median

disposition coefficient is 1.04, which is economically quite large. This coefficient

implies that the median new investor in our data is e1.04 = 2.8 times more likely

to sell a stock whose price is above its purchase price than a stock that has fallen

in value since the time of purchase.

35
 Using the estimated standard errors for each investor, we can classify estimates

as significant or not at any given confidence level. The last two columns of Panel A

show the proportion of investors who have a significantly positive or negative

disposition coefficient at the 10 percent level. Over our entire sample period, 41.8

percent of investors have a disposition coefficient that is statistically greater than

zero. Panel B gives summary statistics for the other coefficients in the hazard

model. None of the controls is statistically significant in the cross-section.

Before we can consider whether investors learn to avoid the disposition effect,

we need to argue that the effect is in fact a behavioral bias. Theoretically, there is

no particularly well accepted model, either rational or irrational, that produces the

disposition effect. It is difficult to imagine a rational model that can produce the

effect. It is not particularly difficult to think of a model in which the probability of

selling a stock increases in the stock’s unrealized return, but it is difficult to think

of a model in which the probability of selling rises dramatically as soon as the

return becomes positive—that is, a model that predicts a ‘kink’ at zero, as show in

Figure 2.3.

It is also difficult to design a model in which traders with the disposition ef-

fect have the characteristics that our traders have. In particular, one necessary

condition for disposition to be a behavioral bias is that investors with more dispo-

sition have inferior investment performance. If disposition is unrelated to invest-

ment performance, investors with the effect would have little incentive to learn to

avoid it. To get a sense of how returns vary with disposition, we examine average

investor returns across quintiles of the disposition coefficient. In this sort, the dis-

36
position coefficients are always estimated one year before the average returns are

calculated, so disposition coefficients and average returns are not mechanically

correlated in any way. For each quintile, Figure 2.4 graphs the average return

earned by investors over different horizons from the purchase date. Returns are

substantially higher in the lowest disposition quintile than in the highest dispo-

sition quintile. For example, in the 30 days following a purchase, a stock’s price

increases 46 bp on average when bought by an investor in the lowest disposition

quintile, compared to a decline of 54 bp if purchased by an investor in the high-

est disposition quintile. The differences between high- and low-quintile average

returns range from 17 bp at the 10-day horizon to 131 bp at the 45-day horizon.

These differences are both economically and statistically large. These results are

consistent with the claim that individuals with high disposition effect coefficients

have relatively poor investment performance. They are also consistent with the

disposition effect being a behavioral bias that investors want to learn to avoid.

Another necessary condition for disposition to be a behavioral bias is that dis-

position is a somewhat stable, predictable attribute of a particular investor. We test

this conjecture by estimating the disposition effect at the investor level in adjacent

time periods. Each set of estimates comes from a completely disjoint dataset. Any

trades that are not closed at the end of the first period are considered censored

in the model estimated with first period data. Therefore, any trades that are not

closed at the end of the first period are completely ignored in the model estimated

with second period data. We explore the stability of disposition coefficients by es-

timating the rank correlation of account-level disposition coefficients over the two

37
periods, testing whether the rank correlation is significantly different from zero.

We estimate the rank correlation between an investor’s disposition coefficient in

year t and their coefficient in year t − 1 to be 0.364, suggesting that there is a fair

degree of persistence in the individual’s disposition coefficient. This correlation is

extremely statistically significant.

Taken together, Figures 2.3 and 2.4, Table 2.2 and the correlation in the previous

paragraph provide strong evidence that the disposition effect is a widespread and

economically important behavioral bias that is present in each year of our study.

Finally, we examine whether more experienced investors are more likely to

avoid the disposition effect, testing our main prediction related to disposition.

Again at this point, we simply regress disposition coefficients on our experience

variables, those variables squared, and some control variables. Columns 3 and 4

of Table 2.3 present our results for the disposition learning regressions. To reduce

the weight given to disposition coefficients that are not estimated very precisely,

we estimate the regressions with weighted least squares, where the weights are

\
proportional to 1/Var ( βd ) from our hazard regression in Equation (2.3). The base

case (Column 3) shows that disposition declines with experience (β 1 < 0). More-

over, investors tend to slow down in their learning as they gain experience since

β 2 > 0. Frequent traders, investors who trade more securities, and investors who

earned higher returns in the previous year all have lower levels of disposition, but

even with these controls our base results are qualitatively unchanged.

Column 4 indicates that an additional 100 trades reduces the disposition co-

efficient by 0.041, which is similar to the coefficient on Experience in Column

38
3. In other words, a year of experience or 100 trades have approximately the

same effect on disposition. In each of the specifications the estimated YearsTraded

and CumulTrades coefficients are statistically significant at the 1 percent level.

Economically, however, our results suggest that investors learn relatively slowly.

Specifically, the estimates in Column 3 suggest that an additional year of expe-

rience corresponds to a reduction in the disposition coefficient of approximately

0.05. To provide some context for this estimate, note that the unconditional me-

dian disposition coefficient in our sample is 1.04. An extra year of experience

decreases this by about five percent.

As discussed in Section 1.2, we also estimate the yearly hazard models pre-

sented in Table 2.4, which are comparable to the model of Feng and Seasholes

(2005).8 These estimates are from pooled hazard models estimated each year, in

which all individuals are treated as if they were just one person. Experience is

interacted with an indicator for whether the price is above the purchase price,

and the coefficient on this interaction term is interpreted as a learning coefficient.

These models again give evidence of learning. However, the learning coefficient

estimates are quite variable over time (0.149 to 0.063), and they are statistically

insignificant in three of the nine years. Furthermore, the average disposition co-

efficient of an investor is estimated in aggregate to be around 0.65 in this model.

This suggests that, depending on the period chosen, an additional year of experi-
8 These regressions are essentially the same as the empirical strategy employed by Feng and

Seasholes (2005), except we use a proportional hazard regression rather than the parametric
Weibull model. Our indicator variable is the same as their ‘Trading Gain Indicator’ (TGI). The
experience variable we use for this model is the total number of trades placed before the current
trade, rather than CumulTrades, which is the total number of trades placed in previous calendar
years.

39
ence corresponds to a reduction in the disposition coefficient of approximately 10

percent to 25 percent, significantly higher than our regression estimate. Table 2.4

also lists the number of observations available each year, and the fraction of the

observations that are censored, or the purchased stocks that are not sold by the

end of each year.

The low level of the average disposition effect, the high variability in the an-

nual learning estimates, and the high level of learning found both by Feng and

Seasholes (2005) and in our implementation of their model suggest that there are

significant differences between our approach and theirs. One important difference

is that, since we estimate a different disposition coefficient for each individual,

those who place a large number of trades have the same weight in our analysis

as those who place seven or eight trades. In Feng and Seasholes (2005), an indi-

vidual’s weight is proportional to her trading volume. Another difference is that

we estimate learning over a much longer period of time, since Feng and Seasholes

(2005) only have about two years of transactions data. Thus, the year-to-year vari-

ation in the annual estimates in Table 2.4 is much less of a concern for our analysis.

2.3.3 Heterogeneity in learning

To examine if there is significant predictable heterogeneity among individuals

in our sample, we separate investors into a number of different groups by observ-

able characteristics that we believe are related to their financial sophistication.

We examine the average disposition and performance of each of these groups,

confirming that our priors are correct and demonstrating that there is significant

40
heterogeneity in the data. We also estimate the simple learning regression for dis-

position and returns for each group, predicting that the less sophisticated investor

groups will learn faster than the more sophisticated investor groups. Importantly,

we do not classify investors on the basis of the estimated disposition coefficient,

βd , because of concerns about measurement error. That is, the most extreme dis-

position estimates are likely those with the most error, and we would therefore

expect these accounts to see a decrease in disposition in future years, even if these

investors are not really learning. To avoid sorting on measurement error, we fo-

cus instead on observable variables that are ex-ante related to performance and

disposition.

Each row of Table 2.5 displays the mean of the disposition coefficient and the

average returns (or performance) of each group, as well as the regression coeffi-

cient of these variables on YearsTraded and the number of observations used in

the calculations. Results for disposition are shown in Columns 1 and 2, and for

returns in Columns 3 and 4. We consider investors ex-ante likely to be relatively

sophisticated if they trade options, have significant wealth, are men, or have had

relatively good past performance.9 Looking at the table, it is clear that the means

for each of our sophistication subgroups is significantly different, and each change

in the mean across subgroups is of the sign we expect. We also expect that less so-

phisticated investors will learn faster than more sophisticated investors. In each

pair of rows of the table this prediction is confirmed. In most cases there is a clear
9 We classify investors who make excess profits that are in the top quarter of the entire market
in the first two years of their trading as ‘winners.’ Our results are not sensitive to alternative
definitions of winners, such as using a one- or three-year classification period, or above-median
excess returns.

41
difference between the unsophisticated investors, who learn to avoid the disposi-

tion effect at a rate of about 10 percent per year, and sophisticated investors, for

whom the learning coefficient is often insignificant.

It is also plausible that investors learn more when the market in general is not

doing well. During periods of high market returns, investors’ incentives to learn

about their biases could be reduced if they attribute their success to their ability,

similar to the behavior modeled in Zingales and Dyck (2002) in the context of me-

dia and bubbles. Thus, we should find that investors are more likely to learn when

the markets are not doing well rather than when they are. To test this we define

the state of the market as an ‘up-market’ if the excess return on a broad Finnish

index is positive in a given year, and as a ‘down-market’ if the excess return is

negative. In the last two rows of the table we re-estimate our base regressions for

each of the two states of the market and find that, as we hypothesized, individuals

learn to avoid the disposition effect primarily when the market is not doing well.

These results are consistent with Prediction 3 and suggest that there is substan-

tial heterogeneity both in initial ability (performance and disposition) and in rates

of learning. It is unsophisticated investors and investors who start out with poor

returns who learn most. These results suggest that considering heterogeneity in

our learning estimates will be important.

We control for time-invariant unobserved individual characteristics by includ-

ing individual fixed effects in our learning regressions. We also control for market

returns and any other time-varying features of performance or disposition by in-

42
cluding year fixed effects in the regressions. Specifically, we estimate:

yi,t+1 = αi + β 1 Experiencei,t + β 2 Experience2i,t + δXi,t + γt + ei,t , (2.5)

where αi is the individual specific fixed effect, γt is the time fixed effect and Xi,t

are other controls. The performance results, reported in Columns 1 and 2 of Ta-

ble 2.6, suggest that an investor with one year of experience will earn 22 bp more

than an inexperienced investor over a 30-day horizon. Column 2 indicates that a

similar increase in returns comes from an additional 400 trades. The disposition

results appear in Columns 3 and 4. While YearsTraded is no longer significant in

these regressions, Column 4 suggests that 100 trades reduces the disposition co-

efficient by approximately 0.03. Comparing these estimates with those reported

in Table 2.3, we find that the learning estimates after controlling for individual

fixed effects fall roughly by one-half. Consistent with Prediction 4, this suggests

that though investor performance improves and disposition declines with expe-

rience, accounting for individual heterogeneity reduces the estimates by about 50

percent.

2.3.4 Survivorship effects

In this section, we explore our data for potentially important survivorship ef-

fects. For these effects to be important, there must be significant attrition in the

data. In other words, many new investors should cease trading within a short

period of time. The model of Mahani and Bernhardt (2007), which suggests that

43
attrition is driven by investors learning about their inherent ability, also predicts

that those traders who remain will trade more intensely over time.

We first present some overall attrition evidence by examining the rate at which

investors who are in our sample in one year (having placed 7 round-trip trades in

one year) fail to place any trades during the rest of our sample period. Since the

rest of the sample period changes from year to year, the earlier years of our sam-

ple period provide more reliable estimates of true exit rates than the later years.

Figure 2.5 shows that attrition is a significant feature of our data. Approximately

25 percent of those traders who enter the sample in one year fail to ever trade

again. Of traders who trade for two or three years, about 5 percent permanently

exit the sample. This is consistent with our Prediction 5.

To check whether trading intensity changes over time, we estimate regressions

with both the number of trades placed and total trade value as dependent vari-

ables, and experience and year dummy variables as explanatory variables. Includ-

ing year dummies ensures that market-wide changes in stock characteristics will

not contaminate our results. We plot the results of these regressions in Figure 2.6.

The plot clearly shows that, conditional on survival, trading intensity increases

over time. This supports the part of Prediction 5 that has to do with trading inten-

sity, which is consistent with the model of Mahani and Bernhardt (2007).

In some of our last sets of results, we carefully account for a survivorship effect

by estimating our learning regressions using the method proposed by Wooldridge

(1995), which is a standard Heckman (1976) correction modified to account for

individual time invariant heterogeneity. Results from the selection model, with

44
two-step efficient estimates of the parameters and standard errors, are given in

Table 2.7. The first-stage selection model uses 30,218 observations, while the

second-stage regression (in first differences) use only 6,511 observations in the

performance regression and 8,818 observations in the disposition regressions.

We estimate the first-stage regression for each year and construct inverse Mills

ratios for each year. For brevity, we only report one set of pooled first-stage esti-

mates in Column 1 of Table 2.7. Results for each of the years are qualitatively sim-

ilar to those reported. We find strong evidence that as investors get bad returns

they cease trading. In particular the estimate on R̄t−1 is positive and significant.

The estimate is also economically meaningful and suggests that, keeping other

explanatory variables at their mean levels, a decrease in returns of one standard

deviation increases the probability that the individual will cease to trade next pe-

riod by around 15 percent. This is strong evidence for H2. As low ability investors

trade, they learn about their inherent ability and cease trading. More successful

investors continue to trade actively.

The other coefficient estimates reported in the first column of Table 2.7 also

seem sensible: investors are more likely to remain in the sample and trade if they

hold relatively diversified portfolios and have relatively more trading experience.

Importantly, coefficient estimates for both of our instruments are statistically sig-

nificant and of the predicted sign. Specifically, inheriting shares increases the

probability that the individual will continue trading the next period and higher

variability in past performance increases the probability that the individual ceases

to trade. Both the instruments also have an economically significant impact. For

45
instance, keeping other variables at mean levels, inheriting shares increases the

probability that the investor will continue trading in the next period by 5 percent.

In line with Prediction 6, we find that accounting for selection has a significant

impact on our learning estimates. Column 2 uses performance as the dependent

variable in a regression of the form of Equation (2.4). Comparing the estimates in

Table 2.7 to the simple model reported in Table 2.3, the coefficient on YearsTraded

is no longer statistically significant, and the coefficient on cumulative trades is re-

duced by about 90 percent.10 When we use disposition as the dependent variable

in Column 3, the coefficient is reduced by slightly more than 50 percent. The joint

tests of statistical significance of the inverse Mills ratios in Columns 3 and 4 also

show that accounting for sample selection is important. Unreported results in

which YearsTraded and CumulTrades are included in separate regression models

yield almost the same coefficients, but regressions that include both variables are

reported for brevity.

Our results suggest that accounting for selection is important and significantly

affects inferences about learning. Investor heterogeneity and survivorship effects

account for something on the order of one-half to three-quarters of the learning es-

timates found in simple and aggregate models. This translates directly into slower
10 The coefficients on YearsTraded in these regressions must be viewed with caution. Because
the change in YearsTraded is always equal to exactly one year, the coefficient on YearsTraded can
only be identified if we leave the fixed effect for one year out of the regression. We leave the fixed
effect for 1997 out of the regression, so the coefficient we report can be thought of as the learn-
ing coefficient for that year. For robustness, we also estimate standard Heckman selection models
without either individual or year fixed effects. The coefficients on YearsTraded in these models
are 0.31 in the performance regression and -0.019 in the disposition regression. Both of these coef-
ficients are marginally statistically significant (at 10 percent). The magnitude and significance of
these coefficients is consistent with our other results.

46
learning than that inferred from simpler models. Taking the disposition-learning

coefficient, for example, 100 trades corresponds to an improvement of about 0.04

in the simplest model, an improvement of about 0.03 in the model with individual

fixed effects, and an improvement of about 0.02 in the survivorship/fixed effects

model. Roughly speaking, if it takes about 100 trades to improve about 4 percent

in the simple model, it takes about 200 trades to achieve the same improvement

after adjusting for survivorship and individual heterogeneity.

Our estimates suggest that the fraction of learning that is driven by investors

learning about their inherent ability (i.e., H2) by trading is large. After adjusting

for this type of learning, the portion of learning that is due to investors learning

to improve their ability over time (i.e., H3) is significantly different from zero, but

not excessively large. In other words, there is support for H3, though accounting

for H2 significantly reduces estimates of how quickly investors become better at

trading. Overall, our findings are consistent with the three hypotheses that were

outlined in Section 2.1.

2.3.5 Other Tests

In this section we conduct some additional tests related to our main predic-

tions. First, it is possible that the performance improvement with experience that

we find is not entirely due to learning of the types we considered but rather due

to a change in the risk preferences of investors over time. To address this possi-

bility, we estimate our learning regressions with risk-adjusted returns (or 30-day

alphas) instead of raw returns, and we regress the average factor betas of stocks

47
purchased by investors on experience and our control variables. Our regressions

also control for survivorship and individual and year fixed effects. The results

of our regressions appear in Table 2.8. The table clearly shows that risk-adjusted

returns improve with experience, with a coefficient that is actually larger than the

coefficient we estimate for raw returns. Looking at the coefficients on average

factor betas makes it clear why this is the case. With more experience, investors

are actually both improving raw returns and taking less risk, or purchasing stocks

with lower factor betas. This result is particularly strong for the market (RMRF)

and size (SMB) factor betas. Thus, it appears very unlikely that our raw return

learning results are driven by changes in risk preferences with experience.

Second, in unreported results, we substitute the market return for each stock’s

return to see if individuals learn to time the market. If investors are learning to

identify good times to buy then the market as a whole will tend to increase after

their purchases; if instead they are learning to select stocks, we will not find evi-

dence of learning when we look only at market returns. In fact, we find that the

coefficient estimates on experience variables are insignificant, which suggests that

performance improved because investors became better at stock selection. Third,

we also conduct the tests on survivorship using the Wooldridge (1995) method,

taking data on investors who resume trading after ceasing to trade for a few years

(the tests reported in the last section had dropped such investors, using only ob-

servations in two consecutive years). Including these investors increases the sam-

ple by around 250 observations in the second stage but does not affect the nature

of the results reported. Fourth, all of the results on disposition remain qualita-

48
tively unchanged if we include a ‘December dummy’ in (2.3) or remove partial

sales from our sample. This rules out tax-motivated selling or rebalancing as pos-

sible explanations for the disposition effect. Finally, in all the fixed effect regres-

sions that control for individual heterogeneity, we cluster the standard errors at

the individual level and find that our results are unaffected.

2.4 Conclusion

We examine learning in a large sample of individual investors in Finland dur-

ing the period 1995–2003. We correlate performance and disposition with investor

experience and investor survival rates to determine whether and how investors

learn by trading. We find that performance improves and the disposition effect

declines as investors become more experienced, suggesting that investors learn

by trading. We differentiate between investors learning about their inherent abil-

ity by trading and learning to improve their ability over time by accounting for

investor attrition. We find that a substantial part of this learning occurs when

investors stop trading after learning about their inherent ability rather than con-

tinuing to trade and improving their ability over time. By not accounting for

investor attrition and heterogeneity, the previous literature significantly overes-

timates how quickly investors become better at trading.

Our results suggest a number of interesting implications. First, since investors

who continue trading learn slowly and there is great deal of turnover in the in-

vestor population, it is likely that behavioral biases are an important feature of

49
financial markets. Agents do not learn fast enough to make it impossible for bi-

ases to affect asset prices. Second, while it would be wonderful to know how

quickly investors who cease to trade would learn if they chose to continue trad-

ing, we have no way to estimate this speed. If we assume that those who continue

trading learn more quickly than those who cease to trade, policy makers might

enhance welfare by devising screening mechanisms, or tests that measure and re-

veal inherent investing ability. Allowing unskilled investors to learn of their poor

ability without incurring significant costs might be more valuable than encourag-

ing people to become active investors. Third, an open question in the literature

is why there is such high trading volume, particularly among seemingly unin-

formed individual investors. Our results indicate that such trading may be ratio-

nal; investors may be aware that they will learn from experience, and choose to

trade in order to learn. Our results also suggest that differences in the expected

performance of investors may arise from different experience levels. Finally, if

many inexperienced investors begin trading around the same time, their trades

could lead to time-varying market efficiency. Our evidence is therefore consis-

tent with the recent results of Greenwood and Nagel (2007), and the more general

discussion found in Chancellor (2000) and Shiller (2005).

50
 Table 2.1: Summary Statistics

This table presents summary statistics for our data. Panel A includes all individual accounts in our
data that started trading during the sample period. Panel B gives results just for those accounts for
which we are able to estimate at least one disposition coefficient. We only estimate the disposition
coefficient if an individual has placed at least seven round-trip trades in a given year. Number
of trades is the total number of trades placed by an investor during the sample period. Average
portfolio value is the average marked-to-market value of an investor’s portfolio using daily closing
prices.

