Attention‐Induced Trading and Returns: Evidence from Robinhood Users

Date: 2023-06-30
Full text

EXHIBIT
42

Case 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 1 of 51

THE JOURNAL OF FINANCE • VOL. LXXVII, NO. 6 • DECEMBER 2022
Attention-Induced Trading and Returns:
Evidence from Robinhood Users
BRAD M. BARBER, XING HUANG, TERRANCE ODEAN,
and CHRISTOPHER SCHWARZ*
ABSTRACT
We study the inﬂuence of ﬁnancial innovation by ﬁntech brokerages on individual
investors’ trading and stock prices. Using data from Robinhood, we ﬁnd that Robin-
hood investors engage in more attention-induced trading than other retail investors.
For example, Robinhood outages disproportionately reduce trading in high-attention
stocks. While this evidence is consistent with Robinhood attracting relatively inexpe-
rienced investors, we show that it is also driven in part by the app’s unique features.
Consistent with models of attention-induced trading, intense buying by Robinhood
users forecasts negative returns. Average 20-day abnormal returns are −4.7% for the
top stocks purchased each day.
OVER THE PAST HALF-CENTURY, INVESTOR trading has changed signiﬁcantly.
Decades ago, retail investors traded via phone only during market hours,
paying heavy commissions to do so. The 1990s brought about online trad-
ing and signiﬁcantly lower commissions. More recently, the ﬁntech brokerage
Robinhood brought about even more changes, as the ﬁrst brokerage to offer
commission-free trading on a convenient, simple, and engaging mobile app.
How do these changes in the investment landscape affect individual investors’
trading behavior?
*Brad M. Barber is at Graduate School of Management, UC Davis. Xing Huang is at Olin Busi-
ness School, Washington University in St. Louis. Terrance Odean is at Haas School of Business, UC
Berkeley. Christopher Schwarz is at Merage School of Business, UC Irvine. We appreciate the com-
ments of Azi Ben-Rephael (discussant); Charles Jones (discussant); Michaela Pagel (discussant);
Ivo Welch; and seminar and conference participants at the University of Central Florida, the Uni-
versity of Missouri, Erasmus University Rotterdam, Maastricht University, Ohio State University,
Q Group, 3rd Virtual QES NLP and Machine Learning in Investment Management Conference,
FSU SunTrust Beach Conference, the 2021 SFS Cavalcade, China Meeting of the Econometric
Society, WFA 2021, and EFA 2021. We also thank Paul Rowady, Jr. at Alphacution. Finally, we ap-
preciate the helpful comments and guidance of the editor, associate editor, and reviewers. None of
the authors has conﬂicts of interest to disclose. Barber serves on the advisory board of Vert Asset
Management. Odean is a member of the academic advisory board of Matson Money, an advisory
editor to the Financial Planning Review, a member of the academic advisory board of The Journal
of Investment Consulting, and was previously a member of the academic advisory board of Russell
Investments.
Correspondence: Xing Huang, Washington University, Olin Business School, One Brookings
Drive, St. Louis, MO 63130-4899; e-mail: xing.huang@wustl.edu. Christopher Schwarz, University
of California Irvine, Merage School of Business, Irvine, CA 92697-3125; e-mail: cschwarz@uci.edu
DOI: 10.1111/joﬁ.13183
© 2022 the American Finance Association.
3141
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 2 o

3142
The Journal of Finance®
On the one hand, the lack of commissions and the simple trading interface
may reduce the costs and barriers to investing in the stock market. Because
even small costs can reduce stock market participation among less wealthy
households (Vissing-Jorgensen (2002)), Robinhood and similar ﬁntech applica-
tions may increase stock market participation and bring new investors into the
stock market.
On the other hand, gamiﬁcation and simplicity can inﬂuence trading be-
havior in ways that might redound to the detriment of investors. To its app,
Robinhood “added features to make investing more like a game. New members
were given a free share of stock, but only after they scratched off images that
looked like a lottery ticket.”1 New and inexperienced investors may ﬁnd these
features appealing, but some believe that Robinhood overemphasizes the fun
of trading at the cost of sound investment practices. For example, in Decem-
ber 2020, Massachusetts state regulators ﬁled a complaint against Robinhood
citing its “aggressive tactics to attract inexperienced investors” and “use of
strategies such as gamiﬁcation to encourage and entice continuous and repeti-
tive use of its trading application.”2 Indeed, Robinhood users are unusually ac-
tive; in the ﬁrst quarter of 2020, Robinhood users “traded nine times as many
shares as E-Trade customers, and 40 times as many shares as Charles Schwab
customers, per dollar in the average customer account in the most recent quar-
ter.”3 Thus, while Robinhood’s innovations may have had a positive effect on
market participation (and Robinhood’s customer acquisition), their effect on
trading behavior is an open question. With these issues in mind, in this paper
we study the behavior of Robinhood users using data on aggregate Robinhood
user changes at the stock-day level from May 2018 to August 2020.
We ﬁrst conjecture that Robinhood users are more likely to be inﬂuenced
by attention than other investors. Half of Robinhood users are ﬁrst-time in-
vestors4 who are unlikely to have developed their own clear criteria for buying
a stock. Inexperienced stock investors are more heavily inﬂuenced by attention
(Seasholes and Wu (2007)) and by biases that lead to return-chasing (Green-
wood and Nagel (2009)). With turnover rates many times higher than those of
other brokerage ﬁrms, Robinhood users are more likely to trade speculatively.
As a result, a smaller proportion of their trading is motivated by nonspecu-
lative objectives such as saving for retirement, meeting liquidity needs, har-
vesting tax losses, or rebalancing their portfolio. The higher rate of speculative
trading by Robinhood users thus increases the potential for attention-driven
trading.
If Robinhood users are more likely than other investors to be inﬂuenced by
attention, their purchase behavior is more likely to be correlated, that is, they
1 See https://www.nytimes.com/2020/07/08/technology/robinhood-risky-trading.html.
2 See https://www.sec.state.ma.us/sct/current/sctrobinhood/MSD-Robinhood-Financial-LLC-Co
mplaint-E-2020-0047.pdf.
3 See https://www.nytimes.com/2020/07/08/technology/robinhood-risky-trading.html.
4 See
https://blog.robinhood.com/news/2020/5/4/robinhood-raises-280-million-in-series-f-fund
ing-led-by-sequoia.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 3 o

Attention-Induced Trading and Returns
3143
are more likely to herd than other investors.5 This is exactly what we ﬁnd.
We document that 35% of net buying by Robinhood users is concentrated in
10 stocks compared to 24% of net buying by the general population of retail
investors. We then analyze herding episodes by Robinhood users. We deﬁne a
herding episode as a day when the number of Robinhood users owning a par-
ticular stock increases dramatically. In our primary analysis, we focus on the
top 0.5% of positive user changes as a percent of prior-day user count each day.
This represents about 10 herding episodes per day or almost 5,000 episodes
over the 26-month sample period. We show that these herding episodes are pre-
dicted by attention measures (e.g., recent investor interest, extreme returns,
or unusual volume). Finally, we show that during Robinhood outages, retail
trading drops more in high-attention stocks than in other stocks relative to pe-
riods with no outage. The ﬁndings based on Robinhood outages provide strong
evidence that Robinhood users are more likely to engage in attention-induced
trading than other retail investors.
The simplicity of Robinhood’s app is likely to guide investor attention. The
app prominently displays lists of stocks in an environment relatively free of
complex information. For example, besides basic market information, Robin-
hood provides only ﬁve charting indicators, while TD Ameritrade provides
489.6 This streamlined and simpliﬁed interface likely guides the choices of
Robinhood users. Moreover, the Robinhood app makes it very easy to trade on
these guided choices. The basic information display and trading simplicity of-
fered by the app reduces cognitive burdens, which leads investors to rely more
on their intuition and less on critical thinking, or more on System 1 and less
on System 2 thinking (Kahneman (2011)). Of course, the Robinhood app is not
the only channel through which Robinhood users’ attention becomes focused
on a common subset of stocks; for example, many Robinhood users share infor-
mation and opinions on online forums such as Reddit’s WallStreetBets.7
To identify the effect of the app on Robinhood users, we focus on the “Top
Movers” list, which comprises only 20 stocks and changes every day (and
throughout each day). Crucially, this list displays stocks with the largest abso-
lute percentage price changes from the previous-day close. In contrast, many
websites provide separate lists of stocks with the largest daily gains and losses
(e.g., Yahoo! Finance Gainers and Yahoo! Finance Losers), and on these sites
top gainers tend to be more prominently displayed. Moreover, Google search
volume suggests that investors are about twice as likely to look for stocks with
5 We focus on purchase herding because retail investors tend to buy rather than sell attention-
grabbing stocks because retail investors tend to sell what they own, hold few stocks, and engage in
limited short selling (Barber and Odean (2008)). Consistent with this evidence, we ﬁnd that sales
herding events in the Robinhood data are less dramatic (see Section III.F). In contrast, institutions
hold many stocks and engage in short selling. Thus, the institutional herding literature has not
drawn a distinction between purchase and sales herding (e.g., Lakonishok, Shleifer, and Vishny
(1992)).
6 See https://www.stockbrokers.com/compare/robinhood-vs-tdameritrade.
7 See https://www.wsj.com/articles/the-real-force-driving-the-gamestop-amc-blackberry-revolu
tion-11611965586?page=1.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 4 o

3144
The Journal of Finance®
same-day gains than those with same day losses.8 Thus, if the app itself is
driving Robinhood users’ trading, we would expect Robinhood traders to buy
both gainers and losers heavily, while other retail investors will tend to buy
gainers. This is precisely what we ﬁnd: Robinhood users are drawn to trad-
ing both extreme gainers and extreme losers, whereas other retail investors
prefer to buy extreme gainers. While prior work documents that investors buy
extreme winners and losers (Barber and Odean (2008)), our evidence indicates
the Robinhood app affects the intensity of this behavior because of the unique
way Robinhood displays the Top Movers list.
We provide additional evidence that the Top Movers list inﬂuences Robin-
hood user buying behavior by exploiting another unique feature of the list.
Robinhood requires stocks above $300 million in market capitalization to be
displayed in the Top Movers list. We use a sharp regression discontinuity de-
sign to show that Robinhood users are more likely to buy stocks with market
capitalization between $300 million and $350 million that were in the top 20
stocks when sorting on absolute returns than stocks with similar absolute re-
turns but market capitalization between $250 million and $300 million. Thus,
stocks that just miss making the list due to market capitalization below the
$300 million cutoff do not observe the increase in users associated with being
on the Top Movers list. This evidence indicates that the display of information
affects investors’ trading.
Models of attention-induced trading and returns predict that periods of in-
tense buying will be followed by negative abnormal returns (e.g., Barber and
Odean (2008), Pedersen (2022)). We conjecture that the concentrated buying
of Robinhood users, who are susceptible to attention-induced trading, provides
an unusually strong setting to identify the return effects of attention-induced
trading. In our ﬁnal set of analyses, we focus on this return prediction and doc-
ument large negative abnormal returns following Robinhood herding episodes.
Speciﬁcally, the top 0.5% of stocks bought every day lose about 4.7% over the
subsequent month.
The magnitude of the negative abnormal returns increases dramatically as
we identify fewer but more intense herding episodes. To systematically ana-
lyze the relation between the herding intensity and price reversal, we focus on
stocks with a minimum of 100 Robinhood users and identify different sets of
herding episodes by varying the daily percentage increase in users holding the
stock from 10% to 750%. At a 10% increase in users, we observe over 20,000
herding episodes, while at a 750% increase, we observe 45 episodes. The large
negative abnormal returns in the month following these herding events grow
from a statistically signiﬁcant −1.8% when we require a 10% increase in users
(>20,000 events) to an extremely large and statistically signiﬁcant −19.6%
when we require a 750% increase in users (45 events).
8 Google trends indicates that the phrase “top gainers today” (“top stock gainers today”) is
searched more than twice as much as “top losers today” (“top stock losers today”) over the ﬁve
years beginning January 24, 2016.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 5 o

Attention-Induced Trading and Returns
3145
The negative returns that follow purchase herding by Robinhood users are
not simply inventory-based reversals as modeled in Jagadeesh and Titman
(1995) and documented around earnings announcements in So and Wang
(2014). Attention-induced trading can also cause return reversals when in-
vestors tend to buy stocks with strong recent returns. However, the negative
returns following herding episodes are not limited to return reversals. First, in
multivariate analyses, in which we control for return reversals, the negative
abnormal returns following herding episodes remain large and statistically sig-
niﬁcant. Second, and more importantly, we observe negative returns following
a day when we observe a surge in Robinhood users that is preceded by an
overnight drop in the stock’s price, perhaps because aggressive Robinhood in-
vestor buying slows the stock’s response to negative news (Barber and Odean
(2008)). The negative returns that we document following purchase herding by
Robinhood users are also not driven by the bid-ask spread since they persist
when we use quote midpoints to calculate returns.
Given the relatively small size of Robinhood, one might question whether
Robinhood users have the potential to inﬂuence market prices. Robinhood has
$81 billion in assets under custody,9 far less than E*TRADE at $600 billion,
TD Ameritrade at $1.3 trillion, and Charles Schwab at $3.8 trillion.10 How-
ever, trades, not passive positions, move prices. There are a lot of Robinhood
users: 13 million as of May 2020 compared to 12.7 million at Schwab and 5.5
million at E*TRADE at the end of 2019.11 Moreover, as noted above, Robinhood
users are extremely active traders. In June 2020, Robinhood users averaged
4.3 million revenue trades per day (Daily Average Revenue Trades or DARTs),
more than E*TRADE at 1.1 million, TD Ameritrade at 3.8 million, Charles
Schwab at 1.8 million, Interactive Brokers at 1.9 million, or Fidelity at 1.4 mil-
lion.12 Thus, Robinhood users accounted for roughly 30% of daily trades from
the largest brokerage ﬁrms catering to retail investors and have the potential
to move prices. In Q1 2020, Robinhood’s order routing revenue was 21.5% of
that of Robinhood, TD Ameritrade, E*Trade, and Schwab combined, and its
shares traded were 12.4%. Robinhood users also favored trading non-S&P 500
stocks more than did the clients at the three other ﬁrms.13 Thus, by several
measures, Robinhood accounted for a material fraction of retail trading, par-
ticularly for smaller-capitalization stocks. Furthermore, as noted above, Robin-
hood trades may be a good proxy for the actions of other attention-motivated
traders who herd in the same stocks.
9 See p. 26 of Robinhood’s SEC Form S1 IPO ﬁling (https://www.sec.gov/Archives/edgar/data/
1783879/000162828021013318/robinhoods-1.htm).
10 See https://www.businessofapps.com/data/robinhood-statistics/.
11 See https://www.nytimes.com/2020/07/08/technology/robinhood-risky-trading.html.
12 See https://www.bloomberg.com/news/articles/2020-08-10/robinhood-blows-past-rivals-in-re
cord-year-for-retail-investing. Fidelity’s daily trades are for all of 2020, not just June (https://
www.barrons.com/articles/ﬁdelitys-trading-volume-surged-in-the-pandemic-but-its-struggling-
to-boost-revenue-51614702735). Note that brokerages exercise some discretion in measuring
DARTs.
13 Case Study: The Robinhood Effect, Alphacution, July 2021.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 6 o

