Does it Pay to Pay Attention?
Antonio Gargano∗
University of MelbourneAlberto G. Rossi†
University of Maryland
February 7, 2018
We employ a novel brokerage account dataset to investigate which individual investors are the
most attentive, how investors allocate their attention, and the relation between investor attention and
performance. Attention is positively related to investment performance, both at the portfolio return
level and the individual trades level. We provide evidence that the superior performance of high-
attention investors arises because they purchase attention-grabbing stocks whose positive performance
persists for up to six months. Finally, we show that paying attention is particularly profitable when
trading stocks with high uncertainty, but for which a lot of public information is available.
We would like to thank Stijn Van Nieuwerburgh (the editor) and two anonymous referees for their comments andsuggestions. The paper has benefited from the comments made at presentations at the NBER 2016 behavioral meeting,the 2017 Conference on Financial Decisions and Asset Markets at Wharton, the 2017 CEPR European Summer Sympo-sium in Financial Markets – Gerzensee, the 2017 WFA Annual Meeting, the 2017 Barcelona GSE Summer Forum, the2017 Finance Down Under Conference at the University of Melbourne, the 2017 EFA Annual Meeting, the 2017 ITAMFinance Conference, the 2017 FSU SunTrust Beach Conference, the 2017 FMA Asia/Pacific Annual Meeting, the 2017China International Conference in Finance, the 2016 HEC-McGill Winter Finance Workshop, Georgetown University,the R.H. Smith School of Business at the University of Maryland, the University of Melbourne, and the Johns HopkinsCarey Business School. We are grateful to Sumit Agarwal, Santosh Anagol, Daniel Andrei, James Ang, Ilona Babenko,Gurdip Bakshi, Turan Bali, Federico Bandi, Jonathan Berk, Brad Barber (the NBER discussant), Jules van Binsbergen,Peter Bossaerts, John Campbell, Maria Cecilia Bustamante, Yingmei Cheng, Carole Comerton-Forde, Max Croce, Hen-rik Cronqvist, Julien Cujean, Francesco D’Acunto, Zhi Da, Sandeep Dahiya, Alexander David, Douglas Diamond, AllanEberhart, Joey Engelberg, Stephen Figlewski, Laurent Fresard, Nicola Fusari, Pengjie Gao (the FSU SunTrust BeachConference discussant), Xiaohui Gao, Bruce Grundy, Jaehoon Hahn (the CICF discussant), Samuel Hartzmark, MichaelHasler, Xing Huang (the ITAM Finance Conference discussant), Irena Hutton, Ryan Israelsen (the WFA discussant),Pete Kyle, April Knill, Mattia Landoni (the CEPR discussant), Juhani Linnainmaa (the FDU discussant), Dmitry Liv-dan, Mark Loewenstein, Spencer Martin, Rich Mathews, Gonzalo Maturana, David McLean, Vladimir Mukharlyamov,Will Mullins, Carsten Murawski, Federico Nardari, Marina Niessner (the Conference on Financial Decisions and AssetMarkets discussant), Greg Nini, Lubos Pastor, Elena Pikulina, Nagpurnanand Prabhala, Tarun Ramadorai, RodneyRamcharan, Steven Riddiough, Nikolai Roussanov, Shrihari Santosh, Philipp Schnabl, Andrei Shleifer, Stephan Siegel,Kelly Shue, Eric So, David Solomon, Zhaogang Song, Juan Sotes-Paladino, Chester Spatt, Michael Ungeheuer (the EFAdiscussant), Stephen Utkus (the Conference on Financial Decisions and Asset Markets discussant), Laura Veldkamp,Katherine Waldock, Quan Wen, Joakim Westerholm, Russ Wermers, Fernando Zapatero, Yeqin Zeng (the FMA discus-sant) and Eric Zwick for comments and suggestions. Antonio Gargano acknowledges support from the Faculty ResearchGrant funded by the University of Melbourne.∗University of Melbourne, Faculty of Business and Economics, 198 Berkeley Street, Melbourne, VIC 3010. Email:
[email protected]†Smith School of Business, University of Maryland, 4457 Van Munching Hall, College Park, MD 20742. Email:
Introduction
Traditional asset pricing models assume that investors continuously incorporate all the available in-
formation into their investment decisions. In reality, however, attention is a scarce resource and indi-
viduals display limited attention. Rational models argue that attention-constrained investors should
benefit from paying attention and should pay attention up to the point where the improvement in
investment performance equals the costs of information acquisition (see, among others, Peress (2004),
Peng and Xiong (2006), Van Nieuwerburgh and Veldkamp (2009, 2010), Kacperczyk, Van Nieuwer-
burgh, and Veldkamp (2016)). The behavioral literature, on the other hand, shows that investors
tend to be overconfident and are subject to numerous biases, such as the disposition effect, suggesting
that paying more attention may harm rather than improve investor performance (see, among others,
Odean (1998a,b, 1999), Barber and Odean (2001), Guiso and Jappelli (2006)).
While the number of normative models on the behavior of attention-constrained investors has
exploded in recent years, lack of data on individual attention has made it virtually impossible to
test directly many of the theories and the literature is still lacking satisfactory answers to seemingly
simple questions, such as, how often do investors pay attention to their investment portfolio? What
individual characteristics lead certain investors to pay more attention to their investment portfolios
compared to others? How do investors allocate their attention? And, more importantly, does paying
more attention lead to better or worse investment decisions?
We answer these questions using a unique novel brokerage account dataset. The dataset is unique
in that it contains – at the individual investor level – detailed information regarding both investor
attention and trading behavior. For approximately 11,000 accounts, we observe the time-stamp of
investors’ login to the brokerage account website, what webpages they browse within the brokerage
account domain, and how much time they spend on each webpage. We use this information to construct
various measures of attention, such as the number of minutes spent on the brokerage account website,
the number of webpages browsed, and the number of log-ins. The extreme granularity of our data
allows us to even identify what type of information and what stocks investors focus their attention
on. For the same accounts, we also have detailed trading activity information. For every trade placed
1
by each investor, we observe whether it is a buy or a sell, an identifier of the security traded, the
time-stamp of when the trade is ordered and executed, the quantity traded, and the price. Finally,
the dataset contains quarterly holdings information and clients’ biographical characteristics.
We first explore the determinants of investor attention, that is, the relation between investor
characteristics and investor attention. We find a very large heterogeneity in attention across investors
that can be explained by the size and risk of the investors’ portfolios, as well as by investor trading
habits and demographic characteristics. Account holders with higher invested wealth and higher
exposure to small capitalization stocks, growth stocks, momentum stocks, and the overall market,
are more attentive. The same is true for investors that trade more frequently. On the other hand,
those investors that hold a higher fraction of their invested wealth in cash or ETFs are less attentive.
Finally, we find that males pay more attention than females and attention is an increasing function of
investors’ age.
We then analyze how investors allocate their attention. We show the average annualized number of
stocks researched by investors is equal to 31. Additionally, we find that the number of stocks researched
by investors is positively related to the overall amount of time spent on the website, the number of
stocks in their portfolio, and the number of stocks they trade. On the contrary, investors that have
more concentrated and riskier portfolios tend to pay less attention to stock-specific information. In
addition to studying the number of stocks researched, we analyze how investors concentrate their
attention. We find that older investors and investors that pay more attention have less concentrated
attention patterns. Investors that hold fewer stocks, have more concentrated portfolios, trade less,
and have a smaller fraction of their portfolio in cash, allocate their attention in a more concentrated
fashion. Finally, we find that investors pay more attention to local stocks, stocks with a higher weight
in their portfolio, stocks that have a higher squared Sharpe ratio, and stocks of companies with a
larger market capitalization – as predicted by Van Nieuwerburgh and Veldkamp (2009, 2010).
We complete the picture by analyzing what stock characteristics are related to investor attention.
Investors pay more attention to companies with higher R&D expenditures, market-to-book ratios,
market capitalization and leverage, suggesting that investors prefer to research large companies that,
while risky, have high growth potentials. Finally, our results confirm that many proxies used in the
2
literature, such as stocks’ volume, turnover, news, and analyst coverage are indeed positively related
to investor attention – with turnover being the most economically and statistically significant.
We then turn to studying the relation between attention and performance. We find a strong and
positive cross-sectional relation between attention and performance, in that more attentive investors
achieve higher risk-adjusted returns and portfolio Sharpe ratios – even after controlling for covariates
related to investment style. Our baseline results show that a standard deviation increase in overall
attention is associated with a 1.5% (3.36 t-statistic) increase in annualized risk-adjusted returns and a
0.09 (7.08 t-statistic) increase in investors’ annualized Sharpe ratios. The results are consistent when
we use other measures of attention – such as the amount of time spent on the research pages of the
brokerage account website. Finally, the results that use the number of pages browsed or the number
of logins as measures of attention are qualitatively similar, but economically smaller, suggesting that
– compared to seconds – pages and logins may be poorer proxies for the process of information
acquisition.
We also find that, for a given level of attention, those investors that specialize and focus their
research efforts on specific stocks perform better compared to those that evenly allocate their attention
across all the stocks in their information set.
Our cross-sectional analysis cannot distinguish whether investors perform well because they pay
more attention from the alternative hypothesis that investors pay more attention because their port-
folio has been performing well. To disentangle the two effects, we present results for panel regressions
that relate each investor attention to future performance. The additional advantage is that panel
regressions allow us to control for (average) investor skills using fixed effects and overall market con-
ditions using time-effects. At the one-month horizon, we find that a one standard deviation increase
in attention is associated with an annualized increase in portfolio risk-adjusted performance of 0.38%
(2.27 t-statistic). The effect increases to 0.62% (4.21 t-statistic) at the two-month horizon and de-
creases slightly to 0.55% (3.66 t-statistic) at the three-month horizon. This indicates that paying
attention allows individual investors to improve their portfolio performance in the short-term, sug-
gesting that the improvement in performance hinges on attentive investors being able to purchase
(sell) stocks that – in the short run – realize relatively large positive (negative) returns. Controlling
3
for trading fees does not have a significant impact on our results.
Finally, we find that the effect of attention is stronger for those investors that are – on average
– less attentive. In particular, we divide our investors in five quintiles based on their unconditional
attention and show that the relation between attention and future performance is monotonically
decreasing across the five groups. Furthermore, while the coefficients on attention for the first four
groups are positive and significant, the coefficient on attention for the last group is small, negative,
and not statistically different from zero – indicating that there are diminishing marginal returns to
attention.
Because portfolio performance does not necessarily capture investors’ active management, we pro-
vide results on the relation between attention and the performance of individual trades. Attention
is positively related to the future performance of the stocks purchased up to four months after the
trade is placed. The economic magnitudes are large. At the three-month horizon, for example, a
unit-standard deviation increase in attention increases the average annualized adjusted returns of the
stocks purchased by 2.13% (3.57 t-statistic). We find – on the other hand – no discernible effect of
attention on the performance of the stocks sold.
To understand the economic mechanism relating attention to the performance of the stocks traded,
we conduct a number of auxiliary exercises. First, by analyzing the performance of the stocks before
they are traded, we provide evidence that the superior performance of high-attention investors arises
because they purchase attention-grabbing stocks that have appreciated in the recent past and whose
positive performance persists for up to six months. However, the relation between attention and
future stock returns remains significant when we include attention-grabbing proxies as controls in the
performance regressions.
Second, we show that attention is particularly profitable when investors trade stocks with high
market capitalization, trading volume, volatility, number of analysts, dispersion of analyst forecasts,
and news – indicating that it is for the stocks with high uncertainty, but for which a lot of public
information is available, that it pays to pay attention.
Third, we show that Odean (1999)’s result that the stocks sold by individuals outperform the ones
4
purchased disappears for high-attention trades, but is very strong for low-attention trades. For high-
attention trades, the average annualized three-month abnormal return of the stocks purchased equals
3.16%, the one for the stocks sold equals 4.20%, and their difference is statistically insignificant, with
a p-value of 0.29. For low-attention trades, on the other hand, the average annualized three-month
abnormal return of the stocks purchased equals -0.28%, the one for the stocks sold equals 3.08%, and
their difference is statistically significant, with a p-value of 0.00.
1 Related Literature
Our paper is related to the theoretical literature that studies the behavior of investors with informa-
tion capacity constraints. A first strand of this literature focuses on the cross-sectional dimension of
inattention. That is, what sources of information or assets should attention-constrained investors pay
attention to, when facing the problem of allocating their wealth across a number of assets. Theoreti-
cally, Peng and Xiong (2006) show that investors with limited attention capacity engage in category-
learning, in that they do not focus on stock-specific information, but on market and sector-wide
information. Van Nieuwerburgh and Veldkamp (2010) propose a model where, depending on their
preferences, investors behave as “specialized learners” – they focus their attention on a single asset
– or “generalized learners” – they focus their attention on multiple assets. Van Nieuwerburgh and
Veldkamp (2009) predict that investors should pay more attention to the assets that have a higher
squared Sharpe ratio and market capitalization. We contribute to this literature by providing evidence
on how brokerage account investors allocate their attention – that is – how many individual stocks
they focus their attention on, whether they spread their attention evenly among them, and what stock
characteristics drive investor attention.
A second strand of this literature focuses on the time-series dimension of attention and shows
that, if information acquisition is costly, it is optimal to alternate long periods of inaction to brief
spurs of attention, where information is acquired and investment decisions are made (see Gabaix and
Laibson (2002), Abel, Eberly, and Panageas (2007, 2013), Huang and Liu (2007), and Alvarez, Guiso,
and Lippi (2012)). We contribute to this literature by providing empirical evidence on the patterns
5
of investor attention and trading. That is, how often investors pay attention to their investment
portfolio, whether they evaluate their portfolio allocations at equally spaced intervals – as many of the
theoretical models predict – or they alternate periods of high attention to periods of low attention.
Finally, we study what demographic and portfolio characteristics are associated with higher or lower
degrees of attention.
Our paper is also related to the literature that studies the performance of individual investors.
For the most part, the literature has documented the mistakes of individual investors. For example,
Odean (1999) and Barber and Odean (2000) show that – on average – individual investors trade too
frequently and that trading is detrimental to their wealth. Subsequent studies have uncovered sub-
stantial cross-sectional variation among investors’ trading performance. Superior trading performance
has been linked to investors’ IQ (Grinblatt, Keloharju, and Linnainmaa (2012) and Korniotis and
Kumar (2013)), education (Von Gaudecker (2015)), wealth (Calvet, Campbell, and Sodini (2007)),
experience (Korniotis and Kumar (2011) and Nicolosi, Peng, and Zhu (2009)), and portfolio concen-
tration (Ivkovic, Sialm, and Weisbenner (2008)). More recently, Frydman, Hartzmark, and Solomon
(2017) document that investors make better investment decisions when they sell one asset and quickly
buy another one. On reinvestment days, investors display no disposition effect and make better selling
decisions. Our results provide novel empirical evidence on the relation between attention and trading
performance. If investors acquire valuable information while spending time on the trading platform,
we expect their trades to be more profitable as they pay more attention. If investors are incapable
of processing the information they acquire, on the other hand, we expect to find no relation between
attention and performance. Finally, if investors systematically misinterpret the information they ac-
quire, we expect to find a negative relation between attention and performance. The fact that we find
a positive relation between attention and performance, even in the presence of investor fixed-effects,
is an indication that – at least certain – investors are systematically able to understand and exploit
the information they acquire.
We also contribute to the empirical literature that investigates investor attention, its determinants,
and its impact on asset prices. Since its inception, this literature has faced significant challenges in
measuring attention itself, leading researchers to resort to attention proxies such as trading volume
6
((Gervais, Kaniel, and Mingelgrin, 2001)), price limits (Li and Yu (2012) and Seasholes and Wu
(2007)), and news (Yuan (2015) and Barber and Odean (2007)), and making the implicit assumption
that investors are likely to pay attention to stocks that are mentioned in the news or that have
been heavily traded on a given day. More recently, Da, Engelberg, and Gao (2011) propose the
use of Google searches as a direct measure of aggregate attention and – using Google searches –
Vlastakis and Markellos (2012) and Andrei and Hasler (2015) show that aggregate attention varies
as a function of stock market volatility. The advantage of using aggregate Google web-searches over
news is that they identify the information investors actively seek, rather than the information they
are potentially exposed to. A key shortcoming, on the other hand, is that they are not specific to
the individual investor and therefore cannot shed new light on how attention and the process of
information acquisition relates to trading at the individual investor level.
The only studies that obtain direct measures of investor attention at the individual level – and
are therefore closest to ours – are Karlsson, Loewenstein, and Seppi (2009), and Sicherman et al.
(2016). Using a large panel of investors’ logins to 401K accounts as a measure of attention, they
show that investors pay less attention to their investment portfolio after stock market declines. They
also show that investors’ attention vary as a function of portfolio holdings, wealth, and demographic
characteristics – such as age and gender. While it confirms many of the findings in Sicherman et al.
(2016), our work is different along several dimensions. First, we have information on brokerage rather
than 401K accounts. This is important, because 401K investors can only choose among a limited
number of fixed income and equity funds, and are completely unable to purchase or sell individual
stocks. Furthermore, as shown by Agnew, Balduzzi, and Sunden (2003), Madrian and Shea (2001) and
Sialm, Starks, and Zhang (2015), 401K investors display very limited trading, a high degree of inertia,
and extreme asset allocations.1 Second, our measures of attention are not limited to investors’ login,
because – for each investor – we observe what information he/she browses and how much time he/she
spends doing it. This means that we are able to provide results regarding what information investors
pay attention to and how much time they spend thinking about their portfolio decisions. Finally, we
have detailed information regarding investors’ portfolio allocation and trades, meaning that we can
1As reported in Agnew, Balduzzi, and Sunden (2003), the distribution of allocations to stocks across 401K investorsis strongly bimodal: 48 percent of the average annual equity allocations are zero, while 22 percent are 100 percent.
7
relate investors attention to the type of stocks they own and to the performance of the stocks they
purchase and sell.
The rest of the paper is organized as follows. Section 2 describes the web-activity data and
explains how we extract our measures of attention. Section 3 provides summary statistics on investor
portfolio allocation as well as their trading and attention habits. Section 4 studies the characteristics
of attentive investors and analyzes how investors allocate their attention. Section 5 analyzes the
relation between investors’ attention and portfolio performance. Section 6 studies the relation between
investors’ attention and the performance of their trades. Section 7 provides extensions of the baseline
results. Section 8 concludes.
2 Measuring Attention Using Investor Web-Activity
In this section, we discuss the measures of attention used in the paper. We start by presenting the
data we have access to in terms of investors’ web-activity and we show that – once aggregated across
all the investors in our dataset – the information measures we construct have an information content
similar to that of Google’s Search Volume Index. The advantage of our measures, however, is that
they are available for each investor and therefore allow us to study the relation between attention and
investment decisions at the individual level.
2.1 The web activity we observe
Before describing the measures of attention we construct, we provide an example of the web-activity
we observe for each investor in our sample. Table 1 shows the web behavior of a sample account holder
on January 28, 2014.2 To preserve the anonymity of the brokerage account house, we mask the URLs
we have access to and, rather than reporting the full string characterizing each URL, we only report
the content of the webpage each URL is associated with.
The table shows that the account holder had a total of four sessions over the day. The first connec-
2In compliance with the US privacy law, no Personally Identifiable Information (PII) was provided by the brokeragehouse. For example, each account holder has been anonymized using a numeric account identifier.
8
tion occurred at 7:58:05 am (Central Time) and, after logging to the homepage, the investor checked
his/her balances and positions as well as his/her stocks’ watchlist. The balances and positions page
does not report only the amount of wealth invested in each stock, but also important stock informa-
tion such as past returns; historical and forecasted earnings and dividends; key accounting information
such as price-to-book and price-to-earnings ratios; basic facts such as revenues and institutional own-
ership; and a summary of analysts’ opinions as well as recent news. The watchlist is instead a webpage
containing information on all the stocks and other assets the investor decides to pay attention to.
The second session starts only thirty minutes later, two minutes after the markets open in New
York. This second session is much longer, approximately forty minutes, and entails more actions. The
investor first checks the research page of the website detailing the latest news on the SPX, the S&P
500 index. Right after, he/she connects to the page displaying his stocks’ watchlist to quickly switch
to the research page relative to the V IX. The rest of the session is dedicated to assessing balances
and positions of his/her trading account and the watchlist.
The third session occurs right around lunch time, it lasts a little more than eighteen minutes, and
it involves a look at the watchlist as well as the balances and positions pages. Finally, the last session
of the day occurs at 14:28:54 pm, 30 minutes before the markets close. It lasts only forty seconds and
involves just a quick look at his/her balances and positions and the watchlist.
2.2 From links to attention
The web activity information presented in Table 1 is a small example of the web-activity data we were
granted access to. In particular, as part of a large SQL relational database described in more details
in Appendix A, the brokerage house gave us access to a web-activity table containing the web pages
visited – within its website domain – for approximately 11,000 randomly chosen accounts over the
period January 2013 – June 2014. The dataset is very granular and contains in excess of 17 million
observations. Each observation contains a unique numeric account identifier, the URL of the webpage
visited within the brokerage account domain, the date and time of the first click on the page, the
number of seconds spent on the page, and the number of the web-sessions within the trading day.
Next, we explain how we use this data to construct a variety of attention measures.
9
2.2.1 Overall attention measures
The first set of attention measures is related to the overall attention paid by investors to their brokerage
account. The first measure is the number of seconds spent on the brokerage account website over a
given time interval. The second measure computes the total number of pages visited by the investor.
Finally, the third measure is the number of investor logins to the brokerage account website.
Using three measures is important for robustness purposes. One may prefer the number of pages
visited rather than the number of seconds spent on the website, because it is possible for some in-
dividuals to stay logged into the brokerage account website and leave it in the background, while
performing other activities. On the other hand, one may prefer the number of seconds rather than
the number of pages visited, because it is unlikely for someone to understand the content of a stock
report if he/she only spends one or two seconds reading it. Considering the number of logins is also
important, because it potentially allows us to distinguish between the extensive and intensive margin
of investors’ attention. For example, an investor that connects multiple times a day, but stays con-
nected for a few seconds, is probably looking for very different information compared to an investor
that connects once or twice, but spends an hour or longer on the trading platform.
To show that the attention measures we extract from our data are related to the ones that have
been proposed in the literature, we present in Figure 1 the time-series – at the weekly frequency – of
the total number of seconds aggregated across all investors, and four attention proxies:3 the Google
Search Volume Index for the word “S&P 500” (Panel A), the total number of news pertaining to
stocks in the S&P 500 (Panel B), the trading volume on the S&P 500 (Panel C), and the State Street
Investor Confidence Index (Panel D).4 Each panel reports our measure of attention as a red dashed
line and one of the alternative measures as a solid blue line. The plots clearly show that there is a
very tight relation between our attention measure and both the Google’s Search Volume Index for the
S&P 500 (78.2% correlation, statistically significant at the 1% level) – as well as the trading volume
on the S&P 500 (66.1% correlation, statistically significant at the 1% level). The news variable is also
quite related to our attention measure (43.1% correlation, statistically significant at the 1% level),
3The results for the total number of pages visited and the total number of logins are very similar.4The total number of news pertaining to the stocks in the S&P 500 Index are obtained from Capital IQ, while we
downloaded the S&P 500 trading volume from Yahoo! Finance.
10
while the correlation with the confidence index is relatively low – at only 27.7% – and is statistically
insignificant.
2.2.2 Categorical attention measures
The second set of attention measures also uses seconds, pages and number of logins, but focuses on
the various sections of the website visited by the investors. The brokerage account website has a
hierarchical structure, whereby – for example – all the web-pages related to specific tickers like SPX
and V IX fall under the “Research” category. By parsing all the URLs and categorizing them, we
can obtain many other categories such as “Home Page”, “Balances and Positions” and “Watchlist.”
To maintain parsimony, we classify all the available URLs into 14 categories: Balances and Positions,
Research, Trading, Homepage, Account, Watchlist, History Statement, Bank, Mail, Tax, Help, Search,
Retirement, and Client.5
Panel A of Table 2 reports the daily total number of hours spent across investors for the top
six categories. The most viewed section of the website is – on average – “Balances and Positions.”
This page does not only contain information regarding investors’ performance and portfolio weights,
but also important stock information such as past returns; historical and forecasted earnings and
dividends; key accounting information such as price-to-book and price-to-earnings ratios; basic facts
such as revenues and institutional ownership; and a summary of analysts’ opinions as well as recent
news. The average time spent across all investors is 787 hours per day, but this aggregate measure has
a large standard deviation equal to 412, so the total number of hours spent ranges from 516 to 930
hours for the 25th and 75th percentiles, respectively. Research is the second most popular category,
with 483 hours per day, followed by Trading and Homepage. The remaining categories are much less
visited by the investors. For example, the average number of hours spent on Watchlist is only 46 per
day.
Focusing on the various sub-categories is important because it allows us to discern whether looking
at different types of information leads to different trading behavior and performance.
5These categories are the ones that the brokerage account website is divided into and allow us to categorize 99% ofthe URLs.
11
2.2.3 Stock attention measures
The third and final set of attention measures uses only information associated with the “Research”
URLs and focuses on the tickers researched by investors. Panel B of Table 2 reports the Top 20
companies and ETFs researched by the investors in our dataset over the time-period October 1, 2013
– June 10, 2014.6
To compute the table, we first sum the number of minutes spent on each ticker across all the
account holders in our dataset and over the full sample. We then report the rank of each company
(or ETF) according to the number of minutes (first column), pages (second column) and visits (third
column).