Panel A: Entire Sample (322,454 accounts)

Mean 25th Pctl Median 75th Pctl
Number of years with trades 1.9 1.0 1.0 2.0
Number of securities traded 3.5 1.0 1.0 3.0
Number of Trades 15.4 1.0 3.0 8.0
Average value of shares traded, EUR 3,447 808 1,653 3,310
Average portfolio value, EUR 11,588 1,470 2,794 5,856
Age in 1995 39.3 27.0 39.0 51.0
Gender (1=female) 0.39
Trades options (1=yes) 0.03

Panel B: Accounts with Disposition Estimates (11,979 accounts)

Mean 25th Pctl Median 75th Pctl
Number of years with trades 4.4 3.0 4.0 6.0
Number of securities traded 22.3 12.0 18.0 28.0
Number of trades 222.3 68.0 117.0 224.0
Average value of shares traded, EUR 5356 1855 3235 5759
Average portfolio value, EUR 58828 5102 11483 26147
Age in 1995 35.3 27.0 34.0 44.0
Gender (1=female) 0.15
Trades options (1=yes) 0.17

51
 Table 2.2: Disposition Estimates

This table reports a number of summary statistics for our estimates of the disposition effect. βd is
the coefficient in the hazard regression,
n o
λi,j (t| xi,j (t)) = φi (t) exp βdi I ( Ri,j (t) > 0) + βri R̄ M,t + βsi σM,t + βV
i VM,t .

We estimate this model for each account-year with seven or more round-trip trades. Panel A re-
ports on the cross-section of coefficients estimated each year. The columns labeled ‘positive’ and
‘negative’ report the proportion of investors with a statistically significant coefficient, where signif-
icance is measured at the 10% level using standard errors obtained from the maximum likelihood
estimation of the hazard model. Panel B provides information for all the variables in the hazard
model, where βr , βs , and βV are the coefficients on 5-day moving averages of market returns,
market returns squared, and market volume, respectively.

Panel A: All Accounts with Disposition Estimates

βd estimate Significant at 10%
Year N Obs 10th Pctl Median 90th Pctl Positive Negative
1995 25 -0.34 0.72 2.37 28.0% 4.0%
1996 89 -0.38 0.99 2.39 38.2% 2.2%
1997 248 -0.91 0.96 2.31 33.1% 3.2%
1998 695 -0.69 1.05 2.66 35.4% 2.6%
1999 1958 -0.51 1.08 2.57 40.4% 1.7%
2000 5961 -0.37 1.00 2.54 41.5% 1.5%
2001 3732 -0.29 1.01 2.53 42.7% 1.3%
2002 2649 -0.31 1.10 2.70 45.6% 1.1%
2003 2685 -0.33 1.09 2.53 41.6% 1.3%
All years 18042 -0.36 1.04 2.57 41.8% 1.5%

Panel B: Hazard Function Estimates

Variable Mean t-stat 10th Pctl Median 90th Pctl
βr -0.04 -0.27 -0.65 0.03 0.93
βs 0.01 0.46 -0.05 0.00 0.04
βV 0.25 9.14 -2.62 0.24 3.08
βd 1.32 6.81 -0.36 1.04 2.57

52
 Table 2.3: Simple Learning Model: Estimates at Individual Level

This table presents results for regressions of the form

yi,t+1 = α + β 1 Experiencei,t + β 2 Experience2 + δXi,t + ei,t

where the dependent variable is either the investor’s average 30-day return following purchases
( R̄i,t+1 ) or the investor’s disposition coefficient (βdi,t+1 ). Experience is measured by either years of
experience (YearsTraded) or cumulative number of trades placed (CumulTrades). Xi,t is a vector
of controls including the number of trades placed by the individual in a given year (NumTrades),
the number of securities held by the individual in a given year (NumSec), and the individual’s
average total daily portfolio value (PortVal). Data are from the period 1995 to 2003. Standard
errors are in parentheses, and ∗∗∗ , ∗∗ and ∗ denote significance at 1%, 5% and 10%, respectively.

Dependent Variable
R̄i,t+1 βdi,t+1
(1) (2) (3) (4)
YearsTradedt 0.414 -0.050
(0.160)∗∗∗ (0.014)∗∗∗

YearsTraded2t -0.043 0.007

(0.027) (0.002)∗∗∗

CumulTradest (÷102 ) 0.110 -0.041

(0.043)∗∗∗ (0.003)∗∗∗

CumulTrades2t (÷104 ) -0.0005 0.0002

(0.0003)∗ (0.00002)∗∗∗

NumSect 0.096 0.104 -0.011 -0.012

(0.012)∗∗∗ (0.012)∗∗∗ (0.001)∗∗∗ (0.001)∗∗∗

NumTradest -0.031 -0.043 0.0006 0.005

(0.005)∗∗∗ (0.007)∗∗∗ (0.0005) (0.0006)∗∗∗

PortValt (÷106 ) 0.104 0.099 -0.006 -0.003

(0.061)∗ (0.058)∗ (0.003)∗∗ (0.003)

R̄t−1 -0.005 -0.005

(0.001)∗∗∗ (0.001)∗∗∗

Observations 13404 13404 17715 17715

R2 (%) 1.6 1.7 1.1 1.6

53
 Table 2.4: Simple Learning Model: Disposition Estimates at Aggregate Level

This table presents learning estimates from pooled proportional hazards models, using a method similar to that of Feng and Seasholes (2005). Each
year we pool the trades of all investors, treating them as if they were just one individual and estimating one learning coefficient for the entire
population. The model is,

λ(t) = φ(t) exp{ β 1 I ( Ri,j (t) > 0) + β 2 Exper + β 3 Exper2 + β 4 Exper[ I ( Ri,j (t) > 0)] + β 5 Exper2 [ I ( Ri,j (t) > 0)]},

where Exper is measured in years since first placing a trade and I ( Ri,j (t) > 0) is an indicator variable that takes a value of one when a stock has
increased in price since its purchase date. For brevity, only the β 4 coefficient estimates are reported. Estimated β 5 coefficients are all insignificant. We
also report the number of trades, or observations, considered by the model (in thousands of trades), the percentage of observations censored (trades

54
not closed by the end of the year), and the number of accounts contributing observations to the model. Standard errors are in parentheses, and ∗∗∗ ,
∗∗ and ∗ denote significance at 1%, 5% and 10%, respectively.

Year 1995 1996 1997 1998 1999 2000 2001 2002 2003
β4 -0.074 -0.093 -0.096 -0.149 -0.031 -0.013 -0.063 -0.069 -0.123
(0.039)∗ (0.062) (0.031)∗∗∗ (0.022)∗∗∗ (0.024) (0.011) (0.011)∗∗∗ (0.012)∗∗∗ (0.012)∗∗∗

Trades 6.6 14.5 26.6 44.6 99.6 251.9 161.9 115.6 131.8
Censored 16% 19% 19% 19% 17% 16% 17% 18% 20%
Accounts 384 713 1360 2412 4953 11585 7445 5416 6056
 Table 2.5: Heterogeneity in Learning
This table reports both means and simple learning coefficient estimates from regressions of the
form,
yi,t+1 = α + β 1 YearsTradedi,t + β 2 YearsTraded2i,t + δXi,t + γt + ei,t ,

conditioned on a number of variables that ex-ante might be correlated with trader sophistication.
The variable of interest is either the disposition coefficient (βdi,t+1 ) or returns ( R̄i,t+1 ). For brevity
we only report β 1 coefficients in the table. We classify investors as ‘Trades options’ if they trade
in options at any point during our sample. Similarly, investors are classified as ‘wealthy’ if they
are in the top 25th percentile of average portfolio value. We define the state of the market as ‘up’
if the excess return on a broad market index is positive, and ‘down’ if it is negative. Finally, we
classify investors who make excess profits that are at or above the 75th percentile of all investors
in the first two years of their trading as ‘winners.’ We include year dummies in all the regressions.
Data are from the period 1995 to 2003. Standard errors are in parentheses and ∗∗∗ , ∗∗ and ∗ denote
significance at 1%, 5% and 10% respectively. All group means are significantly different at 1%
level.

Dependent Variable
βdi,t+1 R̄i,t+1
Classification Mean YearsTraded Mean YearsTraded N Obs
No options trades 1.17 -0.096 -0.81 0.501 26698
(0.024)∗∗∗ (0.184)∗∗∗

Trades options 0.99 -0.031 0.10 0.233 8211

(0.032) (0.299)

Not wealthy 1.14 -0.101 -1.18 0.389 26176

(0.023)∗∗∗ (0.169)∗∗

Wealthy 1.11 -0.015 1.08 0.191 8733

(0.043) (0.339)

Females 1.24 -0.140 -0.38 0.482 4703

(0.047)∗∗∗ (0.129)∗∗∗

Males 1.11 -0.078 -0.62 0.361 30206

(0.020)∗∗∗ (0.169)∗∗

Not winners 1.15 -0.840 -0.62 0.732 32456

(0.021)∗∗∗ (0.161)∗∗∗

Winners 0.96 -0.039 -0.03 -0.641 2453

(0.074) (0.703)

Down market 1.14 -0.120 -0.70 0.619 27074

(0.022)∗∗∗ (0.179)∗∗∗

Up market 1.08 0.028 1.77 0.114 7835

(0.077) (0.471)

55
 Table 2.6: Learning with Individual Fixed Effects
This table reports estimates of regressions of the form

yi,t+1 = αi + β 1 Experiencei,t + β 2 Experience2i,t + δXi,t + γt + ei,t ,

where Experience is measured by either years of experience (YearsTraded) or cumulative number

of trades placed (CumulTrades). The dependent variable in models (1) and (2) is individual i’s
average return in the following year, R̄i,t+1 , and in models (3) and (4) it is the individual’s dispo-
sition coefficient, βdi,t+1 . Xi,t is a vector of controls including the number of trades placed by the
individual in a given year (NumTrades), the number of securities held by the individual in a given
year (NumSec), and the individual’s average total daily portfolio value (PortVal). We also include
year dummies and individual-specific intercepts in each regression. Data are from the period 1995
to 2003. Standard errors are in parentheses and ∗∗∗ , ∗∗ and ∗ denote significance at 1%, 5% and
10% respectively.

Dependent Variable
R̄i,t+1 βdi,t+1
(1) (2) (3) (4)
YearsTradedt 0.249 -0.010
(0.401) (0.038)

YearsTraded2t -0.022 0.010

(0.038) (0.003)∗∗∗

CumulTradest (÷102 ) 0.058 -0.030

(0.059)∗∗∗ (0.005)∗∗∗

CumulTrades2t (÷104 ) -0.0009 0.0002

(0.0004)∗∗ (0.00003)∗∗∗

NumSect 0.003 0.002 -0.008 -0.008

(0.024) (0.024) (0.002)∗∗∗ (0.002)∗∗∗

NumTradest -0.021 -0.029 -0.0007 0.0003

(0.010)∗∗ (0.010)∗∗∗ (0.0009) (0.0009)

PortValt (÷106 ) 0.618 0.587 -0.067 -0.065

(0.602) (0.601) (0.602) (0.601)

R̄t−1 -0.002 -0.002

(0.001)∗ (0.001)∗

Observations 13404 13404 17715 17715

Adjusted R2 (%) 45.7 45.9 38.7 38.6
Year Fixed Effects Yes Yes Yes Yes
Individual Fixed Effects Yes Yes Yes Yes

56
 Table 2.7: Learning with Survival Controls
This table reports estimates of selection model regressions with the fixed effects modification developed by Wooldridge
(1995). The regressions are of the form

∆yi,t+1 = β∆Xi,t + ρ2 d2t λi,t + . . . + ρ T dTt λi,t + ei,t ,

where λi,t are the inverse Mills ratios from a year t cross-sectional probit model (the selection model) and d2t , . . . , dTt are
time dummies. Including these variables in the learning regression accounts for the impact of the selection equation. The
joint test of ρt = 0 for t = 2, . . . , T is a test of whether survivorship bias is a concern. The probit model reported in Column
1 is estimated with data from all of the years of the sample pooled together, but the inverse Mills ratios in the second stage
estimates are estimated separately each year. The regressions in Columns 2 and 3 are estimated with all the variables in first
differences, except the inverse Mills ratios. These first differences add fixed effects to the model. Experience is measured by
years of experience (YearsTraded) and cumulative number of trades placed (CumulTrades). The dependent variable in the
second stage is either the individual’s average return in the following year, R̄i,t+1 , or the individual’s disposition coefficient,
βdi,t−1 . Xi,t is a vector of controls including the number of trades placed by the individual in a given year (NumTrades),
the number of securities held by the individual in a given year (NumSec), and the individual’s average total daily portfolio
value (PortVal). R̄t , the investor’s average return, is excluded from the return regression due to the econometric problems
that arise when lagged dependent variables are included in fixed effects models. Additional variables in the first stage
selection model are I (Inherit = 1), an indicator variable for whether the investor inherited shares in year t, and σRt , the
standard deviation of the investor’s returns. Data are from the period 1995 to 2003. Standard errors are in parentheses, and
∗∗∗ , ∗∗ and ∗ denote significance at 1%, 5% and 10% respectively.

First Stage Second Stage Second Stage

(All Years) (First Difference) (First Difference)
Dependent variable Insamplei,t+1 =1 R̄i,t+1 βdi,t+1
CumulTradest (÷102 ) 0.355 0.010 -0.019
(0.021)∗∗∗ (0.004)∗∗∗ (0.002)∗∗∗

CumulTrades2t (÷104 ) -0.002 -0.0003 0.0002

(0.0001)∗∗∗ (0.0002) (0.00003)∗∗∗

YearsTradedt 0.376 0.894 -0.099

(0.0196) (1.766) (0.204)

YearsTraded2t 0.028 -0.015 0.008

(0.007)∗∗ (0.021)∗∗∗ (0.006)

NumSect 0.187 0.034 -0.010

(0.001)∗∗∗ (0.009)∗∗∗ (0.002)∗∗∗

NumTradest 0.963 -0.100 0.002

(0.062)∗∗∗ ( 0.012)∗∗∗ (0.003)

PortValt (÷106 ) 0.003 0.220 -0.007

(0.003) (0.221) (0.007)

R̄t 0.002 -0.003

(0.001)∗∗∗ (0.002)∗

I (Inherit = 1) 0.191
(0.074)∗∗∗

σRt -0.617
(0.061)∗∗∗

Joint Test of ρ = 0 F(8, 6490) = 3.91 F(7, 8799) = 5.29

(from 1996-2003) Pr>F = 0.0000 Pr>F = 0.0000
Observations 30218 6511 8818
Year Fixed Effects Yes Yes
Individual Fixed Effects Yes Yes

57
 Table 2.8: Risk Taking and Experience

This table reports the results of fixed effect selection model estimates of regressions of various
performance and risk measures on experience measures. The method of these regressions and
their associated first-stage estimates are described in Table 2.7. The dependent variables include
each investors average 30-day risk adjusted return (alpha), and each investors average beta coef-
ficient on four factors—RMRF, SMB, HML, and UMD. These betas are estimated in the standard
way, as described in the text. Each regression includes the control variables and inverse Mills
ratios described in Table 2.7, but only the experience variables coefficients are reported in this ta-
ble. Standard errors are in parentheses, and ∗∗∗ , ∗∗ and ∗ denote significance at 1%, 5% and 10%
respectively.

Dependent Variables
Coefficient 30-day α β on RMRF β on SMB β on HML β on UMD
YearsTraded 0.58 -0.034 -0.016 0.0544 0.0073
(0.41) (0.0108)*** (0.0039)*** (0.0435) (0.0250)

YearsTraded2 0.02 0.00246 0.00163 0.0037 -0.00076

(0.02) (0.0012)* (0.0005)*** (0.0028) (0.0019)

CumulTrades (÷102 ) 0.05 -0.01 -0.002 -0.002 -0.016

(0.02)*** (0.002)*** (0.0009)*** (0.004) (.005)***

CumulTrades2 (÷104 ) -0.0002 0.00005 0.00001 0.00002 0.00008

(0.0001)** (0.00001)*** (0.000006)** (0.00002) (0.00002)***

Dep Var Mean 0.07 0.663 0.028 -0.086 -0.054

Standard Dev 3.30 0.288 0.103 0.431 0.440

58
 Figure 2.1: Participation By Year
This graph shows the number of accounts that place one or more trades in each year, including
both accounts that exist at the beginning of the sample and new accounts. There is considerable
variation in the number of accounts placing trades over time, from a low of around 54,196 in 1995
to a high of 311,013 in 2000.

Figure 2.2: Returns Persistence

This figure plots the average 30-day returns earned by investors in years following their first pur-
chase. Investors are grouped into quartiles in their first year of trading, and we then calculate
average returns for each quartile in subsequent years. Returns are calculated using the approach
discussed in the text, and are demeaned by calendar year, which removes the impact of any year
fixed effects. Raw returns are reported here, but the results for risk-adjusted returns are not quali-
tatively different.

59
 Figure 2.3: The Disposition Effect in Aggregate
This graph shows how the propensity to sell a stock depends on the stock’s return since purchase.
Each line plots the regression coefficients from one hazard regression modeling the conditional
probability of selling a stock. The coefficients correspond to dummy variables for return ‘bins’
ranging from [−10, −9) percent to [9, 10) percent. In each year, there is a pronounced kink near
zero, and the hazard increases rapidly for positive returns.

60
 Figure 2.4: Returns by Disposition Quintile
This figure shows average 10-, 20-, 30-, and 45-day returns following a purchase for each dis-
position quintile. Returns are calculated using the approach discussed in the text. Raw returns
are reported here, but the results for risk-adjusted returns are not qualitatively different. Returns
earned by the lowest quintile (1) are higher than those earned by the highest quintile (5).

Figure 2.5: Proportion of Accounts Who Exit

This graph shows the percentage of investors who exit in 1 year, 2 years, . . . , or 6 years follow-
ing the first year in which they place at least seven trades. Exit is defined as placing no further
trades during our sample period. Within each year group, the first bar indicates the percentage of
investors who exited in the 1st year after placing a trade, the second bar indicates the percentage
of investors who exited in the 2nd year, and so on. Data are missing for later years because we do
not know how many investors stopped trading after 2003.

61
 Figure 2.6: Trading Intensity and Experience
This figure shows how trading intensity changes with experience. Intensity is measured as the
number of trades placed (dark blue) and the total value of trades placed (light blue; in 10,000’s
of EUR). Results are demeaned by calendar year to adjust for year fixed-effects. We report results
beginning in the investor’s first full calendar year in the sample.

62
 Chapter III

Who Trades with Whom?

Models of market microstructure posit the existence three types of financial

market participants: informed investors, uninformed “noise traders,” and market-

makers (Glosten and Milgrom 1985, Kyle 1985). An important question in eco-

nomics, dating back at least to Keynes (1936) is to what extent prices are dis-

torted by noise traders.1 Friedman (1953) argues that rational, informed investors

quickly exploit arbitrage opportunities caused by mispricing. But De Long, Shlei-

fer, Summers, and Waldmann (1991) show how noise traders can have long-term

effects on prices, and Shleifer and Vishny (1997) and Abreu and Brunnermeier

(2003) explain why arbitrageurs may be unable to take advantage of known mis-

pricing. If noise traders distort prices, then prices must move in response to their

trading. This chapter examines that possibility.

Who are the proverbial noise traders? While they may have an exogenous

“liquidity” motive for trade, Black (1986, p. 531) defines noise trading as “trad-

ing on noise as if it were information.” The literature provides considerable evi-

1 Keynes (1936, p. 155) describes investing as a “battle of wits to anticipate the basis of conven-

tional valuation a few months hence, rather than the prospective yield of an investment over a
long term of years. . . ”

63
dence that individual investors play this role.2,3 For example, individual investors

make remarkably poor investment decisions. In the data used in this chapter,

stocks heavily bought by individuals underperform stocks heavily sold by 6.8%

over the subsequent year. In sharp contrast, stocks heavily bought by institutions

outperform stocks heavily sold by 9.0% (see Table 3.1). There are two possible

explanations for these results, which have been previously documented in other

data. One, offered by Barber, Odean, and Zhu (2006) and Hvidkjaer (2006), is

that individual investors push prices away from fundamentals; the subsequent

reversal to fundamental value leads to the documented poor performance. The

second, supported by Kaniel, Saar, and Titman (2008) and Campbell, Ramadorai,

and Schwartz (2008), is that well-informed institutions buy undervalued stocks

from individuals and sell overvalued stocks to individuals, and prices subse-

quently move toward fundamentals. Under the first explanation, individuals dis-

tort prices; under the second, institutional demand for trading is met by individ-

uals whose trading supplies liquidity.4

In this chapter, I use the complete daily trading records for all trading in Fin-

land over a nine-year period to examine these competing hypotheses. First, in

contrast to the existing literature, I identify how much trading occurs not only be-
2 See Barberis and Thaler (2005), especially Section 7 and the references therein.
3 Throughout this chapter, I use the terms “individual” and “household” investors interchange-

ably. In all cases, I am referring to individual traders, and not institutional investors or professional
traders.
4 The first story in particular also requires a considerable amount of herding among individual

investors. I provide estimates on the degree of in-group correlation for individual investors below.
Several coordinating mechanisms could explain this herding. For example, individuals may notice
prices falling and choose to purchase because they think it’s a good buying opportunity. This may
be thought of as a price-based mechanism. Another example would be retail brokers acting to sell
inventory by encouraging clients to buy shares.

64
tween individuals and institutions, but also within each group. I document that in-

stitutions are about half as likely to trade with individuals as they are to trade with

other institutions. This is particularly interesting given the herding that has been

documented among both individuals (Odean 1998) and institutions (Wermers

1999), which would mean that trading between groups is very common—if all

institutions are buying, they cannot buy from other institutions. Second, I show

that prices move consistently when institutions trade with individuals. In particu-

lar, when individuals buy shares from institutions, prices decline; and when they

sell shares to institutions, prices increase. Of course, this implies that prices fall

when institutions sell shares to individuals, and rise when institutions buy shares

from individuals. In other words, institutions move prices and individuals supply

liquidity to meet the trading needs of institutions. I confirm these results at daily,

weekly, and monthly horizons, and with vector autoregressions. Third, I show

that when prices move as a result of trading among individuals, they are more

likely to revert than after trading between individuals and institutions or between

two institutions. This price reversion is consistent with individuals being unin-

formed. Moreover, I show that these price reversions occur because institutions

subsequently trade with individuals in a direction that moves prices back toward

previous levels.