3146
The Journal of Finance®
Two additional ﬁndings indicate that the negative returns we observe are
due at least in part to the trading of Robinhood users and other attention-
motivated retail investors. First, we expect the inﬂuence of investors to be most
pronounced among small-cap stocks. In the cross section, we ﬁnd negative re-
turns following Robinhood herding events for stocks with market cap under $1
billion, but not for stocks with market cap over $1 billion. Second, we expect
larger effects during periods with heightened retail trading. In the time series,
retail trading increased signiﬁcantly at Robinhood and elsewhere during the
COVID-19 period (i.e., after March 13, 2020), and the negative return effect
following Robinhood herding events is more pronounced in this period. Irre-
spective of whether the negative returns we document result from the trading
of Robinhood users and other attention-motivated retail investors, these nega-
tive returns lead to trading losses for many investors.
Savvy investors might exploit the trading patterns and negative returns that
we document. To proﬁt from Robinhood herding events, an investor would need
to sell the stock short (or, equivalently, purchase put options that the option
seller would hedge by shorting). Thus, if investors are exploiting Robinhood
user herding, we would expect to see increased short interest around Robin-
hood user herding events. We do indeed ﬁnd a marked increase in short selling
for stocks involved in Robinhood herding events even after controlling for re-
turns and news.
While the Top Movers list and possibly other features of the Robinhood app
inﬂuence Robinhood users, many drivers of attention will affect Robinhood
users and other retail investors similarly. We show that this is the case us-
ing a measure of retail investor herding developed by Barber, Lin, and Odean
(2021, BLO). BLO classify herding events as Trade and Quote (TAQ) retail
trades in the top quintile of retail standardized order imbalance (SOI) and
top decile of abnormal retail volume. For each day in our Robinhood sample,
we modify the BLO methodology to identify the same number of TAQ herding
events as we identify using our Robinhood herding measure. We ﬁnd that 27%
(1,317 of 4,884) of the Robinhood and BLO herding events are identical. Like
Robinhood herding events, these events are, on average, followed by negative
returns. However, the 21% (1,008 of 4,884) of the Robinhood herding events
that are accompanied by retail selling on TAQ also earn negative returns, and
the subsequent negative returns for non-Robinhood herding events are smaller
in magnitude than those related to Robinhood events. These results show that
the actions of Robinhood users are a good proxy for attention-induced trading.
Our study is of particular interest given the unique data set of the retail
investors (Robinhood users) that we analyze. To our knowledge, four papers
use the same. Welch (2022) analyzes the holdings and performance of Robin-
hood users. He concludes that Robinhood users principally held stocks with
large persistent past volume and do not underperform with respect to stan-
dard academic benchmark models.14 In a JP Morgan report, Cheng, Murphy,
14 We too ﬁnd that the aggregate performance of Robinhood users is not reliably different from
zero using standard asset pricing technology. See Table IAI.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 7 o

Attention-Induced Trading and Returns
3147
and Kolanovic (2020) show that users are drawn to stocks that attract investor
attention and that changes in stock popularity predict returns. Unlike these
studies, we document poor returns following extreme attention-driven herding
events by Robinhood users. Ozik, Sadka, and Shen (2021) use the Robintrack
data to analyze the sharp increase in retail trading and the effect on bid-ask
spreads during the pandemic period. They ﬁnd an increase in trading of stocks
with COVID-19 media coverage, which they attribute to an attention-grabbing
effect. They also document that the increase in retail trading generally lowered
stock bid-ask spreads and that price impact of trades. Similarly, Eaton et al.
(Forthcoming) use Robinhood outages to study the effect of retail trading on
market quality and ﬁnd that these negative shocks to Robinhood participation
reduce market volatility and improve liquidity. In contrast to these studies,
we study the attention-induced trading of Robinhood investors and show that
herding episodes by Robinhood investors reliably predict negative returns.
In summary, we provide two main contributions to the academic literature.
First, we present evidence that the Robinhood app inﬂuences investors’ behav-
ior. Speciﬁcally, we show that the prominently featured Top Movers list (which
displays only 20 stocks, sorts stocks on absolute [rather than signed] percent-
age returns, and changes regularly) affects Robinhood users. This ﬁnding ﬁts
into the emerging literature that emphasizes that the display of information
affects investor behavior. Changes in the display of price information affect in-
vestors’ willingness to sell winners versus losers in individual stocks and mu-
tual funds (Frydman and Wang (2020), Loos, Meyer, and Pagel (2020)). News
that investors consume about stocks often conﬁrms their prior beliefs (Cook-
son, Engelberg, and Mullins (Forthcoming)), and its prominence affects the
incorporation of information (Fedyk (2019)). Displaying return performance
for index funds can lead investors to prefer high-fee funds (Choi, Laibson,
and Madrian (2010)), prominently featuring expense information can lead in-
vestors to prefer low-fee funds (Kronlund et al. (2021)), and the prominence of
mutual fund lists affects investors’ fund choices (Kaniel and Parham (2017),
Hong, Lu, and Pan (2019)).15 Our results, and this emerging literature, indi-
cate that disclosure alone is not sufﬁcient to assure good investor outcomes—
how information is displayed can both help and hurt investors. Furthermore,
while the recent literature on complexity in ﬁnance emphasizes its dark side
(Carlin (2009)), our results suggest that simple user interfaces are not neces-
sarily the solution to problems that arise from complexity—both complexity
and simplicity can lead investors astray.
Second, we contribute to the literature that documents the effects of
attention-induced trading on returns. Using Robinhood trading as a proxy for
attention-induced herding episodes, we ﬁnd strong support for the return pre-
dictions of attention-based models of trading. Speciﬁcally, we link episodes of
15 Da et al. (2018) ﬁnd that recommendations of an advisory ﬁrm followed by many Chilean
pension investors generate correlated fund ﬂows and inﬂuence market returns. These attention-
induced effects on the active choice of mutual funds are different from the stickiness of default
options (e.g., Cronqvist and Thaler (2004)), which might result from inertia or a view that defaults
are an implicit recommendation.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 8 o

3148
The Journal of Finance®
intense buying by Robinhood users to negative returns following the herding
episodes. Our focus on trading rather than events is distinct from extant lit-
erature that focuses on events such as Jim Kramer’s stock recommendations
(Keasler and McNeil (2010), Bolster, Trahan, and Venkateswaran (2012), En-
gelberg, Sasseville, and Williams (2012)), the WSJ Dartboard Column (Barber
and Loefﬂer (1993), Liang (1999)), Google stock searches (Da, Engelberg, and
Gao (2011), Da et al. (2020)), and repeat news stories (Tetlock (2011)). Perhaps
as a result of our focus on trading behavior rather than events, the return re-
versals that we document are much larger and more widespread than those
documented in prior studies.16 Moreover, the extant literature documents
negative returns subsequent to positive return events, while we ﬁnd nega-
tive returns following intense buying that coincides with or follows negative
returns.
The paper proceeds as follows. Section I provides a detailed description of
our data set. Section II presents our empirical analysis of Robinhood investors’
attention-induced trading and the role of the app’s unique features. Section III
shows the return effect of Robinhood users’ attention-induced trading. Sec-
tion IV discusses the short-trading activities associated with the intense buy-
ing by Robinhood users. Section V concludes.
I. Data and Methods
In this section, we describe the main Robintrack data set that keeps track of
how many Robinhood users hold a particular stock over time and our methods
for identifying extreme herding events by Robinhood users.
A. Robintrack Data
The primary data set for our analysis comes from the Robintrack website
(https://robintrack.net/), which scrapes stock popularity data from Robinhood
between May 2, 2018, and August 13, 2020.17 Robinhood discontinued the
reporting of stock popularity data on August 13, 2020. The Robintrack data
set contains cross-sectional snapshots of user counts for individual securi-
ties (e.g., 645,535 Robinhood users held Apple stock at 3:46 pm ET on Au-
gust 3, 2020).18 Our main results include all Robintrack securities since we do
16 See Table IAII for a summary of these studies. The biggest magnitude of return reversal is
−4.6% for 39 events over three years (Barber and Loefﬂer (1993)). In a widely cited study, Da,
Engelberg, and Gao (2011) fail to ﬁnd robust evidence of price reversals following spikes in Google
search volume.
17 About 11 dates during the sample period are missing user data, four in January 2019 and
seven in January 2020. For 16 dates on which we observe users, no observations were recorded
between 2 and 4 pm ET.
18 The Robintrack data are generally reported every hour at approximately 45 minutes after
the hour. The data from Robinhood have some lag. For example, the user count on Robintrack for
Apple at 3:46 pm is from sometime before 3:46 pm. Based on analysis of open data, the lag is likely
between 30 and 45 minutes. The Robinhood app appears to update data every 15 minutes.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
se 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 9 o

Attention-Induced Trading and Returns
3149
not have strong priors about what types of securities will experience herding
events.19
We merge the Robintrack data to the Center of Research in Security Prices
(CRSP) and TAQ data by using the ticker on Robintrack. The CRSP database
provides daily returns, closing and opening prices, closing bid-ask spreads, and
market capitalization. We use the TAQ database to identify retail buys and
sells using the Boehmer et al. (2021) algorithm. This algorithm relies on the
observation that retail trades often receive price improvement in fractions of a
penny and are routed to a Financial Industry Regulatory Authority (FINRA)
trade reporting facility (TRF). The algorithm therefore identiﬁes retail buys as
trades reported to a FINRA TRF (exchange code “D” in TAQ) with fractional
penny prices between $0.006 and $0.01, and retail sells as trades reported to a
FINRA TRF with fractional penny prices between $0.00 and $0.004.20
In Figure 1, Panel A, we see that the total number of Robinhood user-stock
positions grew from about 5 million at the beginning of our sample period to
more than 42 million at the end. In May 2020, Robinhood reported having
13 million users, which translates to about three stock positions per user.21
The red vertical line in the ﬁgure denotes the date on which the COVID-19
national emergency was declared in the United States (March 13, 2020). A
clear increase in Robinhood users can be seen after this date. In Figure 1,
Panel B, we plot the total number of TAQ retail trades for comparison. We ﬁnd
that retail trading also increased during the pandemic period. Of course, some
Robinhood trades are part of these retail trades.
The Robintrack data do not allow us to identify individual trades, but they do
allow us to analyze changes in user positions in a particular stock. The analog
to this Robinhood user change variable in TAQ is net retail buying in a stock. In
Figure 1, Panel C, we plot the ﬁve-day moving average of the daily sum across
stocks of the absolute value of Robinhood user changes (green line) and the
ﬁve-day moving average of the daily sum across stocks of the absolute value of
TAQ net buying, that is, the number of retail buys minus the number of retail
sells (blue line). Both measures of trading activity follow similar trajectories,
with a marked increase in the pandemic period.
We present additional summary statistics in Table I, Panel A, across stock-
day observations. The main variable, users_close, measures the total number
of users in a stock prior to the close of trading (4 pm ET) but after 2 pm on the
same day. The key variables in our analysis of herding events are based on the
daily changes in users_close (userchg) or the ratio of users_close on consecutive
days (userratio). For descriptive purposes, we also report users_last, which is
19 The results are similar for common stocks and other securities, though U.S. common stocks
represent 70% of all herding episodes, stocks with non-U.S. headquarters 13%, and American De-
positary Receipts (ADRs) 10%.
20 We also use TAQ to calculate returns in July and August 2020 since CRSP data were not
available through August 2020 at the writing of this draft.
21 See “Robinhood Has Lured Young Traders, Sometimes with Devastating Results” in The
New York Times, July 8, 2020 (https://www.nytimes.com/2020/07/08/technology/robinhood-risky-
trading.html).
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 10
51

3150
The Journal of Finance®
Figure 1. Robinhood and TAQ retail trading. Panel A plots the total number of Robinhood
stock holdings. Panel B plots the total number of TAQ Retail Trades. Panel C plots the total
number of absolute Robinhood user changes (green) and absolute TAQ net buys (blue) as a ﬁve-
day moving average. The red vertical line depicts the date when the COVID national emergency
was declared in the United States (March 13, 2020). (Color ﬁgure can be viewed at wileyonlineli-
brary.com)
the last reported user count for a stock on each day, regardless of the time of
reporting.
The mean stock has a bit more than 2,000 users, though the median user
count is 160. User changes are generally small—the interquartile range of
userchg is 2 and that of userratio is 0.01. Similarly, TAQ net buying is gen-
erally quite small, with an interquartile range of −7 to +10. In Table I, Panel
B, we present descriptive statistics across days. The average day has 7,211
stock holdings and just under 15 million user positions.
B. Herding Events
While we ﬁnd that user changes are generally small, a number of extreme
user change events are in our sample. These extreme events likely occur
because Robinhood users are new to markets and more willing to specu-
late. These extreme events are also likely a good proxy for the behavior
of investors who are unduly inﬂuenced by attention-grabbing events. To
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 11
51

Attention-Induced Trading and Returns
3151
Table I
Summary Statistics
Panel A presents summary statistics across stock-day observations. Panel B sums variables by day and then averages across days. The variables are
users_close (last observed user count for a stock prior to the 4 pm ET close), users_last (last user count of the day), userchg (daily change in users_close),
userratio (users_close(t)/users_close(t −1)), prc (closing price), size (market cap in millions), ret (daily return), openret (overnight return), dayret
(daytime return), daily_buys (number of TAQ daily retail buys), daily_sells (number of TAQ daily retail sells), net_buys (daily_buys −daily_sells),
taq_retimb (net_buys/(daily_buys + daily_sells)), and #stocks (number of stocks with users reported on Robintrack).
Variable
N
Mean
SD
min
p25
p50
p75
max
Panel A: Stock-Day Observations
users_close
3,952,749
2,064.27
15,422.64
0.00
35.00
160.00
674.00
990,059.00
users_last
4,067,791
2,061.48
15,419.80
0.00
35.00
160.00
673.00
990,587.00
userchg
3,851,419
9.46
245.07
−19,643.00
−1.00
0.00
1.00
85,193.00
userratio
3,745,652
1.01
0.32
0.00
1.00
1.00
1.01
263.67
prc
3,765,043
52.47
1,828.84
0.04
10.38
23.71
46.81
344,970.00
size ($mil)
3,625,145
5,674.62
29,154.96
0.00
108.25
498.94
2,416.15
1,581,165.00
ret (%)
3,764,157
0.04
4.21
−91.79
−0.98
0.02
0.99
897.73
openret (%)
3,674,652
0.10
2.64
−88.60
−0.40
0.03
0.55
563.90
dayret (%)
3,696,846
−0.05
3.41
−87.59
−0.91
0.00
0.78
841.18
daily_buys
3,586,637
200.69
1,112.99
0.00
9.00
34.00
117.00
185,930.00
daily_sells
3,586,637
178.97
875.49
0.00
8.00
34.00
115.00
113,152.00
net_buys
3,586,637
21.72
367.93
−30,246.00
−7.00
0.00
10.00
86,640.00
taq_retimb
3,585,659
0.01
0.35
−1.00
−0.14
0.00
0.16
1.00
Panel B: Daily Observations (Summed Variable Averaged across Days)
#stocks
549
7,211.01
741.42
5,805.00
6,559.00
7,199.00
8,054.00
8,131.00
users_close (mil.)
549
14.86
9.93
1.32
8.27
11.83
15.22
42.14
users_last (mil.)
549
14.89
9.93
5.58
8.27
11.83
15.23
42.16
userchg (000)
535
68.11
112.57
−48.83
14.70
23.74
54.11
810.40
daily_buys (000)
549
1,276.35
739.49
387.97
824.86
923.12
1,297.50
3,952.49
daily_sells (000)
549
1,137.28
585.45
360.04
778.65
872.51
1,181.31
3,174.90
net_buys (000)
549
139.07
171.91
−67.98
38.44
63.30
133.80
1,005.84
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 12
51

3152
The Journal of Finance®
Table II
Summary Statistics for Robinhood Herding Events
This table provides summary statistics for stock-days associated with herding events, deﬁned as
securities in the top 0.5% of positive user change ratio on day t and a minimum of 100 users on
day t −1. The variables are users_close (last observed user count for a stock prior to the 4 pm
ET close), users_last (last user count of the day), userchg (daily change in users_close), userratio
(users_close(t)/users_close(t −1)), prc (closing price), size (market cap in millions), ret (daily re-
turn), openret (overnight return), dayret (daytime return), daily_buys (number of TAQ daily re-
tail buys), daily_sells (number of TAQ daily retail sells), net_buys (daily_buys −daily_sells), and
taq_retimb (net_buys/(daily_buys + daily_sells)).
Variable
N
Mean
SD
min
p25
p50
p75
max
users_close
4,884
2,487.02
7,573.30
116.00
353.00
774.50
1,914.50
154,351.00
users_last
4,884
2,605.28
7,887.32
118.00
367.00
810.00
2,007.50
156,826.00
userchg
4,884
1,103.72
3,514.05
16.00
119.00
288.50
803.00
85,193.00
userratio
4,884
1.99
1.69
1.10
1.37
1.56
1.98
44.71
prc
4,712
163.89
6,973.02
0.12
3.34
8.95
21.39
341,000.00
size ($mil)
4,299
2,231.98
11,532.14
0.11
45.29
375.29
1,229.05
468,894.20
ret (%)
4,711
14.02
52.58
−91.79
−7.10
4.88
20.56
874.84
openret (%)
4,707
10.99
39.19
−88.60
−0.94
2.04
11.61
563.90
dayret (%)
4,710
3.43
30.68
−74.69
−7.79
−0.06
8.02
595.19
daily_buys
4,675
3,498.46
9,536.55
0.00
252.00
774.00
2,669.00
162,678.00
daily_sells
4,675
2,710.01
7,108.48
0.00
205.00
615.00
2,149.00
110,145.00
net_buys
4,675
788.44
2,962.92
−8,873.00
7.00
98.00
487.00
56,504.00
taq_retimb
4,675
0.11
0.17
−0.75
0.01
0.09
0.20
1.00
identify Robinhood herding events, we ﬁrst identify stocks with an increase
in users (i.e., userratio(t) > 1) and at least 100 users entering the day (i.e.,
users_close(t −1) ≥100). We then sort these stocks based on the day t userratio
and identify the top 0.5% of stocks as Robinhood herding stocks, which we de-
note with the indicator variable rh_herd. This procedure results in a sample of
4,884 herding events (about nine per day on average) for 2,301 unique tickers.
Table II presents descriptive statistics on the sample herding events across
stock-day observations. The average stock in these episodes has about 2,500
users and experiences an increase in users of 1,100. The return on the stock on
the day of such episodes is on average 14%, with most of the return occurring
at the open of trading—the mean opening return (openret) is 11%. Despite
these large positive mean daily returns, however, we also observe stocks (about
1/3) with large negative returns on these herding event days. As we point out
below, the appeal of large negative stocks may be due in part to the Robinhood
app, which highlights “Top Movers” for the day based on absolute rather than
signed returns, that is, unlike many stock lists that focus on the most positive
movers for the day, the Robinhood app focuses users’ attention on stocks with
extreme returns. We tend to observe a large retail order imbalance in TAQ on
these days as well, which is not surprising since Robinhood trades are a subset
of TAQ trades.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 13
51