Starting from the first column, the table shows that individual investors focus on companies that
belong to the consumer-tech space. Interestingly, while we find tech giants such as Facebook and
Apple ranked first and second, respectively, we also find companies that have much smaller market
capitalization, such as Twitter (7), AT&T (6), Verizon (10), Tesla (11), Sirius XM (15) and Netflix
(19), ranked higher or similarly to much larger firms such as Microsoft (14). All in all, it is remarkable
that 11 out of the top 20 stocks researched by investors are in the technology space. As expected, also
large conglomerates and banks populate the list. For example, among the companies listed we find
Bank of America (3), Ford (4) and General Electric (9). Finally, we find four ETFs in the list, i.e.
SPDR S&P 500 ETF (5), SPDR Gold Trust (17), SPDR Dow Jones ETF (18) and Market Vectors
ETF (20).
The results in the second and third columns are similar, indicating that number of minutes, pages
and visits capture similar patterns of behavior.
3 Summary Statistics
The source of the proprietary data used in this study is a large US discount broker. Unlike the discount
brokers described in Odean (1999), today’s discount brokers operate online and – while providing access
6The sample is dictated by data-availability, i.e. the complete URLs that include ticker information are available fromOctober 1, 2013 until the end of our sample.
12
to analysts’ research, market news, and a large number of tools to help investors with their trading –
they charge very small trading fees (less than $10 per trade). We relegate to Appendix A the detailed
description of the various databases employed in our study, and proceed to present the key summary
statistics of our final dataset below.
Table 3 reports the summary statistics for the subset of accounts (approximately 11,000) for which
we have web-activity information.7,8 Starting from the biographical traits, the first row of Panel A
shows that the average age of account holders in our dataset is approximately 51, the second row
shows that 73% of the account holders are males. While the average and median age are very much
aligned with the ones of previous studies, our dataset has a slightly higher percentage of women –
27% in our study compared to 21% in Barber and Odean (2001), for example. The average number
of accounts per client is 1.34. This occurs because – while 80% of the clients have only one account –
20% of the clients have more than two accounts. Finally, as of June 2014, the average account age is
8.55, which suggests that the average account holder in our sample is quite experienced.
Next, we turn to the portfolio characteristics of account holders as of March 31, 2014. As reported
in Panel B, the average (median) household has a portfolio value that equals $94,000 ($18,000),
indicating that the distribution is heavily skewed to the right. Cash holdings average $16,000 and
their distribution is also heavily skewed to the right. Conditional on having at least one stock in the
portfolio, the average account holds 6.51 stocks worth $82,000 and the median counterpart is four
stocks for a total of $15,000.9 The median values for stock holdings are in line with those in Barber
and Odean (2000), who report that the median household holds 2.61 stocks worth $16,210.
Panel C of Table 3 reports summary statistics for the trading behavior of account holders. The
results are computed as follows. In the first step, we compute the results for each account holder using
his/her full time-series, eliminating the days when the stock markets are closed – mainly week-ends
and holidays – as individuals tend to connect much less at those times. In the second step, we compute
7Even though these accounts were randomly selected from the full set of accounts of the brokerage account house,we allay concerns regarding their representativeness by recomputing the statistics in Table 3 using the entire universeof accounts of the brokerage account house (approximately 3.5 millions). The results, reported in Table Online I, showthat the biographic and portfolio characteristics statistics are very similar across the two samples.
8Because our brokerage house has millions of clients, our sample is likely to be representative of the population of USstock investors.
9All dollar figures have been rounded to the nearest thousand upon request of the data provider.
13
cross-sectional results across account holders. The first row shows that, on average, investors trade
on 3% of the days, with 50% of the investors not trading at all, and 1% of the investors trading more
than 44% of the days. For those investors placing at least two trades, the number of days between
trades averages 47, resulting in approximately one trade every two months. The median counterpart
is 25 days. Conditioning on placing a trade, the average investor places 1.72 trades per day, the
median being 1.31. This indicates that investors tend to cluster their trades, consistent with the idea
that – once they decide to re-optimize their investment positions – investors like to make multiple
transactions on the same day. Finally, the average trade size is $16,000, in line with Barber and Odean
(2000) who report an average trade size of $13,707 ($11,205) for stock purchases (sales).
Panel D reports summary statistics for the attention behavior of account holders. The average
percentage of days with logins (across investors) equals 17% – which is almost six times larger than
the trades’ frequency. The most active 1% of the investors logs-in 96% of the days, while the median
investor logs-in 6% of the days. The number of days between logins also shows that login and trading
behaviors are quite different in terms of magnitude. The average number of days between logins
averages 27.51 across account holders, while the median value is 11.20. Note that these numbers are
not only much smaller than their trades counterparts, but they are also computed using information
for twice as many account holders, i.e. those accounts that do not trade at all over our sample.
Conditional on logging-in, investors re-visit their trading account several times within a given day
– the average is 10.61 and the median is 7.33. This indicates that there is a large degree of clustering
in the visits we observe and that investors do not log-in at regular intervals. Once they decide to pay
attention, investors seem to spend a substantial amount of time on the trading platform. In particular,
the average number of minutes spent on the website – conditional on logging-in – equals 29, while
the median value is 8. As expected, the shortest 1% of the sessions lasts only 18 seconds, while the
longest 1% of the sessions can be as long as 366 minutes (more than 6 hours).
To help the visualization of the cross-sectional variation in investor attention and to display how
the within-investor attention varies over time, we use a heat-map graph in Figure 2. The figure is
constructed as follows. For each account holder, we generate a time-series from January 2013 through
June 2014, eliminating the days when the stock markets are closed – mainly week-ends and holidays –
14
as individuals tend to connect much less at those times. We then compute, for each investor, the daily
number of minutes spent on their investment account.10 Finally, to ease the visualization, we sort the
accounts by the total number of minutes over the full sample, so that the more active accounts are at
the top of the figure. Figure 2 uncovers considerable heterogeneity in behavior across accounts. At the
top, we find the more attentive investors that consistently log-in and spend about one to two hours
per day on their account. At the very bottom, on the other hand, we find those individuals that rarely
log-in. The figure also highlights some heterogeneity in individual accounts’ behavior over time. For
example, the horizontal lines of “colder” colors – that appear in multiple parts of the figure – identify
periods where a given investor pays more attention than usual to his/her investment portfolio. The
opposite holds true for the horizontal lines of “warmer” colors, even though these are harder to discern.
The first eight days of the sample are characterized by lighter colors for all clients. This is because
the company was introducing and testing the tracking system over that week and the system was not
fully functional. We do not include these observations in the computation of the results reported in
the rest of the paper.
Taken together, these findings represent a new benchmark for the models of optimal inattention,
that often imply inattention intervals much longer than the ones we observe. For example, the model
of Gabaix and Laibson (2002) implies an inattention interval of approximately 12 months, the one by
Abel, Eberly, and Panageas (2007) an inattention interval of 8 months. With values slightly greater
than one month, the only model that predicts inattention intervals in line with the ones we find in
the data is the one by Abel, Eberly, and Panageas (2013). Our findings also show that it is crucial
for inattention models to capture the asynchronicity between attention and trading, as featured in
Abel, Eberly, and Panageas (2013) and Alvarez, Guiso, and Lippi (2012). Finally, the clustering of
attention, whereby investors pay attention to their portfolios for several days or weeks and then decide
to be inattentive for months (or even years) is difficult to reconcile with standard models of optimal
inattention – that predict instead regular inattention intervals.
10For ease of visualization, we winsorize the number of minutes at the 95-th percentile, which results in the numberof winsorized minutes to range from zero to 150. The summary statistics reported in Tables 3 are computed usingnon-winsorized data.
15
4 Attention Allocation
We start this section by analyzing the determinants of investor attention. That is, we explore what
are the observable traits that characterize more and less attentive investors. The second part of this
section is dedicated to providing new evidence on how investors allocate their attention. We first
provide novel facts regarding the degree of investor specialization in information acquisition. We
then relate the number of stocks investors pay attention to, both inside and outside their investment
portfolio, to their characteristics. Finally, we explore how investors allocate their attention among the
stocks they track and the characteristics of the stocks they research.
4.1 Who are the most attentive investors?
In Van Nieuwerburgh and Veldkamp (2010), information acquisition involves two dimensions: how
much attention to allocate, and how to allocate attention across the assets considered. Thanks to
the richness of our data, we are able to study empirically the process of information acquisition at
the individual investor level. In this subsection, we focus on the first dimension and study the cross-
sectional determinants of investor attention. We use the number of minutes investors spend on the
brokerage account website – and its various sections – as a measure of attention and relate it to
investor holdings, trading, and demographic characteristics. We estimate the following cross-sectional
regression at the account level:
Attentioni = α+ x′i β + εi, for i = 1, . . . , N, (1)
where Attentioni is the (log) total number of minutes spent over the 18 months period on the brokerage
account website by account holder i and N is the total number of accounts in our dataset.11
We divide the conditioning variables into three groups. The first group comprises demographic
variables: Male, a male dummy variable; Brokerage, a brokerage account dummy; Age, the age of the
investor; and Account Age, the age of the account. The second category comprises portfolio holdings
11More precisely, Attentioni is computed as Attentioni = log(1 + minutesi). We use the same formula for all thelogged attention measures in the paper.
16
variables as of December 31, 2013: Portfolio Value, the total value of the invested portfolio; Fr.
in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of wealth held in cash, traded funds
and mutual funds, respectively. The third category comprises regressors related to portfolio risk and
trading activity: Beta Mkt, Beta SMB, Beta HML, and Beta MOM, the loadings on the Market,
Small-Minus-Big, High-Minus-Low, and Momentum factors – computed using daily returns over the
full sample; and N. of Stocks Traded, the number of stocks traded over the period.
To ease the economic interpretation of our estimates, we de-mean and standardize all regressors
so that they have unit variance.12 The results for this cross-sectional regression are reported in Table
4. The first column reports the coefficient of each regressor, the second its t-statistics, and the third
its economic magnitude. This last column is computed as follows. We first compute the number
of minutes spent on the brokerage account website by the base-case investor, which equals 820.91
minutes, or 13.6 hours. We then multiply eβk − 1 to this base number for each coefficient βk where
k = 1, . . . ,K.13 For example, the economic magnitude of the Brokerage dummy is computed as
820.91 · (e0.590−1) = 660.11, meaning that brokerage account holders spend 660.11 additional minutes
– or 11.1 hours – compared to non-brokerage account holders, for a total of 820.91+660.11=1,481.02
minutes. This result is quite remarkable, as it shows that brokerage account investors spend twice as
much time on their account compared to IRA, and other account-type holders.
The remaining regressors associated with demographic characteristics show that, first, males pay
more attention than women to their investment portfolios – by an average of 299.14 minutes. Second,
investors pay more attention to their investment portfolios as they get older. The effect is also quite
strong, as a standard deviation increase in age is associated with an increase in attention of 277.52
minutes. Third, investors’ attention does not increase with the amount of time investors have had
their account open.
12We do not standardize the dummy variable regressors Male and Brokerage.13Because we are estimating a log-linear regression, the economic interpretation of each coefficient βk is computed as
eβk − 1, and is interpreted as the percentage change in the number of minutes spent on the brokerage account websitewhen the k-th regressor increases by one standard deviation. Finally – because all regressors have been de-meaned – thenumber of minutes spent on the brokerage account website by non-male and non-brokerage account holders, that areaverage in terms of all the other conditioning variables, represent our “base-case investor.” The number of minutes thebase-case investor spends on the brokerage account website can be computed as eα̂ ·N−1∑N
i=1 eε̂i , where α̂ and ε̂i are
obtained from Equation (1), see Duan (1983).
17
Turning to the regressors associated with portfolio characteristics, we find that investors with
higher wealth pay more attention, while attention is negatively related to the fraction of the portfolio
invested in cash or in exchange traded funds. Finally, the fraction of wealth invested in mutual funds
does not seem to be related to attention in any significant manner.
In terms of portfolio performance and risk, we find that investors that have greater exposure to the
aggregate stock market, as measured by the beta of their portfolio with respect to the market factor,
pay more attention. The same is true for those investors that are more exposed to the SMB and MOM
factors, indicating that those investors that invest in small caps and momentum stocks are, overall,
more attentive to their portfolios. The opposite holds for HML exposure, where we find that investors
that focus on growth stocks, as opposed to value stocks, are more attentive. The final regressor in this
category is the number of trades undertaken by the account holders over our sample. As expected,
the more trades they place, the greater the degree of attention they pay to their investment portfolio.
The coefficient for this regressor is both statistically and economically very significant: a standard
deviation increase in the number of trades is associated with an increase in attention of 1,030.81
minutes (17.2 hours) over the sample.
Table Online II shows that the results are similar when using Research and Balances and Positions
as measures of attention, and Table Online III shows that our results are robust to using number of
pages and logins.
4.2 How Do Investors Allocate Their Attention?
The second dimension of information acquisition – as modeled by Van Nieuwerburgh and Veldkamp
(2009, 2010) – involves allocating attention across the assets considered. In this respect, we provide
novel empirical evidence along two dimensions. We first focus on the number of assets investors decide
to pay attention to – i.e. how many stocks they have on their “radar.” We then analyze how investors
allocate their attention among the assets on their radar. To conduct this analysis, we focus on the
stock-specific attention measures discussed in Section 2.2.3.
18
4.2.1 How Many Stocks do Investors Pay Attention to?
If investors did not have attention constraints, they would continuously acquire information regarding
all the stocks potentially available. Because of their limited time and resources, however, investors
generally collect information regarding a rather small number of assets, and some investors do not
collect stock-specific information at all. For example, out of the 10,768 accounts in our dataset, only
4,970 collect stock-specific information over the time-period October 1, 2013 – June 10, 2014. Among
the investors that collect stock-specific information, the average investor acquires information for a
total of 31 stocks annually. The distribution, however, is very skewed to the right, with the median
investor looking at 8 stocks and the 75th, 95th, and 99th percentiles of the distribution equal to 25,
128, and 376, respectively.
To understand whether there is a systematic relation between investor characteristics and the
number of stocks in their radar, we estimate the following baseline cross-sectional regression:
N Stocks Researchedi = α+ x′i β + εi, for i = 1, . . . , N, (2)
where N Stocks Researchedi is the (log) number of stocks searched by investor i, N is the total
number of investors included in the analysis, and x′i is the same vector of investor characteristics
contained in Equation (1), augmented with four regressors: Total Attention, the total amount of time
spent by the investor on the website, see Section 2.2 for details; Stock Herfindahl, the Herfindahl
index of each investor’s portfolio holdings as of December 31, 2013; N. Stocks, the total number of
stocks held as of December 31, 2013; and N. of Stocks Traded, the number of stocks traded during the
sample. We also estimate modifications of the baseline specification, where we focus, respectively, on
the attention paid to the stocks inside or outside investors’ portfolios. As in Section 4.1, we de-mean
and standardize all regressors so that they have unit variance.
The results of the three specifications – reported in Panel A of Table 5 – suggest the following.
First, across all specifications, we find that the number of stocks researched by investors is positively
related to the overall amount of time spent on the website (Total Attention), the number of stocks in
their portfolio (N. Stocks), and the number of stocks they trade (N. of Stocks Traded). On the contrary,
19
investors that have more concentrated (Herfindahl) and riskier (Risk Idio) portfolios research fewer
stocks. We also find that older investors tend to acquire information regarding a larger number of
stocks. Overall, these results suggest that the investors that have broader portfolio allocations (and
trade more often) are also the ones that monitor the stocks they own and explore new investment
opportunities more actively. In terms of economic magnitudes, the effect of the various covariates
is quite large. For example, the coefficient on Total Attention in Panel A.I. equals 0.423, indicating
that a one-standard deviation increase in the number of stocks held by investors is associated with a
e0.423−1 = 52.6% increase in the number of stocks researched. For the average investor that researches
31 stocks, this translates in an additional 16.3 stocks researched.
Second, we find that certain covariates have differential impact on the number of stocks researched
inside and outside the portfolio. For example, male account holders seek new investment opportunities
more actively – compared to females – as exemplified by the coefficients on Male reported in Panel
A.II. We do not, however, observe any difference between males and females for the stocks inside the
investment portfolios.
Third, as shown in Panel A.III., the value of the investment portfolio is positively related to the
number of stocks in the portfolio monitored by the investors. Along the same lines, we find that the
greater the fraction of wealth invested in Cash, ETF or Mutual Funds, the lower the number of stocks
in the portfolio tracked by the investors. These results are in stark contrast with the ones reported
for the same regressors in Panel A.II., where the regression coefficients are positive – rather than
negative – and statistically significant for the fraction of wealth invested in Cash. Economically, this
positive coefficients suggests that it is the investors with free cash that actively seek new investment
opportunities.
Finally, our results suggest that greater exposures to the Market, SMB, HML and MOM factors
are negatively related to the attention paid by investors to new investment opportunities. The same
regressors have instead an insignificant – or positive – effect on the attention spent by investors
monitoring their current holdings.
The results reported in Panel A of Table 5 compute coefficient estimates exploiting variation across
investors. If we take the relation between the number of stocks held and Total Attention as an example,
20
they show that those investors that pay more attention to their investment portfolio also track more
stocks. We also compute panel regressions with time and account fixed effects, which exploit within-
investor variation. Our findings, reported in Panel A of Table Online IV, are that investors tend to
increase the number of stocks they track as they pay more attention and as they trade more.
4.2.2 How do Investors Allocate Attention Among the Stocks they Follow?
We address the second question, i.e., how do account holders allocate their attention among the assets
on their radar, using three distinct approaches. The first exploits only cross-sectional variation. We
estimate a regression very similar to the one in Equation (2), but replace the dependent variable
N Stocks Researchedi with Herfindahl Researchedi: the Herfindahl index of the stocks researched
by investor i. This quantity is computed as∑Kk=1 ω
2k−1/K
1−1/K if K > 1 and 1 if K = 1, where ωk is the
fraction of the total stock-specific attention of investor i – measured in seconds – allocated to stock k
over the full sample and K is the total number of stocks searched. The measure is bounded between
0 – attention allocated evenly across all assets – and 1 – all attention concentrated in one asset.
The results, reported in Panel B of Table 5, indicate that older investors and investors with greater
total attention have less concentrated attention patterns, as shown by the negative coefficients on Age
and Total Attention. We also find a positive effect for Herfindahl Stocks, and a negative effect for N.
Stocks, Fr. in Cash, and N. of Stocks Traded, meaning that investors that hold fewer stocks, more
concentrated portfolios, trade less, and have have smaller fraction of their portfolio in cash allocate
their attention in a more concentrated fashion. Finally, we find that investors whose portfolios have
higher idiosyncratic risk have more concentrated patterns of attention across the stocks they search.
The second approach exploits within-investor variation in a panel regression setting. The results,
reported in Panel B of Table Online IV, indicate that the more investors pay attention to their
investment portfolio and the more they trade, the more evenly they allocate attention. None of the
other covariates related to portfolio characteristics and risk are related to how investors allocate their
attention.
In the third exercise, we study which stocks investors pay attention to whenever they decide to
21
acquire stock-specific information. We follow Hartzmark (2014) and, for every investor, we focus only
on those days when stock specific information is acquired. We then estimate the following regression:
Att Dummyi,j,t = αi + x′i,j,t β + εi,j,t, (3)
where Att Dummyi,j,t is a dummy equal to 1 if investor i researches stock j on date t and is equal
to 0 for all those stocks that are in investor i’s investment opportunity set on date t, but are not
researched.
Because it is not possible for us to know what is the investment opportunity set of every account-
holder at any given point in time, we consider three alternatives. The first includes only the stocks
held in the investment portfolio at time t (Panel A). The second uses all the stocks researched by the
investor over our sample (Panel B). The third includes all the stocks researched by the investor as
well as all the stocks in the S&P 500 index at time t (Panel C).
For each investment opportunity set, we estimate two specifications. The first uses only covariates
motivated by the theoretical models in Van Nieuwerburgh and Veldkamp (2009, 2010), which predict
that investors should pay more attention to the stocks that have a higher squared Sharpe ratio and
market capitalization. Our vector x′i,j,t therefore includes: Squared Sharpe Ratio, the squared realized
Sharpe ratio for each stock in the investment opportunity set, computed over the previous 21 days;14
and Size, the market value of the stock – computed as the product of the price and shares outstanding.
In the second specification, we additionally include Dummy Close, an indicator variable identifying
whether the headquarter of stock j is located within a 250 miles radius from investor i, as in Ivkovic
and Weisbenner (2005).15 In panel A, we also include Portfolio Weight, the weight of stock j in
investor i’s portfolio as of time t. Finally, we include account-holder fixed-effects and we standardize
the regressors to ease the interpretation of their coefficients.
Depending on the definition of investment opportunity set, the coefficients have a different inter-
14Computing the squared Sharpe ratio over the previous 42 or 63 months leaves the results unchanged.15We use the standard formula for computing the distance in miles between account holder i and firm j, as follows:
dist(i, j) = arccos(cos(lati)cos(longi)cos(latj)cos(longj) + cos(lati)sin(longi)cos(latj)sin(longj) + sin(lati)sin(latj)
)r,
where lat/long are the latitudes/longitudes of location i and location j (expressed in radians), respectively, and r denotesthe radius of the Earth (approximately 3,963 miles).
22
pretation. If we take the coefficient on Squared Sharpe Ratio as an example, in Panel A the coefficient
measures the relation between a standard deviation increase in the squared Sharpe ratio of a stock
and the probability that an investor focuses on it rather than on another stock he/she holds in his/her
investment portfolio. In panels B and C the interpretation is similar, but the comparison is with the
stocks that the investor tracks over the sample (Panel B) or all the stocks in the S&P 500 index in
addition to the ones the investor tracks over the sample (Panel C).
In line with the predictions of Van Nieuwerburgh and Veldkamp (2009, 2010), the results – reported
in Table 6 – indicate that investors pay more attention to the stocks that have a higher squared
Sharpe ratio as the coefficients on Squared Sharpe Ratio are positive and significant across all panels
and specifications. Size also has a positive coefficient, but is significant only in Panels B and C and
marginally insignificant (t-statistic of 1.45) in the first specification of Panel A.
The remaining regressors have coefficients that follow intuition. The coefficient on Dummy Close in-
dicates that investors spend more attention researching the stocks that are located nearby. This finding
is important, because it uncovers the potential mechanism associated with the proximity/familiarity
results in Massa and Simonov (2006), Ivkovic and Weisbenner (2005), Huberman (2001), and Grinblatt
and Keloharju (2001). While the literature has shown that investors hold a more than proportional
amount of local stocks in their portfolio, our results suggest that proximity/familiarity operates at the
attention level as well. This means that investors include local stocks in their portfolios because they
pay attention to local stocks, as opposed to the alternative hypothesis that investors pay attention
uniformly across all stocks (near and far) and include in their portfolio only the ones geographically
close, because proximity gives them informational advantages. Finally, the coefficient on Portfolio
Weight suggests that – among the assets they hold – investors pay more attention to the assets that
have a higher weight in their portfolio.
4.3 Which Stocks Do Investors Pay Attention to?
While the results in Section 4.2 show that idiosyncratic factors such as geographic proximity affect
investor attention to specific stocks, the summary statistics in Section 3 provide prima facie evidence
that investor attention is correlated as investors tend to focus on large and well-known companies such
23
as Facebook, Apple, Bank of America, and Ford – among others. In this section we dig deeper and
analyze what stock characteristics are related to investor attention. We consider the sample October
1, 2013 to June 10, 2014 to construct the overall attention paid across all investors to each individual
ticker, and estimate the following baseline specification:
Minutesj = α+ x′j β + εj , for j = 1, . . . , J, (4)
where Minutes is the (log) number of minutes spent researching stock j over the eight months period
and J is the total number of stocks for which we have both attention and conditioning information.
We divide our conditioning variables in three groups. The first group comprises attention proxies:
Analyst, the number of analysts covering the stock; News, the total number of news associated with
the stock;16 Volume, the stock’s volume; and Turnover, computed as the stock’s volume divided by
the shares outstanding.
The second category comprises regressors related to stocks’ risk and performance. Alpha, Beta
Mkt, Beta SMB, Beta HML, and Beta MOM are the intercept and the loadings on the Market, SMB,
HML and MOM factors. The coefficients are estimated for each stock j using daily data over the
sample. Return, Variance, Skewness and Kurtosis, are the realized returns, variance, skewness and
kurtosis for each stock j, computed using daily returns. For the higher moments we follow Amaya
et al. (2015) and use the following expressions:
RV arj =N∑i=1
r2i,j , RSkewj =
√N
∑Ni=1 r
3i,j
RV ar3/2j
, RKurtj =N
∑Ni=1 r
4i,j
RV ar2j
,
where ri,j is the daily return of stock j on day i and N is the number of trading days in the sample.
The third category comprises regressors related to stock characteristics: MktBook, the ratio of
market value of assets and the book value of assets;17 R&D, computed as the ratio of research and
16We obtained the total number of news for each stock from Capital IQ.17The market value of assets is computed as the sum of the closing stock price (COMPUSTAT item: PRCCQ) mul-
tiplied by the common shares (COMPUSTAT item: CSHPRQ), debt in current liability (COMPUSTAT item: DLCQ),long-term debt (COMPUSTAT item: DLTTQ), and preferred stocks (COMPUSTAT item: PSTKQ), minus deferredtaxes and investment tax credit (COMPUSTAT item: TXDITCQ).
24
development expenses (COMPUSTAT item: XRDQ) and sales (COMPUSTAT item: SALEQ);18
Profitability, the ratio between operating income before depreciation (COMPUSTAT item: OIBDPQ)
and total assets (COMPUSTAT item: ATQ); Age, the number of years a company has been in the
CRSP dataset; Size, the market value of assets – computed as the product of the price and shares
outstanding; Leverage, computed as the sum of long-term debt and debt in current liabilities, divided
by the market value of assets; and Inst. Investors, the fraction of the shares outstanding owned by
institutional investors.
As in Section 4.1, we de-mean and standardize all the regressors so that they have unit variance.