My data include the complete daily trading records of all households, finan-

cial institutions and other entities that trade stocks on the Helsinki stock exchange

between 1995 and 2003. There are three notable features of these data that make

them particularly well-suited to examining the relation between trading and price

65
changes. First, the data include account identifiers that classify the investor as

a household, financial institution, nonfinancial corporation, government agency,

nonprofit organization, or foreigner. Therefore, there is no need to estimate an

investor classification as there is in datasets available for the U.S. Second, whereas

data available in the U.S. are either available quarterly or from proprietary data-

sets covering small samples of traders and/or short time periods, the Finnish data

record all transactions placed each day by each investor. This allows me to analyze

the interaction of investors at a high frequency without relying on the estimation

technique developed by Campbell, Ramadorai, and Schwartz (2008). Third, the

data cover a nine-year period for the entire Finnish stock market. The sample in-

cludes both the “bubble” period in technology stocks during which many Finnish

stocks rose dramatically, as well as periods before and after this rise. This helps

ensure that the results are generally applicable to a variety of market conditions,

and not driven by rare events.

The poor performance of individual investors documented by Odean (1998,

1999) and Barber and Odean (2000, 2001) can result either because they trade

with better-informed institutional investors, or because they push prices above

or below fundamentals and subsequently lose money in the ensuing correction.

Barber, Odean, and Zhu (2006) and Hvidkjaer (2006) present evidence that the

trading of individual investors moves prices, which then slowly revert. Because

of constraints on the data available for the U.S. market, these authors adopt a

clever strategy to identify the trading of individual investors: they examine the

imbalance of buyer- and seller-initiated transactions for small quantities of trades

66
and classify this as the trading of individuals. Barber, Odean, and Zhu show that

this order-imbalance is correlated with the order-imbalance among a sample of

investors at a discount brokerage firm. A potential problem with this approach is

that either individuals or institutions could be initiating these small trades. The

identification strategy employed by these authors, however, relies on the assump-

tion that individuals initiate these trades. This could raise concerns about the

generality of the results found in these papers. Indeed, in contrast to these results,

Kaniel, Saar, and Titman (2008), Campbell, Ramadorai, and Schwartz (2008), and

Linnainmaa (2007) all find that individual investors supply liquidity to meet in-

stitutional demand for immediacy.5

The remainder of the chapter is organized as follows. Section 3.1 develops the

hypotheses to be tested in the chapter. In Section 3.2, I motivate my classification

method and show how it is implemented. I present my results in Section 3.3.

Section 3.4 concludes.

5 The large literature examining the performance of institutional investors also highlights the
importance of using high-frequency data. Increases in institutional ownership are associated with
price increases at a quarterly or annual frequency (Nofsinger and Sias 1999, Cai and Zheng 2004).
However, it is difficult to tell from quarterly data whether institutional trading within the quarter
leads, lags, or is contemporaneous with returns. This makes it difficult to determine causality,
particularly since fund managers follow momentum strategies (Carhart 1997, Daniel, Grinblatt,
Titman, and Wermers 1997, Wermers 1999). Sias, Starks, and Titman (2006) develop a technique
to extract the covariance of returns and institutional trading at a monthly or weekly level from
quarterly data. They find that institutional trading leads returns. Griffin, Harris, and Topaloglu
(2003), examine daily trading in Nasdaq 100 securities during a ten-month period beginning in
May, 2000. They focus on the relation between returns and the buy-sell imbalance of individuals
and institutions—not the total amount of trading within and between groups examined in this
chapter.

67
3.1 Hypotheses

As discussed above, the low returns earned by stocks following high levels of

buying by individuals could arise either from individuals pushing prices above

fundamental value, or by institutions selling overvalued stocks to individuals. To

differentiate between these alternatives, I develop and test a number of hypothe-

ses.

While researchers typically think of liquidity provision as submitting a limit

order that gives others the option to trade, Kaniel, Saar, and Titman (2008) note

that practitioners think of a buy order placed when prices are falling, or a sell

order placed when prices are rising, as supplying liquidity, regardless of whether

the trader submits a limit or market order. This is the sense in which I use the term

“liquidity provision” in the chapter. Individual investors may not set out to pro-

vide liquidity to institutions, actively posting limit buy and sell orders and taking

the spread as compensation for their services; rather, they may respond to price

changes caused by institutional trading and end up supplying liquidity. One way

this can occur is if individuals have “latent” limit orders—prices at which they

plan to buy or sell in the future—and these orders get triggered by price move-

ments. For example, individuals who suffer from the disposition effect are more

likely to sell a stock after seeing its price rise.6 Grinblatt and Han (2005) show that

momentum in stock returns can be caused by disposition-prone investors behav-

ing this way.

6 The disposition effect is the well-documented reluctance of investors to sell stocks below
their purchase price. The behavior was first described by Shefrin and Statman (1985). See Seru,
Shumway, and Stoffman (2007) and references therein.

68
 Before stating the hypotheses, it is useful to consider possible price paths sur-

rounding a trade, as shown in the stylized examples in Figure 3.1. The figure

shows four price paths following a trade at time t0 . In the top two graphs, the

trade is buyer-initiated. The bottom two graphs depict seller-initiated trades. The

left two graphs show trade between an informed buyer and an uninformed seller,

while the right two graphs show trade between an uninformed buyer and an in-

formed seller. When the trade is initiated by an uninformed trader, prices sub-

sequently revert, as seen in the northeast and southwest quadrants. If the trade

initiator is informed, however, no such reversion takes place. This price reversion

is a feature of models with asymmetric information: in contrast to the perma-

nent price impact of informed trades, uninformed trading causes immediate price

changes to compensate liquidity providers, but expected future cash flows have

not changed.7 The reversion can stem from bid-ask “bounce,” and is critical to

the estimation of liquidity measures such as Roll’s (1984) spread and Pastor and

Stambaugh’s (2003) liquidity factor.

Given the poor performance of individual investors discussed above, the ques-

tion studied in the chapter is whether individuals demand liquidity and actively

move prices (top right of Figure 3.1) or supply liquidity as prices move (bottom

left of the figure). If institutions are more likely to be informed, and individuals

provide institutions liquidity in the sense defined above, then this stylized exam-
7 See Glosten and Milgrom (1985), Kyle (1985), Easley and O’Hara (1987), Campbell, Grossman,

and Wang (1993), Llorente, Michaely, Saar, and Wang (2002), Chordia and Subrahmanyam (2004),
and Avramov, Chordia, and Goyal (2006), among others.

69
ple leads to a number of hypotheses. First, when institutions trade with individu-

als, prices should move. In particular,

Hypothesis 1. When institutions purchase shares from individuals, prices contempora-

neously increase. When institutions sell shares to individuals, prices contemporaneously

decrease.

Price increases accompanying institutional buying, and decreases accompanying

institutional selling, are consistent with institutions demanding liquidity. In con-

trast, evidence against Hypothesis 1 would indicate that institutions supply liq-

uidity. To test this hypothesis, I regress daily returns on a set of variables that

summarize the amount of trading that took place between each investor group

on each day. I also test the relation using weekly and monthly horizons. Details

of the estimation procedure and results of tests of Hypothesis 1 are presented in

Section 3.3.2.

Second, the stylized examples in the Figure 3.1 indicate that prices will change

predictably after trading, depending on which types of investors caused the price

change. In particular,

Hypothesis 2. Price reversion is more likely following days when individuals trade with

other individuals than days when individuals trade with institutions.

Tests of this hypothesis are similar to those of Hypothesis 1, but instead of ex-

amining the contemporaneous relation between returns and trading by different

groups, I investigate how returns change in the period following trading by in-

dividuals and institutions. In particular, I use a regression framework to test

70
whether negative autocorrelation in daily returns is stronger following days when

more trading takes place between two individuals than days when more trading

occurs between individuals and institutions.

If Hypothesis 2 is true, it is also interesting to determine whose trading leads to

price reversion. In particular, if trading comes primarily from individuals trading

among themselves and prices change, we might expect institutions to react to

the price movement by trading in a direction that pushes prices back to previous

levels. That is, we would expect institutions to cause the price reversion by trading

subsequently with individuals. This leads to the third hypothesis:

Hypothesis 3. Institutions react to price changes caused when individuals trade with

each other by subsequently trading with individuals to move prices back toward previous

levels.

To test Hypothesis 3, I examine the relation between institutional trading and the

previous day’s proportion of individual trading interacted with the price changes.

I use a regression framework to test (a) whether institutions are more likely to sell

to individuals following days that have both high returns and more intragroup

individual trading; and (b) whether institutions are more likely to buy from indi-

viduals following days that have both low returns and more intragroup individual

trading. Details of the estimation procedure and results for Hypotheses 2 and 3

are presented in Section 3.3.5.

71
3.2 Data and methods

In this section, I begin by describing the salient features of the data used in

this study. I then discuss the procedures I use to classify investors into different

groups, as this is key to the empirical implementation in the chapter.

3.2.1 Data description

The data used in this chapter are discussed in Chapter I. Summary statistics

relevant for this chapter are presented in Table 3.2. The data include the transac-

tions of nearly 1.3 million individuals and firms, beginning in January, 1995 and

ending in December, 2003. For accounts that existed prior to 1995, opening ac-

count balances are also included, making it possible to reconstruct the total port-

folio holdings of an account on each day. While the dataset includes exchange-

traded options and certain irregular equity securities, I focus on trading in ordi-

nary shares. After excluding trading in very thinly-traded securities,8 there are

more than 37.5 million trades in the data, including 10.7 million trades by 583,518

unique household accounts. The remaining trades are placed by financial insti-

tutions, corporations, and to a much smaller degree, government agencies and

certain other organizations. Because all trading is recorded in the data, it is possi-

ble to construct measures of trader interaction that are not feasible with datasets

that include small samples of the population.

8I exclude securities with fewer than 25,000 trades, or about ten trades per day during the
sample period. This leaves 116 stocks. The additional requirement that daily returns data are
available on Datastream reduces the sample to 106 stocks.

72
3.2.2 Complications

There are a number of complications that arise when working with these data

especially relevant to this chapter. First, transactions are not time-stamped, so it

is not possible to determine the order in which trades took place within a day.

Second, while each transaction is identified as either a buy or a sell, the absence

of quote data makes it impossible to identify which side initiated the trade, as is

typically done with the quote-based algorithm of Lee and Ready (1991). Never-

theless, if prices consistently move in the same direction as an investor’s trading,

this investor must be initiating trades; an investor who sells shares into a rising

market cannot be setting prices. To be more precise, since every transaction re-

quires both a buyer and a seller, there is a sense in which both the buyer and the

seller are setting prices. However, it is commonly understood in the literature

that the initiator of a trade causes the price change; indeed, this is the identifying

assumption of the Lee-Ready algorithm.

Third, the investor group classification is not always correct because of the

existence of so-called “nominee” accounts. I outline my approach in dealing with

these accounts in Section 3.2.2. A related issue is that financial institutions that

act as market-makers are not separately identified in the data. In Section 3.2.2, I

discuss the method I use to separate these accounts from my analysis.

Fourth, there is no direct match between the buy-side and sell-side of a trans-

action. For each trade, at least two observations are recorded in the data: one

purchase and one sale. However, some trades are comprised of shares purchased

73
or shares sold by more than one account, and in these situations there can be more

than two records per trade. For example, if investor A buys 100 shares of Nokia

and investor B sells 100 shares at the same price, they may have traded with each

other, but no link between these transactions is reported in the data. This neces-

sitates using a technique to identify the amount of trading that occurs between

groups, which I discuss in Section 3.2.2.

Nominee accounts

Unlike investors domiciled in Finland, foreign investors are not required to

register with the CSD. Instead, they may trade in an account that is registered

in the name of a “nominee” financial institution. As a result, trading by foreign

individuals can appear to be trading by the nominee institution. This is also true

of American Depository Receipts (ADRs), which many Finnish firms have. For

example, if an individual in the U.S. trades shares of Nokia’s ADR on the NYSE,

it will be classified in the CSD data as a trade by the institution that serves as a

nominee for the ADR.

Since the focus of this chapter is on examining the differences in the price im-

pact of trading among groups of investors, it is important to deal with this mis-

classification in some way. Note, however, that this misclassification makes it

more difficult to find differences in the price impact of trading by different groups,

because the misclassification causes a “mixing” of the group members. Never-

theless, I adopt the following approach to identify accounts that are likely to be

nominee or ADR accounts. First, I require an institution to own at least ten per-

74
cent of the total shares outstanding before considering it to be acting as a nomi-

nee. Then, for each stock and day, I calculate the number of trades and quantity

of shares traded by each institution as a percent of all trading in that stock/day.

I then count the number of days in a month in which the institution accounts for

more than ten percent of both the number of trades placed and the quantity of

shares traded. Finally, I classify as a nominee any institution with ten or more

such days in a month. While the specifics of this procedure are somewhat arbi-

trary, it generally identifies only a few accounts that serve as nominees for each

security, which is expected. Moreover, the results in this chapter are not sensitive

to using a variety of alternative parameters of the classification procedure.

Market-makers

For the main analysis in this chapter, I focus on individuals and institutions

that trade for information or liquidity reasons. Therefore, it is necessary to iden-

tify certain institutions and individuals that trade particularly actively, effectively

acting as market-makers by both buying and selling a given stock on a particular

day. While the motivation for such trading may differ, this type of trading may

be broadly classified as market-making. (Linnainmaa (2007) explores the behav-

ior of individual day traders in this market.) Traders who act as market-makers

are unlikely to be trading based on fundamental or long-term information, and I

therefore exclude this set of traders from my analysis.

To identify market-makers, I simply check whether an account purchased and

sold shares in the same company on the same day. Occasionally, there are cases

75
where an account buys and sells shares, but the amounts traded are different by

orders of magnitude. For example, an account may purchase 5000 shares and sell

5 shares of one stock. It is unclear why this would occur, but it is probably not

correct to call this trader a market-maker. Therefore, I require the amount pur-

chased to be between 10 and 90 percent of the total amount traded by an account

for it to be classified as a market-maker on that day. (The results are not sensi-

tive to this restriction.) If an account is classified as a market-maker on at least

five trading days in a particular month, I treat it as a market-maker for the entire

month. The results in the chapter are unchanged if I exclude only accounts acting

as market-makers on a day-by-day basis, or if I only allow financial institutions to

be classified as market-makers. Note that market-makers are identified after the

nominee accounts are identified, as discussed above in Section 3.2.2.

Identifying investor interaction

Given a classification of investors into groups, as in my data, it is possible to

estimate the amount of trading that occurred between and within groups. For

example, suppose trading on one day for one stock at one particular price is sum-

marized as follows:
Shares Shares
Bought Sold
Group A 250 450
Group B 2000 1800
Group C 250 250
Total 2500 2500

While we cannot be certain how much trade occurred between or within each

group of investors, we may approximate these quantities by assuming that trade

76
occurs in proportion to the amount of buying or selling accounted for by each

group. Of the 250 shares purchased by Group A, we would therefore estimate that

450/2500 = 18% (45 shares) were purchased from other members of Group A,

1800/2500 = 72% (180 shares) were purchased from members of Group B, and

so on. Continuing with this example, we would estimate the amount of trading

within and between Groups A, B, and C as follows:

Seller
Buyer A B C
A 45 180 25
B 360 1440 200
C 45 180 25

This estimation strategy conditions on the actual amount of purchases and sales

by each group; we are not assuming that investors from each group randomly

choose to buy and sell. Instead, we are assuming that, given the actual number of

shares sold by a group, the probability of selling to any other group that bought

shares is proportional to the relative amount of shares purchased and sold by each

group.

This procedure is an approximation technique. In the previous example, it is

possible that members of Group A did not purchase any shares from other mem-

bers of Group A, although the estimate is that 45 shares were traded within this

group. However, the procedure can yield an exact identification of the amount

of trading that occurred between two groups. In the example, as is frequently the

case in the data, one group (B) accounts for much of the purchases and sales. Since

Group B accounts for so much of the trading, much of the trading must have oc-

curred within this group—there simply is not enough selling by Groups A and C

77
to meet the large demand for shares from Group B. In fact, by applying the proce-

dure for each price at which the stock traded in a day and then aggregating to get

a daily measure, I maximize the frequency with which this happens. Especially

for all but the most frequently traded stocks, it is common for groups of investors

to have only purchased or sold shares, but not both, at a particular price; it is

therefore frequently possible to know with certainty exactly how much trading

occurred between groups.9 In my data, on average 28.2% (ranging from 3.4% to

60.3%) of a stock’s trading volume is exactly identified in this manner.

An alternative method to determine how much trading occurs within and be-

tween groups involves establishing bounds on how much trading could have oc-

curred. Referring again to the example, note that Group A investors could have

traded no more than 250 shares with other Group A investors, since Group A only

bought 250 shares in total. Similarly, intragroup trading in Group B could have

been no more than 1800 shares. But since total purchasing by Groups A and C

amounts to only 500 shares, Group B investors must also have sold no fewer than

1800 − 500 = 1300 shares to other Group B investors. Therefore, intragroup trad-

ing for Group B must have been between 1300 and 1800 shares. Since any amount

of trading within these bounds could be used as the estimated amount of trading,

using the mean seems sensible. In the example, this value is quite close to the

value estimated using the technique above—1550 vs. 1440, or 62% vs. 58% of to-

tal trading. When applied to my data, these two methods generally provide very
9 Suppose that institutions purchased 800 shares and sold 1000 shares, and that individuals pur-

chased 200 shares but sold no shares. The institutions must have sold 200 shares to the individuals
and 800 shares to other institutions—there is no ambiguity here, and no estimation is required.

78
similar estimates of trade interaction, so the results reported below are generated

from the first approach, which is somewhat less complicated. The results are not

substantively changed, however, by using the second method.

3.3 Results

I turn now to presenting the results of the chatper. I begin by quantifying the

amount of trading that occurs among investor groups. Institutions and individ-

ual investors account for a large proportion of total trading, and it is certainly

possible that trading by either of these groups could have important prices ef-

fects. I then examine the relation between contemporaneous returns at a daily,

weekly, or monthly horizon and inter- and intragroup trading quantities. Returns

are higher when institutions buy shares from individuals, and lower when insti-

tutions sell shares to individuals, suggesting that institutions move prices. This

indicates that, on average, institutions are informed and individuals are not. To

allow returns and trading quantities to be mutually dependent, I also estimate a

vector autoregression (VAR). The VAR allows me to estimate price impact func-

tions for each of the groups in the data. Next, using transaction prices, I show that

institutions buy stocks at higher prices than do individuals when returns are high,

and sell at lower prices than do individuals when returns are low, which provides

additional evidence at an intraday frequency that institutions set prices. I then

show that price reversion is more prevalent when intragroup trading by individ-

79
uals is high. Moreover, the price reversion is caused by the subsequent trading of

institutions.

3.3.1 Investor interaction

Table 3.3 shows the amount of trading that is accounted for by each of the

four largest groups of traders. The omitted groups are Government agencies,

Nonprofit organizations, and Registered foreigners; for most stocks on most days,

these groups account for a negligible amount of total trading. The total amount of

trading may be calculated as a percent of the number of shares traded (presented

in Row 1) or as a percent of the value of shares traded (Row 2). The combined

trading of individuals, institutions, nominees, and other corporations accounts for

approximately 95% of all trading. About 40% comes from institutions, and about

20% from individuals (“Households”). Individuals account for 20.6% of trading

by number of shares traded, and 17.4% by value, indicating that individuals tend

to trade more heavily in low-priced stocks.

Using the total proportion of trade from each group, it is possible to calculate

the amount of trading we would expect to occur within and between each group.

The results of this exercise, using the percent of trading by shares (Row 1) is re-

ported in the table under the heading “Expected % of all trading.” For example,

since financial institutions account for 38.1% of all trading, 0.3812 =14.5% of all

trading should be between two institutions. Similarly, the amount of trading that

would occur between institutions and individuals is 0.381 × 0.206 × 2 = 0.157,

80
where we multiply by two because either the institution can buy from the indi-

vidual or the individual can buy from the institution.10

The lower panel of Table 3.3 shows the estimated amount of trading that oc-

curred between and within groups, using the technique described above in Sec-

tion 3.2.2. That is, in contrast to the “expected” percentage in the top panel, we

now condition on the actual amount of buying and selling by each group on each

day, rather than total trading volume alone. There are several notable differences

between actual and expected trading. Institutions trade more with each other, and

less with individuals, than expected. As well, individuals trade more with other

individuals than expected. Previous research has documented herding behavior

among both individuals and institutions; in other words, within-group trading is

positively correlated. This means that trading between groups should be quite

common—if all individuals are buying, they cannot trade with each other. The

results presented here show that the previous findings mask an important fact: a

great deal of trading occurs between two individuals or between two institutions.

In these data, individuals are more than twice as likely to trade with other indi-

viduals than their trading volume would suggest, and only 9.3% of trading takes

place between individuals and institutions. Alternatively, ignoring nonfinancial

corporations and ADRs, we see that institutions are about half as likely to trade

with individuals as they are to trade with other each other.