Attention-Induced Trading and Returns
3153
II. Attention and Stock Selection
In the ﬁrst part of our analysis, we document that Robinhood users show
excessively concentrated trading activities, compared with the general pop-
ulation of retail investors as captured by the TAQ data set. We then show
that attention measures strongly predict Robinhood herding episodes and use
the model to forecast the probability of herding episodes for individual stocks.
To convincingly show that Robinhood users are particularly prone to trading
in attention-grabbing stocks, we exploit Robinhood trading outages and doc-
ument sharper drops in retail volume for stocks with a high probability of a
herding episode during these outages. In a ﬁnal analysis, we show that unique
features of the Robinhood platform predict the trading of Robinhood users
more strongly than other retail investors, which suggests that the app itself
guides user decisions.
A. The Concentration of Buying versus Selling
In theory, attention-induced trading should predominantly affect purchase
rather than sale decisions—retail investors can buy any stock that captures
their attention but can only sell stocks that they own (unless they sell short,
which is relatively uncommon among retail investors and not possible on the
Robinhood platform). We expect attention-motivated trading to be common
among Robinhood users. To test this conjecture, we compare the concentra-
tion of buying activity to the concentration of selling activity for Robinhood
users. We also expect the concentration of buying to be greater for Robinhood
users than the general retail investor population. While attention certainly af-
fects the general population of retail investors, we expect other motives to play
a greater role in their trading decisions (e.g., trade to rebalance, harvest tax
losses, diversify, or save/consume rather than engage in attention-motivated
trading). This is especially true since half of Robinhood users are new in-
vestors, who are more subject to attention biases (Seasholes and Wu (2007)).
To empirically investigate these questions, we ﬁrst identify the 10 stocks
with the most Robinhood-buying activity on each day (i.e., largest Robinhood
user changes). We then calculate the concentration of buying among these 10
stocks as the total number of new users for these stocks divided by sum of user
increases for all stocks with an increase in users. Similarly, we identify the
10 stocks with the most TAQ retail buying (i.e., largest net buying based on
number of retail buys minus number of retail sells). We then calculate the con-
centration of buying among these 10 stocks as the total number of net buys for
these stocks divided by sum of net buying for all stocks with net buying (i.e., a
positive order imbalance). These calculations are repeated for each day, yield-
ing a time series of daily measures of buying concentration for Robinhood and
TAQ retail investors. We also perform an analogous calculation for negative
user changes on Robinhood and TAQ net selling.
Figure 2 presents the mean concentration of buying (Panel A) and selling
(Panel B) for Robinhood users (green bars) and TAQ retail trades (white bars).
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 14
51

3154
The Journal of Finance®
Figure 2. The concentration of buying and selling. In Panel A, the ﬁgure depicts the mean
daily percent of net buying that is observed in the stock with ranks from 1 to 10, with the rank of
1 corresponding to stocks with the most net buying. In Panel B, the ﬁgure depicts the mean daily
percent of net selling that is observed in the stock with ranks from 1 to 10, with the rank of 1
corresponding to stocks with the most net selling. Whiskers depict 95% conﬁdence intervals based
on standard errors across days. In Robinhood, net buying is user changes. In TAQ, net buying is
retail buys less retail sells. (Color ﬁgure can be viewed at wileyonlinelibrary.com)
Consistent with the idea that attention has a bigger effect on buying than
selling, we ﬁnd that for both Robinhood and TAQ retail traders, the concentra-
tion of buying (Panel A) is higher than the concentration of selling (Panel B).
However, both buying and selling concentrations are stronger for Robinhood
investors than for the general population of retail traders.22
In Table III, we summarize the mean percentage of trades observed in the
top 10 stocks and calculate mean daily Herﬁndahl-Hirschman (HH) indexes
for buying (Panel A) and selling (Panel B). For Robinhood users (TAQ retail
trades), about 35% (24%) of all net buying is in the top 10 stocks, while 25%
(14%) of selling is concentrated in the top 10 stocks. The HH indexes for buy-
ing are larger than those observed for selling for both Robinhood and TAQ,
which indicates that, in general, the concentration of buying activity is higher
than the concentration of selling activity for retail traders. Moreover, the HH
indexes for Robinhood are greater than those for TAQ for both buying and sell-
ing, which indicates a higher degree of buying and selling concentration for
Robinhood users.
One concern with the analysis above is that Robinhood user changes (new
owners of the stock) do not map perfectly to TAQ net buying. Essentially, we
assume that measures of new owners and net buying generate similar concen-
tration statistics. While we cannot compare the concentration of new owners
and net buying in the Robinhood (or TAQ) data set, we can examine this is-
sue using the discount broker trade and position data of Barber and Odean
(2000). To do so, we use these data to calculate two variables: daily new users
for each stock (from daily positions) and daily order imbalance for each stock
22 This conclusion assumes that there is no bias in the Boehmer et al. (2021) methodology that
would affect concentration measures.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 15
51

Attention-Induced Trading and Returns
3155
Table III
Concentration of Retail Buying and Selling
The sample consists of stocks with a measure of Robinhood user changes and net retail buying
in TAQ. Statistics are based on averages across days. In Panel A, HH_buy is the Herﬁndahl-
Hirschman index for stocks with net buying (sum of squared share of buying) and Top10_buy is
the percentage of all net buying in the 10 stocks with the highest level of net buying. In Panel B,
HH_sell is the Herﬁndahl-Hirschman index for stocks with net selling (sum of squared share of
selling) and Top10_sell is the percentage of all net selling in the 10 stocks with the highest level
of net selling. ***p < 0.01, **p < 0.05, *p < 0.1.
Robinhood User Changes
TAQ Retail Trades
Difference (RH −TAQ)
Panel A: Mean Daily Concentration Measures among Stocks with Net Buying
HH_buy (bps)
246.952***
110.820***
136.133***
(13.091)
(5.575)
(11.433)
Top10_buy (%)
34.621***
23.560***
11.061***
(0.412)
(0.315)
(0.361)
Panel B: Mean Daily Concentration Measures among Stocks with Net Selling
HH_sell (bps)
172.760***
48.566***
124.194***
(20.069)
(2.357)
(20.055)
Top10_sell (%)
25.370***
14.076***
11.294***
(0.369)
(0.218)
(0.384)
(from trades). The correlation between these two series is 87%; they gener-
ate concentration statistics in the broker data that differ by less than 0.5% at
the individual stock level. Given the differences in concentration that we ob-
serve in Table III and Figure 2, we conclude that the differences are not due to
measurement differences and therefore Robinhood users’ purchase activity is
indeed more concentrated than the general population of retail traders.
B. Attention Proxies and Robinhood Herding Events
If attention is guiding the trading decisions of Robinhood investors, we ex-
pect proxies for investor attention to be strong predictors of Robinhood herding
episodes. Here, we examine the relation between a set of attention measures
and the herding episodes we study. While this analysis is interesting on its
own, the model also identiﬁes stocks that are at high risk of a herding episode,
which is useful for our analysis of retail trading during Robinhood outages
below. To begin, we estimate a linear probability model by regressing the ex-
treme herding episode indicator on a set of attention measures, including ex-
treme absolute lagged returns, lagged abnormal volume, lagged user change,
lagged level of users, lagged abnormal Google search volume, lagged abnor-
mal news coverage, and lagged earnings announcement. We process Google
search volume index (SVI) following Da, Engelberg, and Gao (2011) and Niess-
ner (2015). We construct the news coverage variable by counting the daily num-
ber of news articles written on the ticker based on data obtained from Thomson
Reuters MarketPsych Indices (TRMI). All abnormal measures on day t −1 are
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 16
51

3156
The Journal of Finance®
Table IV
Determinants of Robinhood Herding Indicator Variable
(Top 0.5% of Positive Percentage User Change)
The table examines the determinants that predict the Robinhood herding indicator variable (top
0.5% of positive user change ratio on day t and a minimum of 100 users on day t −1) using a linear
probability model. rh_herd(t −1) is the lagged Robinhood herding indicator. Extreme absolute
return (t −1) is an indicator that equals one if the absolute return is ranked in the top 20 on day
t −1. Abnormal Vol (t −1) is the logarithm of the ratio of stock market volume on day t −1 to the
average volume from days t −21 to t −2. User Change (t −1) is the change in Robinhood users
from day t −2 to day t −1. ln(Users(t −1)) is the logarithm of Robinhood users before the market
closes on day t −1. Abnormal SVI (t −1) is the logarithm of the ratio of Google search volume on
day t −1 to the average Google search volume from days t −21 to t −2. Abnormal News (t −1) is
the logarithm of the ratio of news article count on day t −1 to the average news article count from
days t −21 to t −2. Earnings Announcement (t −1) is an indicator that equals one if the ﬁrm has
an earnings announcement on day t −1. Robust standard errors are clustered by day and stock
level. ***p < 0.01, **p < 0.05, *p < 0.1.
rh_herd(t)
(1)
(2)
(3)
(4)
rh_herd(t −1)
0.106***
0.103***
0.102***
0.102***
(0.005)
(0.005)
(0.005)
(0.005)
Extreme absolute return (t −1)
0.0505***
0.0494***
0.0485***
0.0486***
=1: top 20 absret; =0: otherwise
(0.003)
(0.003)
(0.003)
(0.003)
Abnormal Vol (t −1)
0.000551***
0.000453***
0.000368***
0.000348***
= ln(Vol(t −1)/AvgVol(t −21,t −2))
(0.000)
(0.000)
(0.000)
(0.000)
User change (t −1) (in 000s)
0.00806***
0.00738***
0.00733***
(0.001)
(0.001)
(0.001)
ln(Users(t −1))
0.000193***
0.000170***
0.000166***
(0.000)
(0.000)
(0.000)
Abnormal SVI (t −1)
0.000149***
0.000142***
= ln(SVI(t −1)/AvgSVI(t −21,t −2))
(0.000)
(0.000)
Abnormal news (t −1)
0.00195***
0.00107***
= ln(News(t −1)/AvgNews(t −21,t −2))
(0.000)
(0.000)
Earnings announcement (t −1)
0.00875***
(0.001)
Observations
3,792,584
3,792,584
3,792,584
3,792,584
R2
0.022
0.022
0.023
0.023
computed as the logarithm of the ratio of the value on day t −1 to the aver-
age from day t −21 to t −2. The lagged herding indicator is also included in
the regression to capture persistence in the herding episodes. Robust standard
errors are clustered by day and stock level.
Table IV presents the results. We ﬁnd evidence of persistence in the herd-
ing episodes. The coefﬁcient on the lagged herding indicator is positive and
statistically signiﬁcant. A stock that is heavily bought by Robinhood investors
is 10% more likely to experience another episode the next day. This is not
surprising—the herding episode itself may generate discussion and attract at-
tention through media or social media platforms and lead to additional herding
the next day. Moreover, consistent with our results that Robinhood investors
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 17
51

Attention-Induced Trading and Returns
3157
respond to the Top Movers list, we ﬁnd that if a stock’s absolute return is
ranked in the top 20, the probability of this stock being heavily bought the
next day increases by about 5%, signiﬁcant at 1% level. In addition, the other
attention measures all show a higher probability of heavy buying activities on
Robinhood.
Taken together, the results show that the extreme herding episodes of Robin-
hood investors are persistent and can be predicted by various attention mea-
sures. We next use the predicted values from these speciﬁcations to capture the
attention-driven component of the extreme herding episodes in our analyses of
Robinhood outages.
C. The Effect of Robinhood Outages on Retail Trading
In this section, we build on the linear probability model of the prior section
to establish the importance of Robinhood user trading in attention-grabbing
stocks. To do so, we exploit three unexpected trading outages on the Robin-
hood user platform. These outages allow us to estimate the impact of Robin-
hood trading on retail trading in general. More importantly, the outages allow
us to demonstrate a sharper drop in retail trading for stocks that are good can-
didates for the herding events we study or that are popular among Robinhood
users. This evidence supports our claim that Robinhood users are more likely
to engage in attention-induced trading than other retail investors.
To identify Robinhood outages, we review the incident history on Robinhood
websites. Three outages affected equity trading, on March 2, March 3, and
June 18, 2020. The most prolonged outage occurred on March 2 and lasted
virtually the entire trading day (posted as under investigation at 9:38 am ET
and as resolved at 2:13 am ET on March 3). On March 3, there was another
outage at 10:04 am ET posted as under investigation with service partially
restored at 11:35 am ET and fully restored at 11:55 am ET. The third outage
occurred on June 18 and began at 11:39 am ET (posted as under investigation)
with improvement at 12:43 pm ET (posted as “starting to see improvement”)
and resolution at 1:08 pm ET.
To estimate the economic impact of Robinhood trading, we use these out-
ages, which are arguably exogenous events that prevent Robinhood users from
trading but have no effect on retail investors who trade using other brokers.23
Speciﬁcally, we measure the proportion of retail trading relative to all trading
at hourly intervals during the trading hours (i.e., 9:30 am to 4:00 pm ET, with
the ﬁrst interval spanning 9 to 10 am ET). Retail trades are identiﬁed in TAQ
as in Boehmer et al. (2021).
Figure 3 shows the mean proportion of retail trading for the 50 most popular
stocks on Robinhood during these key outage events. Outages are depicted
23 One might be concerned that unusually heavy volume caused the Robinhood outage. We do
not think this is a major concern since the March 2 full-day outage has volume that ranks 12th
out of the 21 days centered on March 2. March 3 ranks 9th during the same period. June 18 is the
lowest-volume day during the 21 days centered on June 18.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 18
51

3158
The Journal of Finance®
Figure 3. The effect of Robinhood outages on retail trade. In Panels A to C, the sample
consists of the 50 most popular Robinhood stocks on February 28, 2020. In Panel A, March 2,
2020, is the day of a Robinhood outage (red bars). In Panel B, March 2, 2020, is the day of a
Robinhood outage (white bar). A second shorter outage occurred around 10:00 am on March 3,
2020, and lasted for a bit more than an hour (white bar). In Panel C, March 2, 2020, is the day of
a Robinhood outage (white bar). Robinhood tweeted all systems were fully restored at 11:55 am
on March 3, 2020 (black bar). In Panel D, the sample consists of the 50 most popular Robinhood
stocks as of June 17, 2020. The outage occurs between 11:30 and 12:30 on June 18, 2020 (white
bar). (Color ﬁgure can be viewed at wileyonlinelibrary.com)
with red bars. In Panel A, the March 2 full-day outage has the lowest percent
of retail trade. In Panel B, we see that mean retail trading during the 10 am
hour was low on both March 2 (full-day outage) and March 3 (intraday outage).
In Panel C, we see that mean retail trading during the noon hour was low on
March 2 (full-day outage) but high when trading resumed on Robinhood fol-
lowing an early outage on March 3. In Panel D, we see that mean retail trading
between 11:35 am and 12:40 pm ET is low on June 18 relative to other days.
To more formally test for differences, we estimate the following regression
for the March 2 day-long outage:
RetailPercit = a + bOutaget + μtod + μstock + eit,
(1)
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 19
51