The economic interpretation of the coefficients in Table 7 therefore parallels the one of Table 4.
The first column reports the coefficient of each regressor, the second its t-statistic, and the third its
economic magnitude.
The covariates that have been adopted as attention proxies in the literature are the ones that have
the greatest economic impact. For example, starting from a baseline of 45.24 minutes, a standard
deviation increase in Turnover is associated with a 59.07 minutes increase in attention. Large are also
the magnitudes associated with Volume and Analyst. News are significant as well, but their economic
impact is smaller.19
Among the second set of regressors, we note that investors pay more attention to stocks that
have higher exposure to the market and momentum factors. This indicates that investors spend
time researching stocks that have higher systematic risk and have a positive exposure to momentum.
The coefficient on SMB is negative and marginally significant, while the coefficient on HML is not
significant. Once exposures to market factors are controlled for, stock returns, variance, skewness and
kurtosis do not have a significant impact on attention.
In terms of firm characteristics, investors pay attention to stocks that have high growth potentials
– as evidenced by the positive coefficients on MktBook and R&D – and that are well-established
18Following Frank and Goyal (2003), we replace R&D with 0 if the entry related to research and development expenseis missing.
19This is mainly due to the fact that Turnover, Volume, Analyst and News are correlated with each other. As expected,the exclusion of three out of four covariates boosts dramatically the economic and statistical significance of the fourthcovariate. It is nonetheless interesting to note that News has a positive and statistically significant impact – even aftercontrolling for the other three regressors.
25
but relatively risky – as evidenced by the positive coefficient on Size and Leverage. We also find a
very strong and negative relation between attention and Inst. Investors, suggesting that individual
investors pay attention to stocks’ that are different, compared to the ones owned by institutional
investors. Finally, Profitability and Age do not have a significant coefficient.
Table Online V shows the results are robust when we use Pages and Visits as measures of attention.
5 Attention and Performance: Evidence from Portfolio Returns
We now turn to the set of results relating investor attention and portfolio performance. Investors that
spend more time acquiring information are likely to receive more trading signals. If these investors are
able to correctly process these signals, we would expect them to achieve superior portfolio performance.
On the other hand, if they systematically misinterpret these signals, we would expect them to perform
poorly. To estimate what relation holds in the data, we propose two strategies. The first strategy uses
cross-sectional regressions and addresses the question of whether investors that pay more attention
perform better. The second strategy implements panel regressions and uses within-individual variation
over time to assess whether changes in investor attention are associated with performance differentials.
In both cases, we find a positive relation between attention and investment performance.
5.1 Cross-sectional regressions
In Table 8 we evaluate the relation between investor attention and risk-adjusted performance, where
the former is measured as overall attention and the latter is measured as the DGTW abnormal return
of the investor portfolio.20 We estimate the following cross-sectional regression:
AV G DGTW Reti = α+ β Attentioni + x′i γ + εi for i = 1, . . . , N, (5)
where AV G DGTW Reti is the annualized average percentage DGTW abnormal return of investor i
over the sample, Attentioni is the total attention spent on the brokerage account website by account
20The DGTW-adjusted portfolio return is computed as in Equation (1) on Page 7 of Wermers (2003) and in Equation(1) on Page 1,041 of Daniel et al. (1997).
26
holder i over the sample and xi is a subset of the covariates that explain the cross-section of investor
attention – as shown in Section 4.1.21 All regressors are standardized so that they have zero mean
and unit standard deviation.
The results highlight a statistically and economically significant relation between attention and
performance. The coefficient in the first specification equals 1.531% and is significant at the one
percent level. The results imply that a one standard deviation increase in attention is associated with
an increase in the investor’s DGTW abnormal return of 1.53%, which is economically rather large.
However, while the coefficient on attention is positive and significant, the low R-Squared of 0.1%
indicates there is a lot of cross-sectional heterogeneity in performance that attention does not explain.
The corresponding coefficient estimates for research attention (Table Online VI, Panel A) and
balances and positions attention (Table Online VI, Panel B) are 0.988% and 1.267%, respectively. A
common feature of the results reported in this section – as well as the subsequent ones – is that bal-
ances and positions attention has sometimes a stronger impact on performance, compared to research
attention. While surprising at first sight, the results may be driven by the fact that investors obtain
important research information also on the balances and positions portion of the website. The infor-
mation includes stocks’ past returns; historical and forecasted earnings and dividends; key accounting
information such as price-to-book and price-to-earnings ratios; basic facts such as revenues and in-
stitutional ownership; and a summary of analysts’ opinions as well as recent news. Research pages,
on the other hand, contain in-depth analyst reports and balance sheet information for each company.
Also, because investors spend almost double the time on balances and positions pages compared to
research pages – as shown in Table 2 – it may be that the amount of time spent on the balances
and positions pages is a more precise estimate of the process of information acquisition, particularly
for those investors that acquire information also through other sources such as Yahoo Finance or
Bloomberg, for example.
In order to control for possible differences in information acquisition capabilities, Specification
2 in each panel includes demographic characteristics as control variables. Adding these regressors
21Compared to the covariates in Table 4 of Section 4.1, we do not include the regressors related to risk-factor exposures,that is, the loadings on the market, Small-Minus-Big, High-Minus-Low and Momentum factors, because these exposuresare directly controlled for by the DGTW procedure.
27
leaves the effect of attention on investment performance largely unchanged for all attention measures,
indicating that the relation between these additional regressors and performance is largely orthogonal
to that of attention. Among the newly added covariates, the ones significant at the 5% level are age
and the brokerage dummy. The first has a negative coefficient – indicating that younger investors
achieve a better risk-adjusted performance on their investment portfolio. The second shows instead
that brokerage accounts seem to outperform IRA and other accounts, on average, possibly because
investors in these accounts are actively seeking underpriced stocks rather than holding broad and
diversified passive portfolios.
Finally, Specification 3 adds – as further controls – regressors related to portfolio size, portfolio
allocation, and trading activity. The inclusion of these additional covariates strengthens the effect of
attention on performance by 35-50%, depending on the specification: the overall attention coefficient
increases to 2.326%, the research attention coefficient increases to 1.776% and the balances and po-
sitions coefficient increases to 2.007%. All coefficients remain statistically significant at the 1% level,
indicating that, even after controlling for portfolio composition and trading frequency, the attention
is significantly related to the performance of the investors in our sample.
As for the effect of the additional controls, we find a negative relation between portfolio size and
number of trades and performance. Interestingly, we find that including the number of trades as
a control variable increases the effect of attention on performance, suggesting that paying attention
attenuates the negative effects of excessive trading documented in the literature, see Barber and Odean
(2000).
Table 8 reports results in terms of annualized abnormal percentage returns. To compute the dollar
gains per hour spent on the platform for the average investor, we combine the information contained
in Tables 3 and 8. A one standard deviation increase in attention is associated with an increase in
abnormal returns of 1.53%, 1.60% or 2.33%, depending on the specification. From Table 3, we know
that the average portfolio size in our dataset is $94, 000. The standard deviation of Overall attention is
equal to 88 hours. Using this information, we can compute the dollar gains associated with each hour
of attention. For the first specification, this equals: $94, 000× 1.53%/88 = $16.3. For specifications 2
and 3, the dollar gains per hour equal $94, 000 × 1.60%/88 = $17 and $94, 000 × 2.33%/88 = $24.9.
28
The dollar gains for each hour of attention therefore range from $16 to $25. The economic significance
of these quantities depends on the opportunity cost of each individual investor. The average American
has an annual income of $41, 000. This converts to $41, 000/52/40 = $19.7 per hour. The average
American that invests in the stock market, however, has an annual income of $52, 000, which translates
to $52, 000/52/40 = $25 per hour.22
As a first robustness exercise, we implement the calendar-time approach of Barber and Odean
(2000). We first sort the investors into five groups according to attention (from Q1 to Q5). We then
compute weekly value-weighted and equally-weighted returns for each group and estimate Carhart
4-factor regressions on these portfolios. We report the results in Table 9. Going from low- to high-
attention and focusing on the equally-weighted returns reported in Panel A, the five groups have
the following annualized alpha estimates 0.03, 0.02, 0.02, 0.03, and 0.06, with t-statistics equal to
0.96, 0.82, 0.96, 1.15, and 2.02. The alphas increase in significance as we move from the second
to the fifth portfolio and only the top portfolio is significant at the 5% level, confirming that risk-
adjusted performance is positively related to attention. Finally, we form a portfolio long in the most
attentive investors (Q5) and short in the least attentive investors (Q1) and find it has an alpha of
0.04, statistically significant at the 5% level. The value-weighted returns in Panel B are aligned with
the equally-weighted returns. The top two portfolios have significant outperformance – at the 10%
significance for Q4 and the 5% significance for Q5. The long-short portfolio in the last column also
displays an alpha of 0.04, statistically significant at the 1% level.
As a second robustness exercise, Table Online VII shows the results that use number of pages
(Panel A) or logins (Panel B) as measures of attention. While the results are qualitatively similar,
they are economically smaller, suggesting that – compared to seconds – pages and logins may be
poorer proxies for the process of information acquisition.
As a third robustness exercise, we use the full sample Sharpe ratio as a measure of risk-adjusted
performance. The results, reported in Table Online VIII, highlight a statistically and economically
significant relation between attention and performance. For overall attention, the coefficient in the
first specification equals 0.090 and is significant at the one percent level. The corresponding coefficient
22The annual income figures are taken from Khorunzhina (2013), and use 2007 PSID data.
29
estimates for research and balances and positions are 0.108 and 0.137, respectively. The results imply
– taking balances and position attention as an example – that a one standard deviation increase in
attention is associated with an increase in the investor’s Sharpe ratio of 0.137. Economically, this
quantity is quite large, because the average Sharpe ratio across the investors in our sample is 1.316.23
The results reported so far show that investor attention is positively related to risk-adjusted perfor-
mance, even after controlling for investor characteristics and investment style. As we detail in Section
4.1 and the associated Table 4, account holders with higher exposure to small capitalization stocks,
growth stocks, momentum stocks, and the overall market, are more attentive. Jointly, therefore, the
results reported in Table 4 and Table 8 show that attention has an indirect and a direct effect on in-
vestor performance. The indirect effect is the one related to the type of investment decisions attentive
investors make compared to inattentive investors. The direct effect, on the other hand, is the effect of
attention on performance, controlling for style and other characteristics. Taking the relation between
momentum exposure, attention and risk-adjusted performance as an example, our results in Table
4 uncover the indirect effect of attention on performance, in that more attentive investors tend to
follow momentum strategies – buying (selling) stocks that have increased (decreased) in price over the
previous 12 months. The results in Table 8, on the other hand, uncover the direct effect of attention
on performance, in that attention retains a positive effect on performance even after controlling for
momentum using the DGTW procedure.
5.2 Panel data regressions
The results reported so far potentially suffer from a reverse-causality problem. It may not be that
investors perform better because they pay more attention, but that investors pay more attention
because they have been performing better.
The cross-sectional regressions cannot distinguish between these two alternative hypotheses, be-
cause they compute performance and attention over the full sample. To separate the two effects, we
next estimate panel regressions that identify the effect of attention on performance from time-series
23While this average Sharpe ratio may seem rather large, note that – for the same period – the market Sharpe ratio(computed using returns and volatility on a NYSE/AMEX/NASDAQ value-weighted index) has been 2.62.
30
variations in attention – measured at the individual level. Our baseline specification is:
DGTW Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k, (6)
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time
interval t : t+k and k equals 21, 42 and 63 days – depending on the specification;24,25 Abn Attentioni,t
is account-holder i abnormal attention at time t, computed as the difference between the (log) attention
on day t and the (log) average attention over the previous 21 business days; finally, αi and βm represent
account-holder fixed effects and monthly time-effects, respectively.
Following Da, Engelberg, and Gao (2011), we use abnormal attention – rather than attention – to
capture time trends and other low frequency seasonalities in investors’ attention. This is important,
because investors tend to pay more attention to their investment portfolios right after opening their
accounts and they tend to lose interest subsequently. Furthermore, even though time-effects are
included in the regressions, they are unlikely to capture a large part of the time-variation in investors’
attention, because individual investors hold relatively few stocks. It may therefore be that certain
investors pay a lot of attention in certain periods – because the stocks they own (or they want to
purchase) are in the news – and other investors in other periods. Finally, to ease the comparison
across the specifications, we standardize Abn Attention so that it has zero mean and unit standard
deviation.
To guarantee that our results are not specific to the type of clustering used in the computation of
the standard errors, we report two sets of t-statistics. The first – in round brackets – are computed
using standard errors that are double-clustered by account-holder and time, see Petersen (2009). The
second – in square brackets – are computed using Driscoll and Kraay (1998) standard errors. We
base our discussion on the double-clustered standard errors, because they result in more conservative
estimates in our data.
Reported in Table 10 are the results based on the overall attention measure. Across all horizons, the
24The DGTW-adjusted portfolio return is computed as in Equation (1) on Page 7 of Wermers (2003) and in Equation(1) on Page 1,041 of Daniel et al. (1997).
25The results are economically similar if we use simple returns or market-adjusted returns.
31
estimates highlight a positive and significant relation between abnormal attention and performance.
Statistically, the results are significant at the 5% level for the one month horizon and the 1% level
at the two- and three-month horizons. Economically, the effect of overall abnormal attention is quite
large. At the one-month horizon, we find that a one standard deviation increase in attention is
associated with an annualized increase in portfolio DGTW-adjusted performance of 0.032 ·12 = 0.38%
per year. The effect increases to 0.103 · 6 = 0.62% at the two-month horizon and decreases slightly to
0.138 · 4 = 0.55% at the three-month horizon. This indicates that paying attention allows individual
investors to improve their portfolio performance in the short-term, suggesting that the improvement
in performance hinges on attentive investors being able to purchase (sell) stocks that – in the short
run – realize relatively large positive (negative) returns.
Panels A and B of Table Online IX report the results for research attention and balances and
positions attention, respectively. The results show that, while the effect of balances and positions
attention is economically stronger than overall attention, the effect of research attention is somewhat
weaker. For example, the annualized effect for the two-month horizon specification equals 0.135 · 6 =
0.81% for balances and positions attention and 0.078 · 6 = 0.47% for research attention. In both cases
the coefficients are statistically significant at the 5% level. Across all panels, the coefficients are always
more significant if we use Driscoll and Kraay (1998) standard errors.
The baseline results reported in Table 10 use a window of 21 days to compute abnormal attention.
For robustness, in Tables Online X and Online XI we compute abnormal attention using windows of
42 and 63 days, respectively. In all cases, the results become stronger if we use a longer window to
compute abnormal attention.
The results reported so far use number of seconds as a measure of attention. Table Online XII
shows that the results reported above are generally less significant when we measure attention using
number of pages or logins instead of seconds, suggesting – again – that the latter two measures may
be poorer proxies for the process of information acquisition.
The results in Table 10 do not account for trading fees. Barber and Odean (2000) show that
transaction costs can substantially reduce the post-fee returns of individual investors, as the average
total cost of a round-trip trade was approximately 4% in 1999. Thanks to technological progress
32
and increased competition among brokers, trading fees are much lower nowadays: it is possible to
trade many ETFs without incurring any fee and the majority of the brokers charge less than $10
for on-line trades – a very small amount compared to the average trade size of $16,000 reported in
Table 3. Furthermore, brokers often offer their customers promotions whereby trading fees are waived
for a certain number of trades and/or a certain period of time. Because we do not observe these
promotions in our dataset, we assume that every trade is charged the maximum fee advertised by
our broker and re-estimate Equation (6) using abnormal DGTW returns adjusted for trading fees.
Compared to Tables 10 and Online IX, the coefficients in Table Online XIII are only slightly smaller,
but still strongly significant, indicating that controlling for trading fees does not affect our results.
Finally, while the results in Table 10 use monthly time-effects, we find that the results are virtually
unchanged when we use daily time-effects – as we show in Table Online XIV.
6 Attention and Performance: Evidence from Investor Trades
In an effort to understand the mechanism through which higher investor attention relates to superior
portfolio performance, in this section we analyze how attention affects the profitability of investors’
active management decisions, that is, their trades. We first relate the performance of the stocks bought
and sold by individual investors to the attention they pay to their investment portfolio in the month
preceding each trade. We take each individual’s trade as the unit of observation and we measure
performance as the risk-adjusted returns of the stock over the one-, two- all the way to twelve-month
period after the trade has been placed. The adjustment for risk is performed using the DGTW model,
see Daniel et al. (1997), but the results are similar if we use simple or market-adjusted returns.
We then provide results showing that the positive effect of attention is – at least in part – due to
the fact that high-attention individuals purchase attention-grabbing stocks that have appreciated in
the recent past and whose positive performance persists for up to six months.
33
6.1 Baseline parametric results
We start by presenting pooled regression estimates in Table 11, testing whether there is a positive
relation between investor attention and stock performance. We separate the buys and the sells, and
estimate the regressions:
DGTW Ret Buysi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k, (7)
DGTW Ret Sellsi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k, (8)
where DGTW Ret Buysi,j,t:t+k (DGTW Ret Sellsi,j,t:t+k) are the cumulative abnormal returns of
security j bought (sold) by investor i over the time interval t : t+k, computed using the DGTW model;
Attentioni,j,t is the (log) number of seconds spent on the brokerage account website by investor i over
the month preceding the trade in stock j that occurs at time t. Note that – to make the results more
interpretable – we scale Attentioni,j,t so that it has zero mean and unit variance. Finally, cumulative
abnormal returns are computed at the one-, two-, three-, four-, five-, six- and twelve-month horizons.
Panel A focuses on the returns of the buys. We find a positive relation (statistically significant at
the 5% level) between the total attention investors spend on the trading platform and the performance
of their trades at the two-, three-, four-, and five-month horizons. The relation is also positive, but only
statistically significant at the 10% level, at the one month horizon. The results are also economically
significant. At the three-month horizon, for example, a unit-standard deviation increase in attention
increases the average annualized adjusted returns of the stocks purchased by 0.533% · 4 = 2.13%.
The results for the two specialized measures of attention are very much in line with the overall
attention results. For example, at the three-month horizon the effect equals 0.489% · 4 = 1.96%
and 0.408% · 4 = 1.63% for research and balances and positions attention, respectively. The results
for research attention and balances and positions are reported in Table Online XV. Compared to
overall attention, research attention seems to have a stronger effect at the one-month horizon and a
slightly smaller effect at the five-month horizon. Finally, the results for balances and positions are
somewhat weaker, as the results are significant at the two- and three-month horizons, but they just
miss significance at the four- and five-month horizons.
34
Panel B reports the performance of the sells. For overall attention, we find that none of the
coefficients is significant at the 5% level, and only one coefficient is significant at the 10% level –
indicating that there is virtually no relation between overall attention and the performance of the
stocks sold by investors. Once again, the results for research attention and balances and positions
attention (Table Online XV) are very much aligned. The only exception is the relation between
research attention and returns at the one- and two-months horizons, which is positive and significant
at the 5% level.
Overall, the results in Table 11 show that there is a positive effect of attention on the stocks
purchased, but not on the stocks sold. This suggests both that the time spent on the brokerage
account website is dedicated to searching for new stocks and that the investors are able to translate
the information they acquire into profitable trades.26
6.2 Non-parametric results
To corroborate the parametric results reported above, we now report evidence based on portfolio
sorts. Before analyzing the effect of attention on trading performance, we first confirm the findings
in Odean (1999) that – unconditionally – the stocks purchased by individual investors systematically
underperform the ones sold. Panel A.I. of Table 12 reports the average cumulative abnormal returns at
the one-, two-, three-, four-, five-, six- and twelve-month horizons after stock purchases and sales. For
every horizon, we also report the difference in performance between the two groups and a bootstrap
p-value testing whether the difference is equal to zero.27 We find that, at every horizon, the stocks
sold outperform the ones purchased and the difference is statistically significant for three out of seven
horizons. For example, the average annualized three-month abnormal return of the stocks purchased
equals 0.37% · 4 = 1.48%, the one for the stocks sold equals 0.91% · 4 = 3.64%, and their difference
is statistically significant with a p-value of 0.01. While in line with Odean (1999), our results are
somewhat weaker, possibly due to our sample covering the 2013-2014 bull market.
26Table Online XVI shows that the results are similar, but less significant, when we measure attention using number ofpages or logins instead of seconds. Table Online XVII shows that the results are economically and statistically strongerwhen we use market-adjusted, rather than DGTW-adjusted, abnormal returns.
27All p-values in this section are computed using the bootstrap procedure suggested by Barber, Lyon, and Tsai (1999)with 10,000 bootstrap iterations.
35
Panel A.II. conditions the trades on the overall attention paid by investors to their investment
portfolio in the month preceding each trade. We divide the buys in two groups – low and high – and
report their performance. We do the same for the sells. The results indicate that overall attention
is strongly related to the performance of investor trades. For low-attention trades, we find that the
Odean (1999)’s bias is very strong. Low-attention buys strongly underperform the low-attention sells
in a statistically significant fashion: the average annualized three-month abnormal return of the stocks
purchased equals −0.07% ·4 = −0.28%, the one for the stocks sold equals 0.77% ·4 = 3.08%, and their
difference is statistically significant, with a p-value of 0.00. On the other hand, even though the bias is
not overturned for high-attention trades, high-attention buys do not underperform the high-attention
sells in a statistically significant fashion: the average annualized three-month abnormal return of
the stocks purchased equals 0.79% · 4 = 3.16%, the one for the stocks sold equals 1.05% · 4 = 4.2%
and their difference is statistically insignificant, with a p-value of 0.29. While these results indicate
that the behavioral bias first discovered by Odean (1999) is smaller for high attention trades, it is
difficult to relate this behavioral bias to portfolio performance. This is because the returns that affect
the performance of the investors are the ones of the stocks they hold in their portfolio. Even if the
performance of the stock an investor sells is superior to the one he/she buys, it is not clear that he/she
will underperform the market as a result. This is simply because both stocks may be delivering a return
superior to that of the market in the subsequent period. To assess the relation between attention and
performance, it is preferable to focus on overall portfolio performance, as we do in Tables 8, 9, and
10.
We also find that the high-attention buys strongly outperform the low-attention buys up to four
months and the difference becomes insignificant thereafter. Economically, the difference in perfor-
mance is large: the average annualized three-month difference between high- and low-attention pur-
chases equals 3.16%−(−0.28%) = 3.44% and is statistically significant, with a p-value of 0.00. Finally,
we find no significant impact of overall attention on the performance of the stocks sold as low- and
high-attention sells have similar performance.
The results reported above show that, the more time investors spend on their trading account,
the higher the performance of the stocks they buy. In an effort to uncover the main drivers of the
36
results, we show below that the more investors pay attention to their brokerage account the more they
purchase attention-grabbing stocks that have appreciated in the recent past and continue to appreciate
for up to six months, and this seems to be the source of their outperformance.
6.3 Understanding the results
While on the brokerage account website, investors may be looking at a wide range of information, such
as firms’ accounting data, news, and analysts’ reports – for example. One obvious piece of information
that may attract investors’ attention is the past performance of the securities they are considering
buying or selling. Moskowitz, Ooi, and Pedersen (2012) document time-series momentum, that is,
the positive predictability from a security’s own past returns that persists for up to 12 months. To
test whether time-series momentum can explain our results, we conduct two complementary exercises.
First, we estimate the relation between investors’ attention and the past performance of the stocks they
trade using a non-parametric symmetric nearest neighbor approach. Second, we repeat the analysis
contained in Panel A of Table 12, but focus on the performance of the stocks before the trades occur,
rather than after. The results of this analysis are reported in Panel B of Table 12.
Starting from the unconditional results, Panel B.I. shows that investors tend to both buy and sell
stocks that have appreciated in the past, as shown in Odean (1999). Panel B.II. sharpens the results,
as it uncovers a very strong relation between the attention spent by investors and the risk-adjusted
returns of the stocks they trade. For example, at the one-month horizon, Panel B.II of Table 12 shows
that the average past performance of the high attention buys (sells) equals 3.13% (3.55%), the past
performance of the low attention buys (sells) equals 0.70% (1.19%), and the difference in performance
between the high and low attention buys (sells) is statistically significant, with a p-value equal to
0.00 (0.00). For the same one-month horizon, Panel A of Figure 3 displays a very strong positive and
virtually monotonic relation between log attention and past performance.28 The remaining panels of
Figure 3 and the remaining columns of Table 12, Panel B.II., show that the pattern is very robust
across horizons.
28These plots exclude the top 5 percentiles of log-attention, because the data is more sparse in that region of thesupport, and the nearest neighbor estimates more erratic.
37
Past returns are just one attention-grabbing proxy. Another proxy proposed in the literature is
abnormal volume, see Gervais, Kaniel, and Mingelgrin (2001). Gervais, Kaniel, and Mingelgrin (2001)
find evidence that stocks experiencing unusually high volume tend to appreciate for up to 5-6 months.
Finally, we consider the number of news. Past returns, abnormal volume, and number of news are all
attention-grabbing proxies, but have slightly different flavors. The first relates to the past returns of
the stock. The second relates to how much the stock is traded. The third relates to how much the
company is covered in the news. To assess which one better explains our attention results, we first
estimate univariate regressions of the attention spent before any given trade on each proxy. We then
include all proxies together and conduct a horse-race between them.
We report the results in Table Online XVIII separately for Buys (Panel A) and Sells (Panel B). To
ease the comparison among the estimated coefficients, we standardize the regressors so that they have
zero mean and unit variance. The results in Panel A show that, the higher the attention of a stock
purchase, the more it relates to an attention-grabbing stock. The coefficient is positive and significant
for all proxies. When we include all proxies in the same specification, we find that the past abnormal
returns is the most significant regressor with a t-statistic of 4.53. The coefficient on abnormal volume
has a t-statistic of 2.32 and a coefficient that is half the magnitude. Finally, past news is negative
and barely significant. The results in Panel B are similar, but abnormal returns and abnormal volume
have a similar relation to the attention associated with the stocks sold.