10 The table is lower-diagonal because it shows the expected amount of total trading—not sep-

arated into buying and selling. A matrix showing the expected proportion of buying and selling
separately would be symmetric, with the off-diagonal elements equal to one-half of the reported
off-diagonal elements.

81
 What explains the discrepancy between the actual and expected amount of

trading among individuals? There are several possibilities, each of which could

partially contribute to the result. First, institutions can arrange large block trades

with each other away from the regular limit-order book, so for a fixed amount of

trading, less volume will take place between individuals. Second, as in the model

of Easley and O’Hara (1987), informed investors might only trade when there has

been an information event. If institutions tend to be informed, they will be less

likely to trade when no information event has occurred, and any trading by indi-

viduals will tend to be with other individuals. Indeed, Easley, Engle, O’Hara, and

Wu (2002) find that uninformed orders are clustered in time, but also that unin-

formed investors avoid trading when informed investors are likely to be present.

A related explanation is offered by Barber and Odean (2006), who document that

individual investors are more likely to trade following events that get media at-

tention.

It is worth noting at this point that the amount of trading coming from each

group determines how much power I will have to find a relation between group

trading and price changes. For example, if almost all trading came from institu-

tions, then it would be difficult to find a relation between price changes and the

trading of any group because returns (the left hand side variable in my regres-

sions) would vary, but the proportion of trading from each group (the right hand

side variable) would not. The percent of trading reported in Table 3.3 suggests

that trading is sufficiently spread among different groups to provide adequate

power for my tests.

82
 The results in Table 3.3 are calculated using data aggregated from all stocks

in the sample. In the Finnish market, a few stocks account for much of the trad-

ing volume and market capitalization. Notably, the largest stock, Nokia, makes

up 36% of the total stock market capitalization on average during the sample pe-

riod (ranging daily from 16% to a high of 64% at one point in 2000). On average,

Nokia accounts for 57% of the value of daily trading (ranging from 3% to 95%).

To confirm that the reported results apply to Finnish stocks in general and are not

driven by Nokia or a few other large stocks, I first calculate a time-series average

of the daily interaction estimates for each stock, and then examine cross-sectional

statistics for these means. This procedure puts equal weight on each stock’s aver-

age interaction estimates. These results are reported in Table 3.4. Across the 106

stocks in my sample, trading among institutions and among individuals accounts

for an average of 14.7% and 19.3%, respectively, confirming that intragroup trad-

ing is important across stocks. Trading among individuals ranges from 0.1% to

60.6% of all trading; the smaller numbers come from large actively-traded stocks

in which trading is dominated by institutions, while the larger numbers come

from smaller, less-liquid stocks. It is interesting to note that the range for trading

between institutions and individuals is much narrower, indicating that it is gener-

ally an important part of trade volume, although it never accounts for more than

14.8% of the total.

The data presented in this section show that individuals account for approx-

imately one-fifth of trading in the Finnish stock market. This is certainly suffi-

ciently large for the trading of individual investors to have substantial price ef-

83
fects. About half of individual investors’ trading is with other individuals, and

half is with institutions. In the next section, I examine which type of trading is

most associated with price changes.

3.3.2 Daily returns

To understand how prices are determined by the interaction of different in-

vestors in the market, I examine the relation between returns and the proportion

of trading within and between investor groups. In particular, I estimate the re-

gression

Ri,t = α + βRi,t−1 + γ1 Inst/Insti,t + γ2 Inst/Indi,t

+ γ3 Ind/Insti,t + γ4 Ind/Indi,t + ei,t , (3.1)

where the notation A/Bi,t represents the proportion of trading that is accounted

for by investors from Group A purchasing shares from investors in Group B. “Ind”

and “Inst” denote individuals and institutions, respectively. For brevity, I focus in

the remainder of the chapter only on the trading of individuals and institutions.11

Since not all investor groups are included in the regression, the A/B trade vari-

ables do not sum to one, and including an intercept does not result in perfect

collinearity.
11 The main conclusions are not altered if I include the other investor groups. In particular, the
price impact estimates for nonfinancial corporations and nominee accounts lie between those of
individuals and institutions, and the estimated coefficients for individuals and institutions are
basically unchanged if these additional groups are included.

84
 I begin by estimating the regression at a daily horizon, and later confirm that

similar results obtain using weekly or monthly data. If trading between members

of Group A and Group B (possibly the same group) leads to price changes, then

contemporaneous returns will be positive or negative, and the estimated γ coef-

ficient for the relevant combination of trade will be significant. If no group has a

consistent effect on prices, then the four trade variables will be economically small

and statistically insignificant. The lag return, Ri,t−1 , is included to control for the

known autocorrelation in daily returns.

I estimate regression (3.1) using the approach of Fama and MacBeth (1973) in

two ways: averaging results from time-series regressions by stock or from cross-

sectional regressions by date. The latter approach results in a time-series of es-

timates, which may be autocorrelated. Therefore, I use the robust standard er-

ror calculation of Newey and West (1987) with five lags, which corresponds to

one week of trading.12 The adjustment for autocorrelation is unnecessary for the

cross-section of estimates obtained from the first approach. Estimating the regres-

sion separately along each dimension allows me to check whether the results are

driven by a cross-sectional relation, a time-series relation, or both. Using a panel

regression with clustered standard errors yields similar results, but could suffer

from a bias due to the lagged dependent variable, so I do not report these esti-

mates.
12 With 2184 daily observations, the rule of thumb noted by Greene (2005, p. 267) calls for
√4
b 2184c = 7 lags, and the significance level of all results is unchanged by using the longer lag
length.

85
 Coefficient estimates for regression (3.1) are presented in Table 3.5, and provide

an interesting picture of how prices are affected by the interaction of individual

and institutional investors in the market. Panel A reports results where the A/B

trade variables are calculated as the proportion of total trade volume each day,

and Panel B uses trade variables calculated as the proportion of daily turnover

(shares traded divided by number of shares outstanding) each day. Results from

the stock-by-stock regressions are denoted “FM by stock,” and the day-by-day

regressions are denoted “FM by date.”

Consider first the “FM by stock” regressions in Panel A. When institutions

trade with other institutions, there is no significant change in price. As discussed

above in Section 3.2.2, a large portion of trading comes from institutions trading

with other institutions and households trading with other households. To date,

the literature has not been able to address the importance of price changes that oc-

cur during this trading. The results in Table 3.5 bear directly on this question. The

estimated coefficient is 0.0092, with a t-statistic of 1.35. Of the 106 stock-by-stock

regressions, the Inst/Inst coefficient is significantly negative in 8% of regressions

and significantly positive in 11% of regressions, using a 5% significance level. Sim-

ilarly, trading between individuals does not yield significant results, although 27%

of stocks yield statistically negative coefficients. That is, there appears to be little

in the way of price changes when trading takes place within each investor group.

In sharp contrast, when institutions purchase from individuals, prices increase.

The estimated coefficient of 0.0316 is highly significant, and fully 61% of the re-

gressions yield statistically positive results; none of the coefficients is statistically

86
negative. And when individuals buy from institutions, prices fall: the coefficient

of −0.0423 is again highly significant, and 72% of the cross-sectional regressions

yield statistically positive coefficients, with none significantly negative. The cross-

sectional average of the stock-level standard deviation in the trade variables is

reported in the table under the heading “C.S. Std. Dev.” These values can be

used to ascertain the economic significance of the coefficient estimates. A one

standard deviation increase in the Inst/Ind variable (that is, institutions buying

shares from individuals) increases daily returns by 0.316 × 0.1151 = 0.0036 or 36

basis points (bps). This is economically large relative to an average daily return of

approximately 10 bps. Similarly, a one standard deviation increase in the Ind/Inst

variable leads to a decrease in returns of 0.0423 × 0.1145 = 48 bps.

The day-by-day cross-sectional (“FM by date”) regressions deliver similar re-

sults, suggesting that the result is not driven solely by either time-series or cross-

sectional relations. The effect of trading between households in this specification

is to reduce contemporaneous returns. This result is not robust, however, which

can be seen by looking at the results in Panel B. Across both panels and both

specifications, the result that is consistent is that prices increase when individuals

sell shares to institutions, and prices decrease when individuals buy shares from

individuals. This evidence strongly supports Hypothesis 1.

3.3.3 Vector autoregressions

Another approach to examining the relation between group trading and subse-

quent price changes is to estimate a vector autoregression (VAR) as in Hasbrouck

87
(1991). There are a number of benefits to this approach. First, allowing the trade

variables and returns to depend on lags of each other provides a way to examine

potentially complicated dynamics among the variables. Second, the lag structure

of the VAR allows me to plot contemporaneous price impact and subsequent price

changes. These plots are the empirical analogue of the stylized price paths shown

in Figure 3.1.

Let yi,t ≡ (Ind/Indi,t , Inst/Insti,t , Ind/Insti,t , Inst/Indi,t , Ri,t )0 . As above, the

notation A/Bi,t denotes the proportion of trading in stock i on date t that comes

from Group A purchasing shares from Group B, and “Ind” and “Inst” denote

individuals and institutions, respectively. The reduced-form VAR for each stock,

i, is
p
yi,t = ∑ Φk yi,t−k + ei,t . (3.2)
k =1

In order to allow returns to depend contemporaneously on the trade variables, I

estimate a dynamic structural VAR (see Hamilton (1994, Section 11.6)). In partic-

ular, triangular factorization of the error covariance matrix, Σ ≡ E(et e0t ), yields a

lower-diagonal matrix, A0 , with ones on the principal diagonal such that A0 ΣA00 =

Σd where Σd is a diagonal matrix with all positive elements. Multiplying both

sides of equation (3.2) by A0 gives the dynamic structural VAR

p
A0 yi,t = ∑ Ak yi,t−k + ηi,t , (3.3)
i =1

88
where Ak = A0 Φk and ηt = A0 et . The shocks in this system are uncorrelated,

since E(ηt η0t ) = E(A0 et e0t A00 ) = Σd . Moreover, since A0 is lower-diagonal, this

specification allows each variable in yi,t to depend on contemporaneous realiza-

tions of the variables that precede it in the vector:

p
Ind/Indi,t = ∑ A1k yt−1 + ηi,1t , (3.4a)
i =1
p
Inst/Insti,t = ∑ A2k yt−1 − a21Ind/Indi,t + ηi,2t , (3.4b)
i =1
p
Ind/Insti,t = ∑ A3k yt−1 − a31Ind/Indi,t − a32Inst/Insti,t + ηi,3t , (3.4c)
i =1
p
Inst/Indi,t = ∑ A4k yt−1 − a41Ind/Indi,t − a42Inst/Insti,t − a43Ind/Insti,t + ηi,4t ,
i =1

(3.4d)

and

p
Ri,t = ∑ A5k yt−1 − a51Ind/Indi,t − a52Inst/Insti,t
i =1

− a53 Ind/Insti,t − a54 Inst/Indi,t + ηi,5t . (3.4e)

Here amn denotes the (m, n)th element of A0 , and A jk denotes the jth row of Ak .

Since the order in which the variables appear in the yi,t vector determines which

variables are allowed to affect other variables contemporaneously, there is a strong

theoretical reason to put returns last, but the order of the trade variables is less

obvious. I therefore confirm that all the results below are unaffected by permuting

the order of the trade variables.

89
 Methods to estimate VARs in panel data are not well-developed. Therefore, in

the spirit of Fama and MacBeth (1973), I separately estimate the dynamic struc-

tural VAR for each stock and then take cross-sectional means of coefficient es-

timates. Statistical significance is determined from the cross-sectional standard

errors of these means. I choose ten lags (p = 10) by examining the Akaike Infor-

mation Criterion for the VAR. While a lower-order VAR fits well for some stocks, I

fit the same model to all stocks to ease comparison of results. Estimating a model

with five lags yields results that are substantially the same as those reported here.

Estimation of (3.2) yields estimates of the Φk matrices, and triangular factoriza-

tion of the estimated error covariance matrix gives an estimate of A0 , which is

then used to calculate estimates of the Ak coefficient matrices. This is repeated for

each of the 106 stocks in the sample, and Table 3.6 summarizes the results from

these regressions for k = 0, 1, 2. For brevity, higher-order lags and the constant

term are not reported.

Controlling for complex serial correlations does not alter the results reported

above. The contemporaneous effect (k = 0) on returns of the Ind/Inst and Inst/Ind

variables are very close to those reported in the previous table. In this specifica-

tion, there is also evidence that returns are negative when individuals trade with

each other: the Ind/Ind variable is negative and significant. The magnitude of

this effect is considerably smaller than the effect of intergroup trading.

Looking at k = 1, the negative autocorrelation in daily returns that is consis-

tent with bid-ask bounce is quite prevalent in these data. The amount of within-

and between-group trading is positively autocorrelated up to three lags. For ex-

90
ample, at the first lag (k = 1), the autocorrelation coefficients range from 0.0964

for the Inst/Inst variable to 0.1493 for the Ind/Inst variable, and all are significant

at the 1% level. This pattern holds for higher-order lags and for all of the variables

except returns.

Empirical price paths

The coefficient estimates from the VAR can be used to construct impulse re-

sponse functions, which are the empirical analogue to the stylized price paths

presented earlier in Figure 3.1. That is, I calculate the effect of a one standard de-

viation impulse to each of the elements of the orthogonalized shocks, ηi,t , one at a

time. For example, to see the effect on returns of a shock to Inst/Ind I use equa-

tions (3.4d) and (3.4e) to estimate the increase in returns caused by a one standard

deviation increase in ηi,4t .

Results from applying this procedure to the return equation of the VAR are pre-

sented in Figure 3.2. The figure plots the price impact function for the variables

Inst/Ind (solid line), Inst/Inst (short dash), Ind/Ind (long dash), and Ind/Inst

(dash-dot). Time is measured in trading days, so the ten lags that are plotted cor-

respond to two weeks of trading. The graph begins at time −1 to show the price

impact in the first day relative to the previous day’s close. A one standard devia-

tion innovation in Inst/Ind increases the contemporaneous return by 21 bps, and

there is no evidence of subsequent reversion. Similarly, a one standard deviation

innovation in Ind/Inst leads to a return of −28 bps, and there is no subsequent

reversion. In sharp contrast, a one standard deviation innovation in Ind/Ind is

91
associated with a return of −12 bps, but prices subsequently revert. Comparing

these empirical price impact functions to the stylized examples in Figure 3.1 in-

dicates that when institutions purchase shares from individuals or sell shares to

individuals they are informed traders demanding liquidity from individuals. This

result is clearly not consistent with individual investors actively moving prices.

3.3.4 Alternate horizons

Weekly and monthly results

Table 3.7 presents results for estimation of regression (3.1) over different trad-

ing horizons. Panel A shows results when the trading percentage variable and

returns are calculated over weekly horizons, and Panel B shows results calculated

at a monthly horizon. At these longer horizons, the results remain consistent with

what was found in the daily regression. Daily returns are contemporaneously

higher when institutions purchase shares from individuals and lower when they

sell shares to individuals. There is little or no price effect from intragroup trading

by individuals or institutions. As in the daily results, the price impact of institu-

tions is stronger when they sell to individuals than when they buy from individ-

uals.

It is interesting to note that the magnitude of the effect on returns (in absolute

terms) is greater when institutions sell to households than when institutions buy

from households, especially at the monthly horizon. That is, institutions appear to

have larger price impact when they sell to individuals than when they buy from

92
individuals. Campbell, Ramadorai, and Schwartz (2008) find a similar asymme-

try, and suggest that it could stem from the reluctance of institutions to use short

sales.13 This asymmetry is not apparent in the daily results presented earlier, per-

haps because daily returns are small and can be affected by institutions with small

positions, but short-sale constraints then become binding at longer horizons.

Intraday results: Evidence from trading prices

A potential concern with the daily results presented above relates to the timing

of trades within the day. It is possible that trading by individuals moves prices,

and that institutions subsequently trade at those new prices, but that institutional

trading does not actually affect prices. For example, suppose individuals trade in

the morning at prices above the previous day’s close, and when institutions see

prices increasing they act as momentum traders and decide to buy shares. Their

buying, however, could occur at prices that are not higher than the prices set by

individual trading. In this situation, we would find that institutions purchase

shares on days when prices rise, but institutions did not cause the price change.

The strength and robustness of the results above suggest that this scenario

is unlikely. When individuals purchase shares, same-day returns are typically

lower than when they sell shares. Nevertheless, an additional test to rule out this

possibility is in order. Unfortunately, the transactions in the dataset are not time-
13 Their argument is as follows: if institutions wish to increase their exposure to a particular risk
factor, they may buy stocks that load strongly on that factor. If their buying causes price increases
in one stock, they can purchase shares of another stock that also loads on the factor, thereby reduc-
ing their overall price impact. However, if an institution wishes to decrease its exposure to a factor
and is reluctant or unable to use short sales, it can only sell stocks it currently owns. This forces
the institution to sell more aggressively the stocks it owns, and could lead to larger price impact.

93
stamped, so it is not possible to examine the order in which trades were placed

and the path prices took within the day. However, we do observe trade prices for

each transaction, so it is possible to compare the prices at which institutions and

individuals purchase and sell shares, and the relation between these prices and

contemporaneous returns.

To understand this test, suppose that on a particular day, a stock trades only at

two prices, $10 and $11. Suppose further that both Groups A and B bought at $10

and sold at $11, but only A bought at $11. If the closing price is $11, it can only be

because Group A moved the price; Group B did not purchase any shares at $11,

so they could not have caused the price to move up to that level. That is, since

Group B bought at a lower price than did Group A—and prices increased—it is

not possible for Group B to have caused prices to move. This suggests that we can

test whether one group moves prices by examining the relation between returns

and the difference in the purchase or sale prices of individuals and institutions.

The point of this exercise is not to determine trading profits within a day, since we

are not comparing one group’s purchase and sale prices; rather, we are looking at

whether one group purchased stocks at higher prices than did another group on

days when prices rose, or sold stocks at lower prices on days when prices fell.

This test also allows us to differentiate intraday price movements from close-

to-open price movements. For example, suppose the opening price is above the

previous day’s close, but then remains flat during the day. If most trading just

happened to come from institutions buying shares from individuals, the earlier re-

gressions would find that institutions buy from individuals when prices increase—

94
but the price change happened entirely when the market was closed. This intra-

day test, however, would find that individuals and institutions both bought at the

same price on a day when returns increased.

Of course, stocks typically trade at more than two prices on one day, so I

look instead at the average price at which institutions and individuals buy or sell

shares. Specifically, I test the relation using Fama-MacBeth estimation of the re-

gressions

I H
b̄i,t − b̄i,t = β 1 · 1{ Ri,t >0} + β 2 · 1{ Ri,t <0} + ei,t (3.5)

and

I H
s̄i,t − s̄i,t = β 1 · 1{ Ri,t >0} + β 2 · 1{ Ri,t <0} + ei,t , (3.6)

where b̄ and s̄ denote the average purchase and sale prices, respectively, for stock

i on date t, and superscript I and H denote institutions and households, respec-

tively. The indicator function, 1{·} , takes a value of one only when the condition

in curly brackets is true; otherwise it is zero. Intercepts are excluded to prevent

perfect collinearity. In a second set of regressions, I allow the purchase or sale

price difference to vary with the magnitude of the return:

I H
b̄i,t − b̄i,t = α + β 1 · 1{ Ri,t >0} | Ri,t | + β 2 · 1{ Ri,t <0} | Ri,t | + ei,t (3.7)

and

I H
s̄i,t − s̄i,t = α + β 1 · 1{ Ri,t >0} | Ri,t | + β 2 · 1{ Ri,t <0} | Ri,t | + ei,t . (3.8)

95
Excluding intercepts in this second set of regressions is unnecessary. In each of

these regressions, it is important to scale the prices so that the magnitude of the

estimates may be compared across stocks, which is key to the Fama-MacBeth ap-

proach. Therefore, I scale the price difference by each stock’s average trade price

for that day. Scaling by the average closing price in the previous week yields

similar results, as does using share-weighted prices instead of a simple average.

Estimates of regressions (3.5)–(3.8) are presented in Table 3.8. The results are

not consistent with individuals moving prices. Focusing on Column 1, on days

when returns are positive, institutions purchase shares at higher prices than indi-

viduals purchase shares, but on days when prices are negative, there is no statis-

tical difference between the purchase prices. Economically, there is clearly a large

difference between the coefficients on positive- and negative-return days (25 bps

compared to 2 bps). Similarly, on days when returns are negative, institutions sell

at lower prices than do individuals (Column 3). On the sell side, the economic

magnitude of the difference is not as large as it is for purchase prices, but there is

still a very large statistical difference. Columns 2 and 4 present results for simi-

lar regressions, but here we allow the effect size to vary with the level of returns.

The statistical differences are even stronger in this specification. The economic

magnitude of these estimates is not particularly large: when a stock has a positive

return of 1%, institutions purchase shares at about 6.5 bps above individual pur-

chase prices. Nevertheless, the magnitude of the difference between positive and

negative return coefficients is both economically and statistically large.

96
 These results convincingly show that prices move in response to institutional

demand. Price changes cannot be attributed to the trading of individuals: when

prices rise, institutions buy shares at higher prices than do individuals, and when

prices fall, institutions sell shares at lower prices than do individuals. In either

case, prices are clearly not being pushed up or down by individuals. Combined

with the evidence presented in Section 3.3.2 for daily, weekly, and monthly hori-

zons, these results provide strong support for Hypothesis 1.