Attention-Induced Trading and Returns
3159
where RetailPercit is the percent of trades that are retail trades on TAQ during
hour t for stock i, and Outaget is an indicator variable that takes a value of one
on March 2. As controls, we include time of day ﬁxed effects (μtod) and stock
ﬁxed effects (μstock). The key coefﬁcient estimate, b, measures the percentage
point decline in the percentage of total trading during the Robinhood outage
period. Standard errors are double clustered by day and stock.
For the intraday outage on March 3, we estimate the regression
RetailPercit = a + bOutaget + cRepairt + μt + μtod + μstock + eit.
(2)
For this episode, Outage is an indicator variable that takes a value of one
between 10 and 11 am on March 3, and Repair is an indicator variable that
takes a value of one between noon and 1 pm (the hour after systems are fully
restored).
For the intraday outage on June 18, RetailPercit is measured at ﬁve-minute
intervals to estimate
RetailPercit = a + bOutaget + cPartialt + dRepairt + μt + μtod + μstock + eit.
(3)
In equation (3), Outage takes a value of one for intervals beginning at 11:35
am to 12:35 pm, Partial takes a value of one for the intervals beginning at
12:40 to 1:00 pm, and Repair takes a value of one for the intervals beginning
at 1:05 to 2:00 pm.
Table V summarizes the results. We estimate models for all stocks, the 50
most popular Robinhood stocks, and the 50 highest-attention stocks in columns
(1) to (3), respectively. To identify the 50 highest-attention stocks, we use the
ﬁtted values from the linear probability model that predicts the Robinhood
herding events (column (3), Table IV). In Panel A, we see that the full-day
outage on March 2 reduces trading for all stocks by 0.723 percentage points
(ppt) (p < 0.001), which represents 6% of the average fraction of retail trading
(12.10%) during this period. For the 50 most popular Robinhood stocks, retail
trading declines by 5.227 ppt (p < 0.001), which represents 36% of the average
fraction of retail trading (14.60%) for these stocks. For the 50 high-attention
stocks, we observe a 4.948 ppt decline in trades, which represents a 28% decline
in the typical level of retail trades for these stocks (17.68%).
For the intraday outage of March 3, we observe similar patterns and magni-
tudes during the outage period. However, we also observe detectable rebounds
in trading in the ﬁrst hour after the outage was resolved, suggesting that
the outage generated some pent-up demand to trade among retail investors.
For the intraday June 18 outage, we observe similar patterns but somewhat
smaller magnitudes.24
24 In Figure IA1, we show ﬁve-minute mean trading during the 11:35 am to 12:40 pm interval.
Trading volume increases noticeably at the end of this interval. We do not know if this is random
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 20
51

3160
The Journal of Finance®
Table V
The Effect of Robinhood Outages on Percent Retail Trade
The dependent variable is the proportion of TAQ trades that are identiﬁed as retail trades per
period. Outage is an indicator variable that takes a value of one at the time of an outage. In Panels
B and C, Repair is an indicator variable that takes a value of one for the hour after systems are
fully operational. In Panel C, Partial is an indicator variable that takes a value of one for the period
in which systems are partially restored. Column (1) presents results for all stocks; column (2) for
the 50 most popular Robinhood stocks, and column (3) for the 50 high-attention stocks (based on
ﬁtted values of model 3, Table IV). In Panel A, the Robinhood outage is for the full day on March
2, 2020, and observations are stock-hours. In Panel B, the Robinhood outage starts March 3, 2020
at 10:15 am, with all systems back online sometime between 11 am and noon; observations are
stock-hours. The data set for Panels A and B spans February 18 to March 17. In Panel C, the
Robinhood outage starts June 18, 2020, at 10:39 am, with systems improvement at 12:43 pm and
fully restored at 1:08 pm. The data set for Panel C spans June 4 to July 2. Outage is an indicator
variable that takes a value of one for the time intervals between 11:35 am and 12:35 pm, June 18;
observations are stock-ﬁve-minute periods. Robust standard errors are double clustered by day
and ticker. ***p < 0.01, **p < 0.05, *p < 0.1.
All Stocks
50 Popular Stocks
50 High Attention Stocks
Panel A: March 2 Outage (All Day)
Outage
−0.723***
−5.227***
−4.948***
(0.203)
(0.338)
(0.340)
Observations
1,090,382
8,395
8,319
R2
0.393
0.745
0.618
Day FE
NO
NO
NO
Ticker FE
YES
YES
YES
Time of Day FE
YES
YES
YES
Panel B: March 3 Outage (Late Morning)
Outage
−1.720***
−6.610***
−5.122***
(0.172)
(0.712)
(0.550)
Repair
1.581***
3.742***
1.652***
(0.133)
(0.131)
(0.124)
Observations
1,038,326
7,997
7,821
R2
0.397
0.764
0.614
Day FE
YES
YES
YES
Ticker FE
YES
YES
YES
Time of Day FE
YES
YES
YES
Panel C: June 18 Outage (Late Morning)
Outage
−0.678***
−3.042***
−0.752**
(0.067)
(0.455)
(0.350)
Partial
0.476***
1.005***
0.973**
(0.091)
(0.312)
(0.346)
Repair
0.731***
1.543***
0.458***
(0.092)
(0.190)
(0.145)
Observations
9,443,439
81,814
64,652
R2
0.231
0.621
0.176
Day FE
YES
YES
YES
Ticker FE
YES
YES
YES
Time of Day FE
YES
YES
YES
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 21
51

Attention-Induced Trading and Returns
3161
In summary, the analysis of outages shows that Robinhood users account
for as much as 6.6% of total trades in stocks (and one-third of retail trades)
in the 50 most popular Robinhood stocks. Perhaps more importantly for our
purposes, the analysis reveals that Robinhood users are particularly active in
high-attention stocks, accounting for as much as 6.5% of total trading in these
stocks (and more than one-fourth of total retail trading). The latter result lends
credibility to our assumption that the analysis of Robinhood trading is a good
proxy for attention-motivated trading. These magnitudes are also consistent
with the June 2020 DARTs data suggesting that Robinhood represents ap-
proximately 30% of retail trades.
D. The Robinhood User Interface and Stock Selection
One potential driver for the excessively concentrated trading on Robinhood
could be the coordination of common signals. Given individuals’ aversion to
complexity (Oprea (2020), Umar (2020)), Robinhood employs a simple plat-
form design to reduce the cognitive hurdles to ﬁnancial decision making. Its
simpliﬁed interface is in striking contrast with traditional brokerage ﬁrms,
which provide investors a rich set of indicators and research tools. For exam-
ple, besides basic market information, Robinhood provides only ﬁve charting
indicators, while TD Ameritrade provides 489.25 Presented with a large vari-
ety of stimuli, investors using traditional investing products are likely to have
heterogeneous responses given the limited capacity and highly ﬂexible alloca-
tion of human attention (Kahneman (1973)). In contrast, the reduced number
of stimuli on Robinhood makes it easy for investors to focus their attention and
is likely to generate coordinated attention-induced responses.
In this section, we analyze the effect of Robinhood’s lists on investor’s trad-
ing. These lists are displayed prominently on the platform and are easily ac-
cessible under “News/Popular Lists.” The two most prominently displayed lists
are the “Most Popular” list and the Top Movers list. We focus on the Top Movers
list because it is short and constantly changing, while the Most Popular list
is long and largely static. The Top Movers list includes stocks with the day’s
largest percent gains and losses since the market close the previous day.
D.1. Top Movers’ Absolute Return Feature
The default sorting of the Top Movers list is based on absolute returns and
thus mixes top gainers and top losers.26 This feature differs from almost all
other media accounts (e.g., Wall Street Journal, Yahoo! Finance, CNBC, etc.)
or perhaps a result of Robinhood systems being partially operational before the posted time on
their website.
25 See https://www.stockbrokers.com/compare/robinhood-vs-tdameritrade.
26 Figure IA2 provides an example of the Top Movers list on October 8, 2020, as shown on the
website. The initial screen shows four top-ranked stocks by absolute returns, which includes three
top gainers and one top loser.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 22
51

3162
The Journal of Finance®
that also report the top movers, but separate the top gainers and top losers
rather than mix them together.
We exploit this unusual feature of Robinhood’s Top Movers list and compare
the buying activity between Robinhood investors and general retail investors
measured by TAQ data. Given the top gainers and top losers are displayed
in the same list on Robinhood, we would expect the degree of availability for
the two groups of stocks to be similar for Robinhood investors. As a result, the
buying activity of Robinhood investors would not differ much between top gain-
ers and top losers. In contrast, the top losers are less prominently displayed on
other media accounts, where they are reported separately from the top gainers.
Accordingly, we would expect general retail investors to respond less strongly
to top losers than to top gainers.
Figure 4 presents graphical evidence for the comparison. We measure the
buying activity of Robinhood investors by intraday user change27 and the buy-
ing behavior of retail investors by TAQ intraday retail net purchases (i.e., num-
ber of buyer-initiated retail trades minus seller-initiated retail trades).28 In
Panel A, we rank stocks based on absolute overnight returns from the market
close of day t −1 to the market open of day t. We measure buying activity on
day t,29 subsequent to ranking. The graph on the left shows how the buying ac-
tivity of Robinhood users on day t varies with the rank of the 20 stocks with the
highest absolute value of overnight returns (i.e., the 20 highest Top Movers);
the graph on the right shows buying activity for retail investors. We plot the
mean Robinhood intraday user change and TAQ net retail buying for the top
20 movers separately for stocks with positive returns (top gainers) and stocks
with negative returns (top losers).
Stocks with bigger absolute price changes are bought more by both Robin-
hood users and retail investors. This is consistent with evidence of attention-
based buying documented in Barber and Odean (2008). Robinhood investors
respond similarly to top gainers and losers, while other retail investors buy top
gainers much more aggressively than top losers. This is consistent with our hy-
pothesis that the attention of Robinhood users is directed to both top winners
and top losers because both appear on the Top Movers list, while the attention
of other investors is directed more to stocks that appear on top gainers lists. In
a robustness test, Panel B sorts top movers on the daily close-to-close return
and shows similar patterns.
27 Since the algorithm by Boehmer et al. (2021) cannot identify retail trades at the open auction,
for this analysis we exclude the user change at the open on Robinhood to make the Robinhood user
change more comparable with TAQ net retail buying.
28 Since Robinhood does not allow short selling, to make TAQ net retail buying more comparable
with the Robinhood user change, we remove short trades from TAQ following Boehmer and Song
(2020). The results are similar with or without short trades removed from TAQ.
29 We do not have the actual Top Mover list as it appears on Robinhood but use the return ranks
as a proxy for whether a stock is likely to appear on the list. Accordingly, we only rank stocks with
market cap greater than $300 million, since Robinhood only ranks stocks above this size threshold
to create the Top Mover list.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 23
51

Attention-Induced Trading and Returns
3163
Figure 4. Mean Robinhood intraday user change versus TAQ intraday net retail buying
on top mover rankings. The panels present the mean Robinhood intraday user change and TAQ
intraday net retail buying of day t against top mover rankings based on different measures. The
rankings are sorted for stocks with market cap above $300 million at the market open of day
t. Panel A sorts top movers on absolute overnight returns of day t; Panel B sorts top movers
on absolute daily returns of day t. RH intraday user change measures the change in Robinhood
users from ﬁrst time stamp that excludes market open trades to the last time stamp before the
market closes. TAQ intraday net retail buying is TAQ retail buying minus TAQ retail selling
(with short trades removed following Boehmer and Song (2020)) during market trading hours.
The error bar represents the 90% conﬁdence interval for the mean. (Color ﬁgure can be viewed at
wileyonlinelibrary.com)
To more formally test whether the difference is statistically signiﬁcant be-
tween Robinhood intraday user change and TAQ intraday net retail buying,
we estimate
NetBuyit = β0 + β1Scoreit + β2IRit<0 + β3Scoreit × IRit<0 + αt + εit,
(4)
where the dependent variable is Robinhood or TAQ buying activity for stock i
on day t, and Scoreit assigns a score to each rank of top movers. For expositional
ease, we assign 20 to the stock with highest absolute returns and 1 to the stock
with the 20th highest absolute returns, so scores increase with the absolute
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 24
51

3164
The Journal of Finance®
returns. The variable IRit<0 is an indicator that equals one if the stock return
is negative. For each day, we include only the top 20 movers in the regression.
Day ﬁxed effects are included, and robust standard errors are clustered at the
daily level. We hypothesize that while other retail investors are less likely to
buy top losers than top gainers, Robinhood investors are not.
In Table VI, columns (1) and (2) sort top mover scores on absolute overnight
returns. The buying activity increases with top mover scores in both columns.
This result indicates that, for both Robinhood investors (column (1)) and gen-
eral retail investors (column (2)), attention is affected by the ranks within the
top movers, with higher ranks making stocks more salient.
The key difference in the buying activity of Robinhood investors and general
retail investors is reﬂected by the coefﬁcient on the indicator for negative re-
turns (or, top losers). For general retail investors, the top losers garner much
less buying activity than the top gainers. Within the same rank, the TAQ net
retail buying decreases by 133.2 trades for a top loser versus a top winner,
which is similar to the decrease associated with the rank of a top gainer drop-
ping by 10 (≈133.2/13.41). In addition, the interaction of the top rank and
negative return is negative indicating the magnitude of negative return effect
is larger for the more extreme returns. This pattern differs from that of Robin-
hood investors. If anything, Robinhood buying activity is slightly stronger for
the top losers than for the top gainers and the interaction effect is positive.
Columns (3) and (4) sort top mover scores on absolute daily returns and ﬁnd
similar results for the general population of retail traders.
An alternative explanation for the differences between Robinhood and TAQ
shown in Figure 4 and Table VI could be a different combination of positive
feedback traders versus contrarians between Robinhood and general retail in-
vestors. Under this explanation, investors simply respond to extreme price
movements rather than the information display. Here, there are two distinct
types of traders: positive feedback traders who respond positively to extreme
past returns and contrarians who respond negatively to extreme past returns.
If the two groups of investors are evenly distributed in Robinhood while pos-
itive feedback traders dominate among general retail investors, we would ob-
serve the patterns shown in Figure 4 and Table VI. To address this concern
and provide additional evidence of the app’s inﬂuence on Robinhood users, in
the next section we explore another unique feature of the Top Movers list.
D.2. Market Cap Requirement for Top Movers
In this section, we exploit a regression discontinuity design to further es-
tablish the causal impact of the app interface on investors’ trading behavior.
Ideally, one would consider a regression discontinuity design that exploits the
threshold between the 20th and 21st top mover stocks. However, as we do not
observe the actual ranking of price swings and the ranking is likely to change
throughout the day, the noise potentially introduced by using approximated
rankings may be too large to result in clean identiﬁcation.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 25
51

Attention-Induced Trading and Returns
3165
Table VI
The Effect of Top Movers on Robinhood Intraday User Change versus TAQ Intraday Net Retail Buying
The table examines how the top mover rankings affect Robinhood (RH) intraday user change and TAQ intraday net retail buying. The rankings are
sorted for stocks with market cap above $300 million at the market open of day t. The stocks are sorted on absolute overnight return (columns (1)
and (2)) and absolute daily return (columns (3) and (4)), respectively. The sample requires both Robinhood user change and TAQ net retail buying
be available and only includes the top 20 stocks for each day. RH intraday user change measures the change in RH users from ﬁrst time stamp that
excludes market open trades to the last time stamp before the market closes. TAQ Intraday Net Retail Buying is TAQ retail buying minus TAQ retail
selling (with short trades removed following Boehmer and Song (2020)) during market trading hours. Top mover score assigns a score for each rank,
with 20 for the highest absolute return and 1 for the 20th highest. Negative return is an indicator variable that equals one if the stock return is
negative. Regressions include day ﬁxed effects. Robust standard errors are clustered by day. ***p < 0.01, **p < 0.05, *p < 0.1.
Top Mover Score is Sorted on:
Absolute Overnight Return (t)
Absolute Return (t)
RH Intraday User
Change (t)
TAQ Intraday Net
Buy (t)
RH Intraday User
Change (t)
TAQ Intraday Net
Buy (t)
(1)
(2)
(3)
(4)
Top mover score
7.198***
13.41***
11.95***
17.89***
=20: highest absret; =1: 20th highest absret
(0.510)
(1.701)
(0.539)
(1.671)
Negative return
46.49***
−133.2***
2.980
−86.90***
=1: top mover return is negative; =0, otherwise
(9.913)
(29.950)
(10.252)
(28.705)
Top mover score × Negative return
3.462***
−6.591***
−1.472**
−6.245***
(0.736)
(2.369)
(0.724)
(2.222)
Day FE
Yes
Yes
Yes
Yes
Observations
10,247
10,247
10,342
10,342
R2
0.233
0.211
0.272
0.192
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 26
51