To test whether the relation between attention and performance remains significant after the
inclusion of attention-grabbing proxies, we re-estimate the results in Table 11 including abnormal
volume and news as additional controls. We find that the coefficient estimates and their significance
remain unchanged.
These results – in conjunction with the ones contained in Table 11 and Panel A of Table 12 –
paint the following picture. Investors that pay a lot of attention to their trading accounts tend to
trade attention-grabbing stocks that have appreciated greatly in the past. The good past performance
persists into the future for the purchases, as we find that the high-attention buys significantly outper-
form the low-attention buys up to four to five months. The outperformance is relatively short-lived,
however, as we find it disappears for horizons greater than five months.
38
7 Extensions and Additional Analyses
In the previous sections we showed that higher attention is related to superior performance. We now
provide a set of additional analyses to further understand our main findings. We start by providing
extensions of the cross-sectional performance regressions and show a positive relation between special-
ization in information acquisition and investment performance. We continue by providing extensions
of the baseline panel results and show that the positive relation between attention and performance
is mainly driven by the positive effect of attention for those investors that, in general, are rather inat-
tentive. Finally, we provide extensions of the baseline trading results and show that paying attention
is particularly profitable when trading stocks that have high market capitalization, trading volume,
volatility, number of analysts, dispersion of analyst forecasts, and news.
7.1 Specialization in information acquisition and investment performance
The baseline cross-sectional performance regressions show there is a positive relation between atten-
tion and investment performance. In this section, we dig deeper and assess whether the degree of
specialization in information acquisition also has an impact on investment performance.
We focus only on those individuals that acquire stock specific information over the sample and we
measure the degree of specialization using the Herfindahl index of the stocks researched. For investor
i, this quantity is computed as∑Kk=1 ω
2k−1/K
1−1/K if K > 1 and 1 if K = 1, where ωk is the fraction of the
total stock-specific attention of investor i – measured in seconds – allocated to stock k over the full
sample and K is the total number of stocks researched. The measure is bounded between 0 – attention
allocated evenly across all assets – and 1 – all attention concentrated in one asset. We then add this
regressor to the baseline specification in Equation (5) and estimate:
AV G DGTW Reti = α+β1 Attentioni
+β2 Herfindahl Researchedi + x′i γ + εi for i = 1, . . . , N, (9)
whereHerfindahl Researchedi is the normalized Herfindahl index of the stocks researched by investor
39
i over the sample and the remaining regressors are the same as in Equation (5).
The regression estimates, reported in Table 13, suggest two main findings. First, even when we
estimate our results on the subset of investors that collect stock-specific information on the brokerage
account website, the effect of the covariate Attention is positive and significant. This is true if we use
overall, research or balances and positions attention.
Second, the more investors specialize their attention, i.e., the higher the Herfindahl Index of the
stocks researched, the superior the performance. The results highlight that, for a given level of
attention, those investors that specialize and focus their research efforts on specific stocks perform
better, compared to those that evenly allocate their attention across the stocks in their information
set. Economically, the magnitudes are large. For overall attention (Panel A), the β2 coefficient equals
1.861% and is significant at the one percent level. The corresponding coefficient estimates for research
(Panel B) and balances and positions (Panel C) are 2.231% and 1.639%, respectively.
As a final exercise, we estimate the effect of paying attention for those that display more and
less specialization in information acquisition. To this end, we divide the investors for which we have
stock-specific attention in two groups on the basis of the Herfindahl index of the stocks they research
and re-estimate Specification 5 separately for the two groups. The coefficient of interest β equals 2.40
for the less specialized individuals and 3.92 for the more specialized individuals. If we repeat the
dollar gains per hour calculations associated with Table 8, we find that the effects are economically
quite large. For the low-specialization investors, they equal $94, 000 × 2.40%/88 = $25.6. For the
high-specialization investors, they equal $94, 000× 3.92%/88 = $41.66.29
7.2 Panel portfolio results by investor type
To assess whether the effects estimated in Table 10 are driven by those investors that – on average
– pay a lot of attention or by those that rarely pay attention, we divide our account-holders in five
quintiles based on their overall total attention over the sample and – for each group – we re-estimate
Equation (6). Table 14 presents the results for the 21, 42 and 63 days horizons in panels A, B, and
29The fact that both the coefficients on low- and high- specialization individuals are higher than the correspondingcoefficient of 2.33 reported in Table 8 is an indication that investors that look at stock-specific information benefit morefrom paying attention compared to those that don’t.
40
C, respectively.
Moving from left to right, the results for columns Q1 through Q5 report estimates for account-
holders with greater degrees of overall attention. For the shortest horizon of 21 days (Panel A),
only the investors with low to medium attention (Q1 through Q3) seem to benefit from paying more
attention. The effect is instead not statistically significant for high attention investors (Q4 and Q5).
Furthermore, the coefficients are monotonically decreasing as we move from low attention to high
attention individuals. The results in panels B and C display similar patterns. The coefficients on
abnormal attention decrease as average attention increases and become negative for the investors that
pay the most attention. Also, the coefficients for the first four groups are all consistently significant
at the 5% level. For the last group, on the other hand, we never find a significant effect of attention
on performance.
Taken together, the results in Table 14 uncover diminishing returns to abnormal attention. Ab-
normal attention for those investors that rarely pay attention to their account seems to improve their
portfolio allocation. On the other hand, those that tend to spend a lot of time on their brokerage
account do not seem to benefit from additional time researching stocks and worrying about their
portfolio positions.
7.3 Investor trades results controlling for stock characteristics
To sharpen the results in Table 11 and understand under what circumstances it pays the most to
pay attention, we condition the performance of the trades on the characteristics of the stocks traded.
We use the volume of the stock on the day of the trade, its volatility, and the disagreement between
analysts’ earnings-per-share forecasts as alternative measures of valuation uncertainty. We use the
market capitalization of the company, the number of analysts, and the number of news as proxies
for the amount of public information available when placing a trade. For each conditioning variable
we divide the trades in two groups: the low group contains the trades associated with firms whose
characteristic is below the median; the high group contains the trades associated with firms whose
characteristic is above the median. We then re-estimate Equations (7) and (8) separately for each
group.
41
The results, displayed in Table 15, are based on the 3-month horizon and are reported separately
for stock purchases (Panel A) and sales (Panel B).30 A clear pattern emerges from Panel A. Attention
is associated with higher future returns when account holders trade companies with high valuation
uncertainty, that is, stocks with high volume, volatility, and analysts’ disagreement. The same is
true for companies with larger amount of public information, that is, stocks with greater market
capitalization, number of analysts, and number of news.
Statistically, all coefficients in the high group are significant at the 1% level, while only one coeffi-
cient is significant in the low group – the one associated with volatility. The two groups also differ in
terms of economic significance. For example, a one standard-deviation increase in attention leads to
a 0.764% · 4 = 3.06% increase in future adjusted returns for stocks with high market capitalization,
while for stocks with low market capitalization the increase is 0.261% · 4 = 1.04%.
While weaker than the results for the buys, attention seems to be related to the performance of
the sells for stocks with high uncertainty and high levels of public information. For two conditioning
variables out of five – market capitalization and number of analysts – the coefficients are significant at
the 1% level for the high group. Furthermore, for high analysts’ disagreement and news, the coefficients
on attention are significant at the 5% and 10% level, respectively. The fact that higher attention is
related to greater future performance after stock sales, combined with the known fact that investors
rarely realize losses, possibly suggests that – in situations of high uncertainty – paying high attention
leads investors to sell their stocks too early, while they are still appreciating. None of the coefficients
for the low group are significant at the 1% level and only one coefficient – the one for volatility – is
significant at the 10% level.
8 Conclusions
We use a novel brokerage account dataset to study how individual investors pay attention to their
investment portfolio and how investor attention relates to investment performance.
We find a very large heterogeneity in attention across investors that can be explained by the size and
30The results are qualitatively the same when we use 1-month to 5-month return horizons.
42
risk of the investors’ portfolios, as well as by investor trading habits and demographic characteristics.
We also uncover a large heterogeneity in the degree of specialization in information choice across
investors. On average, investors research 31 individual stocks per year. The number of stocks they
research is positively related to the overall amount of time they spend on the website, the number of
stocks in their portfolio, and the number of stocks they trade. We also find that investors that hold
fewer stocks, have more concentrated portfolios, trade less and have a smaller fraction of their portfolio
in cash allocate their attention in a more concentrated fashion. Finally, we find that investors pay
more attention to local stocks, stocks with a higher weight in their portfolio, stocks that have a higher
squared Sharpe ratio, and stocks of companies with a larger market capitalization – as predicted by
Van Nieuwerburgh and Veldkamp (2009, 2010).
We study the relation between investor attention and investment performance both at the portfolio
returns level and at the individual trades level. We find a strong and positive cross-sectional relation
between attention and performance, in that more attentive investors achieve higher portfolio risk-
adjusted returns and Sharpe ratios, even after controlling for covariates related to investment style.
Using panel regressions that control for investor skills using fixed effects, we also show that periods of
higher attention are related to superior future portfolio performance. We find similar results when we
focus on the performance of individual trades. Attention is positively related to the future performance
of the stocks purchased up to four months after the trade is placed. We find – on the other hand – no
discernible effect of attention on the performance of the stocks sold.
To understand the economic mechanism relating attention and trading profitability, we conduct
a number of auxiliary exercises. First, by analyzing the characteristics of the stocks before they are
traded, we show that the superior performance of high-attention investors arises because they purchase
attention-grabbing stocks that have appreciated in the recent past and whose positive performance
persists for up to six months. Second, we show that attention is particularly profitable when investors
trade stocks with high market capitalization, trading volume, volatility, number of analysts, dispersion
of analyst forecasts, and news – indicating that it is for the stocks with high uncertainty, but for which
a lot of public information is available, that it pays to pay attention.
Our study contributes to the literature that studies the behavior of individual investors with
43
information capacity constraints, in that we provide empirical evidence on how investors allocate their
attention. We also contribute to the literature that studies the performance of individual investors.
While the majority of the literature focuses on investor mistakes, we show that paying attention relates
to better investment decisions.
44
References
Abel, A. B., J. C. Eberly, and S. Panageas. 2007. Optimal inattention to the stock market. American
Economic Review 92:244–9.
———. 2013. Optimal inattention to the stock market with information costs and transactions costs.
Econometrica 81:1455–81.
Agnew, J., P. Balduzzi, and A. Sunden. 2003. Portfolio choice and trading in a large 401(k) plan.
American Economic Review 93:193–215.
Alvarez, F., L. Guiso, and F. Lippi. 2012. Durable consumption and asset management with transac-
tion and observation costs. American Economic Review 102:2272–300.
Amaya, D., P. Christoffersen, K. Jacobs, and A. Vasquez. 2015. Does realized skewness predict the
cross-section of equity returns? Working Paper .
Andrei, D., and M. Hasler. 2015. Investor attention and stock market volatility. Review of Financial
Studies 28:34–72.
Barber, B. M., J. D. Lyon, and C.-L. Tsai. 1999. Holding size while improving power in tests of
long-run abnormal stock returns. Journal of Finance 54:165–2020.
Barber, B. M., and T. Odean. 2001. Boys will be boys: Gender, overconfidence, and common stock
investment. Quarterly Journal of Economics 116:261–92.
———. 2002. Online investors: Do the slow die first? Review of Financial Studies 15:455–87.
Barber, M. B., and T. Odean. 2007. All that glitters: The effect of attention and news on the buying
behavior of individual and institutional investors. Review of Financial Studies 21:785–818.
———. 2000. Trading is hazardous to your wealth: The common stock investment performance of
individual investors. Journal of Finance LV:773–806.
Calvet, L. E., J. Y. Campbell, and P. Sodini. 2007. Down or out: Assessing the welfare costs of
household investment mistakes. Journal of Political Economy 115:707–47.
45
Da, Z., J. Engelberg, and P. Gao. 2011. In search of attention. Journal of Finance LXVI:1461–99.
Daniel, K., M. Grinblatt, S. Titman, and R. Wermers. 1997. Measuring mutual fund performance
with characteristic-based benchmarks. Journal of Finance LII:1035–58.
Driscoll, J. C., and A. C. Kraay. 1998. Consistent covariance matrix estimation with spatially depen-
dent panel data. Review of economics and statistics 80:549–60.
Duan, N. 1983. Smearing estimate: a nonparametric retransformation method. Journal of the Amer-
ican Statistical Association 78:605–10.
Frank, M. Z., and V. K. Goyal. 2003. Testing the pecking order theory of capital structure. Journal
of Financial Economics 67:217–48.
Frydman, C., S. M. Hartzmark, and D. H. Solomon. 2017. Rolling mental accounts. Review of
Financial Studies Forthcoming.
Gabaix, X., and D. Laibson. 2002. The 6d bias and the equity-premium puzzle. In NBER Macroeco-
nomics Annual 2001, Volume 16, 257–330. MIT Press.
Gervais, S., R. Kaniel, and D. H. Mingelgrin. 2001. The high-volume return premium. Journal of
Finance LVI:877–919.
Grinblatt, M., and M. Keloharju. 2001. How distance, language, and culture influence stockholdings
and trades. The Journal of Finance 56:1053–73.
Grinblatt, M., M. Keloharju, and J. Linnainmaa. 2012. IQ, trading behavior, and performance. Journal
of Financial Economics 104:339–92.
Guiso, L., and T. Jappelli. 2006. Information acquisition and portfolio performance. Working Paper .
Hartzmark, S. M. 2014. The worst, the best, ignoring all the rest: The rank effect and trading
behavior. The Review of Financial Studies 28:1024–59.
Huang, L., and H. Liu. 2007. Rational inattention and portfolio selection. Journal of Finance 62:1999–
2040.
46
Huberman, G. 2001. Familiarity breeds investment. Review of financial Studies 14:659–80.
Ivkovic, Z., C. Sialm, and S. Weisbenner. 2008. Portfolio concentration and the performance of
individual investors. Journal of Financial and Quantitative Analysis 43:613–56.
Ivkovic, Z., and S. Weisbenner. 2005. Local does as local is: Information content of the geography of
individual investors’ common stock investments. Journal of Finance 60:267–306.
Kacperczyk, M., S. Van Nieuwerburgh, and L. Veldkamp. 2016. A rational theory of mutual funds’
attention allocation. Econometrica 84:571–626.
Karlsson, N., G. Loewenstein, and D. Seppi. 2009. The ostrich effect: Selective attention to informa-
tion. Journal of Risk and Uncertainty 38:95–115.
Khorunzhina, N. 2013. Structural estimation of stock market participation costs. Journal of Economic
Dynamics and Control 37:2928–42.
Korniotis, G. M., and A. Kumar. 2011. Do portfolio distortions reflect superior information or psy-
chological biases? Review of Economics and Statistics 93:244–65.
———. 2013. Do portfolio distortions reflect superior information or psychological biases? Journal
of Financial and Quantitative Analysis 48:1–45.
Li, J., and J. Yu. 2012. Investor attention, psychological anchors, and stock return predictability.
Journal of Financial Economics 104:401–19.
Madrian, B. C., and D. F. Shea. 2001. The power of suggestion: Inertia in 401(k) participation and
savings behavior. Quarterly Journal of Economics 116:1149–87.
Massa, M., and A. Simonov. 2006. Hedging, familiarity and portfolio choice. Review of Financial
Studies 19:633–85.
Moskowitz, T. J., H. Y. Ooi, and L. Pedersen. 2012. Time series momentum. Journal of Financial
Economics 228–50.
47
Nicolosi, G., L. Peng, and N. Zhu. 2009. Do individual investors learn from their trading experience?
Journal of Financial Markets 317–38.
Odean, T. 1998a. Are investors reluctant to realize their losses? The Journal of Finance LIII:1775–98.
———. 1999. Do investors trade too much? American Economic Review 89:1279–98.
———. 1998b. Volume, volatility, price, and profit when all traders are above average. The Journal
of Finance 53:1887–934.
Peng, L., and W. Xiong. 2006. Investor attention, overconfidence and category learning. Journal of
Financial Economics 80:563–602.
Peress, J. 2004. Wealth, information acquisition, and portfolio choice. Review of Financial Studies
17:879–914.
Petersen, M. A. 2009. Estimating standard errors in finance panel data sets: Comparing approaches.
Review of financial studies 22:435–80.
Seasholes, M. S., and G. Wu. 2007. Predictable behavior, profits, and attention. Journal of Empirical
Finance 14:590–610.
Sialm, C., L. T. Starks, and H. Zhang. 2015. Defined contribution pension plans: Sticky or discerning
money? Journal of Finance LXX:805–38.
Sicherman, N., G. Loewenstein, D. Seppi, and S. Utkus. 2016. Financial attention. Review of Financial
Studies 29:863–97.
Van Nieuwerburgh, S., and L. Veldkamp. 2010. Information acquisition and under-diversification. The
Review of Economic Studies 77:779–805.
———. 2009. Information immobility and the home bias puzzle. Journal of Finance .
Vlastakis, N., and N. R. Markellos. 2012. Information demand and stock market volatility. Journal
of Banking & Finance 36:1808–21.
48
Von Gaudecker, H.-M. 2015. How does household portfolio diversification vary with financial literacy
and financial advice? The Journal of Finance LXX:489–507.
Wermers, R. 2003. Is money really ‘smart’? New evidence on the relation between mutual fund flows,
manager behavior, and performance persistence .
Yuan, Y. 2015. Market-wide attention, trading, and stock returns. Journal of Financial Economics
548–64.
49
Appendix A Data Construction
In this appendix, we describe the auxiliary datasets we use in our study. As mentioned in Section 2.2,
our data source gave us access to data structured as a SQL relational database. Besides the Web-
activity table described at length in Section 2, this study uses four additional tables named Trades,
Clients, Accounts, and Account Holdings, respectively. Finally, the study also uses information from
a variety of standard data sources. Further details are provided below.
Trades
The Trades table includes the record of all the trades made by account holders over the period January
2010 through June 2014. This table contains information regarding 3,528,001 accounts. The Trades
table has a total of 197,870,535 observations and each observation contains the following information:
Acct id, a unique numeric account identifier; Client id, a unique numeric client identifier; Ord ts, the
time-stamp of the trade order; Exec ts, the time-stamp of the trade execution; Scrty type descr, denot-
ing whether the security traded is a stock, a bond, an option or a mutual fund; Cusip id, the CUSIP
code of the security traded; Action, denoting whether the action is a buy or a sell; Trd prncpl amt, the
amount traded; Trd qty, the number of stocks traded; Chnl, denoting whether the trade is web-based
or phone based. In our sample, 99% of the trades are web-based, showing how much the broking
business has changed compared to the Barber and Odean (2002) data.
The Trades table also contains information on the type of account associated with a trade. This
variable is named Acct type and takes on 61 values such as 401K, 403B, IRA, and many others.
In much of the analysis performed in the paper, we group individual brokerage accounts and Joint
Tenants with Right of Survivorship (JTWROS) accounts as one broad category that comprises 57.6%
of all accounts in our data.
Clients
The Clients table contains information on the characteristics of the clients. This file has 2,812,877
observations. Comparing the total number of accounts reported above to the total number of clients
50
shows that many clients have more than one account. In fact, 83.9% of the clients have only one
account, 2.5% of them have two accounts, and 3.5% of them have three or more accounts. Clients that
have multiple accounts usually have an individual account and an IRA account. Out of the Clients
table, we use the following variables: client id, a unique numeric client identifier; and Client age, the
age of the account holder.
Accounts
The Accounts table includes data on the characteristics of the accounts. This file has 3,528,001
observations and, for each observation, we make use of the following variables: Acct id, a unique
numeric account identifier; Client id a unique numeric client identifier; Gender, the gender of the
account holder; Stnd pstl cd, the zip-code of the account-holder; Acct open dt, the account opening
date; and Acct close dt, the account closing date.
Account Holdings
The Account Holdings table includes quarterly holdings for every account in the dataset. This file
has 194,438,993 observations and the following variables: Acct id, a unique numeric account identifier;
mkt close dt, the date of the holdings snapshot; Cusip id, the CUSIP code of the security held by the
account holder; Scrty type descr, denoting whether the security held is a stock, a bond, an option, or
a mutual fund; Qty, the quantity held; and Amt, the dollar value of the quantity held. The Account
Holdings table also contains cash-holdings information. We construct account holdings at the daily
frequency by merging the Account Holdings table with the Trades table.
Additional Data Sources
Stock market information such as prices, returns and trading volumes – among others – is obtained
from CRSP, CRSP OTC and CRSP Mutual Funds. Stocks’ accounting information is obtained from
COMPUSTAT. Benchmark returns are obtained from the Fama-French website and by constructing
DGTW returns at the daily frequency. News concerning stocks are obtained from Capital IQ. Analysts’
51
information is obtained from I/B/E/S. Finally, information regarding zip-codes’ latitude and longitude
is obtained from Census.
Construction of the Final Dataset
The final dataset is obtained in two steps. In the first, we merge the contents of the Web-activity,
Trades, Clients, Accounts and Account Holdings tables using the acct id identifier. In the second, we
merge the resulting dataset with the additional data sources using either the stocks’ Cusip id or ticker
as identifiers.
52
Table 1. Example of Web Activity Within the Brokerage Account Website
Timestamp Masked URL Duration Session
28jan2014 07:58:05 Homepage 00:00:13 128jan2014 07:58:18 Balances and Positions 00:00:08 128jan2014 07:58:26 Watchlist 00:00:06 1
28jan2014 08:32:16 Research / Stocks Overview 00:00:02 228jan2014 08:32:18 Research / Ticker Symbol=SPX 00:05:33 228jan2014 08:37:51 Watchlist 00:00:06 228jan2014 08:37:57 Watchlist / Refresh 00:00:27 228jan2014 08:38:24 Research / Stocks Overview 00:00:01 228jan2014 08:38:25 Research / Ticker Symbol=VIX 00:00:47 228jan2014 08:39:12 Watchlist 00:00:04 228jan2014 08:39:16 Watchlist / Refresh 00:06:16 228jan2014 08:45:32 Balances and Positions 00:00:05 228jan2014 08:45:37 Watchlist 00:00:29 228jan2014 08:46:06 Watchlist / Refresh 00:26:00 228jan2014 09:12:06 Balances and Positions 00:00:04 228jan2014 09:12:10 Watchlist 00:02:38 2
28jan2014 12:59:46 Balances and Positions 00:00:18 328jan2014 13:00:04 Watchlist 00:17:53 3
28jan2014 14:28:54 Homepage 00:00:12 428jan2014 14:29:06 Balances and Positions 00:00:05 428jan2014 14:29:11 Watchlist 00:00:23 4
This table displays the web activity – within the brokerage account website – of an account holder, on January 28, 2014.Timestamp includes the date, hour, minute and second of the first click on the webpage; Masked URL is the maskedURL of the webpage browsed by the investor. Duration is the number of seconds spent on the webpage, and Sessionis the web-session number within the trading day. The URLs presented in the table have been masked to preserve theanonymity of the brokerage account house.
53
Table 2. Summary Statistics of Investor Attention
Panel A. Total daily number of hours spent across all accounts onthe top six sections of the brokerage account website
Mean St. Dev p.25 p.50 p.75
Balance and Positions 787 412 516 669 930
Research 438 281 260 381 523
Trading 415 263 213 353 561
Homepage 370 143 271 346 438
Account 61 47 35 53 76
Watchlist 46 21 35 42 52
Panel B. Rank of the Top 20 Companies and ETFs Researchedby Brokerage Account Holders
Rank by Rank by Rank byMinutes Pages Visits
Facebook 1 2 2Apple 2 1 1Bank of America 3 6 6Ford 4 5 5SPDR S&P 500 ETF Trust 5 12 21AT&T 6 4 4Twitter 7 11 16General Electric 8 3 33D System 9 10 10Verizon 10 9 8Tesla Motors 11 8 9Gilead Sciences 12 19 27JC Penney 13 15 19Microsoft 14 7 7Sirius XM 15 18 17Amazon 16 14 15SPDR Gold Trust 17 26 32SPDR Dow Jones Industrial Average ETF 18 23 41Netflix 19 13 11Market Vectors ETF Trust 20 33 43
This table reports summary statistics of investor attention. Panel A reports summary statistics of the total daily numberof hours spent across all account holders on various sections of the brokerage account website. For each section, we reportthe mean (Mean), the standard deviation (St.Dev), and the 25th, 50th, and 75th percentiles of the daily number of hours– all computed in the time-series dimension. Panel B reports the top 20 companies and ETFs researched by brokerageaccount holders. The three columns show the rank based on the number of minutes, pages and visits, respectively.
54
Tab
le3.
Su
mm
ary
Sta
tist
ics
Pan
el
A.
Ch
aracte
ris
tics
of
Cli
ents
Pan
el
B.
Portf
oli
oC
haracte
ris
tics
Mean
Med
ian
St.
Dev
Mean
Med
ian
St.
Dev
Age
50.9
351
15.9
1P
ort
folio
Valu
e$94,0
00
$18,0
00
$368,0
00
Gen
der
0.7
31
0.4
4C
ash
Hold
ings
$16,0
00
$1,0
00
$72,0
00
Nu
mb
erof
Acc
ou
nts
1.3
41
0.6
8S
tock
Hold
ings
$82,0
00
$15,0
00
$341,0
00
Acc
ou
nt
age
8.5
57.5
25.5
4N
um
ber
of
Sto
cks
6.5
14
8.7
7
Pan
el
C.
Trad
ing
Beh
avio
r
Percenti
les
Mean
St.