3.3.5 Returns following trade

The second hypothesis to be tested is that price reversion is more likely fol-

lowing days when trading is dominated by intragroup individual trading. It is

possible that when the bulk of trading is between individuals, without much in-

stitutional trading, prices are pushed away from fundamental values. If this is the

case, we might expect prices to revert in subsequent trading. To examine this, I

estimate the regression

Ri,t = α + β 1 Ind/Indi,t−1 Ri,t−1 + β 2 Inst/Insti,t−1 Ri,t−1

+ γ1 Ri,t−1 + γ2 Ind/Indi,t−1 + γ3 Inst/Insti,t−1

+ γ4 Ind/Insti,t−1 + γ5 Inst/Indi,t−1 + ei,t , (3.9)

for stock i on date t. “Inst” denotes financial institutions, and “Ind” denotes indi-

vidual investors. Pairs of groups represent purchasing of shares by the first group

97
from the second group. For example, “Ind/Inst” is the proportion of trading ac-

counted for by individuals buying shares from institutions.14

If returns tend to revert after days when individuals trade with other individ-

uals and returns are either high or low, β 1 should be negative. As shown in the

results presented in Table 3.9, this is precisely what we find. The negative rela-

tion appears in both the cross-section and time-series Fama-MacBeth regressions.

Moreover, there is no such effect for intragroup institutions trading—β 2 is not sig-

nificantly different from zero in either set of regressions. Consistent with the VAR

results presented above, the estimates in the last two lines of Table 3.9 show that

days when individuals buy shares from institutions—which we previously found

to be days when prices decline—are followed by further price declines. And when

institutions buy shares from individuals—which are days when prices increase—

prices tend to increase further on the following day. That is, prices revert less than

usual when there is more intragroup trading; they revert more when individu-

als trade with other individuals or when institutions trade with other institutions.

These results strongly support Hypothesis 2, that price reversion is more likely

after intragroup trading by individuals than after trading between the groups.

The price reversion that we observe must be caused by trading between or

within the groups. In particular, it is possible that institutions react to the price

movements caused by individuals by subsequently purchasing (selling) under-

priced (overpriced) shares from individuals. To investigate this, I examine the

14 In
the remainder of the chapter I use level of the trade variables and not the unexpected com-
ponents from the VAR in section 3.3.3. All reported results remain unchanged if I instead use the
residuals from each of the trade variable equations.

98
relation between institutional trading with individuals on date t and intragroup

trading by individuals on date t − 1. Specifically, I estimate the regressions

Inst/Indi,t = α + β 1 Ind/Indi,t−1 Ri,t−1 + β 2 Inst/Insti,t−1 Ri,t−1

+ γ1 Ri,t−1 + γ2 Ind/Indi,t−1 + γ3 Inst/Insti,t−1

+ γ4 Ind/Insti,t−1 + γ5 Inst/Indi,t−1 + ei,t , (3.10)

and

Ind/Insti,t = α + β 1 Ind/Indi,t−1 Ri,t−1 + β 2 Inst/Insti,t−1 Ri,t−1

+ γ1 Ri,t−1 + γ2 Ind/Indi,t−1 + γ3 Inst/Insti,t−1

+ γ4 Ind/Insti,t−1 + γ5 Inst/Indi,t−1 + ei,t , (3.11)

for stock i on date t. The dependent variable in regressions (3.10) and (3.11) is

either Inst/Indi,t or Ind/Insti,t —that is, the proportion of trading accounted for

by institutions purchasing shares from individuals or the proportion accounted

for by individuals purchasing shares from institutions.

Suppose trading on date t − 1 came largely from individuals trading with other

individuals, and that returns were positive. If this trading moved prices above

fundamentals, then we would expect institutions to be less likely to buy shares

from individuals on date t and more likely to sell shares to individuals on date

t. That is, we would expect β 1 to be negative in regression (3.10) and positive in

regression (3.11). As shown in the estimation results presented in Table 3.10, this

99
prediction is borne out by the data. Institutions are less likely to buy shares from

individuals (Panel A) and more likely to sell shares to individuals (Panel B) if more

trading on t − 1 occurred between two individuals and prices increased. Com-

bined with the results in Table 3.9, these results indicate that if prices are moved

by individuals, institutions subsequently trade with individuals in a direction that

leads prices to revert. This evidence provides strong support for Hypothesis 3.

3.4 Conclusion

This chapter studies the relation between returns and trading by individual

and institutional investors. I show that trading between individuals and insti-

tutions is relatively uncommon, but that these trades consistently lead to price

changes. In contrast to the recent work of Barber, Odean, and Zhu (2006) and

Hvidkjaer (2006), I find very little evidence of price effects from the trading of

individual investors. In addition, I show that when prices change as a result of

individual investors trading with other individuals, price reversion is more com-

mon than when they trade with institutions. Moreover, this reversion is caused

by institutions subsequently trading with individuals in a direction that pushes

prices back toward previous levels.

There are a number of reasons why it is important to understand what type of

trading leads to price changes. First, if prices are determined primarily by traders

with poor information or are contaminated by beliefs of investors suffering from

behavioral biases, then features of the return series, such as volatility and covari-

100
ance with macroeconomic variables like consumption may be spurious. This is of

obvious importance to research in asset pricing, which seeks to explain relations

among these variables. Second, a large literature seeks to understand the per-

formance of mutual funds. Part of the challenge in assessing this performance is

the difficulty of determining whether trading by mutual funds causes price move-

ments. I find that the trading of mutual funds and other institutions moves prices,

so findings of positive correlation between institutional trading and returns at low

frequencies cannot be taken as evidence of good performance. However, my re-

sults do not rule out the possibility of further price movements allowing funds to

earn profits.

In contrast to previous papers, the results presented here are calculated with

daily data. At this relatively high frequency, it is apparent that price movements

are generally caused by the trading demands of institutions. And if prices do

move when individuals trade with each other, they quickly revert. This suggests

that while there may be short-term price effects caused by individual investors,

prices are unlikely to be affected by such distortions at longer horizons.

101
 Table 3.1: Stock Returns Following Trading by Institutions and Households
This table presents average returns to stocks in the year subsequent to trading by institutions and
individuals, separated by the rank of each group’s buying pressure. In June and December of each
year, firms are assigned to one of three portfolios for each group based on the proportion of buy
volume for that group, Bi,t /( Bi,t + Si,t ). The table reports the returns to these portfolios over the
subsequent year. t-statistics are presented in parentheses. Statistical significance at the 10%, 5%
and 1% levels is denoted by † , ∗ , and ∗∗ , respectively.

Institutions Households

Most sold -0.0011 0.1175

(-0.04) (4.34)∗∗

Most bought 0.0892 0.0492

(3.12)∗∗ (1.85)†

102
 Table 3.2: Summary Statistics
This table provides summary statistics for the data. The sample period is 1995–2003. The “Nomi-
nees/ADRs” group is identified using the technique explained in Section 3.2.2. “Other” includes
Government Agencies, Nonprofit Organizations, and Registered Foreigners. The number of ac-
counts and total number of trades are shown in the first two columns, respectively. The remaining
columns present averages for per-trade, per-day, or per-account data.

Number of: Average of:

Shares Value Trades Securities
Accounts Trades
per trade per trade per day traded
(MM) (000s) (000s) (000s)

Financial institutions 920 14.0 3.6 77.8 6.4 33

Households 583,518 10.7 1.2 7.3 4.9 4
Nominees / ADRs 47 7.4 4.2 65.5 3.4 169
Nonfinancial corporations 29,186 4.6 3.8 57.3 2.1 8
Other 12,269 0.8 1.6 30.8 0.4 2
All 625,940 37.5 2.9 47.8 3.4 43

103
 Table 3.3: Trader Interaction
This table shows the amount of trading (in percent) that occurs in total and between the different
investor groups. The expected amount of trading between groups is calculated assuming random
interaction in proportion to the percent of total trades based on number of shares traded presented
in the first row. The actual amount of trading is estimated using the technique discussed in the text.
The total number of shares traded between each group is calculated for each stock/day during the
period 1995–2003. The percentages do not sum to 100% because the trading of certain groups that
account for very little total trading volume is omitted.

Financial Households Nominees / Nonfinancial

Institutions ADRs Corporations

Percent of trading (by shares) 38.1 20.6 15.2 20.9

Percent of trading (by value) 40.9 17.4 16.1 20.5

Expected % of trading with:

Financial institutions 14.5
Households 15.7 4.2
Nominees / ADRs 11.6 6.2 2.3
Nonfinancial corporations 15.9 8.6 6.4 4.4

Estimated % of trading with:

Financial institutions 20.6
Households 9.3 9.1
Nominees / ADRs 10.1 3.5 5.7
Nonfinancial corporations 13.9 8.1 4.9 6.2

104
 Table 3.4: Trader Interaction—Cross-Sectional Statistics
This table shows the average amount of trading (buying and selling, in percent) that occurs between
the different investor groups. The amount of trading is estimated for each stock/day using the
technique discussed in the text. The time-series average is then calculated for each stock, and the
table reports cross-sectional statistics for those means (N = 106 stocks).

Group 1 Group 2 Mean Std. Dev. Min. Max.

Financial institutions Financial institutions 14.7 10.5 0.5 41.4

Households 5.9 2.6 1.6 14.8
Nominees / ADRs 4.0 3.8 0.0 14.4
Nonfinancial corporations 5.2 2.6 0.3 13.2

Households Households 19.3 17.1 0.1 60.6

Nominees / ADRs 2.6 1.9 0.0 8.6
Nonfinancial corporations 6.1 4.4 0.5 17.1

Nominees / ADRs Nominees / ADRs 4.2 4.5 0.0 26.4

Nonfinancial corporations 1.6 1.1 0.0 4.6

Nonfinancial corporations Nonfinancial corporations 4.5 2.2 0.5 15.9

105
 Table 3.5: Returns and Group Interaction—Daily
This table presents the results of the regression

Ri,t = α + βRi,t−1 + γ1 Inst/Insti,t + γ2 Inst/Indi,t + γ3 Ind/Insti,t + γ4 Ind/Indi,t + ei,t ,

where the notation A/Bi,t represents the proportion of trading that is accounted for by investors from Group A purchasing shares from investors in Group B. “Ind” and “Inst” denote
individuals and institutions, respectively. Panel A reports results where the A/B trade variables are calculated as the proportion of total trade volume each day, and Panel B uses trade
variables calculated as the proportion of daily turnover (shares traded divided by number of shares outstanding) each day. The regression is estimated using the Fama-MacBeth (“FM”)
approach in two ways: averaging results from time-series regressions by stock or from cross-sectional regressions by date. t-statistics for the FM by date regressions are calculated using
standard errors robust to heteroscedasticity and autocorrelation using the Newey and West (1987) adjustment with five lags. The columns labeled “Sig. at 5%” report the percentage
of regressions in which the coefficient is significantly negative or positive at the 5% level. Cross-sectional and time-series average standard deviations of the independent variables are
reported in the columns labeled “C.S. Std. Dev” and “T.S. Std. Dev,” respectively. Statistical significance at the 10%, 5% and 1% levels is denoted by † , ∗ , and ∗∗ , respectively.

FM by stock FM by date
Sig. at 5% C.S. Sig. at 5% T.S.
Estimate t-stat Neg. Pos. Std. Dev Estimate t-stat Neg. Pos. Std. Dev
Panel A: Proportion of Daily Trading Volume
Intercept 0.0022 3.72∗∗ 0% 44% 0.0011 3.16∗∗ 20% 26%
Ri,t−1 -0.0634 -5.89∗∗ 52% 17% 0.0011 -0.0837 -11.46∗∗ 28% 11% 0.0009
Buyer: Seller:

106
Institutions Institutions 0.0092 1.35 8% 11% 0.1481 0.0003 0.50 4% 4% 0.1691
Households 0.0316 5.36∗∗ 0% 61% 0.1151 0.0189 16.09∗∗ 2% 11% 0.1180

Households Institutions -0.0423 -4.74∗∗ 72% 0% 0.1145 -0.0172 -12.70∗∗ 11% 3% 0.1130
Households -0.0025 -0.24 27% 0% 0.2002 -0.0023 -4.41∗∗ 9% 6% 0.2306
Panel B: Proportion of Daily Turnover
Intercept -0.0020 -8.45∗∗ 42% 3% -0.0008 -3.06∗∗ 24% 19%
Ri,t−1 -0.0791 -8.86∗∗ 58% 7% 0.0011 -0.1338 -19.87∗∗ 36% 7% 0.0009

Buyer: Seller:
Institutions Institutions 0.0026 4.57∗∗ 0% 42% 0.9767 0.0011 3.69∗∗ 4% 10% 0.7566
Households 0.0481 4.29∗∗ 1% 76% 0.5838 0.0390 19.62∗∗ 2% 27% 0.3728

Households Institutions -0.0339 -3.24∗∗ 48% 5% 0.2920 -0.0167 -10.15∗∗ 20% 6% 0.2782
Households 0.0479 2.54∗ 2% 69% 0.3204 0.0033 2.27∗ 14% 18% 0.4697

Number of observations 134196 134196

Number of FM regressions 106 (stocks) 2184 (days)
 Table 3.6: Returns and Group Interaction—VAR Results
This table presents coefficient estimates from the dynamic structural VAR(10) in equation (3.3),

10
A0 yi,t = ∑ Ai yi,t−k + ηt ,
k =1

where yi,t = (Ind/Indi,t , Inst/Insti,t , Ind/Insti,t , Inst/Indi,t , Ri,t )0 . The notation A/Bi,t denotes the
proportion of trading in stock i on date t that comes from Group A purchasing shares from Group
B, and “Ind” and “Inst” denote individuals and institutions, respectively. A0 is the lower diagonal
matrix from the triangular factorization of the error covariance in the reduced-form VAR (equa-
tion (3.2)). Separate regressions are estimated for each of the 106 stocks in the sample. The table
reports cross-sectional averages of coefficient estimates for k = 0, 1, 2. Standard errors, calculated
from the cross-sectional distribution of the coefficient estimates, are reported in parentheses. For
brevity, the constant terms and lags of order three and higher are omitted. Statistical significance at
the 10%, 5% and 1% levels is denoted by † , ∗ , and ∗∗ , respectively.

Elements of yi,t−k
Equation Ind/Ind Inst/Inst Ind/Inst Inst/Ind Return
k=0
Ind/Ind – – – – –

Inst/Inst -0.1303 – – – –
(0.0355)∗∗
Ind/Inst 0.0592 -0.0041 – – –
(0.0255)∗ (0.0054)
Inst/Ind 0.0859 -0.0074 -0.1091 – –
(0.0325)∗∗ (0.0087) (0.0057)∗∗
Return -0.0161 0.0059 -0.0433 0.0344 –
(0.0046)∗∗ (0.0041) (0.0070)∗∗ (0.0064)∗∗

k=1
Ind/Ind 0.1489 -0.0201 0.0064 0.0002 -0.0135
(0.0074)∗∗ (0.0093)∗ (0.0073) (0.0069) (0.0213)
Inst/Inst 0.0227 0.0964 0.0415 0.0488 -0.0303
(0.0200) (0.0057)∗∗ (0.0089)∗∗ (0.0063)∗∗ (0.0174)†
Ind/Inst 0.0061 0.0181 0.1493 -0.0083 -0.0560
(0.0077) (0.0060)∗∗ (0.0062)∗∗ (0.0039)∗ (0.0130)∗∗
Inst/Ind 0.0113 0.0339 -0.0045 0.1451 0.0231
(0.0067)† (0.0094)∗∗ (0.0035) (0.0116)∗∗ (0.0131)†
Return 0.0047 0.0044 0.0019 -0.0048 -0.0680
(0.0057) (0.0026)† (0.0024) (0.0020)∗∗ (0.0107)∗∗

k=2
Ind/Ind 0.0773 -0.0093 0.0054 0.0013 0.0191
(0.0066)∗∗ (0.0133) (0.0073) (0.0063) (0.0163)
Inst/Inst -0.0157 0.0521 0.0005 0.0249 0.0107
(0.0242) (0.0044)∗∗ (0.0094) (0.0091)∗∗ (0.0157)
Ind/Inst 0.0119 0.0096 0.0760 -0.0103 -0.0516
(0.0051)∗∗ (0.0044)∗ (0.0057)∗∗ (0.0037)∗∗ (0.0132)∗∗
Inst/Ind 0.0004 0.0110 0.0056 0.0696 0.0567
(0.0056) (0.0045)∗∗ (0.0039) (0.0050)∗∗ (0.0119)∗∗
Return 0.0023 0.0011 0.0010 -0.0006 -0.0260
(0.0025) (0.0015) (0.0018) (0.0030) (0.0055)∗∗

107
 Table 3.7: Returns and Group Interaction—Weekly and Monthly
This table presents the results of the regression

Ri,t = α + βRi,t−1 + γ1 Inst/Insti,t + γ2 Inst/Indi,t + γ3 Ind/Insti,t + γ4 Ind/Indi,t + ei,t ,

where the notation A/Bi,t represents the proportion of trading that is accounted for by investors
from Group A purchasing shares from investors in Group B. “Ind” and “Inst” denote individuals
and institutions, respectively. Panels A and B report results for regressions using data aggregated
into weekly and monthly observations, respectively. The regression is estimated using the Fama-
MacBeth (“FM”) approach in two ways: averaging results from time-series regressions by stock
(Columns 3 and 4), or from cross-sectional regressions by date (Columns 5 and 6). t-statistics for
the FM by date regressions are calculated using standard errors robust to heteroscedasticity and
autocorrelation using the Newey and West (1987) adjustment with five lags. Statistical significance
at the 10%, 5% and 1% levels is denoted by † , ∗ , and ∗∗ , respectively.

FM by stock FM by date
Estimate t-stat Estimate t-stat

Panel A: Weekly Horizon

Intercept 0.0057 5.43∗∗ 0.0056 2.77∗∗
Ri,t−1 -0.0355 -3.82∗∗ -0.0713 -5.61∗∗

Buyer: Seller:
Financial institutions Financial institutions 0.0320 1.14 0.0011 0.42
Households 0.1882 3.87∗∗ 0.0630 8.37∗∗

Households Financial institutions -0.2292 -5.59∗∗ -0.0680 -11.07∗∗

Households -0.0356 -0.42 -0.0103 -3.09∗∗
Number of observations 33124 33124
Number of FM regressions 106 (stocks) 456 (weeks)

Panel B: Monthly Horizon

Intercept 0.0117 1.12 0.0290 2.63∗∗
Ri,t−1 0.0001 0.01 -0.0103 -0.33

Buyer: Seller:
Financial institutions Financial institutions -0.0866 -1.13 -0.0197 -0.77
Households 1.1009 5.01∗∗ 0.1689 3.12∗∗

Households Financial institutions -1.7529 -2.97∗∗ -0.2742 -8.52∗∗

Households 2.4445 1.73 -0.0295 -1.49
Number of observations 7680 7680
Number of FM regressions 106 (stocks) 104 (months)

108
 Table 3.8: Returns and Group Interaction—Intraday Price Evidence
This table presents results from Fama-MacBeth regressions
I H
b̄i,t − b̄i,t = β 1 · 1{ Ri,t >0} + β 2 · 1{ Ri,t <0} + ei,t

and
I H
s̄i,t − s̄i,t = β 1 · 1{ Ri,t >0} + β 2 · 1{ Ri,t <0} + ei,t ,
where b̄ and s̄ denote the average purchase and sale prices, respectively, for stock i on date t, and
superscript I and H denote institutions and households, respectively. The indicator function, 1{·} ,
takes a value of one only when the condition in curly brackets is true; otherwise it is zero. Results
for these specifications are presented in Columns (1) and (3), respectively. Columns (2) and (4)
show results for an alternative specification in which the price difference is allowed to vary with
the magnitude of the return, | Ri,t |. For each stock and day, the average price at which individuals
purchase (sell) shares is subtracted from the average price at which institutions purchase (sell)
shares. This difference is used as one observation in a time-series regression for each stock. The
table reports the mean coefficients estimates, with t-statistics (in parentheses) derived from the
cross-sectional distribution of coefficient estimates. Statistical significance at the 10%, 5% and 1%
levels is denoted by † , ∗ , and ∗∗ , respectively.

Dependent Variable
I
b̄i,t H
− b̄i,t I − s̄ H
s̄i,t i,t
(1) (2) (3) (4)

1{ Ri,t >0} 0.0025 -0.0011

(6.46)∗∗ (-1.77)†
1{ Ri,t <0} 0.0002 -0.0016
(0.68) (-6.24)∗∗

1{ Ri,t >0} × | Ri,t | 0.0649 -0.0005

(11.67)∗∗ (-0.09)
1{ Ri,t <0} × | Ri,t | 0.0026 -0.0718
(0.37) (-10.66)∗∗

Number of observations 127501 127501 127501 127501

Number of regressions (stocks) 106 106 106 106

109
 Table 3.9: Returns Following Trade
This table presents results for the regression

Ri,t = α + β 1 · Ind/Indi,t−1 Ri,t−1 + β 2 · Inst/Insti,t−1 Ri,t−1 + γ · Controls + ei,t ,

for stock i on date t. “Inst” denotes financial institutions, and “Ind” denotes individual investors.
Pairs of groups represent purchasing of shares by the first group from the second group. For ex-
ample, “Ind/Inst” is the proportion of trading accounted for by individuals buying shares from fi-
nancial institutions. The regression is estimated using the Fama-MacBeth (“FM”) approach in two
ways: averaging results from time-series regressions by stock (Columns 2 and 3), or from cross-
sectional regressions by date (Columns 4 and 5). t-statistics for the FM by date regressions are
calculated using standard errors robust to heteroscedasticity and autocorrelation using the Newey
and West (1987) adjustment with five lags. Statistical significance at the 10%, 5% and 1% levels is
denoted by † , ∗ , and ∗∗ , respectively.