3166
The Journal of Finance®
To mitigate this concern, we focus on the market cap requirement for the Top
Movers list. Speciﬁcally, Robinhood requires that stocks be above $300 million
in market cap to be displayed in the Top Movers list.30 This feature allows us to
employ a sharp regression discontinuity design to study Robinhood investors’
responses to extreme price movements of stocks with market cap around the
$300 million cutoff. If the interface does not affect investors’ trading behavior,
there should not be any discernible differences in investor responses to top
mover stocks with market cap above or below the $300 million cutoff. Our
assumption is that any differences in investor responses on either side of the
cutoff, after controlling for the effect of market cap, should be due only to the
impact of the Robinhood interface on investors’ trading behavior.
Speciﬁcally, we select top mover stocks with market cap within a small
bandwidth (e.g., $50 million) around the $300 million cutoff for our regres-
sion discontinuity analysis. Given the Robinhood Top Movers list displays the
top 20 stocks with market cap above $300 million that have the largest abso-
lute percentage price moves measured from the previous market close price,
we consider our treatment group for day t as stocks with market cap ∈($300
million, $350 million] that ranked top 20 by absolute day-t overnight returns
(|ROvernight
Treatment, t|) among all stocks with market cap above $300 million. Our con-
trol group then includes stocks with market cap ∈[$250 million, $300 million]
that have absolute day-t overnight returns (|ROvernight
Control, t |) close to stocks in the
treatment group.31 We also vary the bandwidth choices at three other levels
(i.e., $75, $100, and $125 million) to test the robustness of our results.
We exploit a sharp regression discontinuity design. Intuitively, this esti-
mation exploits the discontinuity in information display at the $300 million
market cap threshold and tests for discontinuities in investor buying behavior
around this threshold. We estimate the following pooled, cross-sectional sharp
RD speciﬁcation:
userchgit = β0 + β1Imktcapit>$300m +
N

n=1
βn
2

mktcapit −$300m
n
+
N

n=1
βn
3

mktcapit −$300m
n × Imktcapit>$300m + εit,
(5)
30 According
to
Robinhood
Web
Disclosures
(https://cdn.robinhood.com/disclosures/
WebDisclosures.pdf), “Robinhood uses a proprietary algorithm to display stocks with a mar-
ket cap of more than $300 million that have largest price movements as measured from the
previous market close price” for the Top Movers list.
31 For each stock in the treatment group, we ﬁnd a matched stock that has the closest absolute
return distance with the treated stock among all stocks with market cap ∈[$250 million, $300
million] that satisfy 0.5 ≤|ROvernight
Treatment, t|/|ROvernight
Control, t| ≤2. With the matching, the distributions of
the absolute overnight returns are similar between the treatment group (mean: 0.090, median:
0.073, std dev: 0.059) and the control group (mean: 0.078, median, 0.062, std dev: 0.054). We
implement the same sharp RD test on the absolute overnight returns in Table IAIII and ﬁnd no
discontinuity in the absolute overnight returns at the $300 million threshold.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 27
51

Attention-Induced Trading and Returns
3167
Figure 5. Robinhood intraday user change for top movers around the $300 million mar-
ket cap cutoff. This ﬁgure shows a binned scatterplot for the Robinhood (RH) intraday user
change against market cap for stocks included in the regression discontinuity analysis as described
in Table VII. The red lines are the linear (Panel A), quadratic (Panel B), and cubic ﬁtted lines
(Panel C) estimated separately for market cap above and below the $300 million cutoff. Shaded
areas indicate the 90% conﬁdence interval. (Color ﬁgure can be viewed at wileyonlinelibrary.com)
where userchgit is the net buying activity of Robinhood investors measured
by the intraday user changes on day t, mktcapit is the market cap of stock
i at the market open of day t, and Imktcapit>$300m is an indicator that equals
one for stocks with market cap at the market open of day t greater than $300
million. As controls, we include different polynomial functions of market cap
(N = 1, 2, 3) so that the point estimate on the above-cutoff indicator variable
(β1) is identiﬁed under the assumption that the relation between investor trad-
ing behavior and market cap is not discontinuous exactly at the $300 million
cutoff threshold for reasons besides the app interface.
The regression results in Table VII show a discontinuous increase in Robin-
hood investors’ buying activities when market cap exceeds the $300 million
cutoff. The coefﬁcient estimate (β1) is positive and statistically signiﬁcant. The
results are robust to different bandwidth choices and to including controls for
linear, quadratic, and cubic functions of market cap. Graphical evidence cor-
responding to the three speciﬁcations with the $50 million bandwidth is pre-
sented in Figure 5. As the ﬁgure shows, the intraday user change exhibits a
clear jump at the market cap cutoff of $300 million.
To verify that we are not obtaining spurious estimates of the effects of the
information display using the regression discontinuity design, we conduct a
placebo test exploiting alternative market cap cutoffs at $250 million. For the
placebo test using the $250 million cutoff, we select top mover stocks within a
$50 million bandwidth around the $250 million cutoff (i.e., market cap ∈($200
million, $300 million]) and conduct the same regression estimation. Since none
of the stocks in this placebo exercise would be displayed on the Top Movers list
by Robinhood, the coefﬁcient estimate on the treatment effects (β1) should be
zero. As shown in the regression results in the Internet Appendix Table IAIV,
the coefﬁcient estimate (β1) is indeed statistically insigniﬁcant for the $250
million cutoff.32
32 The Internet Appendix may be found in the online version of this article.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 28
51

3168
The Journal of Finance®
Table VII
Regression Discontinuity in Robinhood Intraday User Change for
Top Mover Stocks
This table estimates a sharp regression discontinuity regression that exploits the discontinuity in
Robinhood (RH) intraday user changes around the market cap cutoff of $300 million. Speciﬁcally,
we use RH intraday user changes at day (t) as the dependent variable, and include an indicator
that equals one for stocks that have market cap at the market open of day t greater than $300
million. We include different polynomial functions of market cap as controls. Our analysis varies
sample bandwidth from Panels A to D. Panel A uses a sample bandwidth of $50 million (i.e.,
market cap ∈[$250 million, $350 million]). For stocks with market cap above $300 million, we
select stocks that rank top 20 by absolute day-t overnight returns among all stocks with market
cap above $300 million. For stocks with market cap below $300 million, we include matched stocks
with absolute day-t overnight returns close to stocks with market cap above $300 million. Standard
errors are reported in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.
RH Intraday User Change
(1)
(2)
(3)
Panel A: Sample Bandwidth = 50 m
Larger than 300 m
95.71**
115.8**
247.3***
(39.472)
(53.933)
(78.782)
Polynomial order, N
1
2
3
Observations
1,332
1,332
1,332
R2
0.013
0.014
0.020
Panel B: Sample Bandwidth = 75 m
Larger than 300 m
90.57**
109.8*
97.18
(37.123)
(61.131)
(80.490)
Polynomial order, N
1
2
3
Observations
2,068
2,068
2,068
R2
0.006
0.006
0.008
Panel C: Sample Bandwidth = 100 m
Larger than 300 m
77.55**
131.0**
174.8**
(33.300)
(53.648)
(72.724)
Polynomial order, N
1
2
3
Observations
2,782
2,782
2,782
R2
0.011
0.011
0.012
Panel D: Sample Bandwidth = 125 m
Larger than 300 m
49.13*
102.3**
147.7**
(26.718)
(41.947)
(59.496)
Polynomial order, N
1
2
3
Observations
3,406
3,406
3,406
R2
0.016
0.016
0.016
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 29
51

Attention-Induced Trading and Returns
3169
Overall, the results in this section are consistent with the idea that the
Robinhood app’s design impacts the trading decisions of its users.
III. Return Results
Above we ﬁnd that relative to general retail investors, Robinhood users
demonstrate more concentrated buying and selling. Concentrated buying is
likely driven by attention and inﬂuenced by information display on the Robin-
hood interface. On days of extreme buying (i.e., herding events), Robinhood
users could create price pressure (see Coval and Stafford (2007)). In this sec-
tion, we examine the return patterns around such herding events.
A. Event-Time Results
Our ﬁrst analysis in this section examines abnormal returns from day −10
to event day 20 around herding events, where day 0 is the herding day. Ab-
normal returns are calculated as the stock’s return less the value-weighted
CRSP index. In Table VIII, we report the mean abnormal return for each day
and buy-and-hold abnormal returns (BHARs) separately before and after the
event. For example, preevent BHARs are calculated as
BHARiτ =
τ
t=τ−10
(1 + Rit) −
τ
t=τ−10
(1 + Rmt) .
(6)
We also report the percent of returns that are positive.
Standard errors are computed with clustering on the event day since we may
have multiple events on the same day. The statistics underlying the mean daily
returns therefore rely on the reasonable assumption that returns are serially
independent. The longer-horizon abnormal returns are also clustered by event
day, which helps corrects for cross-sectional dependence issues. However, the
standard errors are likely too small because of the overlapping nature of the
returns at longer horizons. We address this econometric concern in the next
section with a calendar-time trading strategy.
The BHARs at longer horizons have the advantage that they accurately
represent the return earned by investors. Cumulative abnormal returns (the
sum of daily abnormal returns) are a positively biased representation of long-
horizon abnormal returns in the presence of temporary price pressure effects
or bid-ask bounce, both of which are likely issues in the stocks with the herding
episodes we study.33
33 To see this, consider a stock that cycles between ask and bid prices of $10 and $11 across
three days, starting at the $10 ask. Assume the market return is zero. The daily returns on day 1
are 10% (1/10) and day 2 are −9.1% (−1/11). The daily abnormal returns are 10% and −9.1%. The
cumulative abnormal return is 0.9% (10% to 9.1%). The BHAR is zero (11/10 × 10/11–1). While
this example uses bid and ask prices for simplicity, the same logic applies to any mechanism that
generates negative serial dependence in returns (e.g., temporary price pressure effects or liquidity
provision).
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 30
51

3170
The Journal of Finance®
Table VIII
Event-Time Abnormal Returns
The table reports abnormal returns around Robinhood user herding events. Abnormal returns
(AR) are computed as the raw return minus the CRSP value-weighted average return. Abnormal
returns are averaged across all events. Buy-and-hold abnormal returns (BHAR) are computed as
the product of one plus the stock’s return through event day t less the product of one plus the
market return for the same period. Standard errors are computed by clustering on event day.
%Positive is the percent of returns that are positive. Herding events are deﬁned as the top 0.5%
of stocks with positive user change ratio on day 0 and a minimum of 100 users on the prior day.
***p < 0.01, **p < 0.05, *p < 0.1.
Event Day
AR
Std. Err.
% Positive
BHAR
Std. Err.
% Positive
Preevent
−10
−0.08%
0.14%
47%
−0.08%
0.14%
47%
−9
−0.05%
0.15%
46%
−0.02%
0.27%
46%
−8
0.38%
0.30%
47%
0.22%
0.41%
46%
−7
−0.06%
0.21%
45%
−0.01%
0.50%
45%
−6
−0.28%**
0.13%
46%
−0.48%
0.42%
44%
−5
0.02%
0.15%
48%
−0.47%
0.45%
45%
−4
0.39%*
0.20%
48%
−0.10%
0.50%
46%
−3
0.51%**
0.21%
48%
0.52%
0.59%
47%
−2
0.40%*
0.22%
49%
0.81%
0.60%
48%
−1
4.59%***
0.52%
56%
5.60%***
0.87%
51%
0
13.95%***
0.92%
63%
22.16%***
1.73%
58%
Postevent
1
−1.23%***
0.24%
42%
−1.23%***
0.24%
42%
2
−0.85%***
0.18%
42%
−2.11%***
0.31%
40%
3
−0.43%***
0.16%
44%
−2.74%***
0.29%
38%
4
−0.35%**
0.15%
43%
−3.15%***
0.31%
37%
5
−0.32%**
0.14%
44%
−3.55%***
0.32%
37%
6
−0.37%***
0.13%
44%
−3.99%***
0.34%
37%
7
−0.14%
0.14%
46%
−4.17%***
0.37%
36%
8
0.17%
0.17%
45%
−4.13%***
0.39%
36%
9
0.07%
0.13%
45%
−4.19%***
0.40%
36%
10
0.15%
0.16%
45%
−4.15%***
0.40%
37%
11
−0.03%
0.14%
42%
−4.33%***
0.40%
36%
12
−0.03%
0.12%
46%
−4.42%***
0.41%
36%
13
−0.06%
0.13%
44%
−4.51%***
0.43%
36%
14
−0.09%
0.14%
45%
−4.68%***
0.44%
35%
15
−0.15%
0.10%
45%
−4.87%***
0.45%
35%
16
0.03%
0.11%
46%
−4.83%***
0.47%
35%
17
−0.04%
0.14%
45%
−4.76%***
0.56%
35%
18
0.18%
0.22%
46%
−4.71%***
0.56%
35%
19
−0.01%
0.11%
46%
−4.80%***
0.57%
35%
20
0.15%
0.16%
46%
−4.74%***
0.58%
35%
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 31
51

Attention-Induced Trading and Returns
3171
Figure 6. Returns around herding events. The panel on the left depicts mean BHARs and
Robinhood user changes from 10 days before to 21 days after herding events. The green line rep-
resents BHARs, whereas the gray bars represent user changes. The panel on the right displays
postevent mean BHARs starting from day 0. Herding events are deﬁned as the top 0.5% of stocks
with positive user change ratio on day 0 and a minimum of 100 users on the prior day. (Color ﬁgure
can be viewed at wileyonlinelibrary.com)
Prior to the herding event, stocks have abnormal returns near zero. A day
or two before the herding event, average returns increase and become statis-
tically signiﬁcant. The stocks then experience an extremely positive return on
the herding day—averaging 14%. Interestingly, many stocks have negative re-
turns the day prior to and on the day of the herding event. This is consistent
with our prior results documenting that extreme negative returns draw the
attention of Robinhood users as well.
The pattern after the herding events is starkly different. Immediately af-
ter the herding event, returns turn signiﬁcantly negative. After just ﬁve days,
stocks in the top 0.5% (rh_herd) experience negative abnormal returns of
−3.5%. By the end of the 20-day period, the return decline totals almost 5%.
These results are economically and statistically signiﬁcant, and are not driven
by just a few stocks as almost two-thirds of rh_herd stocks have negative cu-
mulative returns by the end of the 20 days.
To visualize these return patterns, in Figure 6 we plot the BHARs for our
rh_herd events. Results over the entire period are reported in Panel A, and
BHARs starting at Day 1 are reported in Panel B. The pattern of returns
around herding events is quite clear. Robinhood users are attracted by extreme
return events. Their coordinated buying leads to price pressure and then sub-
sequent poor return performance. We also ﬁnd little evidence of unusual return
movement beyond 20 days (see Internet Appendix Figure IA3).
To systematically analyze the relation between the herding intensity and
price reversal, we analyze stocks with a minimum of 100 Robinhood users and
identify different sets of herding episodes by varying the daily user change
ratio from 1.1 to 8.5 (i.e., from a 10% to 750% increase in users holding the
stock). Most user change ratios are close to 1.0 (see Table I) and a user ratio of
1.1 corresponds to approximately the 98th percentile of the distribution of user
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 32
51

3172
The Journal of Finance®
Figure 7. Herding intensity and price reversal. The panels show how herding intensity and
20-day BHARs vary with the daily user change ratio cutoff to identify herding episodes for stocks
with at least 100 Robinhood users. Panel A plots the number of herding episodes (in log scale)
against the user change ratio cutoff. Panel B plots the 20-day BHARs (%) against the user change
ratio cutoff. The daily user change ratio cutoff varies from 1.1 to 8.5. (Color ﬁgure can be viewed
at wileyonlinelibrary.com)
ratios, so we are analyzing unusual events. For more modest user changes, we
observe no relation between returns and the user ratio. In Figure 7, Panel A,
we plot the number of herding episodes (y-axis in log scale) against the user
change ratio to identify episodes (x-axis). At a user change ratio of 1.1, we
observe over 20,000 herding episodes; at a user change ratio of 8.5, we observe
45 episodes. In Panel B, we show that the mean abnormal return following
these herding events grows from a statistically signiﬁcant −1.8% at a user
change ratio of 1.1 (>20,000 events or about 36 events per day) to an extremely
large −19.6% at a user change ratio of 8.5 or more (45 events or about one
event every 12 days). We thus ﬁnd a clear relation between the magnitude of
the buying intensity and the subsequent reversal.
It is possible that the return magnitudes of Figure 7, Panel B, increase as we
move to more extreme cutoffs of the user change ratio because the identiﬁed
stocks become smaller rather than the effect of the herding event. To rule out
this size-based explanation, we rerun Figure 7 after throwing out stocks with
a market cap greater than $1 billion, which ensures that mean and median
size are the same across the user change ratio cutoffs. We continue to see a
dramatic increase in the observed return effects (see Figure IA4, Panel B).
B. Calendar-Time Trading Strategy
To address the cross-sectional dependence issue underlying event-time anal-
yses and to investigate the returns earned on a trading strategy that follows
the herding episodes, we construct a calendar-time portfolio that invests $1 in
each herding episode stock at the close of the event day and holds the stock
for ﬁve days (without rebalancing). We estimate the daily portfolio abnormal
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 33
51