Dev
1st
25th
50th
75th
99th
Fra
ctio
nof
Days
wit
hT
rad
es0.0
30.0
80
00
0.0
10.4
4
Days
Bet
wee
nT
rad
es46.6
360.7
61.4
110
25.3
958.4
6322
Nu
mb
erof
Tra
des
Per
Day
1.7
21.8
01
11.3
11.8
97.8
0
Dollar
Valu
eof
Tra
des
$16,0
00
$64,0
00
––
––
–
Pan
el
D.
Att
enti
on
Beh
avio
r
Percenti
les
Mean
St.
Dev
1st
25th
50th
75th
99th
Fra
ctio
nof
Days
wit
hL
ogin
s0.1
70.2
40
0.0
20.0
60.2
10.9
6
Days
Bet
wee
nL
ogin
s27.5
147.5
91.1
43.9
511.2
030.0
0247
Nu
mb
erof
Login
sP
erD
ay
10.6
117.6
91.5
04.7
27.3
311.5
961.4
4
Nu
mb
erof
Min
ute
s28.7
4288.6
90.3
03.7
78.0
017.4
5366.3
5
This
table
rep
ort
ssu
mm
ary
stati
stic
sof
the
bio
gra
phic
chara
cter
isti
cs(P
anel
A),
the
port
folio
chara
cter
isti
cs(P
anel
B),
and
the
tradin
g(P
anel
C)
and
att
enti
on
beh
avio
r(P
anel
D)
of
the
bro
ker
age
acc
ount
hold
ers
inour
web
-act
ivit
ydata
set.
For
each
vari
able
inea
chpanel
,w
ere
port
the
sam
ple
mea
n(M
ean
),m
edia
n(M
edian
),and
standard
dev
iati
on
(St.Dev
).T
he
stati
stic
sre
port
edin
Panel
sA
and
Bare
com
pute
dacr
oss
acc
ounts
.T
he
stati
stic
sin
Panel
sC
and
Dare
com
pute
dfirs
tin
the
tim
e-se
ries
dim
ensi
on
at
the
acc
ount-
hold
erle
vel
,co
nsi
der
ing
only
day
sw
hen
the
stock
mark
ets
are
op
en.
They
are
then
com
pute
dcr
oss
-sec
tionally
acr
oss
acc
ount
hold
ers.
All
dollar
valu
esw
ere
rounded
toth
enea
rest
thousa
nd
and
the
per
centi
les
of
“dollar
valu
eof
trades
”hav
eb
een
mask
edup
on
reques
tof
the
data
pro
vid
erfo
rco
nfiden
tiality
reaso
ns.
55
Table 4. Characteristics of Attentive Investors
Coeff t-stat Magnitude
Brokerage 0.590∗∗∗ (13.68) 660.109
Male 0.311∗∗∗ (7.23) 299.143
Age 0.291∗∗∗ (12.62) 277.522
Account Age 0.033 (1.55) 27.720
Portfolio Value 0.164∗∗∗ (2.58) 145.975
Fr. in Cash -0.101∗∗∗ (-5.55) -78.669
Fr. in ETF -0.097∗∗∗ (-4.85) -76.177
Fr. in Mutual Fund -0.005 (-0.25) -4.006
Beta Mkt 0.106∗∗∗ (4.62) 92.019
Beta SMB 0.091∗∗∗ (4.07) 78.478
Beta HML -0.062∗∗∗ (-2.58) -48.974
Beta MOM 0.113∗∗∗ (4.45) 98.496
Number of Stocks Traded 0.813∗∗∗ (40.60) 1,030.815
Constant 4.749∗∗∗ (121.35) 820.91
R-Squared 22.9%N 8,574
This table reports regression results on the relation between the time spent on the brokerage account website and accountholder characteristics. We estimate the following cross-sectional regression:
Attentioni = α+ x′i β + εi for i = 1, . . . , N
where Attentioni is the log of the total number of minutes spent on the brokerage account website by account holderi over the sample period, xi is a vector of covariates associated with account holder i and N is the total number ofaccount holders included in the analysis. What follows is a description of the covariates included. The first groupcomprises demographic variables: Brokerage, a brokerage account dummy; Male, a male dummy variable; Age, theage of the investor; Account Age, the age of the account. The second category comprises portfolio holdings variables:Portfolio Value, the total value of the invested portfolio; N. of Assets, the total number of stocks, mutual funds andexchange traded funds held; Fr. in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of the total wealth in thebrokerage account held in cash, traded funds and mutual funds, respectively. The third category comprises regressorsrelated to portfolio risk and trading activity: Beta Mkt, Beta SMB, Beta HML, and Beta MOM, the loadings on themarket, Small-Minus-Big, High-Minus-Low and Momentum factors computed using the full sample available using dailyreturns; and N. of Stocks Traded, the number of stocks traded over the period. We display the ordinary least squarescoefficient estimates (Coeff), the associated t-statistics (t-stat), and the economic magnitudes (Magnitude) – computedas described in Section 4.1. Standard errors are adjusted for heteroskedasticity. Coefficients marked with ***, **, and *are significant at the 1%, 5%, and 10% level, respectively.
56
Tab
le5.
Invest
or
Att
enti
on
All
ocati
on
:C
ross
-Secti
on
al
Regre
ssio
ns
Pan
el
A.
Nu
mb
er
of
Sto
cks
Pan
el
B.
Herfi
nd
ah
l
A.I
.O
verall
A.I
I.O
uts
ide
A.I
II.
Insi
de
Overall
Coeff
t-st
at
Coeff
t-st
at
Coeff
t-st
at
Coeff
t-st
at
Bro
ker
age
0.0
27
(1.4
4)
0.0
33
(1.6
0)
0.0
02
(0.1
8)
-0.0
08
(-1.1
6)
Male
0.0
45∗∗
(2.5
4)
0.0
54∗∗∗
(2.7
4)
0.0
14
(1.0
9)
-0.0
02
(-0.2
9)
Age
0.0
73∗∗∗
(3.8
6)
0.0
65∗∗∗
(3.1
2)
0.0
65∗∗∗
(4.9
3)
-0.0
12∗
(-1.7
6)
Acc
ou
nt
Age
-0.0
22
(-1.1
6)
-0.0
23
(-1.1
0)
-0.0
14
(-0.9
9)
0.0
08
(1.3
1)
Tota
lA
tten
tion
0.4
23∗∗∗
(15.1
9)
0.4
33∗∗∗
(14.3
9)
0.2
65∗∗∗
(13.8
8)
-0.0
57∗∗∗
(-10.6
9)
Port
folio
Valu
e0.0
02
(0.1
1)
-0.0
27
(-1.1
6)
0.0
39∗∗
(2.1
3)
-0.0
05
(-1.0
8)
Her
fin
dah
lS
tock
s-0
.047∗∗
(-2.0
0)
-0.0
37∗
(-1.7
1)
-0.0
44∗∗
(-2.3
9)
0.0
20∗∗∗
(2.8
6)
N.
Sto
cks
0.1
16∗∗∗
(4.4
3)
0.0
92∗∗∗
(3.4
2)
0.1
35∗∗∗
(5.8
8)
-0.0
37∗∗∗
(-5.3
8)
Fr.
inC
ash
0.0
64∗∗
(2.5
3)
0.0
92∗∗∗
(3.3
1)
-0.0
24
(-1.4
2)
-0.0
28∗∗∗
(-3.0
6)
Fr.
inE
TF
0.0
02
(0.1
2)
0.0
30∗
(1.7
6)
-0.0
51∗∗∗
(-2.6
6)
-0.0
04
(-0.7
4)
Fr.
inM
utu
al
Fu
nd
-0.0
18
(-0.9
9)
0.0
08
(0.4
3)
-0.0
50∗∗∗
(-4.3
0)
-0.0
01
(-0.1
6)
Bet
aM
kt
-0.0
47∗
(-1.6
9)
-0.0
63∗∗
(-2.0
9)
0.0
02
(0.1
1)
0.0
15
(1.3
9)
Bet
aS
MB
-0.0
47
(-1.6
2)
-0.0
84∗∗
(-2.5
0)
0.0
43∗∗
(2.4
4)
0.0
12
(0.9
9)
Bet
aH
ML
-0.0
25
(-0.9
4)
-0.0
33
(-1.1
3)
-0.0
06
(-0.3
8)
0.0
03
(0.3
1)
Bet
aM
OM
-0.0
30
(-1.4
2)
-0.0
37
(-1.6
3)
-0.0
04
(-0.3
3)
0.0
07
(0.8
1)
Ris
kId
io-0
.084∗∗∗
(-5.5
1)
-0.0
91∗∗∗
(-5.1
2)
-0.0
33∗∗∗
(-3.2
5)
0.0
32∗∗∗
(5.3
9)
N.
of
Sto
cks
Tra
ded
0.2
12∗∗∗
(4.6
0)
0.1
81∗∗∗
(4.2
1)
0.2
37∗∗∗
(5.4
2)
-0.0
30∗∗∗
(-3.2
0)
Con
stant
1.4
41∗∗∗
(16.5
1)
1.1
80∗∗∗
(12.3
1)
0.4
58∗∗∗
(7.3
6)
0.4
82∗∗∗
(14.8
0)
R-S
qu
are
d23.0
%18.2
%30.7
%6.9
%N
3,5
86
3,5
86
3,5
86
3,5
86
This
table
rep
ort
sre
gre
ssio
nre
sult
son
the
rela
tion
bet
wee
natt
enti
on
alloca
tion
and
acc
ount
hold
erch
ara
cter
isti
cs.
InP
anel
Aw
ees
tim
ate
the
follow
ing
cross
-sec
tional
regre
ssio
n:
NStocks i
=α
+x′ iβ
+ε i,
fori
=1,...,N,
wher
eNStocks
isth
elo
gnum
ber
of
stock
sre
searc
hed
by
inves
tori,xi
the
vec
tor
of
inves
tor-
spec
ific
cova
riate
s,andN
isth
eto
tal
num
ber
of
inves
tors
incl
uded
inth
eanaly
sis.
The
resu
lts
inP
anel
A.I
use
all
the
stock
sre
searc
hed
by
each
acc
ount
hold
erand
Panel
A.I
I(P
anel
A.I
II)
incl
udes
only
the
stock
sin
side
(outs
ide)
each
inves
tor’
sp
ort
folio.
InP
anel
Bw
ees
tim
ate
the
follow
ing
cross
-sec
tional
regre
ssio
n:
Herfindahl i
=α
+x′ iβ
+ε i,
fori
=1,...,N,
wher
eHerfindahl i
isth
enorm
alize
dH
erfindahl
index
of
the
stock
sre
searc
hed
by
inves
tori
–co
mpute
das
∑ K k=
1ω2 k−1/K
1−1/K
ifK>
1and
1ifK
=1,
wher
eωk
isth
efr
act
ion
of
the
tota
lst
ock
-sp
ecifi
catt
enti
on
alloca
ted
tost
ock
kandK
isth
eto
tal
num
ber
of
stock
sse
arc
hed
.W
hat
follow
sis
ades
crip
tion
of
the
cova
riate
sin
cluded
.T
he
firs
tgro
up
com
pri
ses
dem
ogra
phic
vari
able
s:B
roke
rage
,a
bro
ker
age
acc
ount
dum
my;
Ma
le,
am
ale
dum
my
vari
able
;A
ge,
the
age
of
the
inves
tor;
Acc
ou
nt
Age
,th
eage
of
the
acc
ount.
The
seco
nd
cate
gory
com
pri
ses
To
tal
Att
enti
on
,defi
ned
as
the
log
tota
lnum
ber
of
seco
nds
spen
tby
the
inves
tor
on
the
bro
ker
age
acc
ount
web
site
over
the
full
sam
ple
.T
he
thir
dca
tegory
com
pri
ses
port
folio
hold
ings
vari
able
s:P
ort
foli
oV
alu
e,th
eto
tal
valu
eof
the
inves
ted
port
folio;
Her
fin
da
hl
Sto
cks,
the
norm
alize
dH
erfindahl
index
of
inves
tori’
sst
ock
-hold
ings
–co
mpute
das
∑ J j=
1ω2 j−1/J
1−1/J
ifJ>
1and
1ifJ
=1,
wher
ewj
isth
efr
act
ion
of
wea
lth
alloca
ted
tost
ock
jandJ
isth
eto
tal
num
ber
of
stock
shel
d;
N.
Sto
cks,
the
tota
lnum
ber
of
stock
shel
d;
Fr.
inC
ash
,F
r.in
ET
F,
and
Fr.
inM
utu
al
Fu
nd,
the
fract
ion
of
the
tota
lw
ealt
hin
the
bro
ker
age
acc
ount
hel
din
cash
,tr
aded
funds
and
mutu
al
funds,
resp
ecti
vel
y.T
he
fourt
hca
tegory
com
pri
ses
regre
ssors
rela
ted
toth
ep
erfo
rmance
and
risk
of
inves
tors
’p
ort
folios:
Bet
aM
kt,
Bet
aS
MB
,B
eta
HM
L,
and
Bet
aM
OM
,th
elo
adin
gs
on
the
mark
et,
Sm
all-M
inus-
Big
,H
igh-M
inus-
Low
and
Mom
entu
mfa
ctors
com
pute
dusi
ng
the
full
sam
ple
available
usi
ng
daily
retu
rns;
Ris
kId
io,
the
idio
syncr
ati
cri
skof
the
inves
tor’
sp
ort
folio.
Fin
ally,
the
last
cova
riate
isN
.o
fS
tock
sT
rad
ed:
the
num
ber
of
stock
str
aded
over
the
per
iod.
Wit
hin
each
panel
we
dis
pla
yth
eord
inary
least
square
sco
effici
ent
esti
mate
s(Coeff
)and
the
ass
oci
ate
dt-
stati
stic
(t-s
tat)
.Sta
ndard
erro
rsare
adju
sted
for
het
erosk
edast
icit
y.C
oeffi
cien
tsm
ark
edw
ith
***,
**,
and
*are
signifi
cant
at
the
1%
,5%
,and
10%
level
,re
spec
tivel
y.
57
Table 6. Stock Characteristics and Investor Attention
Panel A. Stocks Panel B. Stocks Searched Panel C. Stocks Searchedin the Portfolio over the Sample and S&P 500 Stocks
Squared Sharpe Ratio 0.0085∗∗∗ 0.0085∗∗∗ 0.0031∗∗∗ 0.0031∗∗∗ 0.0001∗∗∗ 0.0001∗∗∗
(8.11) (8.12) (10.06) (10.09) (3.39) (3.44)
Size 0.0046 0.0026 0.0045∗∗∗ 0.0045∗∗∗ 0.0029∗∗∗ 0.0029∗∗∗
(1.45) (0.80) (3.83) (3.82) (15.24) (15.23)
Dummy Close 0.0263∗∗∗ 0.0084∗∗∗ 0.0009∗∗∗
(3.94) (2.92) (4.30)
Portfolio Weight 0.0204∗∗∗
(3.97)
Account Holder FE ! ! ! ! ! !
R-Squared 0.1% 0.5% 0.1% 0.1% 0.3% 0.3%
N 442,833 442,833 3,526,595 3,526,595 30,295,439 30,295,439
This table reports regression results studying which stocks investors pay attention to, whenever they acquire stock-specificinformation. We focus only on days when stock-specific information is acquired and estimate the following regression:
Att Dummyi,j,t = αi + x′i,j,t β + εi,j,t
where Att Dummyi,j,t is a dummy equal to 1 if investor i researches stock j on date t and is equal to 0 for all theother stocks that are in investor i’s investment opportunity set, but are not researched. We construct three alternativeinformation sets for each investor. The first includes only the stocks held in the investment portfolio at time t (PanelA). The second uses all the stocks researched by the investor over our sample (Panel B). The third includes all thestocks researched by the investor as well as all the stocks in the S&P 500 index at time t (Panel C). In each panel, weestimate two specifications. The first isolates the regressors theoretically motivated by the models in Van Nieuwerburghand Veldkamp (2009, 2010): Squared Sharpe Ratio, the squared realized Sharpe ratio for each stock in the investmentopportunity set, computed over the previous 21 days; and Size, the market value of the stock – computed as the productof the price and shares outstanding. The second additionally includes: Dummy Close, an indicator variable identifyingwhether the headquarter of stock j is located within a 250 miles radius from investor i, as in Ivkovic and Weisbenner(2005). We use the standard formula for computing the distance in miles between account holder i and firm j, as follows:
dist(i, j) = arccos(cos(lati)cos(longi)cos(latj)cos(longj) + cos(lati)sin(longi)cos(latj)sin(longj) + sin(lati)sin(latj)
)r,
where lat/long are the latitudes/longitudes of location i and location j (expressed in radians), respectively, and r denotesthe radius of the Earth (approximately 3,963 miles). In panel A, we also include Portfolio Weight, the weight of stock jin investor i’s portfolio as of time t. The estimates include account-holder fixed-effects. Within each panel we display theordinary least squares coefficient estimates (Coeff) and the associated t-statistic (t-stat). Standard errors are clusteredat the account level. Coefficients marked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
58
Table 7. Characteristics of Researched Stocks
Coeff t-stat Magnitude
Analyst 0.441∗∗∗ (12.45) 25.052
News 0.064∗ (1.78) 2.981
Turnover 0.835∗∗∗ (4.45) 59.070
Volume 0.295∗∗∗ (6.82) 15.509
Alpha 0.027 (0.70) 1.228
Beta Mkt 0.254∗∗∗ (5.04) 13.088
Beta SMB -0.050∗ (-1.68) -2.218
Beta HML -0.043 (-1.47) -1.917
Beta MOM 0.073∗∗∗ (2.65) 3.411
Return 0.022 (0.72) 1.027
Variance 0.080 (1.62) 3.759
Skewness -0.043 (-1.54) -1.923
Kurtosis 0.047 (1.55) 2.159
MktBook 0.113∗∗∗ (2.96) 5.393
R&D 0.101∗∗ (2.44) 4.788
Profitability 0.031 (0.68) 1.422
Age -0.007 (-0.30) -0.296
Size 0.201∗∗∗ (8.14) 10.062
Leverage 0.155∗∗∗ (4.72) 7.584
Inst. Investors -0.364∗∗∗ (-8.34) -13.805
Constant 3.206∗∗∗ (68.05) 45.242
R-Squared 45.4%N 2,572
This table reports regression results on the relation between investor attention – computed across all account-holders inour data – and stock characteristics. We estimate the following cross-sectional regression:
Attentionj = α+ x′j β + εj , for j = 1, . . . , J,
where Attentionj is the log total number of minutes spent – across all account-holders – on stock j and xj is a vectorof stock-specific covariates. What follows is a description of the three groups of covariates included in the analysis. Thefirst group comprises regressors we classify as attention proxies: Analyst, the number of analysts covering the stock;News, the total number of news associated with the stock, Volume, the stock’s volume; Turnover, computed as thestock’s volume divided by the shares outstanding. The second category comprises regressors related to stocks’ risk andperformance. Alpha, Beta Mkt, Beta SMB, Beta HML, and Beta MOM are the intercept and the loadings on the market,SMB, HML and MOM factors. The loadings are estimated for each stock j using daily data over the full sample. Return,Variance, Skewness and Kurtosis, are the realized returns, variance, skewness and kurtosis for each stock j – computedusing daily returns. The third category comprises regressors related to stock characteristics: MktBook, the ratio ofmarket value of assets and the book value of assets, R&D, computed as the ratio of research and development expenses(COMPUSTAT item: XRDQ) and sales (COMPUSTAT item: SALEQ); Profitability, the ratio between operating incomebefore depreciation (COMPUSTAT item: OIBDPQ) and total assets (COMPUSTAT item: ATQ); Age, the number ofyears a company has been in the CRSP dataset; Size, the market value of assets computed as the product of the priceand shares outstanding; Leverage, computed as the sum of long-term debt and debt in current liabilities, divided bythe market value of assets; Inst. Investors, the fraction of the shares outstanding owned by institutional investors. Wedisplay the ordinary least squares coefficient estimates (Coeff), the associated t-statistic (t-stat), and the economicmagnitude (Magnitude) – computed as described in Section 4.1. Standard errors are adjusted for heteroskedasticity.Coefficients marked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
59
Table 8. Attention and Portfolio Performance: Cross-Sectional Results
Spec 1 Spec 2 Spec 3
Attention 1.531∗∗∗ 1.600∗∗∗ 2.326∗∗∗
(3.36) (3.46) (4.48)
Brokerage 1.991∗∗ 1.905∗∗
(2.29) (2.17)
Male 0.631 0.605(0.71) (0.68)
Age -1.038∗∗∗ -0.964∗∗
(-2.59) (-2.38)
Account Age 0.043 0.056(0.10) (0.13)
Portfolio Value -0.553∗∗
(-2.29)
Fr. in Cash 0.136
Fr. in ETF 0.438
Fr. in Mutual Fund 0.580(1.64)
N. of Stocks Traded -1.320∗∗∗
(-4.26)
Constant 1.920∗∗∗ 0.710 0.935(4.25) (0.92) (1.18)
R-Squared 0.1% 0.3% 0.4%N 8,340 7,476 7,468
This table reports regression results on the relation between portfolio performance and investor attention. We estimatethe following baseline cross-sectional regression:
AV G DGTW Reti = α+ β Attentioni + x′i γ + εi for i = 1, . . . , N
where AV G DGTW Reti is the annualized percentage average DGTW abnormal return of investor i over the sample,Attentioni is the log of the total number of minutes spent on the brokerage account website by account holder i over thesample period, xi is a vector of covariates associated with account holder i and N is the total number of account holdersincluded in the analysis. We include three specifications. The first specification uses attention as the sole covariate.The second specification includes a group of covariates that control for investor demographic characteristics: Brokerage,a brokerage account dummy; Male, a male dummy variable; Age, the age of the investor; Account Age, the age of theaccount. The third specification includes portfolio holdings and trading activity variables: Portfolio Value, the totalvalue of the invested portfolio; Fr. in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of the total wealth in thebrokerage account held in cash, traded funds and mutual funds, respectively; and N. of Stocks Traded, the number ofstocks traded over the period. Displayed are the ordinary least squares coefficient estimates and associated t-statistics.Standard errors are adjusted for heteroskedasticity. Coefficients marked with ***, **, and * are significant at the 1%,5%, and 10% level, respectively.
60
Table 9. Calendar Portfolio Returns Analysis at the Weekly Frequency
Panel A. Equally-Weighted Returns
Q1 Q2 Q3 Q4 Q5 Q5-Q1
Alpha 0.03 0.02 0.02 0.03 0.06∗∗ 0.04∗∗
(0.96) (0.82) (0.96) (1.15) (2.02) (2.63)
MKT 0.84∗∗∗ 0.89∗∗∗ 0.94∗∗∗ 0.89∗∗∗ 0.85∗∗∗ 0.01(17.33) (18.71) (25.03) (21.73) (15.67) (0.62)
SMB 0.03 0.01 0.02 0.06 0.16∗∗ 0.13∗∗∗
(0.46) (0.21) (0.34) (1.06) (2.43) (4.02)
HML -0.07 -0.02 -0.01 -0.03 -0.03 0.03(-1.10) (-0.42) (-0.17) (-0.59) (-0.48) (0.88)
UMD -0.15∗∗∗ -0.10∗ -0.10∗∗ -0.10∗∗ -0.07 0.09∗∗∗
(-2.77) (-1.79) (-2.61) (-2.05) (-1.07) (2.77)
R-Squared 90% 92% 95% 94% 90% 42%
N 74 74 74 74 74 74
Panel B. Value-Weighted Returns
Q1 Q2 Q3 Q4 Q5 Q5-Q1
Alpha 0.06 0.06 0.05 0.06∗ 0.10∗∗ 0.04∗∗∗
(1.61) (1.52) (1.41) (1.73) (2.30) (2.76)
MKT 0.76∗∗∗ 0.82∗∗∗ 0.86∗∗∗ 0.84∗∗∗ 0.83∗∗∗ 0.08∗∗∗
(14.62) (15.38) (16.03) (17.37) (14.11) (3.11)
SMB 0.17∗∗ 0.16∗∗ 0.18∗∗∗ 0.21∗∗∗ 0.31∗∗∗ 0.14∗∗∗
(2.61) (2.23) (3.20) (3.19) (4.14) (3.29)
HML 0.02 0.01 0.05 -0.02 0.01 -0.02(0.42) (0.22) (1.04) (-0.32) (0.10) (-0.40)
UMD -0.11∗∗ -0.13∗∗ -0.08 -0.07 -0.03 0.09∗∗
(-2.07) (-2.26) (-1.59) (-1.21) (-0.42) (2.28)
R-Squared 87% 89% 92% 91% 87% 56%
N 74 74 74 74 74 74
This table displays the results of calendar-time portfolio return regressions. We first sort account holders into five groups(Q1 through Q5) based on the number of seconds spent on the brokerage account website over the full sample. We thencompute, for each group, equally-weighted weekly returns in Panel A and value-weighted weekly returns in Panel B. Q1are the least attentive investors while Q5 are the most attentive. Finally, we run the following time-series regression:
rq,t − rft = αq + β1,q (Mktt − rft) + β2,q HMLt + β3,q SMBt + β4,q MOMt + εt for t = 1, . . . , T & q = 1, . . . , 5,
where rq,t − rft is the annualized excess-return at time t of portfolio q, Mktt − rft is the annualized excess returnon the NYSE/AMEX/NASDAQ index, HMLt, SMBt and MOMt are the annualized returns on the value, size andmomentum factors respectively. The last column of each panel reports performance regressions for a portfolio long inthe most attentive investors (Q5) and short in the least attentive investors (Q1). We display the ordinary least squarescoefficient estimates and the associated t-statistics computed using Newey-West standard errors. Coefficients markedwith ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
61
Table 10. Attention and Performance: Panel Regression Results
Panel A. Overall
21 Days 42 Days 63 Days
Abn. Attention 0.032∗∗ 0.103∗∗∗ 0.138∗∗∗
(t-statistic) (2.27) (4.21) (3.66)
[t-statistic] [3.58] [7.67] [9.06]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.14% 0.27% 0.38%
N 2,573,951 2,377,766 2,187,222
This table reports panel regression results on the relation between portfolio performance and investor attention. Weestimate the following baseline panel regression:
DGTW Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time interval t : t + k;Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the difference between the (log) attentionon day t and the (log) average attention over the previous 21 business days; finally, αi and βm represent account-holderfixed effects and monthly time-effects, respectively. Attention is measured as the number of seconds spent on thebrokerage account website. We report results for different horizons, i.e. k = 21, 42, and 63 days. Displayed are theordinary least squares coefficient estimates and two sets of t-statistics. The first – in round brackets – are computedusing standard errors that are double-clustered by account-holder and time, see Petersen (2009). The second – in squarebrackets – are computed using the Driscoll and Kraay (1998) standard errors. Coefficients marked with ***, **, and *are significant at the 1%, 5%, and 10% level, respectively, according to the double-clustered standard errors.