FM by stock FM by date
Estimate t-stat Estimate t-stat
Ind/Indi,t−1 × Ri,t−1 -0.2534 -4.09∗∗ -0.1840 -3.50∗∗
Inst/Insti,t−1 × Ri,t−1 0.0103 0.36 0.0443 1.26

Controls:
Intercept 0.0011 5.41∗∗ 0.0009 2.52∗
Ri,t−1 -0.0281 -2.63∗∗ -0.0559 -6.62∗∗
Ind/Indi,t−1 -0.0012 -1.31 -0.0009 -1.19
Inst/Insti,t−1 0.0006 1.18 0.0004 0.72
Ind/Insti,t−1 -0.0018 -2.55∗ -0.0033 -5.84∗∗
Inst/Indi,t−1 0.0022 3.09∗∗ 0.0013 2.37∗

Number of observations 151324 151324

Number FM of regressions 106 (stocks) 2184 (days)

110
 Table 3.10: Institutions’ Response to Individual Trading
This table presents results for the regressions

Inst/Indi,t = α + β 1 · Ind/Indi,t−1 Ri,t−1 + β 2 · Inst/Insti,t−1 Ri,t−1 + γ · Controls + ei,t

and
Ind/Insti,t = α + β 1 · Ind/Indi,t−1 Ri,t−1 + β 2 · Inst/Insti,t−1 Ri,t−1 + γ · Controls + ei,t ,
for stock i on date t. “Inst” denotes financial institutions, and “Ind” denotes individual investors. Pairs of groups repre-
sent purchasing of shares by the first group from the second group. For example, “Ind/Inst” is the proportion of trading
accounted for by individuals buying shares from financial institutions. The dependent variable in Panel A is the proportion
of trading that comes from individuals purchasing shares from institutions (Ind/Insti,t ), and in Panel B it is the proportion
of trading that comes from institutions purchasing shares from individuals (Inst/Indi,t ). The regression is estimated using
the Fama-MacBeth (“FM”) approach in two ways: averaging results from time-series regressions by stock (Columns 2 and
3), or from cross-sectional regressions by date (Columns 4 and 5). t-statistics for the FM by date regressions are calculated
using standard errors robust to heteroscedasticity and autocorrelation using the Newey and West (1987) adjustment with
five lags. Statistical significance at the 10%, 5% and 1% levels is denoted by † , ∗ , and ∗∗ , respectively.

FM by stock FM by date
Estimate t-stat Estimate t-stat

Panel A: Dependent Variable=Inst/Indi,t

Ind/Indi,t−1 × Ri,t−1 -0.6939 -3.12∗∗ -1.7931 -4.02∗∗
Inst/Insti,t−1 × Ri,t−1 0.1050 1.20 -0.0723 -0.43

Controls:
Intercept 0.1240 21.45∗∗ 0.1020 48.13∗∗
Ri,t−1 0.0557 1.67† 0.1645 2.34∗
Ind/Indi,t−1 0.0239 2.81∗∗ 0.0519 7.10∗∗
Inst/Insti,t−1 0.0016 0.44 -0.0241 -7.28∗∗
Ind/Insti,t−1 -0.0643 -11.14∗∗ -0.0180 -3.88∗∗
Inst/Indi,t−1 0.2180 22.09∗∗ 0.2780 47.83∗∗

Panel B: Dependent Variable=Ind/Insti,t

Ind/Indi,t−1 × Ri,t−1 0.3886 2.03∗ 0.9371 2.42∗
Inst/Insti,t−1 × Ri,t−1 -0.1958 -2.38 ∗ -0.1951 -1.54

Controls:
Intercept 0.1186 19.21∗∗ 0.0847 42.22∗∗
Ri,t−1 0.0308 0.89 -0.0116 -0.20
Ind/Indi,t−1 0.0416 5.55∗∗ 0.0984 14.26∗∗
Inst/Insti,t−1 0.0106 2.88∗∗ -0.0150 -4.57∗∗
Ind/Insti,t−1 0.2452 24.36∗∗ 0.3032 61.63∗∗
Inst/Indi,t−1 -0.0483 -10.48∗∗ 0.0005 0.13
Number of observations 151324 151324
Number FM of regression 106 (stocks) 2184 (days)

111
 Figure 3.1: Stylized Timeline of Price Path Around Trade
This figure shows alternative price paths following a trade. The trade takes place at time t0 . In
the top two figures, the trade is initiated by the buyer, and the price immediately increases. In the
bottom two figures, the trade is initiated by the seller, and the price immediately decreases. If the
trade initiator is uninformed, prices subsequently revert. If the trade initiator is informed, there is
no such reversion.
Uninformed buyer Informed buyer

Price Price
6 6
 
PP 
PP
Buyer- P
Initiated p0 p0

- -
t0 Time t0 Time

Price Price
6 6

Seller-
Initiated p0 p0
 
PP 
PP
P
- -
t0 Time t0 Time

112
 Figure 3.2: Cumulative Price Impact Functions
This figure plots the accumulated orthogonalized impulse response function for the structural dy-
namic VAR in equation (3.3),
10
A0 yi,t = ∑ Ai yi,t−k + ηt ,
k =1

where y = (Ind/Ind, Inst/Inst, Ind/Inst, Inst/Ind, Return)0 and A0 is the lower diagonal matrix
from the triangular factorization of the error covariance in the reduced-form VAR (equation (3.2)).
The graphs show the cumulative effect on returns (in basis points) of a one standard deviation
shock (ηi,t ) to each of the trade variables during the trading day that occurs between time −1 and
time 0. Returns are calculated using closing prices each day. Inst/Ind (solid) is the proportion
of trading accounted for by institutions purchasing shares from individuals; Ind/Ind (long dash)
represents trading between two individuals; Inst/Inst (short dash) represents trading between two
institutions; and Ind/Inst (dash-dot) represents individuals purchasing shares from institutions.

40

30

20
Cumulative Return (bps)

10

-10

-20

-30

-40
-1 0 1 2 3 4 5 6 7 8 9 10
Trading Days
Inst/Ind Inst/Inst Ind/Ind Ind/Inst

113
 Chapter IV

When Are Individual Investors Informed?

The decisions of individuals lie at the foundation of economic theory. Unfor-

tunately, there are very few real-world settings where these decisions and their

economic consequences are directly observed at the individual level. One setting

that has proved fruitful for examining such decisions and outcomes is financial

markets. In a series of seminal papers, Odean (1998, 1999) and Barber and Odean

(2000, 2001) examined the trading decisions of a large number of investors at a

large discount brokerage firm in the U.S. Subsequently, a large literature has de-

veloped analyzing the behavior of individual investors in a number of settings.1

The main stylized fact that emerges from this literature is that individuals

make poor investment decisions. For example, individuals trade too much, gener-

ating transaction costs without increasing their returns; they make bad choices, as

stocks they sell subsequently earn higher returns than stocks they buy; and they
1 Some representative papers from this literature include Grinblatt and Keloharju (2000, 2001a,
2001b), Benartzi (2001), Goetzmann and Kumar (2008), Ivković and Weisbenner (2005), Ivković,
Sialm, and Weisbenner (2006), and Feng and Seasholes (2005).

114
hold on to their losing stocks but sell their winning stocks, which is a sure way to

lose money in the presence of momentum.

In this chapter, I revisit the question of whether individual investors make

good investment decisions. I find that the results of previous papers are correct,

but incomplete. In particular, I begin by focusing on one of the earlier findings of

this literature: that post-purchase returns are lower than post-sale returns (Odean

1999). I show that while this is indeed the case unconditionally, it is not the case

when we consider those investment decisions that are more likely, ex ante, to be

informed.

Broadly speaking, there are two reasons an investor might buy or sell a stock:

liquidity, or information. As an example of a liquidity-motivated trade, an in-

vestor may have earned income that she wishes to save, and chooses to invest in

equities. Or she might have unexpected costs, and chooses to sell some securities

to generate the needed cash. In both of these examples, the investor either pur-

chases one or more stocks or sells one or more stocks; she does not transfer money

from one stock to another. Suppose instead that the investor chooses to sell one

stock she owns and use the proceeds to purchase another stock. In this case, she

may have received information that causes her to believe the asset she owns is

going to decline in value, or that the stock she chooses to buy is going to increase

in value; in either case, it is clear that she believes the stock she buys will earn a

115
higher return than the stock she sells. That is, she is forced to place an “extra”

trade relative to what she would have done if she were not cash-constrained.2

The economic intuition that comes from these examples suggests a way to

identify which trades are more likely to be informed. If we observe an investor

liquidating their holdings in one stock and then using the proceeds to purchase

another stock, it is more likely that these transactions are motivated by informa-

tion than other transactions. I use this intuition to identify informed trades among

individual investors. When an investor sells shares in one stock and subsequently

purchases shares in another stock, within a short period of time, I call this a “sub-

stitution” trade. Obviously, not all of these trades are perfectly informed, but

the results below provide strong evidence that this classification does capture a

meaningful distinction between informed and uninformed trades. The details of

the classification procedure I employ are provided in section 4.3, below.

I show that the post-purchase and post-sale returns to substitution trades are

significantly better than the returns for non-substitution (“regular”) trades. This

finding suggests that investors are not as bad at making investment decisions as

previously documented. In particular, it is possible to identify, ex ante, a subset of

investment decisions that are likely to be driven by information. In this regard,

this chapter is related to the work of Coval, Hirshleifer, and Shumway (2005), who

find that a subset of investors are able to beat the market consistently, and Seru,
2 Alternatively, she may not believe that the stock will earn a higher return than stocks she

already holds, but wishes to rebalance her portfolio to change the riskiness of her holdings. The
extent to which this is true will make it more difficult for me to find the results I report below.

116
Shumway, and Stoffman (2007), who find that investors learn to avoid behavioral

biases as they become experienced.

The present study is also related to the recent work of Alexander, Cici, and Gib-

son (2007), who find that liquidity-motivated investment choices of mutual fund

managers earn lower returns than trades that are more likely to be motivated by

private information. The authors distinguish between transactions necessitated

by investor fund flows and those that are not, and use a similar argument as I do

here to claim that the latter trades are more likely to be informed. For example,

they note that a fund manager who purchases a stock even when investors are

withdrawing money from the mutual fund is likely to believe strongly that the

stock is undervalued, whereas if the purchase is made in the presence of fund

inflows, it is more likely to be just a reallocation of holdings from excess cash to

stocks. They find that managers do, in fact, beat the market on these trades, and

not when they trade because of investor inflows and outflows.

I also show that the pattern of returns before and after regular trading differs

from that surrounding substitution trades. When substitution trades are domi-

nated by purchases, prices rise and do not subsequently revert. In contrast, prices

respond negatively to regular purchases, and subsequently revert. This provides

additional evidence that there is distinctly different information content in sub-

stitution and non-substitution trades. These results complement the recent work

of Barber, Odean, and Zhu (2006), Hvidkjaer (2006), and Kaniel, Saar, and Titman

(2008), all of whom examine the relation between individual investor order flow

and stock returns.

117
 A particular strength of this chapter is that I document consistent results us-

ing two large datasets from different countries over different time periods. This

adds considerable support the notion that the finding reflects a result of general

economic behavior, and not a quirk of one particular setting, or a statistical fluke.

The remainder of this chapter is organized as follows. In Section 4.2, I discuss

the main datasets used in this study. I outline my methods and present results in

Section 4.3. Section 4.4 concludes.

4.1 Related literature

The present study contributes to the growing literature on the relation between

the trading of individual investors and stock returns.

Obizhaeva (2007) examines the price impact of trades by institutions that are

transitioning from existing positions to a target portfolio. These transitions fre-

quently occur after a change in the portfolio manager, and the author provides

evidence that transition purchases are informed while sales are likely to be liquid-

ity motivated.

Examining quarterly 13(f) filings of institutional stock holdings, Sias, Starks,

and Titman (2006) provide evidence that trading by institutions causes the ob-

served correlation between quarterly changes in institutional ownership and re-

turns. In addition, they find evidence in support of the notion that institutional

investors are better informed than other traders, and that their information is in-

corporated into prices as they trade.

118
 San (2007) also uses 13(f) filings to conclude that individuals earn higher re-

turns than institutions during the period 1981 to 2004. A potential problem with

this approach, however, is the assumption that traders who do not file 13(f) reports

are individuals, while in fact this group would include hedge funds and smaller

money managers.3 Nevertheless, this is the same approach taken by Gompers

and Metrick (2001), who find that aggregate institutional performance is better

than individual performance by 0.67% per year, and that this difference is driven

more by demand shocks caused by institutional preferences for liquid stocks than

by institutions being “smarter” than individuals.

Kaniel, Saar, and Titman (2008) analyze trades placed by individuals in NYSE

stocks during 2000–2003. They find that individuals are contrarian at a monthly

horizon: they purchase stocks after recent declines and sell after recent increases.

Moreover, they document positive excess returns in months following high levels

of individual buying, which is consistent with individuals supplying liquidity to

institutions.

4.2 Data

In addition to the Finish data outlined in Chapter I, I use a sample of trading

records from investors at a large discount brokerage firm in the United States.4
3 The Securities Exchange Act requires institutional investment managers who exercise invest-
ment discretion over securities having an aggregate fair market value on the last trading day of
any month of at least $100 million to file a 13(f) report within 45 days of the end of each calendar
quarter. These institutions are required to report only those positions of greater than 10,000 shares
or $200,000 in market value. See http://www.sec.gov/about/forms/form13f.pdf.
4 I thank Terry Odean for sharing these data.

119
This data set contains 3.1 million observations, covering the trades of 126,488 ac-

counts from January, 1991 to November, 1996. Each observation consists of an

account identifier, a security identifier (CUSIP), trade date, number and price of

shares traded and a buy/sell indicator.

I use the CUSIP to merge the trade data with data from the Center for Re-

search in Security Prices (CRSP). I extract the return series, ret, to calculate post-

transaction returns, as discussed below. This variable includes both the capital

gains and any dividends or other payouts earned by shareholders.

4.3 Methods and Results

In this section I discuss my methods, and present the results. I first examine the

general pattern of returns following purchases and sales by individual investors,

and confirm the results of Odean (1999) in different datasets. These results are

“unconditional” in the sense that I am not classifying the trades in any way before

calculating post-transaction returns. I then explain how I identify those trades

that are more likely to be informed (“substitution” trades), and show that the pat-

tern of returns for these trades is, in fact, quite different. While unconditionally

the returns following a purchase are lower than the returns following a sale, sub-

stitution purchases earn higher returns and substitution sales earn lower returns

than non-substitution (“regular”) trades. This provides significant evidence that

substitution trades are more informed than other trades.

120
4.3.1 Unconditional post-transaction returns

A straightforward way to assess the success of an individual’s trading record is

to examine the return of stocks purchased and sold subsequent to the transaction

date. Therefore, for each transaction, I calculate the post-transaction holding pe-

riod return, Ri,j,h , for investor i’s jth transaction over a horizon of h trading days.

Holding period returns are calculated as the percentage increase in a cumulative

return index that includes dividend payments. To ensure that my results are not

affected by bid-ask bounce, I calculate returns using the closing price on the day

after the transaction.

These post-transaction returns in the U.S. data are summarized in Table 4.1

for purchases and sales, separately. Odean’s (1999) finding that post-sale returns

are higher than post-purchase returns is confirmed in these data, which partially

overlap with the data used in that study. For example, in the 63 trading days

(3 months) following a trade, average returns are 3.34% for purchases and 3.88%

for sales, for a 54 basis point difference. As Odean points out, this implies that

investors would have earned higher returns (approximately 2% in this case) if

they had simply abstained from trading. It is also interesting to note in this sample

that individuals were approximately 20% more likely to purchase shares than to

sell shares in this sample.

The results in Table 4.1 are calculated using the same approach as Odean (1999).

That is, returns are aggregated across all accounts without first aggregating by

each account. Therefore, the post-purchase returns of one investor are being com-

121
pared to the post-sale returns of another investor. If subsets of investors have

different skill levels, and are present in the data over different time periods, this

approach could lead to incorrect inference. In order to rule this out, I repeat the

analysis using two different approaches.

First, I run the regression

Ri,j = αi + β · 1{ j=Buy} + ei,j , (4.1)

where the αi are individual fixed effect and 1{·} is a dummy variable. The regres-

sion is repeated for each post-transaction return horizon. The individual-specific

intercept controls for all time-invariant individual-specific heterogeneity, such as

different levels of skill. Results from these regressions are shown in Table 4.2. Ac-

counting for individual-specific heterogeneity mitigates the differences in returns,

although purchases still significantly underperform sales. For example, at the 84-

day horizon, buys underperform sells by 45 basis points, compared to the 56 basis

points calculated above. Nevertheless, the results remain strongly significant at

conventional levels.

Second, I calculate average buy- and sell-returns for each account, and then

calculate the difference in means for each account before taking the cross-sectional

mean across all accounts. This approach is designed to examine whether any par-

ticular investor’s purchases can be expected to underperform that same investor’s

sales. The results from this approach are shown in Table 4.3, and are again quite

similar to those already reported. Means in Panel A are equally weighted, and

122
in Panel B are weighted by the number of trades made by the account. Panel

C reports equally weighted means excluding all accounts that placed fewer than

ten trades. With the exception of the 5-day return in Panel C, all results are sig-

nificantly negative at the 1% level. The weighted average results in Panel B are

probably the most conservative, and are very close to the coefficient estimates in

Table 4.2.

The results presented in this subsection confirm the findings of Odean (1999).

These results are robust to using various statistical approaches, and are confirmed

in the Finnish data set.5 In the rest of this chapter, I ask whether these results con-

tinue to hold when we differentiate between trades that are likely to be informed

(“substitution” trades) and those that aren’t.

4.3.2 Substitution trades

As outlined above, I define a “substitution trade” as a group of transactions

comprising at least two different stocks, where a purchase of one stock is followed

by a sale of another within two to five trading days. Moreover, the total value of

shares purchased must be “close” to the total value of shares sold, where “close”

is defined as within two to 25 percent of the value traded.6 These are trades where

the investor has made a clear decision to substitute one security for another.

For example, suppose we observe the following transactions for an investor:

5 Inthe interests of space, the various sets of tests for Finnish data are not reported in this
section, but the fact that they are similar to the U.S. results will become clear in the remainder of
this chapter. All results are available from the author upon request.
6 The results presented below are robust to using a longer window of up to 15 trading days.

123
 Transaction Date Stock ID Buy/Sell Value Traded

1 1/15/02 A Buy $800

2 2/12/02 A Sell 600

3 2/13/02 B Buy 585

4 2/18/02 C Buy 500

5 5/21/02 B Sell 700

In this case, transactions 1, 4, and 5 would be classified as “regular” trades, while

transactions 2 and 3 would be classified as one “substitution” trade.

The number of transactions classified as being part of a substitution trade is

reported in Table 4.4 for the U.S. data. The criteria for classification become less

selective as we increase the number of days in the window, or how close the value

of shares purchased and sold is (denoted “Cutoff” in the table), so the number of

substitution trades increases as we move down and to the right in the table. If

we look only at transactions that occur within one of each other and where the

total value of shares purchased is within 1% of shares sold, 2.9% of transactions

are classified as substitution trades. At the other end of the spectrum, 33.1% of

transactions are classified as substitution trades when we allow transactions to

occur over a 15 day period, and having a purchase value within 25% of the value

of shares sold.

It is difficult to gauge how reasonable these numbers are, although some guid-

ance is available from the existing literature. Using a structural model similar to

that of Glosten and Milgrom (1985), Easley, Kiefer, O’Hara, and Paperman (1996)

124
estimate the probability of informed trading in a number of large NYSE-traded

stocks. Their estimation procedure uses only data on the total number of buyer-

and seller-initiated transactions in a stock on each day, and therefore includes the

trades institutions, which are likely to make up the bulk of trading in most large

stocks. They obtain estimates on the order of 15% of trading being informed. Since

I am looking exclusively at trading by individual investors, it seems reasonable to

expect a lower percentage of informed trades in my sample.

For the remainder of this chapter, I use the 2-day/5% criteria for classification.

That is, a trade is classified as a substitution trade if there are no more than two

days between the first transaction and the last transaction, and if the total value of

shares purchased does not differ from the total value of shares sold by more than

5%. While this choice is somewhat arbitrary, I confirm in unreported tests that

the results are not sensitive to this particular choice of parameters. The general

pattern of these unreported results is that they are weaker as we lengthen the

window of classification and increase the difference allowed between purchase

and sale amounts. This makes sense, since more and more trades are likely to be

incorrectly classified as information driven when we use these looser criteria.

4.3.3 Returns to substitution trades

Results presented in the existing literature, and confirmed above, show that

investors on average make the wrong trading decisions: they purchase stocks

that subsequently decrease in value, and sell stocks whose prices subsequently

increase. In this section, I investigate whether this pattern continues to hold up

125
when we consider those trades that are more likely to be informed (that is, the

“substitution” trades). Results from the U.S. data are presented first, and then the

results from Finland.

The key result of this chapter comes from regression tests of post-transaction

returns. In particular, I report below coefficient estimates from the regression

Ri,t = α + β 1 · Buy + β 2 · Substitution + β 3 · Buy × Substitution + ei,t , (4.2)

where R denotes returns, Buy is a dummy variable that takes a value of one for

purchases and zero for sales, and Substitution is a dummy variable that takes a

value of one when a trade is classified as part of a substitution trade, and zero

otherwise. 56% of all transactions are purchases.