Attention-Induced Trading and Returns
3173
return (alpha) by regressing the portfolio excess return on the Fama-French
ﬁve-factor model plus a momentum factor,
Rpt −R ft = α + β

Rmt −R ft

+
K

k=1
ckFk
t + ept,
(7)
where R ft is the daily risk-free return, Rmt is the value-weighted market in-
dex, and Fk
t are the k = 1, K factor returns related to size, value, investment,
proﬁtability, and momentum (taken from Ken French’s online data library).
For completeness, we include results with just the market excess return (i.e.,
the CAPM) and four-factor alphas using market, size, value, and momentum
factors.
The results, reported in Table IX, are broadly consistent with the event-time
analysis above. Over the sample period, the calendar-time portfolio earns an
economically large daily alpha of −55 to −61 bps (columns (1) to (3)), which
is in line with the ﬁve-day event-time market-adjusted return of −3.55%. We
also ﬁnd that the portfolio alphas are more negative during the 2020 pandemic
period, ranging from −78 to −94 bps per day. The Robinhood user changes for
herding stocks are also more dramatic during the pandemic period, with a
mean (median) userratio of 1.75 (1.47) pre-COVID (prior to March 13, 2020)
and 2.56 (1.83) post-COVID. However, the return effects of herding events are
not unique for the pandemic period, as the negative alphas are also sizable
during the pre-COVID period.
C. Regression Results
To further test the return patterns that we explore in event time, we regress
daily stock returns for stock i on day t (Rit) on lags of the key herding variable
(rh_herd) and controls,
Rit = a +
5

k=1
bkrh_herdi,t−k +
5

k=1
cktaq_retimbi,t−k +
5

k=1
dkRi,t−k + μt + eit.
(8)
The {bk} coefﬁcients estimate the impact of herding events on returns one
to ﬁve days after the event. We sequentially add groups of control variables
to assess how they interact with the herding events that we analyze. Robust
standard errors are estimated with clustering by day, which addresses cross-
sectional dependence that affects the event-time graphs of the prior section.
Turning to the control variables, we ﬁrst include lagged retail order im-
balance (taq_retimb), which positively predicts returns at short horizons in
the United States (Barber, Odean, and Zhu (2008), Kaniel, Saar, and Titman
(2008), Kelley and Tetlock (2013), Boehmer et al. (2021)). The retail imbal-
ance variable allows us to assess whether the poor returns we document are a
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 34
51

3174
The Journal of Finance®
Table IX
Calendar Portfolio Returns of Herding Events
The dependent variable is the dollar-weighted-average daily return over the risk-free rate (%). For each rh_herd event (top 0.5% of positive user
change ratio on day t and a minimum of 100 users on day t −1), 1/Price shares are purchased at the end of the herding day. These stocks are held for
ﬁve days before being liquidated. Dollar-weighted average daily returns are a dollar-weighted-average of all stocks held based on the position value
at the end of the prior day. Returns are over the entire period (columns (1) to (3)), prior to March 2020 (columns (4) to (6)), or March 2020 and after
(columns (7) to (9)). The key estimation is the constant, Alpha. Control variables include excess market returns (Rm −R f ), small-minus-big factor
(SMB), high-minus-low factor (HML), momentum factor (MOM), robust-minus-weak operating proﬁtability factor (RMW), and conservative-minus-
aggressive factor (CMA). Robust standard errors are reported in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.
Entire Period
Prior to March 2020
March 2020 and after
Daily Excess Return on Dollar-Weighted Calendar Portfolio of Herding Events (%)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Alpha
−0.612***
−0.569***
−0.552***
−0.535***
−0.526***
−0.511***
−0.938***
−0.818***
−0.777***
(0.112)
(0.110)
(0.108)
(0.118)
(0.118)
(0.118)
(0.305)
(0.284)
(0.273)
Mkt_R f
0.811***
0.709***
0.660***
0.646***
0.512***
0.437***
0.883***
0.725***
0.699***
(0.111)
(0.120)
(0.127)
(0.120)
(0.130)
(0.135)
(0.138)
(0.155)
(0.156)
SMB
0.689***
0.446**
0.467
0.332
0.719**
0.399
(0.218)
(0.219)
(0.308)
(0.304)
(0.317)
(0.384)
HML
0.103
0.355
−0.421*
−0.148
0.298
0.533
(0.176)
(0.217)
(0.231)
(0.269)
(0.311)
(0.394)
MOM
−0.118
−0.229
−0.511**
−0.599**
0.050
−0.085
(0.158)
(0.166)
(0.233)
(0.238)
(0.232)
(0.266)
RMW
−0.968***
−0.981***
−0.961
(0.301)
(0.337)
(0.595)
CMA
−0.773*
−0.703
−0.858
(0.424)
(0.520)
(0.749)
Observations
555
555
555
448
448
448
107
107
107
R2
0.190
0.234
0.256
0.056
0.079
0.098
0.407
0.481
0.496
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 35
51

Attention-Induced Trading and Returns
3175
manifestation of general retail order imbalance predicting returns during our
sample period.
We also include lagged returns (Ri,t−k) to control for the well-documented
tendency for return reversals at short horizons up to one month (French
and Roll (1986), Lo and Mackinlay (1988, 1990), Jegadeesh (1990), Lehmann
(1990), Campbell, Grossman, and Wang (1993)) with these effects stronger in
volatile markets including the 2020 pandemic period (Nagel (2012), Drechsler,
Moreira, and Savov (2020)). These studies speculate that the origins of return
reversals may emanate from overreaction, the rewards to providing liquid-
ity provision, and/or more banal microstructure issues (e.g., prices bouncing
between bid and ask prices). Attention-motivated trading leads to excessive
buying and return reversals following attention-grabbing price increases, so
may help explain the returns associated with short-term contrarian strategies.
Thus, the inclusion of lagged returns may be overcontrolling as both return re-
versals and the negative returns we document may have similar origins in
attention-motivated trading.
Table X presents results for our herding measure based on the top 0.5% of
daily Robinhood user changes. We present results for the full sample of herd-
ing events in columns (1) to (3). The last row of the table presents the summed
coefﬁcients on the herding indicator variables, which can be interpreted as
the ﬁve-day abnormal return after the event. These ﬁve-day return estimates
range from losses of 2.595% to 2.942%. In Table IAV, we use more ﬁne-grained
controls for extreme positive and negative return moves and ﬁnd similar
ﬁve-day abnormal return estimates derived from the lags of the key rh_herd
variable.
We separately analyze herding events that occur following a positive or nega-
tive overnight return using this regression model.34 To analyze herding events
that follow a positive overnight return, the key indicator variable takes a value
of one only if the overnight return (measured from the close on day t −1 to the
open on day t) is positive and the change in users from the close on day t −1
to the close on day t is in the top 0.5% of stocks with positive user changes.
The results are presented in columns (4) to (6). The herding events that are
preceded by positive overnight returns predict somewhat stronger negative
abnormal returns, ranging from −3.025% to −3.909%. We estimate a similar
regression conditioning on negative overnight returns in columns (7) to (9).
Though smaller in magnitude, we continue to observe negative abnormal re-
turns for those herding events preceded by negative overnight returns, with
ﬁve-day abnormal returns ranging from −1.490% to −2.341%.
These analyses provide further evidence that attention-motivated trading
generates predictable poor returns.
34 We condition on overnight returns rather than close-to-close returns because overnight re-
turns generally precede Robinhood user changes, which occur most commonly at or after the open
of trading. If we condition on close-to-close returns, the herding event might cause the positive
returns. When we condition on positive (negative) close-to-close returns, the ﬁve-day abnormal re-
turn estimated in columns (6) and (9) are −3.375% and −1.887%, respectively (both with p < 0.01).
See Table IAVI.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 36
51

3176
The Journal of Finance®
Table X
Regression of Daily Returns on Lagged Robinhood Herding Indicator (Top 0.5% Percentage User Change)
The dependent variable is the daily stock return (%) winsorized at the 0.1% level (ret). In columns (1) to (3), rh_herd(t) is an indicator variable that
equals one if the percentage change in users is in the top 0.5% for stocks with positive user changes on day t and a minimum of 100 users on day
t −1. In columns (4) to (6) (columns (7) to (9)), the rh_herd(t) indicator variable equals one if the prior conditions are met and the overnight return
from the close on day t −1 to the open on day t is positive (negative). Control variables include retail order imbalance from TAQ (taq_retimb), lagged
returns (ret), lags of an indicator variable if the rh_herd measure is missing (rh_herd is set equal to zero), and day ﬁxed effects. Five-day AR (%) is
the sum of the coefﬁcients on the ﬁve lags of rh_herd. Robust standard errors clustered by day are reported in parentheses. ***p < 0.01, **p < 0.05,
*p < 0.1.
All Events
Overnight Return > 0
Overnight Return < 0
ret(t) (%)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
rh_herd(t −1)
−1.336***
−1.039***
−1.110***
−1.770***
−1.132***
−1.276***
−0.514**
−0.970***
−0.894***
(0.160)
(0.171)
(0.174)
(0.215)
(0.256)
(0.260)
(0.209)
(0.226)
(0.229)
rh_herd(t −2)
−0.758***
−0.813***
−0.798***
−1.148***
−1.235***
−1.212***
−0.152
−0.140
−0.152
(0.129)
(0.141)
(0.141)
(0.175)
(0.221)
(0.220)
(0.175)
(0.202)
(0.203)
rh_herd(t −3)
−0.298**
−0.191
−0.183
−0.378**
−0.147
−0.127
−0.268
−0.431**
−0.449**
(0.122)
(0.134)
(0.132)
(0.160)
(0.202)
(0.199)
(0.194)
(0.208)
(0.207)
rh_herd(t −4)
−0.318***
−0.266**
−0.266**
−0.322**
−0.169
−0.178
−0.388**
−0.557***
−0.534***
(0.115)
(0.128)
(0.128)
(0.144)
(0.183)
(0.182)
(0.163)
(0.181)
(0.181)
rh_herd(t −5)
−0.231**
−0.286**
−0.273**
−0.292*
−0.342*
−0.325*
−0.167
−0.243
−0.237
(0.117)
(0.131)
(0.131)
(0.152)
(0.190)
(0.190)
(0.166)
(0.171)
(0.171)
ret(t −1)
−0.046***
−0.037***
−0.045***
−0.036***
−0.047***
−0.038***
(0.012)
(0.012)
(0.012)
(0.012)
(0.012)
(0.012)
ret(t −2)
−0.002
−0.002
−0.001
−0.001
−0.003
−0.003
(0.012)
(0.012)
(0.012)
(0.012)
(0.012)
(0.012)
ret(t −3)
−0.021*
−0.021*
−0.021*
−0.021*
−0.022*
−0.021*
(0.012)
(0.012)
(0.012)
(0.012)
(0.012)
(0.012)
(Continued)
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 37
51

Attention-Induced Trading and Returns
3177
Table X—Continued
All Events
Overnight Return > 0
Overnight Return < 0
ret(t) (%)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
ret(t −4)
−0.017*
−0.017*
−0.017*
−0.017*
−0.018*
−0.017*
(0.010)
(0.010)
(0.010)
(0.010)
(0.010)
(0.010)
ret(t −5)
−0.005
−0.005
−0.005
−0.005
−0.005
−0.006
(0.010)
(0.010)
(0.010)
(0.010)
(0.010)
(0.010)
taq_retimb(t −1)
0.043***
0.043***
0.042***
(0.008)
(0.008)
(0.008)
taq_retimb(t −2)
0.012
0.012
0.010
(0.008)
(0.008)
(0.008)
taq_retimb(t −3)
0.000
0.000
−0.000
(0.008)
(0.009)
(0.009)
taq_retimb(t −4)
0.009
0.009
0.009
(0.008)
(0.008)
(0.008)
taq_retimb(t −5)
0.009
0.009
0.008
(0.008)
(0.008)
(0.008)
Observations
3,656,926
3,652,401
3,312,553
3,656,926
3,652,401
3,312,553
3,656,926
3,652,401
3,312,553
R2
0.196
0.199
0.205
0.196
0.199
0.205
0.196
0.198
0.205
Days
550
550
550
550
550
550
550
550
550
Events
4,428
4,412
4,327
2,809
2,800
2,742
1,619
1,612
1,585
Five-day AR (%)
−2.942***
−2.595***
−2.630***
−3.909***
−3.025***
−3.119***
−1.490***
−2.341***
−2.265***
Std. Err.
(0.322)
(0.350)
(0.348)
(0.435)
(0.522)
(0.520)
(0.423)
(0.447)
(0.451)
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 38
51

3178
The Journal of Finance®
D. TAQ Herding Events versus Robinhood Herding Events
In the sections above, we employ a herding measure based on Robinhood
user changes. In this section, we show that this herding measure is closely
related to the herding measure of BLO that is based on TAQ retail trades.
BLO document that stocks in the top quintile of retail SOI and the top quintile
of abnormal retail volume earn dismal returns over the 2010 to 2019 period.
We adapt the BLO measure to create a sample of TAQ herding events for
each day in the Robinhood period that we analyze in this paper. Speciﬁcally, on
each day, we use the standardized order imbalance (SOI) measure of BLO to
construct quintiles and identify the stocks in the top quintile of retail order im-
balance. Among stocks in this top quintile, we rank stocks based on abnormal
retail volume (deﬁned as retail volume on day t divided by mean volume from
t −20 to t −1). We then create an indicator variable (taq_herd) that takes a
value of one for the N stocks with the greatest abnormal retail volume, with
N equal to the number of Robinhood herding events that we identify on day t.
Thus, for each day in our sample period, we have an equal number of Robin-
hood herding events and TAQ herding events.
By construction, we have an equal number of TAQ and Robinhood herding
events. More importantly, 27% (1,317 of 4,884) of the stock-day observations
are identical. Thus, there is substantial overlap in the herding events iden-
tiﬁed using Robinhood user changes and the BLO method using TAQ retail
trades.
To explore the ability of the two measures to predict future returns, we
modify the regression framework of the prior section to include both herd-
ing measures. In Table XI, column (1) replicates the main results using the
rh_herd variable for easy comparison; column (2) uses the identical speciﬁca-
tion but replaces rh_herd with taq_herd. We ﬁnd that taq_herd variable also
predicts negative abnormal returns, but the magnitudes are smaller (−2.208%
vs. −2.942%). (See Tables IAVII, IAVIII, and Figure IA5 for event-time and
calendar-time analyses of TAQ herding events.)
In column (3), we add an indicator variable that takes a value of one if TAQ
retail order imbalance is positive (taqpos) and interact it with rh_herd. This
analysis shows that the Robinhood herding events reliably predict ﬁve-day ab-
normal returns of −1.506% even when there is no net buying by retail investors
in TAQ. The interaction effects (TAQ buying and Robinhood herding) are large
and statistically signiﬁcant, which suggests that the return effects of herding
events are larger when the general population of retail investors is also buy-
ing. In column (4), we ﬁnd similar results when we control for lagged returns
and order imbalance.
In column (5), we include rh_herd, taq_herd, and their interactions. These
results indicate the rh_herd variable generally has larger predictive ability
(−2.397% vs. −1.213% for ﬁve-day abnormal returns). However, the interaction
effects for the two herding variables are not signiﬁcant, which indicates that
the negative returns for events identiﬁed by both measures have particularly
dismal returns. Recall that 27% (or 1,317) of the events are common, so the
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 39
51