62
Table 11. Trading Performance and Investor Attention: Baseline Regression Results
Panel A. Performance of Buys
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Attention 0.147∗ 0.342∗∗∗ 0.533∗∗∗ 0.383∗∗ 0.441∗∗ 0.197 0.009(1.81) (3.03) (3.57) (2.18) (2.28) (0.93) (0.03)
Constant -0.139∗ 0.152 0.349∗∗ 0.479∗∗∗ 0.622∗∗∗ 0.691∗∗∗ 0.502(-1.70) (1.31) (2.31) (2.58) (3.22) (3.30) (1.66)
R-Squared 0.01% 0.03% 0.05% 0.02% 0.02% 0.00% 0.00%N 24,139 24,103 24,064 24,001 23,947 23,887 23,436
Panel B. Performance of Sells
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Attention 0.103 0.130 0.287∗ 0.119 0.118 -0.027 -0.425(1.23) (1.04) (1.81) (0.66) (0.58) (-0.13) (-1.13)
Constant -0.023 0.376∗∗∗ 0.914∗∗∗ 1.008∗∗∗ 1.090∗∗∗ 1.182∗∗∗ 1.366∗∗∗
(-0.27) (2.97) (5.61) (5.42) (5.15) (5.11) (3.88)
R-Squared 0.01% 0.01% 0.02% 0.00% 0.00% 0.00% 0.01%N 19,435 19,373 19,320 19,257 19,200 19,142 18,777
This table reports regression results on the relation between investor attention and trading performance. We separatethe buys (i.e. stock purchases) and the sells (i.e. stock sales) and estimate the following pooled regressions:
DGTW Ret Buysi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
DGTW Ret Sellsi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
where DGTW Ret Buysi,j,t:t+k (DGTW Ret Sellsi,j,t:t+k) are the cumulative abnormal returns of security j bought(sold) by investor i over the time interval t : t+k, computed using the DGTW model; Attentioni,j,t is the (log) number ofseconds spent on the brokerage account website by investor i over the month preceding the trade in stock j that occurs attime t. Cumulative abnormal returns are computed at the one-, two-, three-, four-, five-, six- and twelve-month horizons.Panel A reports results for stock purchases. Panel B repeats the exercise for stock sales. Displayed are the ordinary leastsquares coefficient estimates and associated t-statistics. Standard errors are adjusted for heteroskedasticity. Coefficientsmarked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
63
Table 12. Trading Performance and Investor Attention: Portfolio Results
Panel A: Future Returns
Panel A.I. Unconditional Results
Return Horizon 1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Buy -0.13% 0.17% 0.37% 0.50% 0.64% 0.70% 0.50%
Sell -0.02% 0.38% 0.91% 1.01% 1.09% 1.18% 1.37%
Buy-Minus-Sell -0.11% -0.21% -0.54%∗∗∗ -0.51%∗ -0.45% -0.48% -0.86%∗∗
p-val (0.33) (0.19) (0.01) (0.05) (0.17) (0.20) (0.04)
Panel A.II. Results Conditional on Overall Attention
Return Horizon 1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Buy-High 0.00% 0.42% 0.79% 0.85% 0.98% 0.88% 0.35%
Buy-Low -0.28% -0.10% -0.07% 0.12% 0.28% 0.51% 0.67%
Sell-High 0.06% 0.38% 1.05% 0.92% 1.00% 0.92% 0.69%
Sell-Low -0.11% 0.37% 0.77% 1.10% 1.18% 1.45% 2.05%
Buy-High Minus Sell-High -0.06% 0.04% -0.26% -0.07% -0.02% -0.04% -0.34%p-val (0.68) (0.84) (0.29) (0.83) (0.96) (0.94) (0.54)
Buy-Low Minus Sell-Low -0.17% -0.48%∗∗ -0.84%∗∗∗ -0.98%∗∗∗ -0.89%∗∗ -0.94%∗∗ -1.39%∗∗∗
p-val (0.31) (0.05) (0.00) (0.00) (0.01) (0.02) (0.01)
Buy-High Minus Buy-Low 0.28%∗∗ 0.53%∗∗∗ 0.86%∗∗∗ 0.73%∗∗ 0.70% 0.36% -0.32%p-val (0.03) (0.00) (0.00) (0.03) (0.24) (0.73) (0.56)
Sell-High Minus Sell-Low 0.17% 0.01% 0.28% -0.18% -0.17% -0.53% -1.36%∗∗
p-val (0.30) (0.97) (0.46) (0.73) (0.83) (0.45) (0.03)
Panel B: Past Returns
Panel B.I. Unconditional Results
Return Horizon 1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Buy 1.95% 3.42% 5.03% 7.14% 8.33% 10.30% 26.75%
Sell 2.38% 3.93% 5.21% 6.97% 8.12% 9.76% 22.08%
Buy-Minus-Sell -0.43%∗∗∗ -0.51%∗∗∗ -0.18% 0.17% 0.21% 0.53%∗ 4.67%∗∗∗
p-val (0.00) (0.00) (0.48) (0.47) (0.44) (0.08) (0.00)
Panel B.II. Results Conditional on Overall Attention
Return Horizon 1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Buy-High 3.13% 4.83% 6.98% 9.99% 11.56% 14.09% 34.50%
Buy-Low 0.70% 1.92% 2.96% 4.11% 4.91% 6.27% 18.51%
Sell-High 3.55% 5.40% 6.92% 9.37% 10.88% 13.17% 30.57%
Sell-Low 1.19% 2.43% 3.46% 4.51% 5.30% 6.29% 13.43%
Buy-High Minus Sell-High -0.42%∗∗∗ -0.57%∗∗ 0.06% 0.62%∗ 0.67%∗ 0.93%∗∗ 3.94%∗∗∗
p-val (0.01) (0.01) (0.85) (0.05) (0.07) (0.03) (0.00)
Buy-Low Minus Sell-Low -0.49%∗∗∗ -0.51%∗∗∗ -0.50%∗∗ -0.40% -0.39% -0.02% 5.08%∗∗∗
p-val (0.00) (0.00) (0.05) (0.17) (0.26) (0.96) (0.00)
Buy-High Minus Buy-Low 2.43%∗∗∗ 2.91%∗∗∗ 4.02%∗∗∗ 5.88%∗∗∗ 6.65%∗∗∗ 7.82%∗∗∗ 15.99%∗∗∗
p-val (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Sell-High Minus Sell-Low 2.36%∗∗∗ 2.97%∗∗∗ 3.46%∗∗∗ 4.85%∗∗∗ 5.58%∗∗∗ 6.88%∗∗∗ 17.14%∗∗∗
p-val (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
This table reports portfolio results on the relation between investor attention and the performance of the stocks before(Panel B) and after (Panel A) they are traded. Each panel reports the average cumulative DGTW-adjusted returns atthe one-, two-, three-, four-, five-, six- and twelve-month horizons. Panel A.I. and Panel B.I. present the unconditionalresults and – for every horizon – report the difference in performance between the stocks purchased and sold as well asthe p-value testing whether the difference in performance is equal to zero. Panel A.II. and Panel B.II. condition on theoverall attention paid by the account holders in the month preceding each trade. We divide the buys in two groups – lowand high – based on overall attention and report the performance of the the low- and high-attention buys, respectively.We repeat the same procedure to compute low- and high-attention sells. For every horizon, we report the difference inperformance between: 1) high-attention buys and high-attention sells; 2) low-attention buys and low-attention sells; 3)high-attention buys and low-attention buys; and 4) high-attention sells and low-attention sells. In each case, we alsoreport the p-value testing whether the difference in performance is equal to zero. All p-values are based on the bootstrapprocedure suggested by Barber, Lyon, and Tsai (1999). Estimates marked with ***, **, and * are significant at the 1%,5%, and 10% level, respectively. 64
Table 13. Attention, Attention Specialization, and Portfolio Performance:Cross-Sectional Results
Panel A. Overall Panel B. Research Panel C. Balancesand Positions
Attention 3.391∗∗∗ 2.965∗∗∗ 2.415∗∗
(0.00) (3.38) (2.26)
Herfindahl Researched 1.861∗∗∗ 2.231∗∗∗ 1.639∗∗
(2.64) (2.77) (2.49)
Brokerage 2.773∗∗ 3.161∗∗ 3.279∗∗
(2.14) (2.53) (2.51)
Male 0.703 0.716 0.908(0.55) (0.57) (0.71)
Age -0.929∗ -0.796 -0.804(-1.81) (-1.57) (-1.56)
Account Age -0.519 -0.592 -0.515(-0.86) (-0.98) (-0.85)
Portfolio Value -0.817∗∗∗ -0.642∗∗ -0.772∗∗∗
(-2.84) (-2.42) (-2.64)
Fr. in Cash 1.324 1.292 1.397(1.04) (1.01) (1.11)
Fr. in ETF 0.556 0.512 0.572(0.95) (0.87) (1.01)
Fr. in Mutual Fund 0.219 0.159 0.150(0.53) (0.39) (0.37)
Number of Stocks Traded -1.181∗∗∗ -0.889∗∗∗ -1.047∗∗∗
(-3.29) (-2.64) (-2.66)
Constant -0.034 -0.716 -0.040(-0.03) (-0.62) (-0.03)
R-Squared 0.6% 0.5% 0.5%N 3,677 3,677 3,677
This table reports regression results on the relation between portfolio performance and investor attention specialization.We estimate the following baseline cross-sectional regression:
AV G DGTW Reti = α+ β1 Attentioni + β2 Herfindahl Researchedi + x′i γ + εi for i = 1, . . . , N
where AV G DGTW Reti is the annualized percentage average DGTW abnormal return of investor i over the sample,Attentioni is the total attention spent on the brokerage account website by account holder i over the sample period,Herfindahl Researchedi is the normalized Herfindahl index of the stocks researched by investor i over the sample period,xi is a vector of covariates associated with account holder i and N is the total number of account holders included inthe analysis. Attention is measured as the log of the total number of minutes spent on the brokerage account website inPanel A, the log of the total number of minutes spent on the Research pages of the brokerage account website in Panel B,and the log of the total number of minutes spent on the Balances and Positions pages of the brokerage account websitein Panel C. Each panel includes a group of covariates that control for investor demographic characteristics, portfolioholdings, and trading activity variables: Brokerage, a brokerage account dummy; Male, a male dummy variable; Age,the age of the investor; Account Age, the age of the account; Portfolio Value, the total value of the invested portfolio;Fr. in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of the total wealth in the brokerage account held incash, traded funds and mutual funds, respectively; and N. of Stocks Traded, the number of stocks traded over the period.Displayed are the ordinary least squares coefficient estimates and associated t-statistics. Standard errors are adjustedfor heteroskedasticity. Coefficients marked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
65
Table 14. Attention and Performance: Panel Regression Results By Groups
Panel A. 21 Days
Q1 Q2 Q3 Q4 Q5
Abn. Attention 0.241∗∗ 0.086∗ 0.070∗ 0.039 -0.023
(t-statistic) (2.24) (1.76) (1.95) (1.55) (-1.09)
Time FE ! ! ! ! !
Acct Holder FE ! ! ! ! !
R-Squared 0.13% 0.13% 0.14% 0.13% 0.18%N 421,489 510,043 531,510 537,771 573,138
Panel B. 42 Days
Q1 Q2 Q3 Q4 Q5
Abn. Attention 0.412∗∗ 0.206∗∗ 0.222∗∗∗ 0.113∗∗∗ -0.012
(t-statistic) (2.41) (2.53) (3.06) (2.82) (-0.35)
Time FE ! ! ! ! !
Acct Holder FE ! ! ! ! !
R-Squared 0.25% 0.25% 0.27% 0.23% 0.32%N 390,144 471,115 490,610 496,060 529,837
Panel C. 63 Days
Q1 Q2 Q3 Q4 Q5
Abn. Attention 0.588∗∗ 0.271∗∗ 0.415∗∗∗ 0.161∗∗ -0.066
(t-statistic) (2.41) (2.49) (3.04) (2.80) (-1.31)
Time FE ! ! ! ! !
Acct Holder FE ! ! ! ! !
R-Squared 0.36% 0.36% 0.36% 0.33% 0.44%N 359,282 433,371 451,065 455,663 487,830
This table reports panel regression results on the relation between portfolio performance and investor attention for fivegroups of account holders (columns Q1 to Q5) based on the total number of seconds spent on the brokerage accountwebsite. Q1 denotes the bottom quintile, i.e. the least attentive group of account holders, while Q5 denotes the topquintile, i.e. the most attentive group of account holders. We estimate the following baseline panel regression:
DGTW Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time interval t : t + k;Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the difference between the (log) attentionon day t and the (log) average attention over the previous 21 business days; finally, αi and βm represent account-holderfixed effects and monthly time-effects, respectively. Attention is measured as the seconds spent on the brokerage accountwebsite. Panels A, B, and C report results for k = 21, 42, and 63 days, respectively. Displayed are the ordinary leastsquares coefficient estimates and associated t-statistics, computed using standard errors that are double-clustered byaccount-holder and time, see Petersen (2009). Coefficients marked with ***, **, and * are significant at the 1%, 5%, and10% level, respectively.
66
Tab
le15.
Tra
din
gP
erf
orm
an
ce
an
dIn
vest
or
Att
enti
on
:R
egre
ssio
nR
esu
lts
Con
dit
ion
ing
on
Sto
ckC
hara
cte
rist
ics
Pan
el
A.
Perf
orm
an
ce
of
Bu
ys
Volu
me
Siz
eV
ola
tili
tyD
isagreem
ent
Nu
m.
An
aly
sts
New
s
Low
Hig
hL
ow
Hig
hL
ow
Hig
hL
ow
Hig
hL
ow
Hig
hL
ow
Hig
h
Att
enti
on
0.3
46
0.7
06∗∗∗
0.2
61
0.7
64∗∗∗
0.3
30∗∗∗
0.7
73∗∗∗
0.2
33
0.8
24∗∗∗
0.3
18
0.7
11∗∗∗
0.0
91
0.8
59∗∗∗
(1.3
7)
(3.9
8)
(0.8
4)
(7.4
6)
(3.1
8)
(2.9
8)
(1.2
3)
(3.6
0)
(1.2
0)
(5.2
3)
(0.3
2)
(5.7
4)
Con
stant
0.1
40
0.5
35∗∗∗
0.2
61
0.4
16∗∗∗
0.1
78∗
0.6
10∗∗
0.7
93∗∗∗
0.4
35∗
0.5
27∗
0.3
87∗∗∗
0.2
88
0.4
23∗∗∗
(0.5
5)
(2.9
5)
(0.8
2)
(4.3
2)
(1.8
1)
(2.1
7)
(4.4
0)
(1.7
0)
(1.9
6)
(2.9
5)
(0.9
9)
(2.5
8)
R-S
qu
are
d0.0
2%
0.1
1%
0.0
1%
0.4
5%
0.1
0%
0.0
6%
0.0
1%
0.0
9%
0.0
1%
0.2
4%
0.0
0%
0.2
2%
N11,4
50
12,6
14
10,8
86
13,1
78
10,8
92
12,3
09
12,2
66
11,1
78
12,2
23
11,6
48
10,4
31
13,6
33
Pan
el
B.
Perf
orm
an
ce
of
Sell
s
Volu
me
Siz
eV
ola
tili
tyD
isagreem
ent
Nu
m.
An
aly
sts
New
s
Low
Hig
hL
ow
Hig
hL
ow
Hig
hL
ow
Hig
hL
ow
Hig
hL
ow
Hig
h
Att
enti
on
0.3
90
0.2
53
0.0
74
0.4
27∗∗∗
0.2
05∗
0.2
19
0.0
49
0.5
15∗∗
0.1
82
0.3
92∗∗∗
0.2
61
0.3
06∗
(1.5
5)
(1.2
5)
(0.2
3)
(4.1
0)
(1.8
8)
(0.7
6)
(0.2
7)
(1.9
6)
(0.6
6)
(2.5
7)
(0.8
6)
(1.8
1)
Con
stant
0.2
91
1.4
99∗∗∗
1.0
98∗∗∗
0.7
67∗∗∗
0.0
23
1.8
92∗∗∗
0.8
52∗∗∗
1.3
82∗∗∗
1.3
05∗∗∗
0.7
12∗∗∗
0.9
15∗∗∗
0.9
15∗∗∗
(1.1
5)
(7.1
1)
(3.3
3)
(7.0
5)
(0.2
0)
(6.2
8)
(4.4
9)
(5.0
0)
(4.5
2)
(4.8
8)
(3.0
7)
(5.2
7)
R-S
qu
are
d0.0
3%
0.0
2%
0.0
0%
0.1
5%
0.0
4%
0.0
1%
0.0
0%
0.0
4%
0.0
0%
0.0
8%
0.0
1%
0.0
3%
N9,4
01
9,9
19
8,7
11
10,6
09
9,0
57
9,6
85
9,9
28
8,8
96
9,7
44
9,4
27
8,2
90
11,0
30
This
table
rep
ort
sre
gre
ssio
nre
sult
son
the
rela
tion
bet
wee
nin
ves
tor
att
enti
on
and
tradin
gp
erfo
rmance
,co
ndit
ionin
gon
the
chara
cter
isti
csof
the
stock
str
aded
.W
ese
para
teth
ebuys
(i.e
.st
ock
purc
hase
s)and
the
sells
(i.e
.st
ock
sale
s)and
esti
mate
the
follow
ing
poole
dre
gre
ssio
ns:
DGTW
RetBuys i,j,t:t+k
=α
+βAttention
i,j,t
+ε i,j,t:t+k,
DGTW
RetSellsi,j,t:t+k
=α
+βAttention
i,j,t
+ε i,j,t:t+k,
wher
eDGTW
RetBuys i,j,t:t+k
(DGTW
RetSellsi,j,t:t+k)
are
the
cum
ula
tive
abnorm
al
retu
rns
of
secu
rityj
bought
(sold
)by
inves
tori
over
the
tim
ein
terv
alt
:t+k,
com
pute
dusi
ng
the
DG
TW
model
;Attention
i,j,t
isth
e(l
og)
num
ber
of
seco
nds
spen
ton
the
bro
ker
age
acc
ount
web
site
by
inves
tori
over
the
month
pre
cedin
gth
etr
ade
inst
ockj
that
occ
urs
at
tim
et.
Cum
ula
tive
abnorm
al
retu
rns
are
com
pute
dat
the
thre
e–m
onth
hori
zon.
Panel
Are
port
sre
sult
sfo
rst
ock
purc
hase
s,and
mea
sure
satt
enti
on
as
over
all
att
enti
on.
The
resu
lts
are
com
pute
dse
para
tely
for
stock
sw
ith
low
and
hig
hva
lues
of
the
condit
ionin
gva
riable
s.T
he
condit
ionin
gva
riable
suse
dare
–fr
om
left
tori
ght:
Vo
lum
e,th
evolu
me
of
the
stock
on
the
day
of
the
trade;
Siz
e,co
mpute
das
the
log
of
the
mark
etpri
cem
ult
iplied
by
the
num
ber
of
share
souts
tandin
g;
Vo
lati
lity
,co
mpute
das
the
realize
dvola
tility
of
the
stock
over
the
pre
vio
us
month
;D
isa
gree
men
t,th
est
andard
dev
iati
on
of
the
analy
sts’
earn
ings-
per
-share
fore
cast
s;N
um
.A
na
lyst
s,th
elo
gof
the
num
ber
of
analy
sts
cover
ing
the
stock
duri
ng
the
quart
erth
etr
ade
takes
pla
ce;
and
New
s,th
enum
ber
of
new
sfr
om
Capit
al
IQov
erth
epre
vio
us
month
.P
anel
Bre
pea
tsth
eex
erci
sefo
rst
ock
sale
s.D
ispla
yed
are
the
ord
inary
least
square
sco
effici
ent
esti
mate
sand
ass
oci
ate
dt-
stati
stic
s.Sta
ndard
erro
rsare
adju
sted
for
het
erosk
edast
icit
y.C
oeffi
cien
tsm
ark
edw
ith
***,
**,
and
*are
signifi
cant
at
the
1%
,5%
,and
10%
level
,re
spec
tivel
y.
67
Fig
ure
1:
This
figure
com
pare
sth
ew
eekly
tota
lnum
ber
of
seco
nds
spen
ton
the
bro
ker
age
acc
ount
web
site
–co
mpute
dacr
oss
all
acc
ount
hold
ers
–to
four
pro
xie
sof
inves
tor
att
enti
on
that
hav
eb
een
pro
pose
din
the
lite
ratu
re:
the
Sea
rch
Volu
me
Index
for
the
word
“S&
P500,”
rep
ort
edin
the
top-l
eft
panel
;th
eto
tal
num
ber
of
new
sp
erta
inin
gto
the
stock
sin
the
S&
P500,
rep
ort
edin
the
top-r
ight
panel
;th
etr
adin
gvolu
me
on
the
S&
P500,
rep
ort
edin
the
bott
om
-lef
tpanel
;and
the
Sta
teStr
eet
Inves
tor
Confiden
ceIn
dex
,re
port
edin
the
bott
om
-rig
ht
panel
.In
each
panel
,our
mea
sure
of
att
enti
on
isre
port
edas
are
ddash
edline,
while
the
alt
ernati
ve
mea
sure
of
att
enti
on
isre
port
edas
ablu
eso
lid
line.
68
Fig
ure
2:T
his
figure
rep
ort
sth
ehea
t-m
ap
of
the
daily
num
ber
of
min
ute
ssp
ent
on
the
bro
ker
age
acc
ount
web
site
by
each
acc
ount
hold
er.
Each
poin
ton
thex
-axis
repre
sents
abusi
nes
sday
,and
each
poin
ton
they-a
xis
repre
sents
an
acc
ount
hold
er.
By
focu
sing
on
one
valu
eon
they-a
xis
and
mov
ing
from
left
tori
ght,
the
reader
can
ass
ess
how
the
att
enti
on
of
an
indiv
idual
inves
tor
vari
esov
erti
me.
Lik
ewis
e,by
focu
sing
on
one
valu
eon
thex
-axis
and
mov
ing
from
top
tob
ott
om
,th
ere
ader
can
obse
rve
how
–fo
ra
giv
enday
–att
enti
on
vari
esacr
oss
acc
ount
hold
ers.
The
acc
ounts
are
sort
edby
the
tota
lnum
ber
of
min
ute
ssp
ent
on
the
bro
ker
age
acc
ount
web
site
over
the
full
sam
ple
,so
that
the
more
act
ive
inves
tors
are
at
the
top
of
the
figure
.T
he
num
ber
of
min
ute
sis
win
sori
zed
at
the
95-t
hp
erce
nti
le,
soth
at
the
num
ber
of
min
ute
ssp
ent
on
the
bro
ker
age
acc
ount
web
site
ranges
from
zero
to150.
69
On
e-M
onth
Th
ree-M
onth
Six
-Month
12-M
onth
Fig
ure
3:
This
figure
rep
ort
sth
ere
lati
on
bet
wee
nin
ves
tors
’att
enti
on
and
the
past
per
form
ance
of
the
stock
sth
eytr
ade
usi
ng
anon-p
ara
met
ric
sym
met
ric
nea
rest
nei
ghb
or
appro
ach
.A
tten
tion
isco
mpute
das
the
log
tota
lnum
ber
of
seco
nds
spen
tby
each
inves
tor
over
the
month
pre
cedin
gea
chtr
ade.
Sto
cks’
past
per
form
ance
isco
mpute
dov
erth
epre
vio
us
month
inth
eto
p-l
eft
panel
,th
epre
vio
us
thre
em
onth
sin
the
top-r
ight
panel
,th
epre
vio
us
six
month
sin
the
bott
om
-lef
tpanel
and
the
pre
vio
us
twel
ve
month
sin
the
bott
om
-rig
ht
panel
.In
each
panel
,th
etr
ades
are
div
ided
into
stock
purc
hase
sand
stock
sale
s.T
he
past
per
form
ance
of
the
stock
spurc
hase
dis
repre
sente
dby
aso
lid
red
line,
while
the
past
per
form
ance
of
the
stock
sso
ldis
repre
sente
dby
adash
edbla
ckline.
70
Tab
leO
nli
ne
I.S
um
mary
Sta
tist
ics
Pan
el
A.
Ch
aracte
ris
tics
of
Cli
ents
Pan
el
B.
Portf
oli
oC
haracte
ris
tics
Mean
Med
ian
St.
Dev
Mean
Med
ian
St.
Dev
Age
50.8
051
16.5
6P
ort
folio
Valu
e$90,0
00
$14,0
00
$422,0
00
Gen
der
0.7
21
0.4
5C
ash
Hold
ings
$15,0
00
$1,0
00
$75,0
00
Nu
mb
erof
Acc
ou
nts
1.2
51
0.6
0S
tock
Hold
ings
$77,0
00
$11,0
00
$352,0
00
Acc
ou
nt
age
8.5
57.6
35.5
1N
um
ber
of
Sto
cks
6.1
83
10.0
9
Pan
el
C.