The interpretation of the regression coefficients is made particularly easy by

the fact that all regressors are dummy variables. As such, the coefficients corre-

spond to changes in the mean value of the dependent variable for different con-

stellations of the dummy variables. The average post-sale return for a trade that

is not classified as a substitution trade (a “regular” trade) is given by α. The β 1

coefficient corresponds to the difference between a regular purchase and a regular

sale. The average post-sale return for a substitution trade is α + β 2 , whereas the

post-purchase return to a substitution trade is α + β 1 + β 2 + β 3 . Note as well that

this difference-in-difference specification implies that the coefficient on the inter-

action term corresponds to the difference in the difference between post-sale and

126
post-purchase returns for regular trades and substitution trades:

β 3 = [ E( Ri,t | Buy, Substitution) − E( Ri,t | Sell, Substitution)]

− [ E( Ri,t | Buy, Regular) − E( Ri,t | Sell, Regular)] .

A test of β 3 = 0 is therefore a test of Hypothesis 1, namely that substitution trades

perform better than regular trades.

U.S. data

Panel A of Table 4.5 shows the results of four separate regressions as in (4.2)

using post-transaction return horizons of 1, 3, 6, and 12 months, respectively,

with the U.S. data. Importantly, standard errors are calculated using the robust

“sandwich” estimator with clustering at the stock or fund level. This approach

allows the return of a stock to be correlated across individuals, which makes

sense because many individuals may hold the same stock at the same time. Using

clustered standard errors drastically cuts the effective degrees of freedom in the

model, and may be overly conservative, since it seems reasonable to argue that

each individual investor’s decision to transact is the appropriate unit of observa-

tion in this setting.

The coefficient on the Buy dummy is reliably negative, which again confirms

that post-purchase returns are lower than post-sale returns. At a one year hori-

zon, a stock increases on average 16.32% following a sale, and 14.3% following a

127
purchase (2.02% less). For all horizons, these differences are highly statistically

significant.

As noted above, the most relevant estimate is the coefficient on the interaction

term, Buy × Substitution. Even with the clustered standard errors, this coefficient

is positive and highly statistically significant at each horizon, indicating that the

performance of purchases relative to sales is significantly improved for substitu-

tion trades compared to regular trades. Moreover, at each horizon the interaction

coefficient is larger than the coefficient on the Buy dummy, so the worse perfor-

mance of purchases than sales for regular trades is actually reversed for substitu-

tion trades. This can be seen in the row labeled β 1 + β 3 : for example, at the one

year horizon, while a regular buy earns on average 2.02% less than a regular sell,

a substitution buy earns 39 basis points more than a substitution sell. The p-values

for the F-statistics associated with these coefficient sums are reported in square

brackets. Except for at the 3-month horizon, these results are not significantly

greater than zero, but a one-sided test against a null hypothesis of β 1 + β 3 < 0 is

strongly rejected at any conventional significance level.

It is interesting to note as well where the improvement of post-transaction re-

turns is coming from for the substitution trades. The Substitution coefficient is

negative, so substitution sales earn lower returns than regular sales; this is consis-

tent with the notion that substitution trades are made by investors with informa-

tion about which stocks to sell. The difference between a Substitution Buy and a

Regular Buy is given in the row labeled β 2 + β 3 . Here the statistical evidence is

much weaker, although it appears that the results are somewhat stronger at the

128
longer horizon. At the 1- and 3-month horizon substitution purchases earn no

more than 9 basis points less than regular purchases, and at the 6- and 12-month

horizons substitution buys earn 26 and 66 basis points more than regular buys.

In other words, the improvement in post-transaction returns for substitution

trades is mainly due to sales, especially at shorter horizons. Investors placing

substitution trades appear to be significantly better than investors placing regular

trades at identifying which stocks to sell, but only somewhat better at choosing

which stocks to purchase. One potential explanation for this asymmetry is that

identifying a stock to sell from the relatively small number stocks currently held

is easier than identifying a stock to purchase from the much larger set of stocks

that the investor does not currently hold. Another possibility is that short sale

constraints prevent negative information from being quickly impounded in stock

prices, so individual investors can receive negative information about a stock they

currently hold that isn’t yet reflected in prices. Positive information, on the other

hand, might be impounded in prices more quickly, giving individual investors

less time to take advantage of information they may receive.

To ensure that the results presented so far are not driven by especially illiquid

stocks, Panel B of Table 4.5 presents estimates from the same regressions, but the

sample is restricted to just those stocks whose price is above $5. The results here

are qualitatively the same as those in Panel A.

Focusing on stocks with prices above $5 also allows us to investigate a lit-

tle more the possibility that short sale constraints make it easier for individual

investors to identify stocks to sell than stocks to buy. Higher-priced stocks are

129
somewhat less likely to have binding short sale constraints, although perhaps only

marginally. Nevertheless, comparing Panel B to Panel A, post-purchase returns

for Substitution buys are higher in the more liquid stocks of Panel B, and post-

sale returns are lower for the more liquid stocks. For example, at the 12-month

horizon, the difference between Regular sales and Substitution sales is 175 basis

points using the entire sample, but 130 basis points for the higher-price sample;

and the difference between substitution buys and regular buys is 66 basis points

for the entire sample, but 82 basis points for the restricted sample. This is consis-

tent with the short sale constraint story: for the stocks in Panel B, which are less

likely to have binding short-sale constraints, investors are not as good at choos-

ing which stocks to sell but better at choosing which stocks to buy than the entire

sample considered in Panel A.

Finland data

Turning to the Finnish dataset, Table 4.6 presents results from the same type of

post-transaction return regressions as those just presented. The results in Finland

are remarkably consistent with those in the U.S.

In particular, there is a strong negative effect of the Buy dummy, so for non-

substitution trades, purchases are followed by lower returns than sales. Again,

the effect is reversed for Substitution trades: the interaction coefficient is highly

significant and economically important. For example, at a one-year horizon, pur-

chases are followed by 13.78% lower returns than sales for Regular trades, but

Substitution purchases are followed by 31 basis points higher returns than Sub-

130
stitution sales, as shown in the row labeled β 1 + β 3 . The difference for regular

trades is strongly statistically significant, and statistically insignificant for substi-

tution trades. And as in the U.S. data, the improvement in returns for substitution

trades in Finland comes primarily from the sell side: Substitution purchases are

followed by somewhat higher returns than regular purchases, although the dif-

ference is insignificant, but substitution sales are followed by much lower returns

than regular sales, as can be seen by examining the substitution dummy coeffi-

cient and the sum of β 2 + β 3 .

The results presented in this section provide strong evidence for the notion that

Substitution trades are motivated by information much more than other trades.

The evidence is remarkably consistent, especially considering the data are from

substantially different periods (1991–1996 in the U.S. and 1995-2003 in Finland),

and the substantially different structure of the two markets (approximately 2 mil-

lion observations and 11,000 stocks in the U.S. and 6 million observations and 149

stocks in Finland).

4.3.4 Returns and Aggregate Buy Ratios

If substitution trades are more likely to be informed, we would expect them

to display different time series properties than other trades. For example, if these

trades contain private information we would expect them to have a permanent

price impact. We would also expect these trades not to be related to past returns

in a way that uninformed trades may be (if, for example, investors without infor-

mation respond to price momentum). I investigate these questions in this section.

131
 I begin by calculating an aggregate Buy Ratio for substitution trades and non-

substitution trades. For each stock and each day (or week), I count the number of

substitution buys and substitution sells. The substitution Buy Ratio is the number

of substitution buys divided by the total number of substitution sells. Similarly,

the buy ratio for regular trades is calculated as the total number of regular pur-

chases divided by the total number of regular trades. I then estimate the regres-

sion

Ri,t+h = α + β · BRi,t + ei,t , for h = −4, . . . , 4, (4.3)

where BRi,t denotes the Buy Ratio for stock i at time t. The substitution Buy Ratio

or the regular Buy Ratio are used in separate regressions.

Estimated regression coefficients for β are reported in Table 4.7. Each cell

presents the results of a separate regression. Results using a one-day aggregation

window are reported in Panel A, and those using one week are reported in Panel

B. Consider first the estimates in the column labeled “0” in Panel A. There is no

significant relation between the regular buy ratio and contemporaneous returns.

In contrast, there is a strong relation between the substitution buy ratio and con-

temporaneous returns. That is, the higher is the substitution buy ratio, the more

we expect prices to increase; no such relation exists for the regular buy ratio. This

is again consistent with the notion that substitution trades are more likely to be

informed; informed purchases push up prices, while uninformed purchases have

no effect.

132
 Further evidence in support of the claim that substitution trades are more in-

formed than regular trades is provided by the next four columns of the table. In

the days following a high regular buy ratio, returns are significantly negative,

suggesting price movements on “day 0” are quickly reversed. No such reversals

are apparent in the days following a high substitution buy ratio, suggesting that

price impact on day 0 is permanent. Looking at the four days preceding the buy

ratio calculation, there is no evidence that the level of substitution buying or sell-

ing is related to previous returns. There is, however, some evidence that regular

purchases are more likely to follow negative return days. This is borne out in the

weekly data, which we turn to next.

Panel B shows coefficient estimates for the same types of regressions, although

the buy ratios are now calculated using all regular or substitution trades within

a week (Monday to Friday). The significant differences between the return re-

sponse of stocks to the two types of buy ratios shown in Panel A are again found

in Panel B. At the weekly horizon, there is still a highly significant price response

to the contemporaneous substitution buy ratio, but also a highly significant price

response in the opposite direction to the regular buy ratio (column “0”). In other

words, weeks in which substitution trades are predominantly purchases are weeks

with high returns, and weeks in which regular trades are dominated by purchases

are weeks with low returns. Moreover, when returns have been low in the past

four weeks, the regular buy ratio is higher, indicating contrarian trading behav-

ior among regular trades. Substitution trading, on the other hand, is momentum

trading in the sense that it is higher when the previous week’s return is higher.

133
And as in the daily results, there is evidence of a price reversal following regular

trading, but no such reversal following substitution trades.

The pattern of results in Table 4.7 provides strong evidence that substitution

trades are fundamentally different from other trades. In particular, the results are

consistent with the substitution trades being informed and regular trades being

uninformed. We would expect informed trading to cause permanent price move-

ments, and uninformed trading to be followed by price reversals. This is precisely

what we see for the substitution and regular trades. We might also expect un-

informed trading to respond to previous price movements, such as if investors

purchase shares after a price decline because these are interpreted as a “buying

opportunity.” We would not expect strong effects of this sort for informed trad-

ing, and again, this is what we find for the regular and substitution trades, respec-

tively.

4.4 Conclusion

This chapter presents a simple method for classifying the trades of individ-

ual investors “informed” or “uninformed.” Starting with the economic argument

that liquidity trading is likely to show up as purchases or sales of stocks and not

moving money from one stock to another, I claim that trades are more likely to be

informed when they involve selling shares of one stock and using the proceeds to

buy a similar dollar value of another stock in a short period of time.

134
 I implement this classification procedure in two large datasets from the U.S.

and Finland, covering 1991–1996 and 1995–2003, respectively. The datasets are

quite different; for example, investors in the U.S. sample trade upwards of 11,000

different securities, while investors in the Finland data trade 149 securities. De-

spite these differences, a remarkably consistent result emerges: substitution trades

do appear to be informed. The post-transaction returns for substitution trades are

significantly better than the post-transaction returns for non-substitution trades.

Moreover, when substitution trades are dominated by purchases, returns are sig-

nificantly positive, indicating a price impact. This price impact is not subse-

quently reversed for substitution trades, whereas there is a significant reversal

for non-substitution trades.

While the results in this chapter provide substantial evidence that some in-

dividual investor trading is informed, they nevertheless represent a first-pass at

identifying informed trading among individual investors. Surely, more sophisti-

cated classification methods using more of the available data are possible. Such

classifications would give allow a better understanding of the interaction between

informed and uninformed investors in financial markets, and have the potential

to greatly improve upon existing measures such as Easley, Kiefer, O’Hara, and

Paperman’s (1996) “PIN” measure. These and other issues are the focus of contin-

uing research.

135
 Table 4.1: Post-Transaction Returns in U.S. Data—Summary Statistics
This table presents summary statistics for returns subsequent to purchases and sales. The unit of
observation is a transaction, and N denotes the total number of observations. Returns are calcu-
lated as the increase in a cumulative return index (including all payouts) beginning the day after
a transaction, and over the horizon specified in each column. 5 days corresponds to one week, 21
days to one month, and 252, 504, and 756 trading days correspond to one, two, and three years,
respectively.

Horizon (in trading days)

5 21 63 84 126 252 504 756

Purchases
Mean 0.0030 0.0129 0.0334 0.0440 0.0646 0.1516 0.3650 0.5834
Std. Dev. 0.0812 0.1454 0.2490 0.2852 0.3540 0.5432 0.9445 1.2487
25th Pctl. -0.0295 -0.0591 -0.1012 -0.1138 -0.1350 -0.1592 -0.1610 -0.1142
50th Pctl. 0.0000 0.0072 0.0214 0.0256 0.0356 0.0863 0.1951 0.3132
75th Pctl. 0.0315 0.0762 0.1434 0.1709 0.2162 0.3464 0.6267 0.9131
N 821,835 810,397 783,176 770,705 734,680 628,436 467,639 344,860

Sales
Mean 0.0036 0.0156 0.0388 0.0496 0.0736 0.1758 0.3954 0.6327
Std. Dev. 0.0675 0.1410 0.2392 0.2757 0.3469 0.5373 0.9081 1.2798
25th Pctl. -0.0268 -0.0534 -0.0917 -0.1037 -0.1175 -0.1259 -0.1182 -0.0689
50th Pctl. 0.0000 0.0091 0.0256 0.0308 0.0437 0.1099 0.2269 0.3543
75th Pctl. 0.0298 0.0750 0.1456 0.1724 0.2180 0.3691 0.6715 0.9574
N 678,394 666,379 637,962 625,747 596,415 507,254 371,564 271,042

136
 Table 4.2: Post-Transaction Returns in U.S. Data—Regression Tests
This table presents regression results for the fixed-effects regression

ri,j = αi + δm · 1{t∈m} + β · 1{ j=Buy} + ei,j ,

where αi is the individual-specific intercept, and 1{·} is an indicator function. t denotes the date
of the transaction, and m ∈ {1, . . . , 70} is a month index number. Post-purchase returns are lower
than post-sale returns if β < 0. The number of observations used in the regression is denoted N
Obs.

Horizon β t-statistic N Obs.

5 0.0005 2.58 1,500,225
21 0.0006 1.99 1,476,772
63 -0.0024 -4.46 1,421,136
84 -0.0032 -5.03 1,396,450
126 -0.0059 -7.21 1,331,093
252 -0.0145 -10.28 1,135,688
504 -0.0108 -3.72 839,202
756 -0.0239 -5.24 615,901

137
 Table 4.3: Post-Transaction Returns in U.S. Data—Cross-sectional Means
This table presents the cross-sectional means of an individual’s average post-purchase return minus
that individual’s average post-sale return. Horizons are measured in trading days, with 5 days
corresponding to one week, and 756 days to approximately three years. Num. Accts is the number
of unique accounts used in the cross-sectional calculation. In Panel A, a simple average is calculated
across all accounts. Panel B restricts the analysis only to those accounts that place at least 10 trades.
Except where otherwise noted, all means are statistically different from zero at the 1% level.

Horizon Num. Accts Mean Std. Dev. t-statistic

Panel A: All Accounts

5 67,862 -0.0007 0.0599 -3.25
21 67,229 -0.0039 0.1161 -8.77
63 65,813 -0.0087 0.1945 -11.54
84 65,243 -0.0091 0.2180 -10.70
126 63,841 -0.0134 0.2682 -12.66
252 58,979 -0.0340 0.3918 -21.10
504 50,525 -0.0347 0.6669 -11.68
756 42,629 -0.0284 0.9752 -6.00

Panel B: Only Accounts With More Than 10 Trades

5 15,356 -0.0004 0.0242 -2.15a
21 15,352 -0.0012 0.0493 -2.99
63 15,333 -0.0035 0.0816 -5.26
84 15,321 -0.0042 0.0937 -5.49
126 15,277 -0.0061 0.1192 -6.37
252 15,101 -0.0179 0.1882 -11.67
504 14,598 -0.0093 0.3928 -2.85
756 13,900 -0.0185 0.6006 -3.63
a Significant at the 5% level.

138
 Table 4.4: Trade Classifications

This table shows the percentage of trades classified as substitution trades. ‘Days’ is the window of
days within which transactions must take place to be classified as substitution trades, and ‘Maxi-
mum Buy/Sell Difference’ is the largest allowed difference between the value of trades sold and
the value of trades purchased. For example, 5.1% of transactions are classified as substitution
trades when we require all transactions in a substitution trade to be within a two-day period and
also that the total value of shares purchased and shares sold be within 2% of each other.

Maximum Buy/Sell Difference

Days 1% 2% 5% 10% 25%
1 2.9 5.0 9.0 13.2 19.9
2 3.0 5.1 9.5 14.1 21.6
3 3.1 5.4 10.1 14.9 23.1
5 3.2 5.6 10.7 15.9 25.1
10 3.6 6.3 12.1 18.5 30.1
15 3.8 6.8 13.1 20.2 33.1

139
 Table 4.5: Post-Transaction Returns in U.S. Data
This table presents the results of the regression

Ri,t = α + β 1 · Buy + β 2 · Substitution + β 3 · Buy × Substitution + ei,t ,

where R denotes returns, Buy is a dummy variable that takes a value of one for purchases and zero for sales, and Substitution is a dummy variable
that takes a value of one when a trade is classified as part of a substitution trade, and zero otherwise. 56% of transactions are purchases. A trade is
classified as a substitution trade if all transactions occur within one trading day of each other, and the total value of shares purchased is within 2%
of the total value of shares sold. Using this classification, 4.7% of transactions are classified as being part of a substitution trade. Stock returns are
calculated as the increase in a cumulative return index (including all payouts) beginning the day after a transaction, and over the horizon specified
in each column. Mutual fund returns are calculated from monthly returns reported in CRSP beginning in the month after the transaction. Standard
errors, which are robust to correlation within stocks and mutual funds, are in parentheses. The number of clusters used in the robust standard
error calculation is reported below the number of observations. β 1 + β 3 is the difference in the post-transaction return for a Substitution Buy and
a Substitution Sell. β 2 + β 3 is the difference in return for a Substitution Buy and a Regular Buy. p-values for these F-tests are in square brackets.
Statistical significance at the 10%, 5% and 1% levels is denoted by ∗ , ∗∗ , and ∗∗∗ , respectively.

140
Post-Transaction Return Horizon
1 month 3 months 6 months 1 year

Panel A: No price restriction

Constant 0.0140∗∗∗ 0.0358∗∗∗ 0.0672∗∗∗ 0.1602∗∗∗
(0.0006) (0.0019) (0.0039) (0.0091)
Buy -0.0022∗∗∗ -0.0046∗∗∗ -0.0083∗∗∗ -0.0200∗∗∗
(0.0005) (0.0012) (0.0021) (0.0040)
Substitution -0.0038∗∗∗ -0.0073∗∗∗ -0.0062∗∗∗ -0.0136∗∗∗
(0.0007) (0.0013) (0.0021) (0.0038)
Buy×Substitution 0.0039∗∗∗ 0.0069∗∗∗ 0.0114∗∗∗ 0.0230∗∗∗
(0.0012) (0.0018) (0.0027) (0.0055)
β2 + β3 0.0002 -0.0004 0.0051∗∗ 0.0094∗∗
[0.8609] [0.7850] [0.0167] [0.0182]
β1 + β3 0.0018 0.0023 0.0031 0.0030
[0.1557] [0.2572] [0.3305] [0.6440]
Number of observations 2,386,439 2,376,519 2,358,667 2,317,528
Number of stocks/funds 11,578 11,506 11,389 11,061
Mean of dependent variable 0.0127 0.0330 0.0624 0.1487
 Table 4.5, continued
Post-Transaction Return Horizon
1 month 3 months 6 months 1 year

Panel B: Price > 5

Constant 0.0131∗∗∗ 0.0349∗∗∗ 0.0664∗∗∗ 0.1618∗∗∗
(0.0006) (0.0020) (0.0041) (0.0096)
Buy -0.0019∗∗∗ -0.0039∗∗∗ -0.0073∗∗∗ -0.0215∗∗∗
(0.0005) (0.0012) (0.0022) (0.0041)
Substitution -0.0031∗∗∗ -0.0062∗∗∗ -0.0048∗∗ -0.0111∗∗∗
(0.0007) (0.0013) (0.0021) (0.0037)
Buy×Substitution 0.0035∗∗∗ 0.0064∗∗∗ 0.0091∗∗∗ 0.0203∗∗∗
(0.0011) (0.0017) (0.0026) (0.0052)
β2 + β3 0.0004 0.0003 0.0043∗∗ 0.0092∗∗
[0.6758] [0.8443] [0.0390] [0.0176]
β1 + β3 0.0016 0.0026 0.0018 -0.0012
[0.1639] [0.1779] [0.5438] [0.8359]
Number of observations 2,169,367 2,163,009 2,150,619 2,120,179
Number of stocks/funds 10,104 10,069 10,015 9,839

141
Mean of dependent variable 0.0119 0.0325 0.0623 0.1494

Panel C: Risk-adjusted returns

Constant 0.0157∗∗∗ 0.0364∗∗∗ 0.0695∗∗∗ 0.1626∗∗∗
(0.0007) (0.0021) (0.0043) (0.0095)
Buy -0.0030∗∗∗ -0.0059∗∗∗ -0.0102∗∗∗ -0.0245∗∗∗
(0.0006) (0.0014) (0.0024) (0.0048)
Substitution -0.0050∗∗∗ -0.0091∗∗∗ -0.0118∗∗ -0.0208∗∗∗
(0.0011) (0.0019) (0.0030) (0.0050)
Buy×Substitution 0.0024∗∗∗ 0.0073∗∗∗ 0.0086∗∗∗ 0.0256∗∗∗
(0.0015) (0.0024) (0.0037) (0.0095)
β2 + β3 -0.0026∗ -0.0017 -0.0032 0.0048
[0.0647] [0.4200] [0.2963] [0.5342]
β1 + β3 -0.0006 0.0014 -0.0016 0.0011
[0.7147] [0.6074] [0.6905] [0.9155]
Number of observations 2,169,367 2,163,009 2,150,619 2,120,179
Number of stocks/funds 10,219 10,104 9,857 8,592
Mean of dependent variable 0.0140 0.0330 0.0637 0.1488
 Table 4.6: Post-Transaction Returns in Finland Data
This table presents the results of the regression

Ri,t = α + β 1 · Buy + β 2 · Substitution + β 3 · Buy × Substitution + ei,t ,

where R denotes returns, Buy is a dummy variable that takes a value of one for purchases and zero for sales, and Substitution is a dummy variable
that takes a value of one when a trade is classified as part of a substitution trade, and zero otherwise. 50% of transactions are purchases. A trade is
classified as a substitution trade if all transactions occur within one trading day of each other, and the total value of shares purchased is within 2%
of the total value of shares sold. Using this classification, 3.2% of transactions are classified as being part of a substitution trade. Stock returns are
calculated as the increase in a cumulative return index (including all payouts) beginning the day after a transaction, and over the horizon specified
in each column. Standard errors, which are robust to correlation within stocks and mutual funds, are in parentheses. The number of clusters used
in the robust standard error calculation is reported below the number of observations. β 1 + β 3 is the difference in the post-transaction return for a
Substitution Buy and a Substitution Sell. β 2 + β 3 is the difference in return for a Substitution Buy and a Regular Buy. p-values for these F-tests are
in square brackets. Statistical significance at the 10%, 5% and 1% levels is denoted by ∗ , ∗∗ , and ∗∗∗ , respectively.