Attention-Induced Trading and Returns
3179
Table XI
Regression of Daily Returns on Lagged Robinhood Herding Indicator
and TAQ Retail Herding Indicator
The dependent variable is the daily stock return (%) winsorized at the 0.1% level (ret). rh_herd(t)
is an indicator variable that equals one if the percentage change in users is in the top 0.5% for
stocks with positive user changes on day t and a minimum of 100 users on day t −1. On each
day, we identify the same number of herding events (N) using TAQ retail trade data to construct
taq_herd(t), an indicator variable that equals one if the stock is among the N stocks with the
greatest abnormal retail volume within the top quintile of standardized retail order imbalance
on day t. In columns (3) and (4), we include an indicator variable for whether TAQ retail order
imbalance is positive (taqpos) and its interaction with rh_herd. Control variables include retail
order imbalance from TAQ (taq_retimb), lagged returns (ret), lags of an indicator variable if the
rh_herd measure is missing (and rh_herd is set equal to zero), and day ﬁxed effects. Five-day AR
(%) is the sum of the coefﬁcients on the ﬁve lags of rh_herd (or taq_herd). Robust standard errors
clustered by day are reported in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.
ret(t) (%)
(1)
(2)
(3)
(4)
(5)
(6)
rh_herd(t −1)
−1.336***
−0.549**
−0.514**
−1.001*** −0.838***
(0.160)
(0.234)
(0.241)
(0.160)
(0.170)
rh_herd(t −2)
−0.758***
−0.306
−0.358*
−0.550*** −0.587***
(0.129)
(0.194)
(0.193)
(0.138)
(0.144)
rh_herd(t −3)
−0.298**
−0.099
−0.086
−0.232*
−0.166
(0.122)
(0.191)
(0.192)
(0.126)
(0.128)
rh_herd(t −4)
−0.318***
−0.356**
−0.356**
−0.369*** −0.337***
(0.115)
(0.159)
(0.162)
(0.112)
(0.120)
rh_herd(t −5)
−0.231**
−0.196
−0.204
−0.246**
−0.277**
(0.117)
(0.182)
(0.189)
(0.118)
(0.127)
taq_herd(t −1)
−1.220***
−0.896*** −0.794***
(0.154)
(0.151)
(0.182)
taq_herd(t −2)
−0.772***
−0.485*** −0.536***
(0.120)
(0.120)
(0.137)
taq_herd(t −3)
−0.180
0.001
0.147
(0.113)
(0.107)
(0.125)
taq_herd(t −4)
0.050
0.173
0.252*
(0.109)
(0.106)
(0.129)
taq_herd(t −5)
−0.087
−0.006
−0.022
(0.100)
(0.106)
(0.124)
rh_herd(t −1) × taqpos(t −1)
−1.021*** −0.759**
(0.300)
(0.310)
rh_herd(t −2) × taqpos(t −2)
−0.589**
−0.567**
(0.249)
(0.259)
rh_herd(t −3) × taqpos(t −3)
−0.260
−0.129
(0.239)
(0.263)
rh_herd(t −4) × taqpos(t −4)
0.051
0.121
(0.213)
(0.223)
rh_herd(t −5) × taqpos(t −5)
−0.042
−0.085
(0.212)
(0.218)
rh_herd(t −1) × taq_herd(t −1)
−0.165
−0.049
(0.399)
(0.407)
rh_herd(t −2) × taq_herd(t −2)
−0.252
−0.230
(0.329)
(0.334)
(Continued)
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 40
51

3180
The Journal of Finance®
Table XI—Continued
ret(t) (%)
(1)
(2)
(3)
(4)
(5)
(6)
rh_herd(t −3) × taq_herd(t −3)
−0.262
−0.239
(0.316)
(0.321)
rh_herd(t −4) × taq_herd(t −4)
0.015
0.020
(0.267)
(0.273)
rh_herd(t −5) × taq_herd(t −5)
0.066
0.044
(0.262)
(0.265)
Observations
3,656,926 3,656,926 3,656,926 3,312,553 3,656,926 3,312,553
R2
0.196
0.196
0.196
0.205
0.196
0.205
Lagged control variables:
ret
NO
NO
NO
YES
NO
YES
taq_retimb
NO
NO
NO
YES
NO
YES
taqpos
NO
NO
YES
YES
NO
NO
RH ﬁve-day AR (%)
−2.942***
−1.506*** −1.518*** −2.397*** −2.204***
RH Std. Err.
(0.322)
(0.456)
(0.461)
(0.310)
(0.336)
TAQ ﬁve-day AR (%)
−2.208***
−1.213*** −0.954***
TAQ Std. Err.
(0.347)
(0.321)
(0.326)
stocks with both a Robinhood and TAQ herding event represent a signiﬁcant
part of the overall sample of Robinhood herding events.
In summary, the Robinhood herding measure that we identify is closely re-
lated to the BLO (2021) TAQ herding measure. The Robinhood herding mea-
sure has stronger ability to predict short-term negative returns than the TAQ
measure. Both measures combined provide the strongest signal that returns
will be negative in the coming days.
E. Subsample Analyses
To test the robustness of these patterns, we conduct a battery of robustness
tests in Table XII. Because Robinhood users can trade with limited capital,
low priced stocks are appealing to them, and thus we include such stocks in
our analysis. One may be concerned that our results are driven by a tendency
to observe closing prices at bid prices on the day of the herding event and
thus on average negative returns the next day. The fact that most of the losses
are observed during the next day (rather than overnight) suggests that bid-
ask bounce is not the main driver of the results. To further address this issue,
we reestimate the results using returns based on quote midpoints and ﬁnd
qualitatively similar results (Panel B). We also ﬁnd reliably negative returns
for stocks with prices in excess of $5, but the return magnitudes are smaller
(Panel C).
We anticipate that the negative returns we document will be present in small
cap stocks but muted or nonexistent in large cap stocks, where retail trading
is less likely to inﬂuence pricing. Consistent with this idea, we ﬁnd that stocks
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 41
51

Attention-Induced Trading and Returns
3181
Table XII
Subsample of Postherding Return Patterns
The table presents the ﬁve-day abnormal return (%) from speciﬁcations (1) to (3) of Table X for
various subsamples. Robust standard errors are in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.
(1)
(2)
(3)
Panel A: Full Sample
Five-day AR (%)
−2.942***
−2.595***
−2.630***
Std. Err.
(0.322)
(0.350)
(0.348)
Panel B: Quote Midpoint Returns
Five-day AR
−2.797***
−2.378***
−2.408***
Std. Err.
(0.322)
(0.357)
(0.354)
Panel C: Only Stock with Prices > $5
Five-day AR
−0.851***
−0.756***
−0.749**
Std. Err.
(0.242)
(0.266)
(0.265)
Panel D: Small Cap (< $1 billion in Market Cap)
Five-day AR
−4.250***
−3.766***
−3.835***
Std. Err.
(0.418)
(0.434)
(0.432)
Panel E: Large Cap (> $1 billion in Market Cap)
Five-day AR
0.14
−0.0348
−0.0508
Std. Err.
(0.534)
(0.517)
(0.521)
Panel F: Other than Common Stocks
Five-day AR
−3.174***
−2.895***
−2.940***
Std. Err.
(0.583)
(0.669)
(0.667)
Panel G: Post-Covid (after March 13, 2020)
Five-day AR
−4.581***
−2.876***
−3.073***
Std. Err.
(0.680)
(0.750)
(0.743)
Panel H: Pre-Covid (before March 13, 2020)
Five-day AR
−2.221***
−2.239***
−2.231***
Std. Err.
(0.337)
(0.327)
(0.330)
with a market cap less than $1 billion generate larger abnormal returns (−3.8
to −4.3%), while larger stocks with more than $1 billion in market cap have
no discernable return pattern (Panels D and E). Our main results include both
stocks and other securities (e.g., ADRs and exchange traded funds [ETFs]). We
observe similar return patterns for these other securities only (Panel F).
We note that there was a large increase in both retail trading and Robin-
hood user holdings during the pandemic period (after March 13, 2020). We
thus anticipate that the magnitude of the attention-induced subsequent poor
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 42
51

3182
The Journal of Finance®
performance will increase during this period. This is precisely what we observe
(Panels G and H).
F. Sales Herding
As noted previously, attention asymmetrically affects buying and selling be-
cause investors can buy any stock but tend to sell only that which they own.
Thus, driven by this theory of attention, our primary analysis focuses on the
buying behavior of Robinhood investors. It is natural to wonder whether de-
tectable return effects arise when we analyze sales herding events. To shed
light on this question, we construct a sales herding variable that is analogous
to our purchase herding variable. Speciﬁcally, we identify the securities in the
bottom 0.5% of negative user changes as a percent of prior-day user count to
construct the indicator variable rh_negherd.
It is worth noting that these sales herding events have user changes that are
much less dramatic than the purchase herding events. For the nearly 4,900
sales herding events, the mean userratio is 0.87 and the median is 0.90 (see
Table IAIX). Consistent with the attention model, the sales herding is less
dramatic than the purchase herding. It is also worth noting that the sales
herding events tend to follow purchase herding events and are more clustered
than purchase herding events. For example, in a linear probability estimation
that regresses the sales herding variable (rh_negherd) on lags of itself and lags
of the purchase herding variable (rh_herd), the coefﬁcients on lagged rh_herd
are all signiﬁcant and economically large—at one- and two-day lags, the coef-
ﬁcients are 0.192 and 0.132 (see Table IAX), orders of magnitude larger than
the baseline probability of 0.0012. Roughly half of the sales herding events are
preceded by a purchase herding event in the prior ﬁve days.
We next analyze the returns on the sales herding stocks using the regression
format of Table X. We ﬁnd that the sales herding events have relatively modest
negative abnormal returns of about 55 to 63 bps over the ﬁve days following
the decrease in users (see Table IAXI), and most of the negative abnormal re-
turns follow sales herding events that occur after negative overnight returns
(columns (6) to (9)). The return effects following sales herding events suggest
that sales herding events are not as likely to create price pressure as purchase
herding events, but rather accelerate the poor performance that follows pur-
chase herding events.
G. Aggregate Investor Experience: Average Purchase and Sales Price
While we have found that herding events predict negative returns going for-
ward, it could be the case that the Robinhood community still proﬁts around
these episodes. Enough investors could purchase the stock before the herding
event for those users’ proﬁts to exceed any losses by later purchasers. To ad-
dress this question, we compare the average purchase and sales prices for all
users. Speciﬁcally, we compute the purchase prices of Robinhood users dur-
ing the event period, τ = −10, +20. Deﬁne uiτ −ui,τ−1 = uiτ as the change in
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 43
51

Attention-Induced Trading and Returns
3183
users for stock i on event day τ. For event j, we calculate the average purchase
price for days when we observe user increases (uit > 0) as
PPrcj = ui,−11Pi,−11 + +20
τ=−10 UIiτuiτPiτ
ui,−11 + +20
τ=−10 UIiτuiτ
.
(9)
The indicator UIiτ equals one on days when the change in users is positive,
uiτ > 0. Note that all users owning stock i at the close on day −11, before the
event period begins, are assumed to have purchased at the closing price on day
−11, that is, Pi,−11.
We next compute the average sales price on days when we observe a decrease
in users. For any shares that are not sold during the event period, we assume
that they are sold at the end of the event window:
SPrcj = ui,20Pi,20 + +19
τ=−10 (1 −UIiτ ) uiτPiτ
ui,20 + +19
τ=−10 (1 −UIiτ ) uiτ
.
(10)
The proﬁtability of event j is then calculated as the ratio of the sales price
and the purchase price:
PrcRatioj = SPrcj
PPrcj
−1.
(11)
The variable PrcRatioj represents the returns earned by the Robinhood com-
munity on event j.
In addition to computing raw returns, we compute returns that adjust for
market movements. For each day during the event, we consider a counterfac-
tual in which the Robinhood community buys or sells the equivalent amount
of capital in an S&P 500 ETF, speciﬁcally, SPY. This provides us with a market
price ratio (MktPrcRatioj) for each stock event that we can use to benchmark
the price ratio for the event stock. We report results in Figure 8.
On average, using raw returns, we ﬁnd that the Robinhood community loses
approximately 4.3% during each herding event. After adjusting for the market
return, losses are about 5.5%. Both results are highly signiﬁcant both statisti-
cally and economically. There are slight differences between the pre-COVID-19
and COVID-19 periods, although both suggest similar outcomes. Overall, these
ﬁndings suggest that extreme herding causes negative wealth outcomes for the
overall Robinhood community.35
IV. Short Trading and Herding Events
In the prior section, we ﬁnd that Robinhood users’ herding can lead to price
pressure on the targeted stocks and subsequent poor performance. Our test
35 In Figure IA6, we show that in the average event, about 60% of investors who buy during
these herding episodes buy at a price that is higher than the observed price 20 days after the herd-
ing event. Sellers fare better, but buyers outnumber sellers by approximately 18.7 to 3.5 million.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 44
51

3184
The Journal of Finance®
Figure 8. Aggregate investor experience: Average proﬁtability. The ﬁgure presents the av-
erage proﬁtability of the Robinhood user community’s actual trades across the herding episodes,
their counterfactual trades on S&P 500 ETF during the herding episodes, and the difference be-
tween the two. For each herding episode, we compute the weighted average purchase (sales) price
of Robinhood users during the event period [−10,20]. Proﬁtability is calculated as (average sales
price/average purchase price −1). Positive proﬁtability indicates that the Robinhood community
proﬁted from that herding episode. The counterfactual trades on S&P 500 ETF assume that the
Robinhood community purchases or sells the equivalent amount of capital in an S&P 500 ETF.
The error bar represents the 95% conﬁdence interval for the average. (Color ﬁgure can be viewed
at wileyonlinelibrary.com)
design uses data available before the market closes, and therefore implies that
the negative returns are tradable. In other words, other market participants
could trade against Robinhood users’ order ﬂow in these herding situations in
an attempt to proﬁt from Robinhood users’ correlated trading. In some sense,
the disclosure of the stock holdings by Robinhood is similar to the disclosure of
positions by mutual funds, hedge funds, and others through Form 13F. Hedge
funds have argued that these disclosures enable other market participants to
proﬁt from their private information and to front-run their positions.36 Re-
searchers have documented that these disclosures are used by other market
participants (e.g., Brown and Schwarz (2020)). Thus, it is somewhat surpris-
ing that Robinhood would voluntarily make holdings data available to other
market participants who might proﬁt from the data at the expense of Robin-
hood users.37
In this section, we evaluate whether other market participants do attempt
to proﬁt from the predictably negative returns that follow Robinhood herding
36 It is not clear whether hedge fund 13Fs contain private information. Grifﬁn and Xu (2009)
ﬁnd that 13F disclosures have no alpha as of their effective date. Brown and Schwarz (2020) ﬁnd no
alpha as of the date ﬁled with the Security and Exchange Commission’s EDGAR. Aiken, Clifford,
and Ellis (2013) ﬁnd that, after controlling for biases, hedge fund returns themselves have little
alpha.
37 Robinhood ceased releasing user data in August 2020, stating that the way the data are some-
times reported by third parties “could be misconstrued or misunderstood” and does not represent
the company’s user base (see https://www.foxbusiness.com/markets/robinhood-to-stop-sharing-of-
apps-popular-stocks).
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 45
51

Attention-Induced Trading and Returns
3185
Figure 9. Short interest changes and Robinhood trading. The ﬁgure depicts average abnor-
mal short trading (%). Short trading is measured using data from FINRA, which provides daily
short trade data as noted by Boehmer and Song (2020). We measure abnormal short volume on
day t as the ratio of short trades on day t to the prior 20-day ([t −20,t −1]) average of the num-
ber of short trades. To control for outliers, we winsorize the top 0.1% of observations. The x-axis
is daily return ranking of day t. Robinhood Herd are the top 0.5% of stocks with a positive user
change ratio on day t and a minimum of 100 users on day t −1. (Color ﬁgure can be viewed at
wileyonlinelibrary.com)
events. To do so, we collect data from FINRA, which provides daily off-exchange
short trade data as noted by Boehmer and Song (2020). Reed, Samadi, and
Sokobin (2020) point out that off-exchange short sales are less informative
than exchange short sales, so our use of off-exchange short data is expected to
underestimate the actions of informed trades. We compute the abnormal short
volume on herding day t as the ratio of the number of short trades on day t
divided by the average number of short trades from days t −20 to t −1. We
also winsorize the top 0.1% to control for extreme outliers.
To begin our analysis, we examine the relation between top stock perfor-
mance, Robinhood herding, and short trading. Speciﬁcally, we calculate the
daily return rank for each stock, as we expect abnormal short trading to be cor-
related with high returns. We then ﬂag whether the stock was also herded into
by Robinhood users any day during the period using our 0.5% cutoff (rh_herd).
We calculate the average change in abnormal short volume for each return
rank for the non-Robinhood herding and Robinhood herding groups separately.
In Figure 9, we plot the average values for ranks 1–5, 6–10, 11–15, 16–20, 21–
25, 26–50, and 51–100.
We ﬁnd two clear patterns. First, abnormal short volume is indeed higher
for stocks that had a high daily return rank. Second, in all cases abnormal
short volume for stocks herded into by Robinhood investors is far higher than
that for stocks not herded into Robinhood investors. This is consistent with
market participants knowing the pattern of Robinhood return reverses and,
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 46
51