Trad
ing
Beh
avio
r
Percenti
les
Mean
St.
Dev
1st
25th
50th
75th
99th
Fra
ctio
nof
Days
wit
hT
rad
es0.0
40.0
80
00.0
10.0
30.4
3
Days
Bet
wee
nT
rad
es55.2
465.4
81
12.5
031.2
070.3
3367
Nu
mb
erof
Tra
des
Per
Day
1.6
51.7
71
11.2
21.7
38
Dollar
Valu
eof
Tra
des
$17,0
00
$69,0
00
––
––
–
This
table
rep
ort
ssu
mm
ary
stati
stic
sof
the
bio
gra
phic
chara
cter
isti
cs(P
anel
A),
the
port
folio
chara
cter
isti
cs(P
anel
B),
and
the
tradin
g(P
anel
C)
beh
avio
rof
the
enti
reuniv
erse
of
acc
ounts
inour
data
,ex
cludin
gth
eones
inth
ew
eb-a
ctiv
ity
data
set.
For
each
vari
able
inea
chpanel
,w
ere
port
the
sam
ple
mea
n(M
ean
),m
edia
n(M
edian
),and
standard
dev
iati
on
(St.Dev
).T
he
stati
stic
sre
port
edin
Panel
sA
and
Bare
com
pute
dacr
oss
acc
ounts
.T
he
stati
stic
sin
Panel
Care
com
pute
dfirs
tin
the
tim
e-se
ries
dim
ensi
on
at
the
acc
ount-
hold
erle
vel
.T
hey
are
then
com
pute
dcr
oss
-sec
tionally
acr
oss
acc
ount
hold
ers.
All
dollar
valu
esw
ere
rounded
toth
enea
rest
thousa
nd
and
the
per
centi
les
of
“dollar
valu
eof
trades
”hav
eb
een
mask
edup
on
reques
tof
the
data
pro
vid
erfo
rco
nfiden
tiality
reaso
ns.
71
Table Online II. Characteristics of Attentive Investors
Panel A. Research Panel B. Balances & Positions
Coeff t-stat Magnitude Coeff t-stat Magnitude
Brokerage 0.407∗∗∗ (8.60) 31.899 0.371∗∗∗ (7.47) 128.979
Male 0.350∗∗∗ (7.63) 26.633 0.261∗∗∗ (5.39) 85.684
Age 0.218∗∗∗ (8.97) 15.446 0.286∗∗∗ (11.28) 94.878
Account Age 0.085∗∗∗ (3.67) 5.625 0.061∗∗ (2.71) 17.922
Portfolio Value 0.117∗∗∗ (2.70) 7.889 0.195∗∗∗ (2.77) 61.645
Fr. in Cash -0.077∗∗∗ (-4.30) -4.724 -0.197∗∗∗ (-9.28) -51.274
Fr. in ETF -0.058∗∗ (-2.33) -3.603 -0.104∗∗∗ (-6.09) -28.270
Fr. in Mutual Fund -0.012 (-0.97) -0.754 0.019 (0.50) 5.360
Beta Mkt 0.095∗∗∗ (4.90) 6.323 0.160∗∗∗ (6.65) 49.906
Beta SMB 0.003 (0.47) 0.222 0.091∗∗∗ (3.51) 27.324
Beta HML -0.041∗ (-1.81) -2.535 -0.061∗∗ (-2.62) -16.877
Beta MOM 0.097∗∗∗ (4.20) 6.458 0.113∗∗∗ (4.13) 34.214
Number of Stocks Traded 0.746∗∗∗ (28.86) 70.350 0.939∗∗∗ (38.92) 447.053
Constant 1.595∗∗∗ (39.32) 63.50 3.212∗∗∗ (75.03) 286.98
R-Squared 16.0% 22.9%N 8,574 8,574
This table reports regression results on the relation between the time spent on the brokerage account website and accountholder characteristics. We estimate the following cross-sectional regression:
Attentioni = α+ x′i β + εi for i = 1, . . . , N
where Attentioni is the total attention spent on the brokerage account website by account holder i over the sampleperiod, xi is a vector of covariates associated with account holder i and N is the total number of account holdersincluded in the analysis. Attention is measured as the log of the total number of minutes spent on the Research pagesof the brokerage account website in Panel A, and the log of the total number of minutes spent on the Balances andPositions pages of the brokerage account website in Panel B. What follows is a description of the covariates included.The first group comprises demographic variables: Brokerage, a brokerage account dummy; Male, a male dummy variable;Age, the age of the investor; Account Age, the age of the account. The second category comprises portfolio holdingsvariables: Portfolio Value, the total value of the invested portfolio; N. of Assets, the total number of stocks, mutualfunds and exchange traded funds held; Fr. in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of the totalwealth in the brokerage account held in cash, traded funds and mutual funds, respectively. The third category comprisesregressors related to portfolio risk and trading activity: Beta Mkt, Beta SMB, Beta HML, and Beta MOM, the loadingson the market, Small-Minus-Big, High-Minus-Low and Momentum factors computed using the full sample available usingdaily returns; and N. of Stocks Traded, the number of stocks traded over the period. Within each panel we display theordinary least squares coefficient estimates (Coeff), the associated t-statistics (t-stat), and the economic magnitudes(Magnitude) – computed as described in Section 4.1. Standard errors are adjusted for heteroskedasticity. Coefficientsmarked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
72
Table Online III. Robustness for Characteristics of Attentive Investors
Panel A. Pages Panel B. Logins
Coeff t-stat Magnitude Coeff t-stat Magnitude
Brokerage 0.412∗∗∗ (11.56) 203.028 0.250∗∗∗ (7.17) 21.565
Male 0.243∗∗∗ (6.91) 109.610 0.211∗∗∗ (6.30) 17.792
Age 0.197∗∗∗ (10.25) 86.925 0.212∗∗∗ (11.71) 17.917
Account Age 0.014 (0.82) 5.702 0.002 (0.15) 0.186
Portfolio Value 0.151∗∗∗ (2.65) 64.985 0.146∗∗∗ (2.85) 11.946
Fr. in Cash -0.095∗∗∗ (-6.56) -36.008 -0.113∗∗∗ (-8.04) -8.105
Fr. in ETF -0.068∗∗∗ (-4.80) -26.319 -0.064∗∗∗ (-4.83) -4.738
Fr. in Mutual Fund 0.012 (0.76) 4.685 0.007 (0.45) 0.496
Beta Mkt 0.103∗∗∗ (5.40) 43.085 0.087∗∗∗ (5.01) 6.888
Beta SMB 0.046∗∗∗ (2.65) 18.712 0.042∗∗ (2.56) 3.218
Beta HML -0.044∗∗ (-2.51) -16.988 -0.039∗∗ (-2.41) -2.905
Beta MOM 0.087∗∗∗ (4.70) 36.370 0.071∗∗∗ (4.21) 5.597
Number of Stocks Traded 0.715∗∗∗ (44.66) 415.940 0.637∗∗∗ (43.08) 67.604
Constant 4.771∗∗∗ (150.75) 398.253 3.244∗∗∗ (108.69) 75.888
R-Squared 24.3% 22.2%N 8,574 8,574
This table reports regression results on the relation between the time spent on the brokerage account website and thecharacteristics of the account holders. We estimate the following baseline cross-sectional regression:
Attentioni = α+ x′i β + εi for i = 1, . . . , N
where Attentioni is the total attention spent on the brokerage account website by account holder i over the sampleperiod, xi is a vector of covariates associated with account holder i and N is the total number of account holdersincluded in the analysis. Attention is measured as the log of the total number of pages browsed in Panel A, and thelog of the total number of logins in Panel B. What follows is a description of the covariates included. The first groupcomprises demographic variables: Brokerage, a brokerage account dummy; Male, a male dummy variable; Age, the age ofthe investor; Account Age, the age of the account. The second category comprises portfolio holdings variables: PortfolioValue, the total value of the invested portfolio; N. of Assets, the total number of stocks, mutual funds and exchangetraded funds held; Fr. in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of the total wealth in the brokerageaccount held in cash, traded funds and mutual funds, respectively. The third category comprises regressors relatedto portfolio risk and trading activity: Beta Mkt, Beta SMB, Beta HML, and Beta MOM, the loadings on the market,Small-Minus-Big, High-Minus-Low and Momentum factors computed using the full sample available using daily returns;and N. of Stocks Traded, the number of stocks traded over the period. Within each panel we display the ordinary leastsquares coefficient estimates (Coeff), the associated t-statistics (t-stat), and the economic magnitudes (Magnitude) –computed as described in Section 4.1. Standard errors are adjusted for heteroskedasticity. Coefficients marked with ***,**, and * are significant at the 1%, 5%, and 10% level, respectively.
73
Table Online IV. Investor Attention Allocation: Panel Regressions
Panel A. Number of Stocks Panel B. Herfindahl
Coeff t-stat Coeff t-stat
Total Attention 0.534∗∗∗ (100.30) -0.034∗∗∗ (-6.68)
Stock Portfolio Value 0.016 (0.28) 0.001 (0.06)
Herfindahl Stocks 0.007 (0.73) -0.017 (-1.59)
N. Stocks -0.007 (-0.17) 0.019 (0.70)
Beta Mkt -0.002 (-0.70) -0.008 (-0.99)
Beta SMB -0.006 (-1.30) 0.004 (0.63)
Beta HML 0.005 (1.61) -0.007 (-1.48)
Beta MOM -0.004 (-1.00) 0.004 (0.49)
Risk Idio 0.001 (0.35) -0.004 (-1.02)
N. of Stocks Traded 0.089∗∗∗ (6.41) -0.019∗∗∗ (-3.75)
Time FE ! !
Account Holder FE ! !
R-Squared 50.2% 1.5%
N 32,852 13,256
This table reports regression results on the relation between attention allocation and the time-varying characteristics ofaccount holders. In Panel A, we estimate the following panel regression:
N Stocks Researchedi,m = αi + βm + x′i,m γ + εi,m, for i = 1, . . . , N & m = 1, . . . ,M
where N Stocks Researchedi,m is the (log) number of stocks searched by investor i in month m, αi identifies account-holder fixed-effects, βm monthly time-effects, and x′i,m is a vector of covariates that varies at the monthly frequency. InPanel B, we estimate the following panel regression:
Herfindahli,m = αi + βm + x′i,m γ + εi,m, for i = 1, . . . , N & m = 1, . . . ,M
where Herfindahli,m is the normalized Herfindahl index of the stocks researched by investor i in month m – computed
as∑K
k=1 ω2k,m−1/K
1−1/Kif K > 1 and 1 if K = 1, where ωk,m is the fraction of the total stock-specific attention allocated to
stock k in month m and K is the total number of stocks searched in month m. What follows is a description of thecovariates included. The first category of covariates contains Total Attention, defined as the log total number of secondsspent by the investor on the brokerage account website over month m. The second category comprises portfolio holdingsvariables: Stock Portfolio Value, the total value of the invested portfolio as of the end of month m−1; Herfindahl Stocks,
the normalized Herfindahl index of investor i’s stock-holdings as of the end of month m−1, computed as∑J
j=1 ω2j,m−1/J
1−1/Jif
J > 1 and 1 if J = 1, where wj,m is the fraction of wealth allocated to stock j at the end of month m−1 and J is the totalnumber of stocks held at the end of month m− 1; N. Stocks, the total number of stocks held at the end of month m− 1;The third category comprises regressors related to the performance and risk of the investors’ portfolios, measured at theend of month m − 1: Beta Mkt, Beta SMB, Beta HML, and Beta MOM, the loadings on the market, Small-Minus-Big,High-Minus-Low and Momentum factors, computed using a 90-day rolling window; Risk Idio, the idiosyncratic risk ofthe investor’s portfolio, computed over month m. Finally, the last covariate is N. of Stocks Traded: the number of stockstraded over month m. Within each panel we display the ordinary least squares coefficient estimates (Coeff) and theassociated t-statistic (t-stat). Standard errors are clustered at the account level. Coefficients marked with ***, **, and* are significant at the 1%, 5%, and 10% level, respectively.
74
Table Online V. Characteristics of Researched Stocks
Panel A. Pages Panel B. Visits
Coeff t-stat Magnitude Coeff t-stat Magnitude
Analyst 0.421∗∗∗ (12.61) 22.874 0.401∗∗∗ (12.51) 14.007
News 0.076∗∗ (2.29) 3.446 0.079∗∗ (2.47) 2.337
Turnover 0.810∗∗∗ (4.58) 54.417 0.748∗∗∗ (4.41) 31.576
Volume 0.274∗∗∗ (6.62) 13.763 0.258∗∗∗ (6.51) 8.350
Alpha 0.000 (0.01) 0.012 0.000 (0.00) 0.002
Beta Mkt 0.233∗∗∗ (4.96) 11.452 0.227∗∗∗ (5.10) 7.247
Beta SMB -0.057∗∗ (-2.08) -2.416 -0.062∗∗ (-2.43) -1.717
Beta HML -0.030 (-1.13) -1.290 -0.036 (-1.44) -1.004
Beta MOM 0.073∗∗∗ (2.94) 3.324 0.064∗∗∗ (2.65) 1.864
Return 0.030 (1.08) 1.331 0.036 (1.39) 1.050
Variance 0.054 (1.17) 2.415 0.046 (1.03) 1.346
Skewness -0.053∗∗ (-2.04) -2.263 -0.045∗ (-1.81) -1.238
Kurtosis 0.047∗ (1.69) 2.108 0.039 (1.47) 1.133
MktBook 0.116∗∗∗ (3.29) 5.356 0.111∗∗∗ (3.37) 3.323
R&D 0.062 (1.64) 2.812 0.066∗ (1.82) 1.929
Profitability 0.028 (0.67) 1.241 0.016 (0.40) 0.454
Age 0.011 (0.54) 0.474 0.028 (1.50) 0.816
Size 0.198∗∗∗ (8.57) 9.561 0.193∗∗∗ (8.71) 6.043
Leverage 0.150∗∗∗ (5.02) 7.059 0.152∗∗∗ (5.38) 4.649
Inst. Investors -0.336∗∗∗ (-8.52) -12.450 -0.320∗∗∗ (-8.57) -7.777
Constant 3.252∗∗∗ (74.74) 43.632 2.873∗∗∗ (69.45) 28.399
R-Squared 48.4% 49.7%N 2,572 2,572
This table reports regression results on the relation between investor attention – computed across all account-holders inour data – and stock characteristics. We estimate the following cross-sectional regression:
Attentionj = α+ x′j β + εj , for j = 1, . . . , J,
where Attentionj is the total attention spent – across all account-holders – on stock j and xj is a vector of stock-specificcovariates. Attention is measured as the log total number of pages in Panel A, and log total number of visits in PanelB. What follows is a description of the three groups of covariates included in the analysis. The first group comprisesregressors we classify as attention proxies: Analyst, the number of analysts covering the stock; News, the total numberof news associated with the stock, Volume, the stock’s volume; Turnover, computed as the stock’s volume divided bythe shares outstanding. The second category comprises regressors related to stocks’ risk and performance. Alpha, BetaMkt, Beta SMB, Beta HML, and Beta MOM are the intercept and the loadings on the market, SMB, HML and MOMfactors. The loadings are estimated for each stock j using daily data over the full sample. Return, Variance, Skewnessand Kurtosis, are the realized returns, variance, skewness and kurtosis for each stock j – computed using daily returns.The third category comprises regressors related to stock characteristics: MktBook, the ratio of market value of assetsand the book value of assets, R&D, computed as the ratio of research and development expenses (COMPUSTAT item:XRDQ) and sales (COMPUSTAT item: SALEQ); Profitability, the ratio between operating income before depreciation(COMPUSTAT item: OIBDPQ) and total assets (COMPUSTAT item: ATQ); Age, the number of years a company hasbeen in the CRSP dataset; Size, the market value of assets computed as the product of the price and shares outstanding;Leverage, computed as the sum of long-term debt and debt in current liabilities, divided by the market value of assets;Inst. Investors, the fraction of the shares outstanding owned by institutional investors. Within each panel we displaythe ordinary least squares coefficient estimates (Coeff), the associated t-statistic (t-stat), and the economic magnitude(Magnitude) – computed as described in Section 4.1. Standard errors are adjusted for heteroskedasticity. Coefficientsmarked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
75
Table Online VI. Attention and Portfolio Performance: Cross-Sectional Results
Panel A. Research Panel B. Balances & Positions
Spec 1 Spec 2 Spec 3 Spec 1 Spec 2 Spec 3
Attention 0.988∗∗ 1.309∗∗∗ 1.776∗∗∗ 1.267∗∗∗ 1.321∗∗∗ 2.007∗∗∗
(2.24) (3.06) (3.83) (2.64) (2.66) (3.59)
Brokerage 2.203∗∗ 2.238∗∗∗ 2.222∗∗ 2.231∗∗
(2.57) (2.60) (2.56) (2.54)
Male 0.669 0.659 0.734 0.734(0.76) (0.74) (0.82) (0.82)
Age -0.934∗∗ -0.832∗∗ -0.971∗∗ -0.892∗∗
(-2.39) (-2.10) (-2.44) (-2.21)
Account Age 0.016 0.021 0.021 0.025(0.04) (0.05) (0.05) (0.06)
Portfolio Value -0.481∗∗ -0.539∗∗
(-2.25) (-2.30)
Fr. in Cash 0.080 0.193(0.11) (0.26)
Fr. in ETF 0.385 0.444(0.89) (1.04)
Fr. in Mutual Fund 0.566 0.557(1.60) (1.58)
N. of Stocks Traded -1.045∗∗∗ -1.225∗∗∗
(-3.51) (-3.81)
Constant 2.004∗∗∗ 0.633 0.778 1.897∗∗∗ 0.451 0.562(4.45) (0.82) (0.99) (4.06) (0.59) (0.71)
R-Squared 0.1% 0.2% 0.3% 0.1% 0.2% 0.3%N 8,340 7,476 7,468 8,340 7,476 7,468
This table reports regression results on the relation between portfolio performance and investor attention. We estimatethe following baseline cross-sectional regression:
AV G DGTW Reti = α+ β Attentioni + x′i γ + εi for i = 1, . . . , N
where AV G DGTW Reti is the annualized percentage average DGTW abnormal return of investor i over the sample,Attentioni is the total attention spent on the brokerage website by account holder i over the sample period, xi is a vectorof covariates associated with account holder i and N is the total number of account holders included in the analysis.Attention is measured as the log of the total number of minutes spent on the Research pages of the brokerage accountwebsite in Panel A, and the log of the total number of minutes spent on the Balances and Positions pages of the brokerageaccount website in Panel B. Each panel contains three specifications. The first specification uses attention as the solecovariate. The second specification includes a group of covariates that control for investor demographic characteristics:Brokerage, a brokerage account dummy; Male, a male dummy variable; Age, the age of the investor; Account Age, theage of the account. The third specification includes portfolio holdings and trading activity variables: Portfolio Value,the total value of the invested portfolio; Fr. in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of the totalwealth in the brokerage account held in cash, traded funds and mutual funds, respectively; and N. of Stocks Traded, thenumber of stocks traded over the period. Displayed are the ordinary least squares coefficient estimates and associatedt-statistics. Standard errors are adjusted for heteroskedasticity. Coefficients marked with ***, **, and * are significantat the 1%, 5%, and 10% level, respectively.
76
Table Online VII. Robustness for Attention and Portfolio Performance:
Cross-Sectional Results
Panel A. Pages Panel B. Logins
Attention 0.635∗∗∗ 0.665∗∗∗ 1.065∗∗∗ 0.597∗∗ 0.640∗∗ 1.011∗∗∗
(2.58) (2.65) (3.76) (2.15) (2.28) (3.26)
Brokerage 2.158∗∗ 2.103∗∗ 2.269∗∗∗ 2.292∗∗∗
(2.48) (2.39) (2.61) (2.61)
Male 0.709 0.676 0.740 0.723(0.80) (0.76) (0.83) (0.81)
Age -0.946∗∗ -0.869∗∗ -0.946∗∗ -0.876∗∗
(-2.41) (-2.18) (-2.40) (-2.19)
Account Age 0.051 0.070 0.059 0.085(0.12) (0.17) (0.14) (0.20)
Portfolio Value -0.543∗∗ -0.532∗∗
(-2.30) (-2.31)
Fr. in Cash 0.130 0.142(0.18) (0.19)
Fr. in ETF 0.417 0.416(0.97) (0.97)
Fr. in Mutual Fund 0.562 0.572(1.59) (1.62)
Number of Stocks Traded -1.214∗∗∗ -1.104∗∗∗
(-3.90) (-3.61)
Constant -1.310 -2.827∗ -4.696∗∗∗ -0.108 -1.722 -2.879∗∗
(-0.91) (-1.92) (-2.96) (-0.09) (-1.41) (-2.25)
R-Squared 0.1% 0.2% 0.3% 0.1% 0.2% 0.3%N 8,340 7,476 7,468 8,340 7,476 7,468
This table reports regression results on the relation between portfolio performance and investor attention. We estimatethe following baseline cross-sectional regression:
AV G DGTW Reti = α+ β Attentioni + x′i γ + εi for i = 1, . . . , N
where AV G DGTW Reti is the annualized percentage average DGTW abnormal return of investor i over the sample,Attentioni is the total attention spent on the brokerage web-site by account holder i over the sample period, xi isa vector of covariates associated with account holder i and N is the total number of account holders included in theanalysis. Attention is measured as the log of the total number of pages browsed in Panel A, and the log of the totalnumber Logins in Panel B. Each panel contains three specifications. The first specification uses attention as the solecovariate. The second specification includes a group of covariates that control for investor demographic characteristics:Brokerage, a brokerage account dummy; Male, a male dummy variable; Age, the age of the investor; Account Age, theage of the account. The third specification includes portfolio holdings and trading activity variables: Portfolio Value,the total value of the invested portfolio; Fr. in Cash, Fr. in ETF, and Fr. in Mutual Fund, the fraction of the totalwealth in the brokerage account held in cash, traded funds and mutual funds, respectively; and N. of Stocks Traded, thenumber of stocks traded over the period. Displayed are the ordinary least squares coefficient estimates and associatedt-statistics. Standard errors are adjusted for heteroskedasticity. Coefficients marked with ***, **, and * are significantat the 1%, 5%, and 10% level, respectively.
77
Table Online VIII. Attention and Investor Sharpe Ratio: Cross-Sectional Results
Panel A. Overall Panel B. Research Panel C. Balances &Positions
Spec 1 Spec 2 Spec 3 Spec 1 Spec 2 Spec 3 Spec 1 Spec 2 Spec 3
Attention 0.090∗∗∗ 0.086∗∗∗ 0.049∗∗∗ 0.108∗∗∗ 0.103∗∗∗ 0.065∗∗∗ 0.137∗∗∗ 0.132∗∗∗ 0.089∗∗∗
(7.08) (6.55) (3.87) (8.70) (8.13) (5.43) (10.61) (9.98) (6.94)
Brokerage -0.059∗∗ -0.049∗∗ -0.054∗∗ -0.048∗∗ -0.055∗∗ -0.049∗∗
(-2.39) (-2.21) (-2.19) (-2.17) (-2.26) (-2.24)
Male -0.046∗ -0.010 -0.049∗ -0.013 -0.049∗ -0.012(-1.79) (-0.44) (-1.93) (-0.58) (-1.93) (-0.56)
Age 0.016 0.025∗∗ 0.018 0.026∗∗ 0.010 0.022∗∗
(1.27) (2.33) (1.42) (2.36) (0.80) (1.98)
Account Age 0.094∗∗∗ 0.063∗∗∗ 0.092∗∗∗ 0.061∗∗∗ 0.092∗∗∗ 0.061∗∗∗
(7.36) (5.60) (7.18) (5.46) (7.20) (5.48)
Portfolio Value 0.073∗∗∗ 0.073∗∗∗ 0.070∗∗∗
(5.36) (5.33) (5.43)
Fr. in Cash -0.150∗∗∗ -0.151∗∗∗ -0.145∗∗∗
(-9.23) (-9.28) (-8.88)
Fr. in ETF -0.027 -0.027 -0.025(-1.15) (-1.17) (-1.09)
Fr. in Mutual Fund 0.017 0.018 0.017(1.41) (1.43) (1.36)
Beta Mkt 0.465∗∗∗ 0.464∗∗∗ 0.462∗∗∗
(30.45) (30.49) (30.23)
Beta SMB -0.186∗∗∗ -0.184∗∗∗ -0.187∗∗∗
(-13.87) (-13.75) (-13.96)
Beta HML -0.007 -0.007 -0.007(-0.56) (-0.55) (-0.50)
Beta MOM 0.305∗∗∗ 0.305∗∗∗ 0.304∗∗∗
(20.77) (20.76) (20.72)
N. of Stocks Traded 0.021∗ 0.018 0.005(1.65) (1.42) (0.38)
Constant 1.308∗∗∗ 1.368∗∗∗ 1.313∗∗∗ 1.309∗∗∗ 1.369∗∗∗ 1.315∗∗∗ 1.298∗∗∗ 1.358∗∗∗ 1.309∗∗∗
(108.03) (57.52) (62.06) (108.26) (57.97) (62.67) (107.10) (57.78) (62.67)
R-Squared 0.6% 1.4% 24.3% 0.9% 1.6% 24.4% 1.4% 2.1% 24.6%N 8,641 8,512 8,500 8,641 8,512 8,500 8,641 8,512 8,500
This table reports regression results on the relation between portfolio performance and investor attention. We estimatethe following baseline cross-sectional regression:
Sharpei = α+ β Attentioni + x′i γ + εi for i = 1, . . . , N
where Sharpei is the Sharpe ratio of investor i over the sample, Attentioni is the total attention spent on the brokerageweb-site by account holder i over the sample period, xi is a vector of covariates associated with account holder i and Nis the total number of account holders included in the analysis. Attention is measured as the log of the total number ofminutes spent on the brokerage account website in Panel A, the log of the total number of minutes spent on the Researchpages of the brokerage account website in Panel B, and the log of the total number of minutes spent on the Balancesand Positions pages of the brokerage account website in Panel C. Each panel contains three specifications. The firstspecification uses attention as the sole covariate. The second specification includes a group of covariates that control forinvestor demographic characteristics: Brokerage, a brokerage account dummy; Male, a male dummy variable; Age, theage of the investor; Account Age, the age of the account. The third specification includes portfolio holdings, portfoliorisk and trading activity variables: Portfolio Value, the total value of the invested portfolio; Fr. in Cash, Fr. in ETF,and Fr. in Mutual Fund, the fraction of the total wealth in the brokerage account held in cash, traded funds and mutualfunds, respectively; Beta Mkt, Beta SMB, Beta HML, and Beta MOM, the loadings on the market, Small-Minus-Big,High-Minus-Low and Momentum factors computed using daily returns over the full sample; and N. of Stocks Traded, thenumber of stocks traded over the period. Displayed are the ordinary least squares coefficient estimates and associatedt-statistics. Standard errors are adjusted for heteroskedasticity. Coefficients marked with ***, **, and * are significantat the 1%, 5%, and 10% level, respectively.