142
Post-Transaction Return Horizon
1 month 3 months 6 months 1 year
Constant 0.0345∗∗∗ 0.1338∗∗∗ 0.1313∗∗∗ 0.0938∗∗∗
(0.0067) (0.0212) (0.0147) (0.0337)
Buy -0.0193∗∗∗ -0.0787∗∗∗ -0.1114∗∗∗ -0.1378∗∗∗
(0.0080) (0.0225) (0.0187) (0.0280)
Substitution -0.0203∗∗∗ -0.0744∗∗∗ -0.1113∗∗ -0.1495∗∗∗
(0.0038) (0.0143) (0.0160) (0.0148)
Buy×Substitution 0.0217∗∗∗ 0.0845∗∗∗ 0.1180∗∗∗ 0.1409∗∗∗
(0.0059) (0.0179) (0.0183) (0.0204)
β2 + β3 0.0014 0.0101 0.0067 -0.0086
[0.0647] [0.4200] [0.2963] [0.5342]
β1 + β3 0.0024 0.0058 0.0066 0.0031
[0.7147] [0.6074] [0.6905] [0.9155]
Number of observations 6,242,331 6,242,331 6,070,921 5,633,849
Number of stocks/funds 149 149 149 149
Mean of dependent variable 0.0248 0.0928 0.0732 0.0214
 Table 4.7: Returns and Aggregate Buy Ratios
This table presents the results of a regression of stock returns in a given period (days or weeks in Panels A and B, respectively) on either the Buy
Ratio formed from aggregate substitution trading, or the Buy Ratio formed from aggregate non-substitution (regular) trading:

Ri,t+h = α + β · BRi,t + ei,t , for h = −4, . . . , 4,

where BR denotes the Buy Ratio, defined as the ratio of the number of purchases to total number of trades. Separate regressions are run for each time
horizon and the Buy Ratio constructed either from regular trades, or from substitution trades, is used, so each cell displays the results of a separate
regression. Statistical significance at the 10%, 5% and 1% levels is denoted by ∗ , ∗∗ , and ∗∗∗ , respectively.

Return Period Relative to Buy Ratio

-4 -3 -2 -1 0 1 2 3 4

Panel A: Days

143
Regular Buy Ratio -0.0001 -0.0002 -0.0001 -0.0001 0.0001 -0.0003 -0.0002 -0.0003 -0.0002
(0.0001) (0.0001)∗∗ (0.0001) (0.0001) (0.0001) (0.0001)∗∗∗ (0.0001)∗∗ (0.0001)∗∗∗ (0.0001)∗∗
Number of Observations 771,153 769,743 769,077 770,579 770,684 770,069 769,491 770,163 771,516

Informed Buy Ratio 0.0002 0.0002 0.0003 0.0000 0.0005 0.0002 0.0003 0.0003 0.0003
(0.0002) (0.0002) (0.0002) (0.0002) (0.0002)∗∗∗ (0.0002) (0.0002)∗ (0.0002) (0.0002)∗
Number of Observations 177,512 177,167 176,984 177,628 177,311 177,042 177,193 177,170 177,546

Panel B: Weeks

Regular Buy Ratio -0.0045 -0.0049 -0.0061 -0.0077 -0.0078 0.0006 0.0001 0.0003 -0.0000
(0.0003)∗∗∗ (0.0003)∗∗∗ (0.0003)∗∗∗ (0.0004)∗∗∗ (0.0004)∗∗∗ (0.0003)∗ (0.0003) (0.0003) (0.0003)
Number of Observations 455,671 456,528 457,383 458,322 459,444 459,698 459,089 458,502 457,867

Informed Buy Ratio 0.0001 -0.0010 -0.0011 0.0024 0.0111 0.0004 -0.0007 0.0009 -0.0000
(0.0006) (0.0006)∗ (0.0007) (0.0008)∗∗∗ (0.0010)∗∗∗ (0.0005) (0.0005) (0.0005)∗ (0.0005)
Number of Observations 98,001 98,161 98,310 98,469 98,745 98,838 98,751 98,665 98,581
 Table 4.8: Returns and Aggregate Buy Ratios
This table presents the results of a regression of weekly stock returns on either the Buy Ratio formed from aggregate substitution trading, or the Buy
Ratio formed from aggregate non-substitution (regular) trading:

Ri (t + h) = α + β 1 · BR R (t) + β 2 · BR I (t) + β 3 Ri (t + h − 1) + ei,t , for h = −3, . . . , 3,

where BR denotes the Buy Ratio, defined as the ratio of the number of purchases to total number of trades. Separate regressions are run for each time
horizon and the Buy Ratio constructed either from regular trades, or from substitution trades, is used, so each cell displays the results of a separate
regression. Statistical significance at the 10%, 5% and 1% levels is denoted by ∗ , ∗∗ , and ∗∗∗ , respectively.

-3 -2 -1 0 1 2 3

144
Regular buy ratio -0.0120∗∗∗ -0.0136∗∗∗ -0.0223∗∗∗ -0.0457∗∗∗ 0.0018 0.0025∗∗∗ 0.0017∗
(0.0014) (0.0017) (0.0023) (0.0035) (0.0013) (0.0008) (0.0009)
Substitution buy ratio 0.0012∗∗ 0.0011∗∗ 0.0036∗∗∗ 0.0039∗∗∗ 0.0019∗∗∗ 0.0008 0.0012∗∗
(0.0005) (0.0005) (0.0006) (0.0008) (0.0005) (0.0005) (0.0006)
Constant 0.0068∗∗∗ 0.0083∗∗∗ 0.0116∗∗∗ 0.0240∗∗∗ -0.0010 -0.0011∗∗∗ -0.0012∗∗
(0.0007) (0.0009) (0.0012) (0.0020) (0.0007) (0.0003) (0.0005)
Lag return 0.0508∗∗∗ 0.0734∗∗∗ 0.0898∗∗∗ 0.0642∗∗∗ 0.0662∗∗∗ 0.0209 0.0126
(0.0113) (0.0201) (0.0228) (0.0202) (0.0240) (0.0284) (0.0109)
Number of observations 36,278 36,376 36,743 38,185 38,032 36,663 36,357
Bibliography

145
 Bibliography

Abreu, D., and M. K. Brunnermeier, 2003, “Bubbles and Crashes,” Econometrica,

71, 173–204.

Alexander, G. J., G. Cici, and S. Gibson, 2007, “Does Motivation Matter When As-
sessing Trade Performance? An Analysis of Mutual Funds,” Review of Financial
Studies, 20, 125–150.

Arrow, K., 1962, “The Ecomomic Implications of Learning By Doing,” Review of

Economic Studies, 29, 155–173.

Avramov, D., T. Chordia, and A. Goyal, 2006, “Liquidity and Autocorrelations in

Individual Stock Returns,” Journal of Finance, 61, 2365–2394.

Barber, B. M., and T. Odean, 2000, “Trading is Hazardous to Your Wealth: The
Common Stock Investment Performance of Individual Investors,” Journal of Fi-
nance, 55, 773–806.

Barber, B. M., and T. Odean, 2001, “Boys Will Be Boys: Gender, Overconfidence,
and Common Stock Investment,” Quarterly Journal of Economics, 116, 261–292.

Barber, B. M., and T. Odean, 2006, “All That Glitters: The Effect of Attention and
News on the Buying Behavior of Individual and Institutional Investors,” Forth-
coming, Review of Financial Studies.

Barber, B. M., T. Odean, and M. Strahilevitz, 2004, “Once Burned, Twice Shy:
Naı̈ve Learning, Counterfactuals and the Repurchase of Stocks Previously
Sold,” Working paper, UC Berkeley.

Barber, B. M., T. Odean, and N. Zhu, 2006, “Do Noise Traders Move Markets?,”
Working paper, UC Davis.

Barberis, N., and R. Thaler, 2005, “A Survey of Behavioral Finance,” in Richard H.

Thaler (ed.), Advances in Behavioral Finance, vol. 2, chap. 1, pp. 1–75, Princeton
University Press.

146
Benartzi, S., 2001, “Excessive Extrapolation and the Allocation of 401(k) Accounts
to Company Stock,” Journal of Finance, 56, 1747–1764.

Black, F., 1986, “Noise,” Journal of Finance, 41, 529–543.

Bolton, P., and C. Harris, 1999, “Strategic Experimentation,” Econometrica, 67, 349–
374.

Cai, F., and L. Zheng, 2004, “Institutional Trading and Stock Returns,” Finance
Research Letters, 1, 178–189.

Campbell, J. Y., 2006, “Household Finance,” Journal of Finance, 61, 1553–1604.

Campbell, J. Y., S. J. Grossman, and J. Wang, 1993, “Trading Volume and Serial
Correlation in Stock Returns,” Quarterly Journal of Economics, 108, 905–939.

Campbell, J. Y., T. Ramadorai, and A. Schwartz, 2008, “Caught on Tape: Institu-

tional Trading, Stock Returns, and Earnings Announcements,” Working paper,
Harvard University.

Carhart, M., 1997, “On Persistence in Mutual Fund Performance,” Journal of Fi-
nance, 52, 57–82.

Chancellor, E., 2000, Devil Take the Hindmost: A History of Financial Speculation,
Plume, New York, N.Y.

Choi, J., D. Laibson, B. Madrian, and A. Metrick, 2007, “Reinforcement Learning

and Investor Behavior,” Working paper, Yale University.

Chordia, T., and A. Subrahmanyam, 2004, “Order Imbalance and Individual Stock
Returns: Theory and Evidence,” Journal of Financial Economics, 72, 485–518.

Coval, J. D., D. A. Hirshleifer, and T. Shumway, 2005, “Can Individual Investors

Beat the Market?,” Working paper, University of Michigan.

Coval, J. D., and T. Shumway, 2005, “Do Behavioral Biases Affect Prices?,” Journal
of Finance, 60, 1–34.

Cox, D. R., 1972, “Regression Models and Life Tables,” Journal of the Royal Statisti-
cal Society, Series B, 34, 187–220.

Cox, D. R., and D. Oakes, 1984, Analysis of Survival Data, Chapman and Hall, New
York.

147
Daniel, K., M. Grinblatt, S. Titman, and R. Wermers, 1997, “Measuring Mutual
Fund Performance with Characteristic Based Benchmarks,” Journal of Finance,
52, 1035–1058.

De Long, J. B., A. Shleifer, L. H. Summers, and R. J. Waldmann, 1991, “The Sur-

vival of Noise Traders in Financial Markets,” Journal of Business, 64, 1–19.

Easley, D., R. F. Engle, M. O’Hara, and L. Wu, 2002, “Time-Varying Arrival Rates
of Informed and Uninformed Trades,” Working paper, Cornell University.

Easley, D., N. M. Kiefer, M. O’Hara, and J. B. Paperman, 1996, “Liquidity, Infor-

mation, and Infrequently Traded Stocks,” Journal of Finance, 51, 1405–1436.

Easley, D., and M. O’Hara, 1987, “Price, Trade Size, and Information in Securities
Markets,” Journal of Financial Economics, 19, 69–90.

Fama, E., and K. French, 1993, “Common Risk Factors in the Returns on Stocks
and Bonds,” Journal of Financial Economics, 33, 3–56.

Fama, E., and J. MacBeth, 1973, “Risk, Return, and Equilibrium: Empirical Tests,”
Journal of Political Economy, 81, 607–636.

Feng, L., and M. Seasholes, 2005, “Do Investor Sophistication and Trading Expe-
rience Eliminate Behavioral Biases in Financial Markets?,” Review of Finance, 9,
305–351.

Frazzini, A., 2006, “The Disposition Effect and Underreaction to News,” Journal of
Finance, 61, 2017–2046.

Friedman, M., 1953, “The Case for Flexible Exchange Rates,” in Essays in Positive
Economics, pp. 157–203, University of Chicago Press.

Genesove, D., and C. Mayer, 2001, “Loss-Aversion and Seller Behavior: Evidence
from the Housing Market,” Quarterly Journal of Economics, 116, 1233–1260.

Glosten, L. R., and P. R. Milgrom, 1985, “Bid, Ask and Transaction Prices in a
Specialist Market with Heterogeneously Informed Traders,” Journal of Financial
Economics, 14, 71–100.

Goetzmann, W. N., and A. Kumar, 2008, “Equity Portfolio Diversification,” Forth-

coming, Review of Finance.

Gompers, P. A., and A. Metrick, 2001, “Institutional Investors and Equity Prices,”
Quarterly Journal of Economics, 116, 229–259.

148
Greene, W., 2005, Econometric Analysis, Prentice Hall, 5th edn.

Greenwood, R., and S. Nagel, 2007, “Inexperienced Investors and Bubbles,” Work-
ing paper, Stanford University.

Griffin, J. M., J. H. Harris, and S. Topaloglu, 2003, “The Dynamics of Institutional

and Individual Trading,” Journal of Finance, 58, 2285–2320.

Grinblatt, M., and B. Han, 2005, “Prospect Theory, Mental Accounting, and Mo-
mentum,” Journal of Financial Economics, 78, 311–339.

Grinblatt, M., and M. Keloharju, 2000, “The Investment Behavior and Perfor-
mance of Various Investor Types: A Study of Finland’s Unique Data Set,” Jour-
nal of Financial Economics, 55, 43–67.

Grinblatt, M., and M. Keloharju, 2001a, “How Distance, Language, and Culture
Influence Stockholdings and Trades,” Journal of Finance, 56, 1053–1073.

Grinblatt, M., and M. Keloharju, 2001b, “What Makes Investors Trade?,” Journal of
Finance, 56, 589–616.

Grossman, S., R. Kihlstrom, and L. Mirman, 1977, “A Bayesian Approach to the

Production of Information and Learning by Doing,” Review of Economic Studies,
44, 533–547.

Hamilton, J. D., 1994, Time Series Analysis, Princeton University Press, Princeton,
NJ.

Hasbrouck, J., 1991, “Measuring the Information Content of Stock Trades,” Journal
of Finance, 46, 179–207.

Heckman, J., 1976, “The Common Structure of Statistical Models of Truncation,

Sample Selection, and Limited Dependent Variables and a Simple Estimator
for Such Models,” Annals of Economic and Social Measurement, 5, 475–492.

Hvidkjaer, S., 2006, “Small Trades and the Cross-Section of Stock Returns,” Work-
ing paper, University of Maryland.

Ivković, Z., C. Sialm, and S. J. Weisbenner, 2006, “Portfolio Concentration and

the Performance of Individual Investors,” Forthcoming, Journal of Financial and
Quantitative Analysis.

149
Ivković, Z., and S. J. Weisbenner, 2005, “Local Does as Local Is: Information Con-
tent of the Geography of Individual Investors’ Common Stock Investments,”
Journal of Finance, 60, 267–306.

Kaniel, R., G. Saar, and S. Titman, 2008, “Individual Investor Trading and Stock
Returns,” Journal of Finance, 63.

Keynes, J. M., 1936, The General Theory of Employment, Interest and Money, Macmil-
lan.

Kyle, A. S., 1985, “Continuous Auctions and Insider Trading,” Econometrica, 53,
1315–1335.

Lee, C., and M. Ready, 1991, “Inferring Trade Direction from Intraday Data,” Jour-
nal of Finance, 46, 733–746.

Linnainmaa, J., 2006, “Learning from Experience,” Working paper, University of

Chicago.

Linnainmaa, J., 2007, “The Limit Order Effect,” Working paper, University of
Chicago.

List, J. A., 2003, “Does Market Experience Eliminate Market Anomalies?,” Quar-
terly Journal of Economics, 118, 41–71.

Llorente, G., R. Michaely, G. Saar, and J. Wang, 2002, “Dynamic Volume-Return

Relation of Individual Stocks,” Review of Financial Studies, 15, 1005–1047.

Mahani, R., and D. Bernhardt, 2007, “Financial Speculators’ Underperformance:

Learning, Self-Selection, and Endogenous Liquidity,” Journal of Finance, 62,
1313–1340.

Newey, W. K., and K. D. West, 1987, “A Simple, Positive Semi-definite, Het-

eroskedasticity and Autocorrelation Consistent Covariance Matrix,” Economet-
rica, 55, 703–708.

Nicolosi, G., L. Peng, and N. Zhu, 2004, “Do Individual Investors Learn from
Their Trading Experience?,” Working paper, Yale University.

Nofsinger, J. R., and R. W. Sias, 1999, “Herding and Feedback Trading by Institu-
tional and Individual Investors,” Journal of Finance, 54, 2263–2295.

Obizhaeva, A., 2007, “Information vs. Liquidity: Evidence From Portfolio Transi-
tion Trades,” Working paper, MIT.

150
Odean, T., 1998, “Are Investors Reluctant to Realize Their Losses?,” Journal of Fi-
nance, 53, 1775–1798.

Odean, T., 1999, “Do Investors Trade Too Much?,” American Economic Review, 89,
1279–1298.

Pastor, L., and R. F. Stambaugh, 2003, “Liquidity Risk and Expected Stock Re-
turns,” Journal of Political Economy, 111, 642–685.

Roll, R., 1984, “A Simple Implicit Measure of the Effective Bid-Ask Spread in an
Efficient Market,” Journal of Finance, 39, 1127–1139.

San, G., 2007, “Who Gains More By Trading—Institutions or Individuals?,” Work-

ing paper, Hebrew University of Jerusalem.

Scholes, M., and J. Williams, 1977, “Estimating Betas from Nonsynchronous

Data,” Journal of Financial Economics, 5, 309–327.

Seru, A., T. Shumway, and N. Stoffman, 2007, “Learning By Trading,” Working

paper, University of Michigan.

Shapira, Z., and I. Venezia, 2001, “Patterns of Behavior of Professionally Managed

and Independent Investors,” Journal of Banking and Finance, 25, 1573–1587.

Shefrin, H., and M. Statman, 1985, “The Disposition to Sell Winners Too Early and
Ride Losers Too Long: Theory and Evidence,” Journal of Finance, 40, 777–790.

Shiller, R. J., 2005, Irrational Exuberance (Second Edition), Princeton University Press,
Princeton, N.J.

Shleifer, A., and R. W. Vishny, 1997, “The Limits of Arbitrage,” Journal of Finance,
52, 35–55.

Shumway, T., and G. Wu, 2006, “Does Disposition Drive Momentum?,” Working
paper, University of Michigan.

Sias, R. W., L. T. Starks, and S. Titman, 2006, “Changes in Institutional Ownership

and Stock Returns: Assessment and Methodology,” Journal of Business, 79, 2869–
2910.

Weber, M., and C. Camerer, 1998, “The Disposition Effect in Securities Trading:
An Experimental Analysis,” Journal of Economic Behavior and Organization, 33,
167–184.

151
Wermers, R., 1999, “Mutual Fund Herding and the Impact on Stock Prices,” Jour-
nal of Finance, 54, 581–622.

Wooldridge, J. M., 1995, “Selection Corrections for Panel Data Models Under Con-
ditional Mean Independence Assumptions,” Journal of Econometrics, 68, 115–
132.

Zingales, L., and A. Dyck, 2002, “The Bubble and the Media,” in Peter K. Cor-
nelius, and Bruce Kogut (eds.), Corporate Governance and Capital Flows in a Global
Economy, Chap. 4, pp. 83–104, Oxford University Press, New York.

152