3186
The Journal of Finance®
Table XIII
Regression of Abnormal Short Trading on Returns and Top
Robinhood Changes
The dependent variable is abnormal short volume. Short volume is measured using data from
FINRA, which provides daily short trade data as noted by Boehmer and Song (2020). We compute
abnormal short volume on day t as the ratio of short trades on day t to the prior 20-day ([t −20,t −
1]) average of the number of short trades. The top 0.1% of observations are winsorized. The key
independent variables are rh_chgratio, which is the percentage change in users from day t −1 to
t, and rh_herd, which is an indicator variable that equals one if the percentage change in users
is in the top 0.5% for stocks with positive user changes on day t and a minimum of 100 users on
day t −1. Control variables include the excess return of day t, dummy variables representing the
return ranks (ret_rank) of day t, and abnormal news coverage, which is the logarithm of the ratio of
news article count on day t to the average news article count from day t −20 to t −1. Coefﬁcients
are in percent. Standard errors are computed using Fama-Macbeth (1973). *** p < 0.01, ** p <
0.05, * p < 0.1.
Abnormal Short Volume (%)
(1)
(2)
(3)
(4)
rh_chgratio
637.14***
551.10***
(12.75)
(11.10)
rh_herd
854.65***
686.57***
(14.5)
(17.25)
Excess return
499.54***
498.83***
(16.32)
(15.95)
Abnormal News
6.64***
7.17***
6.37***
6.85***
(0.13)
(0.14)
(0.12)
(0.13)
Ret_rank_1
858.14***
1049.44***
(47.95)
(36.36)
Ret_rank_2
704.80***
818.70***
(44.31)
(34.22)
Ret_rank_3
568.94***
670.82***
(68.35)
(32.64)
Ret_rank_4
583.67***
635.83***
(35.00)
(31.91)
Ret_rank_5
534.46***
554.07***
(30.95)
(30.50)
Ret_rank_6
465.52***
480.40***
(26.90)
(26.95)
Ret_rank_7
410.78***
482.44***
(39.91)
(28.08)
Ret_rank_8
371.97***
398.30***
(27.60)
(25.01)
Ret_rank_9
381.49***
395.11***
(26.37)
(26.27)
Ret_rank_10
361.47***
385.87***
(24.36)
(24.53)
Ret_rank_11 −25
269.46***
279.53***
(8.20)
(8.55)
Ret_rank_26 −50
145.48***
152.75***
(4.30)
(4.56)
Ret_rank_51 −100
83.45***
86.87***
(2.31)
(2.45)
Observations
2,681,173
2,681,173
2,681,173
2,681,173
Avg. R2
0.162
0.146
0.271
0.257
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 47
51

Attention-Induced Trading and Returns
3187
more broadly, with market participants using the disclosed Robinhood position
information.
We next examine short-interest changes in a multivariate setting. Each pe-
riod, we regress abnormal short trading on either the daily excess return or
the return ranking (ret_rank). We also include abnormal news coverage to
control for attention that news may have brought and caused. Our primary
variables of interest are either the average user change ratio during the period
(rh_chgratio) or herding of the stock by Robinhood users (rh_herd). We average
across periods and obtain standard errors and t-values via Fama and Macbeth
(1973).
In Table XIII, we ﬁnd results consistent with those of Figure 9. A 100% in-
crease (doubling) in the average change ratio leads to at least a 637% increase
in short trading (column (1)), which is both statistically and economically sig-
niﬁcant. If a stock is herded by Robinhood users (columns (2)), abnormal short
volume is approximately 855% higher than if the stock is not herded. Overall,
these results again suggest strongly that market participants examined Robin-
hood ownership data, knew about the subsequent poor performance caused by
Robinhood herding, and traded against Robinhood order ﬂow.
V. Conclusion
The stated mission of the Robinhood brokerage is to “democratize ﬁnance for
all…[and] make investing friendly, approachable, and understandable.” Robin-
hood facilitates “friendly, approachable, and understandable” investing by of-
fering a simple downloadable app that makes trading incredibly easy. The app
displays only a small fraction of the stock-level indicators that other brokerage
platforms provide. Instead, the app highlights easily understood lists of stocks
such as a Top Movers list of stocks with the largest price moves on the current
day.
We argue that the combination of simpliﬁed information display and in-
experience exacerbates attention-driven buying by Robinhood users. Height-
ened attention-driven buying leads to more concentrated trading by Robinhood
users than other retail investors and contributes to buy-side herding events
that are usually followed by negative returns. For example, the top 0.5% of
stocks bought by Robinhood users each day experiences negative average re-
turns of approximately 5% over the next month. More extreme herding events
are followed by negative average returns of almost 20%.
Robinhood has been successful in its stated mission in as much as it has at-
tracted 13 million users (as of 2020). Half of these are ﬁrst-time investors, who
are likely in the long run to beneﬁt from participating in markets. Robinhood
attracts investors by reducing frictions and promoting simplicity. While a lack
of frictions encourages market participation, it also makes speculative trading
easy, which can lead to lower investment returns (Odean (1999), Barber and
Odean (2000), Barber et al. (2009)). However, even in an industry that uses
complexity to obscure risks and costs (Carlin (2009), Henderson and Pearson
(2011); Célérier and Vallée (2017), Gao et al. (2021)), simplicity is not problem
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 48
51

3188
The Journal of Finance®
free. As we show, simply focusing the attention of many investors on a small
number of stocks can promote herding behavior that affects market returns
and redounds to the investors’ detriment. Thus, while it is important that in-
vestors have access to transparent, pertinent information, disclosure alone is
not sufﬁcient to assure good investor outcomes—how information is displayed
inﬂuences decisions in ways that can both help and hurt investors.
Initial submission: February 2, 2021; Accepted: September 12, 2021
Editors: Stefan Nagel, Philip Bond, Amit Seru, and Wei Xiong
REFERENCES
Aiken, Adam L., Christopher P. Clifford, and Jesse Ellis. 2013, Out of the dark: Hedge fund report-
ing biases and commercial databases. Review of Financial Studies 26, 208–243.
Barber, Brad M., Yi-Tsung Lee, Yu-Jane Liu, and Terrance Odean. 2009, Just how much do indi-
vidual investors lose by trading? Review of Financial Studies 22, 609–632.
Barber, Brad M., Shengle Lin, and Terrance Odean. 2021, Resolving a paradox: Retail trades
positively predict returns but are not proﬁtable, Working paper.
Barber, Brad M., and Douglas Loefﬂer. 1993, The “Dartboard” column: Second-hand information
and price pressure. Journal of Financial and Quantitative Analysis 28, 273–284.
Barber, Brad M., and Terrance Odean. 2000, Trading is hazardous to your wealth: The common
stock investment performance of individual investors, Journal of Finance 55, 773–806.
Barber, Brad M., and Terrance Odean. 2002, Online investors: Do the slow die ﬁrst? Review of
Financial Studies 15, 455–488.
Barber, Brad M., and Terrance Odean. 2008, 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.
Barber, Brad M., Terrance Odean, and Ning Zhu. 2008, Do retail trades move markets? Review of
Financial Studies 22, 151–186.
Boehmer, Ekkehart, Charles M. Jones, Xiaoyan Zhang, and Xinran Zhang. 2021, Tracking retail
investor activity. Journal of Finance 76, 2249–2305.
Boehmer, Ekkehart, and Wanshan Song. 2020, Smart retail traders, short sellers, and stock re-
turns, Working paper.
Bolster, Paul, Emery Trahan, and Anand Venkateswaran. 2012, How mad is mad money? Jim
Cramer as a stock picker and portfolio manager. Journal of Investing 21, 27–39.
Brown, Stephen J., and Christopher Schwarz. 2020, Do market participants care about portfolio
disclosure? Evidence from hedge funds’ 13F ﬁlings, Working paper.
Campbell, John Y., Sanford J. Grossman, and Jiang Wang. 1993, Trading volume and serial corre-
lation in stock returns. Quarterly Journal of Economics 108, 905–939.
Carlin, Bruce I. 2009, Strategic price complexity in retail ﬁnancial markets. Journal of Financial
Economics 91, 278–287.
Célérier, Claire, and Boris Vallée. 2017, Catering to investors through security design: Headline
rate and complexity. Quarterly Journal of Economics 132, 1469–1508.
Cheng, Peter, Thomas J. Murphy, and Marko Kolanovic. 2020, Following the robinhood money. JP
Morgan report.
Choi, James J., David Laibson, and Brigitte C. Madrian. 2010, Why does the law of one price fail?
An experiment on index mutual funds. Review of Financial Studies 23, 1405–1432.
Cookson, J. Anthony, Joseph Engelberg, and William Mullins, Forthcoming, Echo chambers. Re-
view of Financial Studies.
Coval, Joshua, and Erik Stafford. 2007, Asset ﬁre sales (and purchases) in equity markets. Journal
of Financial Economics 86, 479–512.
Cronqvist, Henrik, and Richard H. Thaler. 2004, Design choices in privatized social-security sys-
tems: Learning from the Swedish experience. American Economic Review 94, 424–428.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 49
51

Attention-Induced Trading and Returns
3189
Da, Zhi, Joseph Engelberg, and Pengjie Gao. 2011, In search of attention. Journal of Finance 66,
1461–1499.
Da, Zhi, Jian Hua, Chih-Ching Hung, and Lin Peng, 2020, Market returns and a tale of two types
of attention, Working paper.
Da, Zhi, Borja Larrain, Clemens Sialm, and Jose Tessada. 2018, Destabilizing ﬁnancial advice:
Evidence from pension fund reallocations. Review of Financial Studies 31, 3720–3755.
Drechsler, Itamar, Alan Moreira, and Alexi Savov. 2020, Liquidity and volatility, Working paper.
Eaton, Gregory W., Clifton Green, Brian Roseman, and Yanbin Wu, Forthcoming, Retail trader so-
phistication and stock market quality: evidence from brokerage outages. Journal of Financial
Economics.
Engelberg, Joseph, Caroline Sasseville, and Jared Williams. 2012, Market madness? The case of
mad money. Management Science 58, 351–364.
Fama, Eugene F., and James D. MacBeth. 1973, Risk, return, and equilibrium: Empirical tests.
Journal of Political Economy 81, 607–636.
Fedyk, Anastassia. 2019, Front page news: The effect of news positioning on ﬁnancial markets,
Working paper.
French, Kenneth R., and Richard Roll. 1986, Stock return variances: The arrival of information
and the reaction of traders. Journal of Financial Economics 17, 5–26.
Frydman, Cary, and Baolian Wang. 2020, The impact of salience on investor behavior: Evidence
from a natural experiment. Journal of Finance 75, 229–276.
Gao, Pengjie, Allen Hu, Peter Kelly, and Cameron Peng. 2021, Exploited by complexity, Working
paper.
Greenwood, Robin, and Stefan Nagel. 2009, Inexperienced investors and bubbles. Journal of Fi-
nancial Economics 93, 239–258.
Grifﬁn, John M., and Jin Xu. 2009, How smart are the smart guys? A unique view from hedge fund
stock holdings. Review of Financial Studies 22, 2331–2370.
Henderson, Brian J., and Neil D. Pearson. 2011, The dark side of ﬁnancial innovation: A case study
of the pricing of a retail ﬁnancial product. Journal of Financial Economics 100, 227–247.
Hong, Claire Yurong, Xiaomeng Lu, and Jun Pan. 2019, FinTech platforms and mutual fund dis-
tribution, Working paper.
Jegadeesh, Narasimhan. 1990, Evidence of predictable behavior of security returns. Journal of
Finance 45, 881–898.
Jegadeesh, Narasimhan, and Sheridan Titman. 1995, Overreaction, delayed reaction, and contrar-
ian proﬁts. Review of Financial Studies 8, 973–993.
Kaniel, Ron, and Robert Parham. 2017, WSJ category kings–the impact of media attention on
consumer and mutual fund investment decisions. Journal of Financial Economics 123, 337–
356.
Kaniel, Ron, Gideon Saar, and Sheridan Titman. 2008, Individual investor trading and stock re-
turns. Journal of Finance 63, 273–310.
Kahneman, Daniel. 1973, Attention and Effort (Prentice-Hall, Englewood Cliffs, NJ).
Kahneman, Daniel. 2011, Thinking, Fast and Slow (Farrar, Straus and Giroux, NY).
Keasler, Terrill R., and Chris R. McNeil. 2010, Mad money stock recommendations: Market reac-
tion and performance. Journal of Economics and Finance 34, 1–22.
Kelley, Eric K., and Paul C. Tetlock. 2013, How wise are crowds? Insights from retail orders and
stock returns. Journal of Finance 68, 1229–1265.
Kronlund, Mathias, Veronika K. Pool, Clemens Sialm, and Irina Stefanescu. 2021, Out of sight
no more? The effect of fee disclosures on 401(k) investment allocations. Journal of Financial
Economics 141, 644–668.
Lakonishok, Josef, Andrei Shleifer, and Robert W. Vishny. 1992, The impact of institutional trad-
ing on stock prices. Journal of Financial Economics 32, 23–43.
Liang, Bing. 1999, Price pressure: Evidence from the “Dartboard” column. Journal of Business 72,
119–134.
Lehmann, Bruce N. 1990, Fads, martingales, and market efﬁciency. Quarterly Journal of Eco-
nomics 105, 1–28.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 50
51

3190
The Journal of Finance®
Lo, Andrew W., and A. Craig MacKinlay. 1988, Stock market prices do not follow random walks:
Evidence from a simple speciﬁcation test. Review of Financial Studies 1, 41–66.
Lo, Andrew. W., and A. Craig MacKinlay. 1990, When are contrarian proﬁts due to stock market
overreaction? Review of Financial Studies 3(2), 175–205.
Loos, Benjamin, Steffen Meyer, and Michaela Pagel. 2020, The consumption effects of the disposi-
tion to sell winners and hold losers, Working paper.
Nagel, Stefan. 2012, Evaporating liquidity. Review of Financial Studies 25, 2005–2039.
Niessner, Marina. 2015, Strategic disclosure timing and insider trading, Working paper.
Odean, Terrance. 1999, Do investors trade too much? American Economic Review 89, 1279–1298.
Oprea, Ryan. 2020, What makes a rule complex? American Economic Review 110, 3913–3951.
Ozik, Gideon, Ronnie Sadka, and Siyi Shen. 2021, Flattening the illiquidity curve: Retail trading
during the COVID-19 lockdown. Journal of Financial and Quantitative Analysis 56, 2356–
2388.
Pedersen, Lasse H., Forthcoming, Game on: Social networks and markets. Journal of Financial
Economics.
Reed, Adam V., Mehrdad Samadi, and Jonathan S. Sokobin. 2020, Shorting in broad daylight:
Short sales and venue choice. Journal of Financial and Quantitative Analysis 55, 2246–2269.
Seasholes, Mark S., and Guojun Wu. 2007, Predictable behavior, proﬁts, and attention. Journal of
Empirical Finance 14, 590–610.
So, Eric C., and Sean Wang. 2014, News-driven return reversals: Liquidity provision ahead of
earnings announcements. Journal of Financial Economics 114, 20–35.
Tetlock, Paul C. 2011, All the news that’s ﬁt to reprint: Do investors react to stale information?
Review of Financial Studies 24, 1481–1512.
Umar, Tarik. 2020, Complexity aversion when seeking alpha, Working paper.
Vissing-Jørgensen, Annette. 2002, Limited asset market participation and the elasticity of in-
tertemporal substitution. Journal of Political Economy 110, 825–853.
Welch, Ivo. 2022, Retail raw: Wisdom of the Robinhood crowd and the covid crisis. Journal of
Finance 77, 1489–1527.
Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s website:
Appendix S1: Internet Appendix.
Replication Code.
 15406261, 2022, 6, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jofi.13183 by Laurence Rosen - Test , Wiley Online Library on [14/06/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ase 1:21-md-02989-CMA   Document 585-9   Entered on FLSD Docket 06/30/2023   Page 51
51
File and source

File: gov.uscourts.flsd.590042.585.9.pdf
Size: 2,006,937 bytes
SHA-256: ed48d5a3f6a621e4beafa3f0810bc00d662fa33b6d7cd5777c40eb1fb88ce792
Our copy: gov.uscourts.flsd.590042.585.9.pdf
Original: archive.org
Attention‐Induced Trading and Returns: Evidence from Robinhood Users

Full text

File and source

Related records