78
Table Online IX. Attention and Performance: Panel Regression Results
Panel A. Research
21 Days 42 Days 63 Days
Abn. Attention 0.023 0.078∗∗ 0.105∗∗
(t-statistic) (1.37) (2.57) (2.28)
[t-statistic] [1.72] [4.46] [4.57]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.14% 0.27% 0.38%
N 2,573,951 2,377,766 2,187,211
Panel B. Balances and Positions
21 Days 42 Days 63 Days
Abn. Attention 0.046∗∗∗ 0.135∗∗∗ 0.175∗∗∗
(t-statistic) (3.22) (5.64) (5.18)
[t-statistic] [4.80] [8.41] [10.59]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.14% 0.27% 0.38%
N 2,573,951 2,377,766 2,187,211
This table reports panel regression results on the relation between portfolio performance and investor attention. Weestimate the following baseline panel regression:
DGTW Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time interval t : t + k;Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the difference between the (log) attentionon day t and the (log) average attention over the previous 21 business days; finally, αi and βm represent account-holderfixed effects and monthly time-effects, respectively. Attention is measured as the number of seconds spent on the Researchpages of the brokerage account website in Panel A, and the number of seconds spent on the Balances and Positions pagesof the brokerage account website in Panel B. Each panel reports results for different horizons, i.e. k = 21, 42, and 63 days.Displayed are the ordinary least squares coefficient estimates and two sets of t-statistics. The first – in round brackets– are computed using standard errors that are double-clustered by account-holder and time, see Petersen (2009). Thesecond – in square brackets – are computed using the Driscoll and Kraay (1998) standard errors. Coefficients markedwith ***, **, and * are significant at the 1%, 5%, and 10% level, respectively, according to the double-clustered standarderrors.
79
Table Online X. Robustness for Attention and Performance: Panel RegressionResults with Abnormal Attention Computed over a 42 Day Window
Panel A. Overall
21 Days 42 Days 63 Days
Abn. Attention 0.064∗∗∗ 0.139∗∗∗ 0.166∗∗∗
(t-statistic) (3.56) (4.09) (3.18)
[t-statistic] [4.88] [7.37] [8.21]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.38%
N 2,425,736 2,230,535 2,041,229
Panel B. Research
21 Days 42 Days 63 Days
Abn. Attention 0.061∗∗∗ 0.138∗∗∗ 0.171∗∗∗
(t-statistic) (3.05) (3.55) (2.91)
[t-statistic] [3.68] [6.88] [6.27]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.38%
N 2,425,736 2,230,535 2,041,229
Panel C. Balances and Positions
21 Days 42 Days 63 Days
Abn. Attention 0.077∗∗∗ 0.173∗∗∗ 0.212∗∗∗
(t-statistic) (4.33) (5.36) (4.60)
[t-statistic] [5.99] [8.71] [10.14]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.38%
N 2,425,736 2,230,535 2,041,229
This table reports panel regression results on the relation between portfolio performance and investor attention. Weestimate the following baseline panel regression:
DGTW Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time interval t : t + k;Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the difference between the (log) attentionon day t and the (log) average attention over the previous 42 business days; finally, αi and βm represent account-holderfixed effects and monthly time-effects, respectively. Attention is measured as the number of seconds spent on thebrokerage account website in Panel A, the number of seconds spent on the Research pages of the brokerage accountwebsite in Panel B, and the number of seconds spent on the Balances and Positions pages of the brokerage accountwebsite in Panel C. Each panel reports results for different horizons, i.e. k = 21, 42, and 63 days. Displayed are theordinary least squares coefficient estimates and two sets of t-statistics. The first – in round brackets – are computedusing standard errors that are double-clustered by account-holder and time, see Petersen (2009). The second – in squarebrackets – are computed using the Driscoll and Kraay (1998) standard errors. Coefficients marked with ***, **, and *are significant at the 1%, 5%, and 10% level, respectively, according to the double-clustered standard errors.
80
Table Online XI. Robustness for Attention and Performance: Panel RegressionResults with Abnormal Attention Computed over a 63 Day Window
Panel A. Overall
21 Days 42 Days 63 Days
Abn. Attention 0.074∗∗∗ 0.133∗∗∗ 0.103∗
(t-statistic) (3.58) (3.37) (1.70)
[t-statistic] [4.78] [6.93] [4.56]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.39%
N 2,277,370 2,083,511 1,895,070
Panel B. Research
21 Days 42 Days 63 Days
Abn. Attention 0.085∗∗∗ 0.171∗∗∗ 0.183∗∗∗
(t-statistic) (3.85) (3.96) (2.86)
[t-statistic] [4.50] [7.01] [6.30]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.39%
N 2,277,370 2,083,511 1,895,070
Panel C. Balances and Positions
21 Days 42 Days 63 Days
Abn. Attention 0.089∗∗∗ 0.178∗∗∗ 0.165∗∗∗
(t-statistic) (4.31) (4.67) (3.04)
[t-statistic] [5.45] [8.03] [6.27]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.39%
N 2,277,370 2,083,511 1,895,070
This table reports panel regression results on the relation between portfolio performance and investor attention. Weestimate the following baseline panel regression:
DGTW Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time interval t : t + k;Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the difference between the (log) attentionon day t and the (log) average attention over the previous 63 business days; finally, αi and βm represent account-holderfixed effects and monthly time-effects, respectively. Attention is measured as the number of seconds spent on thebrokerage account website in Panel A, the number of seconds spent on the Research pages of the brokerage accountwebsite in Panel B, and the number of seconds spent on the Balances and Positions pages of the brokerage accountwebsite in Panel C. Each panel reports results for different horizons, i.e. k = 21, 42, and 63 days. Displayed are theordinary least squares coefficient estimates and two sets of t-statistics. The first – in round brackets – are computedusing standard errors that are double-clustered by account-holder and time, see Petersen (2009). The second – in squarebrackets – are computed using the Driscoll and Kraay (1998) standard errors. Coefficients marked with ***, **, and *are significant at the 1%, 5%, and 10% level, respectively, according to the double-clustered standard errors.
81
Table Online XII. Robustness for Attention and Performance:Panel Regression Results
Panel A. Pages
21 Days 42 Days 63 Days
Abn. Attention 0.018∗ 0.057∗∗∗ 0.061∗∗∗
(t-statistic) (1.72) (3.59) (2.90)
[t-statistic] [2.13] [3.83] [4.32]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.14% 0.27% 0.38%
N 2,573,951 2,377,766 2,187,211
Panel B. Logins
21 Days 42 Days 63 Days
Abn. Attention 0.007 0.032∗∗∗ 0.024
(t-statistic) (0.79) (2.66) (1.56)
[t-statistic] [0.83] [2.41] [1.76]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.14% 0.27% 0.38%
N 2,573,951 2,377,766 2,187,211
This table reports panel regression results on the relation between portfolio performance and investor attention. Weestimate the following baseline panel regression:
DGTW Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time interval t : t + k;Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the difference between the (log) attentionon day t and the (log) average attention over the previous 21 business days; finally, αi and βm represent account-holderfixed effects and monthly time-effects, respectively. Attention is measured as the number of pages browsed on thebrokerage account website in Panel A, and the number of logins to the brokerage account website in Panel B. Each panelreports results for different horizons, i.e. k = 21, 42, and 63 days. Displayed are the ordinary least squares coefficientestimates and two sets of t-statistics. The first – in round brackets – are computed using standard errors that aredouble-clustered by account-holder and time, see Petersen (2009). The second – in square brackets – are computed usingthe Driscoll and Kraay (1998) standard errors. Coefficients marked with ***, **, and * are significant at the 1%, 5%,and 10% level, respectively, according to the double-clustered standard errors.
82
Table Online XIII. Attention and Performance Net of Trading Fees:Panel Regression Results
Panel A. Overall
21 Days 42 Days 63 Days
Abn. Attention 0.027∗ 0.090∗∗∗ 0.110∗∗∗
(t-statistic) (1.89) (3.70) (2.93)
[t-statistic] [2.92] [6.52] [6.99]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.39%
N 2,492,150 2,299,617 2,113,051
Panel B. Research
21 Days 42 Days 63 Days
Abn. Attention 0.021 0.074∗∗ 0.096∗∗
(t-statistic) (1.31) (2.51) (2.12)
[t-statistic] [1.72] [4.35] [4.34]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.39%
N 2,492,150 2,299,617 2,113,051
Panel C. Balances and Positions
21 Days 42 Days 63 Days
Abn. Attention 0.042∗∗∗ 0.125∗∗∗ 0.158∗∗∗
(t-statistic) (2.97) (5.26) (4.74)
[t-statistic] [4.25] [7.61] [9.75]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 0.15% 0.27% 0.39%
N 2,492,150 2,299,617 2,113,051
This table reports panel regression results on the relation between portfolio performance and investor attention. Weestimate the following baseline panel regression:
DGTW NET Reti,t:t+k = αi + βm + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW NET Reti,t:t+k is the DGTW-adjusted portfolio return – net of trading fees – of account-holder i overthe time interval t : t+ k; Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the differencebetween the (log) attention on day t and the (log) average attention over the previous 21 business days; finally, αi andβm represent account-holder fixed effects and monthly time-effects, respectively. Attention is measured as the numberof seconds spent on the brokerage account website in Panel A, the number of seconds spent on the Research pages ofthe brokerage account website in Panel B, and the number of seconds spent on the Balances and Positions pages of thebrokerage account website in Panel C. Each panel reports results for different horizons, i.e. k = 21, 42, and 63 days.Displayed are the ordinary least squares coefficient estimates and two sets of t-statistics. The first – in round brackets– are computed using standard errors that are double-clustered by account-holder and time, see Petersen (2009). Thesecond – in square brackets – are computed using the Driscoll and Kraay (1998) standard errors. Coefficients markedwith ***, **, and * are significant at the 1%, 5%, and 10% level, respectively, according to the double-clustered standarderrors.
83
Table Online XIV. Attention and Performance: Panel Regression Resultswith Daily Time-Effects
Panel A. Overall
21 Days 42 Days 63 Days
Abn. Attention 0.025∗ 0.097∗∗∗ 0.134∗∗∗
(t-statistic) (1.73) (3.99) (3.28)
[t-statistic] [2.72] [5.92] [8.89]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 1.54% 1.45% 0.98%
N 2,573,951 2,377,766 2,187,211
Panel B. Research
21 Days 42 Days 63 Days
Abn. Attention 0.022 0.081∗∗∗ 0.108∗∗
(t-statistic) (1.32) (2.67) (2.31)
[t-statistic] [1.68] [3.54] [4.86]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 1.54% 1.45% 0.97%
N 2,573,951 2,377,766 2,187,211
Panel C. Balances and Positions
21 Days 42 Days 63 Days
Abn. Attention 0.038∗∗∗ 0.129∗∗∗ 0.170∗∗∗
(t-statistic) (2.62) (5.39) (4.72)
[t-statistic] [4.04] [7.86] [10.39]
Time FE ! ! !
Account Holder FE ! ! !
R-Squared 1.54% 1.45% 0.98%
N 2,573,951 2,377,766 2,187,211
This table reports panel regression results on the relation between portfolio performance and investor attention. Weestimate the following baseline panel regression:
DGTW Reti,t:t+k = αi + βt + γ Abn Attentioni,t + εi,t:t+k for i = 1, . . . , N & t = 1, . . . , T,
where DGTW Reti,t:t+k is the DGTW-adjusted portfolio return of account-holder i over the time interval t : t + k;Abn Attentioni,t is account-holder i abnormal attention at time t, computed as the difference between the (log) attentionon day t and the (log) average attention over the previous 21 business days; finally, αi and βt represent account-holderfixed effects and daily time-effects, respectively. Attention is measured as the number of seconds spent on the brokerageaccount website in Panel A, the number of seconds spent on the Research pages of the brokerage account website inPanel B, and the number of seconds spent on the Balances and Positions pages of the brokerage account website in PanelC. Each panel reports results for different horizons, i.e. k = 21, 42, and 63 days. Displayed are the ordinary least squarescoefficient estimates and two sets of t-statistics. The first – in round brackets – are computed using standard errors thatare double-clustered by account-holder and time, see Petersen (2009). The second – in square brackets – are computedusing the Driscoll and Kraay (1998) standard errors. Coefficients marked with ***, **, and * are significant at the 1%,5%, and 10% level, respectively, according to the double-clustered standard errors.
84
Table Online XV. Trading Performance and Investor Attention: Baseline Regression Results
Panel A. Performance of Buys
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Research 0.393∗∗∗ 0.536∗∗∗ 0.489∗∗∗ 0.339∗ 0.352 0.051 -1.773∗∗∗
(4.55) (4.17) (3.01) (1.70) (1.63) (0.24) (-5.94)
Constant -0.140∗ 0.156 0.363∗∗ 0.489∗∗∗ 0.635∗∗∗ 0.699∗∗∗ 0.536∗
(-1.70) (1.31) (2.35) (2.63) (3.27) (3.33) (1.75)
R-Squared 0.10% 0.08% 0.04% 0.02% 0.01% 0.00% 0.15%N 24,139 24,103 24,064 24,001 23,947 23,887 23,436
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Balances & 0.079 0.230∗ 0.408∗∗∗ 0.276 0.314 -0.001 0.135Positions (1.03) (1.88) (2.63) (1.51) (1.53) (-0.01) (0.41)
Constant -0.133 0.165 0.369∗∗ 0.494∗∗∗ 0.639∗∗∗ 0.700∗∗∗ 0.502(-1.64) (1.37) (2.32) (2.58) (3.28) (3.33) (1.65)
R-Squared 0.00% 0.02% 0.03% 0.01% 0.01% 0.00% 0.00%N 24,139 24,103 24,064 24,001 23,947 23,887 23,436
Panel B. Performance of Sells
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Research 0.196∗∗ 0.344∗∗ 0.312∗ 0.073 0.019 -0.367 -2.618∗∗∗
(2.05) (2.60) (1.72) (0.37) (0.09) (-1.64) (-8.04)
Constant -0.014 0.392∗∗∗ 0.928∗∗∗ 1.012∗∗∗ 1.090∗∗∗ 1.166∗∗∗ 1.246∗∗∗
(-0.16) (3.07) (5.64) (5.41) (5.15) (5.06) (3.58)
R-Squared 0.03% 0.04% 0.02% 0.00% 0.00% 0.01% 0.29%N 19,435 19,373 19,320 19,257 19,200 19,142 18,777
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Balances & 0.060 0.041 0.149 -0.038 0.006 -0.050 -0.158Positions (0.64) (0.31) (0.82) (-0.19) (-0.03) (-0.20) (-0.37)
Constant -0.025 0.375∗∗∗ 0.909∗∗∗ 1.010∗∗∗ 1.089∗∗∗ 1.184∗∗∗ 1.372∗∗∗
(-0.29) (2.98) (5.60) (5.44) (5.16) (5.14) (3.93)
R-Squared 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%N 19,435 19,373 19,320 19,257 19,200 19,142 18,777
This table reports regression results on the relation between investor attention and trading performance. We separatethe buys (i.e. stock purchases) and the sells (i.e. stock sales) and estimate the following pooled regressions:
DGTW Ret Buysi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
DGTW Ret Sellsi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
where DGTW Ret Buysi,j,t:t+k (DGTW Ret Sellsi,j,t:t+k) are the cumulative abnormal returns of security j bought(sold) by investor i over the time interval t : t + k, computed using the DGTW model; Attentioni,j,t is the (log)number of seconds spent on the brokerage account website by investor i over the month preceding the trade in stockj that occurs at time t. Cumulative abnormal returns are computed at the one-, two-, three-, four-, five-, six- andtwelve-month horizons. Panel A reports results for stock purchases, and measures attention as, alternatively, the totalnumber of seconds spent on the research section of the brokerage account website, and the total number of seconds spenton the balances and positions section of the brokerage account website. Panel B repeats the exercise for stock sales.Displayed are the ordinary least squares coefficient estimates and associated t-statistics. Standard errors are adjustedfor heteroskedasticity. Coefficients marked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
85
Table Online XVI. Robustness for Trading Performance and Investor Attention
Panel A. Performance of Buys
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Pages 0.068 0.246∗∗ 0.429∗∗∗ 0.249 0.293 0.117 -0.175(0.81) (1.98) (2.63) (1.28) (1.34) (0.52) (-0.54)
Constant -0.135∗ 0.156 0.354∗∗ 0.485∗∗∗ 0.629∗∗∗ 0.695∗∗∗ 0.510∗
(-1.64) (1.31) (2.38) (2.59) (3.25) (3.30) (1.67)
R-Squared 0.00% 0.02% 0.03% 0.01% 0.01% 0.00% 0.00%N 24,139 24,103 24,064 24,001 23,947 23,887 23,436
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Logins 0.173∗∗ 0.358∗∗∗ 0.551∗∗∗ 0.338∗∗ 0.406∗∗ 0.203 -0.408(2.17) (3.07) (3.56) (1.89) (2.05) (0.92) (-1.31)
Constant -0.140∗ 0.152 0.350∗∗ 0.482∗∗∗ 0.625∗∗∗ 0.692∗∗∗ 0.518∗
(-1.70) (1.28) (2.31) (2.66) (3.22) (3.29) (1.71)
R-Squared 0.02% 0.04% 0.05% 0.01% 0.02% 0.00% 0.01%N 24,139 24,103 24,064 24,001 23,947 23,887 23,436
Panel B. Performance of Sells
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Pages 0.032 -0.007 0.249 0.003 -0.013 -0.075 -0.591(0.37) (-0.05) (1.48) (-0.01) (-0.06) (-0.33) (-1.41)
Constant -0.023 0.376∗∗∗ 0.914∗∗∗ 1.008∗∗∗ 1.090∗∗∗ 1.182∗∗∗ 1.363∗∗∗
(-0.27) (2.97) (5.61) (5.42) (5.15) (5.11) (3.86)
R-Squared 0.00% 0.00% 0.01% 0.00% 0.00% 0.00% 0.02%N 19,435 19,373 19,320 19,257 19,200 19,142 18,777
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Logins 0.104 0.073 0.276∗ -0.026 -0.008 -0.230 -1.110∗∗∗
(1.25) (0.60) (1.70) (-0.14) (-0.04) (-1.04) (-3.15)
Constant -0.024 0.375∗∗∗ 0.911∗∗∗ 1.009∗∗∗ 1.090∗∗∗ 1.185∗∗∗ 1.376∗∗∗
(-0.28) (2.97) (5.60) (5.43) (5.15) (5.13) (3.92)
R-Squared 0.01% 0.00% 0.01% 0.00% 0.00% 0.01% 0.05%N 19,435 19,373 19,320 19,257 19,200 19,142 18,777
This table reports regression results on the relation between investor attention and trading performance. We separatethe buys (i.e. stock purchases) and the sells (i.e. stock sales) and estimate the following pooled regressions:
DGTW Ret Buysi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
DGTW Ret Sellsi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
where DGTW Ret Buysi,j,t:t+k (DGTW Ret Sellsi,j,t:t+k) are the cumulative abnormal returns of security j bought(sold) by investor i over the time interval t : t+k, computed using the DGTW model; Attentioni,j,t is the (log) number ofpages (or visits) spent on the brokerage account website by investor i over the month preceding the trade in stock j thatoccurs at time t. Cumulative abnormal returns are computed at the one-, two-, three-, four-, five-, six- and twelve-monthhorizons. Panel A reports results for stock purchases, and measures attention as, alternatively, the total number of pagesbrowsed, and the total number of logins into the brokerage account website. Panel B repeats the exercise for stock sales.Displayed are the ordinary least squares coefficient estimates and associated t-statistics. Standard errors are adjustedfor heteroskedasticity. Coefficients marked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
86
Table Online XVII. Robustness for Trading Performance and Investor Attention
Panel A. Performance of Buys
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Attention 0.184∗∗ 0.405∗∗∗ 0.639∗∗∗ 0.465∗∗∗ 0.532∗∗∗ 0.210 -0.570∗∗
(2.41) (3.76) (4.67) (2.99) (3.13) (1.14) (-2.08)
Constant -0.195∗∗∗ -0.109 0.043 0.222 0.309∗ 0.059 -1.277∗∗∗
(-2.63) (-1.01) (0.32) (1.41) (1.78) (0.32) (-4.72)
R-Squared 0.02% 0.04% 0.06% 0.03% 0.03% 0.00% 0.01%N 32,876 32,835 32,792 32,724 32,666 32,596 32,106
Panel B. Performance of Sells
1-Month 2-Months 3-Months 4-Months 5-Months 6-Months 12-Months
Attention 0.129∗ 0.136 0.340∗∗ 0.164 0.099 -0.221 -1.203∗∗∗
(1.65) (1.19) (2.36) (1.03) (0.55) (-1.11) (-3.66)
Constant 0.025 0.292∗∗ 0.708∗∗∗ 0.799∗∗∗ 0.864∗∗∗ 0.708∗∗∗ -0.454(0.30) (2.45) (4.70) (4.70) (4.55) (3.43) (-1.43)
R-Squared 0.01% 0.01% 0.02% 0.00% 0.00% 0.00% 0.06%N 25,294 25,224 25,166 25,096 25,036 24,978 24,588
This table reports regression results on the relation between investor attention and trading performance. We separatethe buys (i.e. stock purchases) and the sells (i.e. stock sales) and estimate the following pooled regressions:
Mkt Adj Ret Buysi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
Mkt Adj Ret Sellsi,j,t:t+k = α+ β Attentioni,j,t + εi,j,t:t+k,
where Mkt Adj Ret Buysi,j,t:t+k (Mkt Adj Ret Sellsi,j,t:t+k) are the market-adjusted cumulative abnormal returns ofsecurity j bought (sold) by investor i over the time interval t : t+ k; Attentioni,j,t is the (log) number of seconds spenton the brokerage account website by investor i over the month preceding the trade in stock j that occurs at time t.Cumulative abnormal returns are computed at the one-, two-, three-, four-, five-, six- and twelve-month horizons. PanelA reports results for stock purchases. Panel B repeats the exercise for stock sales. Displayed are the ordinary leastsquares coefficient estimates and associated t-statistics. Standard errors are adjusted for heteroskedasticity. Coefficientsmarked with ***, **, and * are significant at the 1%, 5%, and 10% level, respectively.
87
Table Online XVIII. Attention of Trades and Attention-Grabbing Proxies
Panel A: Buys
Spec 1 Spec 2 Spec 3 Spec 4
Abnormal Returns 0.063∗∗∗ 0.054∗∗∗
(5.17) (4.53)
Abnormal Volume 0.047∗∗∗ 0.026∗∗
(4.43) (2.32)
News 0.039∗∗ -0.033∗
(2.40) (-1.83)
Constant 0.048 0.006 0.013 0.056(1.34) (0.18) (0.37) (1.56)
R-Squared 0.44% 0.23% 0.15% 0.59%
N 24,140 32,467 33,188 24,140
Panel B: Sells
Spec 1 Spec 2 Spec 3 Spec 4
Abnormal Returns 0.063∗∗∗ 0.044∗∗∗
(3.40) (2.86)
Abnormal Volume 0.064∗∗∗ 0.056∗∗∗
(3.40) (3.05)
News 0.008 -0.050∗∗∗
(0.49) (-2.75)
Constant -0.000 -0.020 -0.016 0.017(-0.00) (-0.54) (-0.42) (0.43)
R-Squared 0.36% 0.38% 0.01% 0.81%
N 19,473 25,053 25,450 19,473
This table reports regression results from the following pooled regression:
Attentioni,j,t = α+ x′i,j,t:t−kγ + εi,j,t
where Attentioni,j,t is the (log) number of seconds spent on the brokerage account website by investor i over the monthpreceding the trade in stock j that occurs at time t and xi,j,t:t−k is a vector of regressors that includes: AbnormalReturns, the DGTW abnormal return of stock j over the previous month; Abnormal Volume, the abnormal volume ofthe stock, computed following Gervais, Kaniel, and Mingelgrin (2001); and News, the (log) number of news over theprevious month. All variables are standardized so that they have zero mean and unit variance. Panel A reports resultsfor stock purchases while Panel B repeats the exercise for stock sales. Displayed are the ordinary least squares coefficientestimates and associated t-statistics. Standard errors are adjusted for heteroskedasticity. Coefficients marked with ***,**, and * are significant at the 1%, 5%, and 10% level, respectively.
88