Understanding Retail Investors: Evidence from China

Understanding Retail Investors: Evidence from China

Charles M. Jones, Donghui Shi , Xiaoyan Zhang and Xinran Zhang*

First v ersion: September 2019
This version: May 2022

Abstract

Using comprehensive account -level data from 2016 to 201 9, we examine retail investor trading
behavior in the Chinese stock market . We separate millions of retail investors into fi ve groups by
their account sizes and document strong heterogeneity in their trading dynamics and performance.
Retail investor s with smaller account sizes cannot predict future price movements correctly, in the
sense that they buy future losers and sell future winners . Their trading patterns are momentum ove r
daily horizons, but become c ontrarian over weekly horizons . These investors fail to process public
news and display behavioral properties such as overconfidence and gambling preferences. In sharp
contrast, retail investors with larger account balances predict future returns correctl y, display
contrarian trading patterns, incorporate public news in their trading, and their return predictive
power are stronger in stocks which are more attractive to investors with behavioral biases .

Keywords: Retail i nvestors , Chinese stock market, Return predictabilit y, Information content.
JEL code: G12, G14, G15

• Charles M. Jones is with Columbia Business School , Donghui Shi is with Fanhai International
  School of Finance, Fudan University , Xiaoyan Zhang is with PBC School of Finance , Tsinghua
  University, and Xinran Zhang is with School of Finance, Central University of Finance and
  Economics . Xiaoyan Zhang acknowledges financial support from the National Natural Science
  Foundation of China [Grant 71790605]. We thank Laruen Cohen , Ron Kaniel , Zhiguo H e, Hao
  Zhou, Utpal Bhattacharya , and seminar participants at Tsinghua PBC School of Finance , Renmin
  University, Shanghai Jiaotong University , Fudan University , Shanghai University of Finance and
  Economics, and conference au diences at the 2019 CIFFP , 2021 CFRC , 2021 CICF for helpful
  comments and suggestions . All remaining errors are our own. Corresponding author: Xiaoyan
  Zhang , PBC School of Finance , 43 Chengfu Road , Beijing , China, 100083 ,
  zhangxiaoyan@pbcsf.tsinghua.edu.cn .

Understanding Retail Investors: Evidence from China

First v ersion: September 2019
This version: May 2022

Abstract

Using comprehensive account -level data from 2016 to 2019, we examine retail investor trading
behavior in the Chinese stock market. We separate millions of retail investors into five groups by
their account sizes and document strong heterogeneity in their t rading dynamics and performance.
Retail investors with smaller account sizes cannot predict future price movements correctly, in the
sense that they buy future losers and sell future winners. Their trading patterns are momentum over
daily horizons, but become c ontrarian over weekly horizons . These investors fail to process public
news and display behavioral properties such as overconfidence and gambling preferences. In sharp
contrast, retail investors with larger account balances predict future returns c orrectly, display
contrarian trading patterns, incorporate public news in their trading, and their return predictive
power are stronger in stocks which are more attractive to investors with behavioral biases.

Keywords: Retail investors, Chinese stock mar ket, Return predictability, Information content.
JEL code: G12, G14, G15
1
Retail investors are important participants in financial market s, and many studies are devoted
to understand ing their trading motives, their performance s, and their roles in information
transmission and price discovery . However, these studies provide seemly conflicting results. For
instance, Barber and Odean (2000 , 2001 , 2008 ) document behavioral bias es exhibited b y retail
investors, such as over -confidence and overtrading, and as a result, retail investors make sub -
optimal investment choices. Later studies, such as Barber , Odean, and Zhu (2008), Kaniel et al.
(2008), Kelley and Tetlock (2013), and Boehmer et al. (2021), suggest that retail investors correctly
predict future stock returns and trade accordingly , which indicates that retail investors might know
something about future stock price movements. Most recently, interest s have shifted to a new
generation of retail investors, who trade at zero -commission trading platforms such as Robinhood .
Barber et al. (2021), Eaton et al. (2021) and Welch (2021) find that Robinhood investors perform
well, demand liquidity and engage more in attention -induced trading. How can we reconcile the
conflicting results from previous studies? One possibility is that retail investors are not born equal,
so the above -mentioned empirical results could be dominated by subgroups of retail investors .
However, d ue to data limitation s, few previous studies directly examine the heterogeneity of retail
investors.
China ’s equity market, the second largest in the world, provides an ideal setting for studying
retail investors and their heterogeneity . According to the annual report of the Shanghai Stock
Exchange, retail investors contribute 85% of daily trading volume s on the exchange , while 2
institutional investors only contribute 15%. The dominance of retail trading in this market clearly
bring s retail investors to the center stage. Behind the high trading volume s are tens of millions of
retail investors in China, account ing for the largest population of retail investors in the global
capital market. Given the dominant role and the large population of Ch inese retail investors, it is
crucial for researchers, regulators and practitioners to understand Chinese retail investors ’
investment choices and how these choices affect information transmission and price discovery .
With account -level data from one main stock exchange, we examin e the rich cross -section of
retail investors in China , which greatly helps us to investigate the heterogeneity of retail investors
and how the ir trading interacts with stock returns and information flows. We obtain account -level
trading and holding s data from 2016 to 2019 for over 53 million retail accounts. To comply with
regulatory re quirements, all Chinese retail accounts are categorized into five groups by account
balances: less than 100,000 CNY (RT1), between 100,000 and 500 ,000 CNY (RT2), between
500,000 and 3,000,000 CNY (RT3), between 3,000,000 and 10,000,000 CNY (RT4), and greater
than 10,000,000 CNY (RT5). The five group s account for 58.7%, 28.6%, 10.9%, 1.4% and 0.4 %
of total number of accounts , respectively. With addit ional gender and age information , we find the
majority of Chinese retail investo rs are young or middle -aged male s with account size s below 500k
CNY .
With this rich cross section of ret ail investor data , we first examine whether (some) retail
investors are informed about future price movement s, in the sense that whether their buy and sell 3
activit ies can predict future stock returns . If the market is perfectly efficient, stock price s would
follow random walks, and trading would not predict future returns. If the marke t is not perfectly
efficient, and if some investors have value -relevant information for future stock prices, their order
flows would positive ly predict future returns . On the other han d, if some investors have
information disadvantage or fail to incorporate timely information into their trades, their order
flows might negatively predict future returns. Using daily retail order imbalance s from each retail
group , we predict fut ure stock r eturns at horizons ranging from o ne to 60 days. The smaller retail
investors, RT1 -RT4, predict next -day returns with negative coefficients. That is, t he prices of
stocks they buy experience negative returns the next day, while the ones they sell experience
positive returns. In contrast, the largest r etail investors , RT5, positively predict next -day returns ,
indicating that they buy and sell stocks in directions consistent with future price movements. When
we look at longer horizons, the above -mentioned pred ictive patterns persist for at least 12 weeks .
These patterns are also quite robust when we form long -short strategies on order flow information,
and for subsets of stocks with differences in size, value, liquidity and share price level s.
Previous literatu re provides multiple explanations for the trading motives for retail investors,
such as order flow persistence, liquidity provision, behavioral biases and information
(dis)advantages. These explanations also naturally connect with the predictive pattern of retail
order flows for future returns. We adopt the two -stage decomposition procedure in Boehmer et al .
(2021) to examine whether these hypotheses can explain the trading activities of different retail 4
investor groups, and how these trading activitie s contribute to the predictive patterns for future
returns.
Our results show that order flows from all retail investors display persistence. Order flows
from smaller retail investors show momentum pattern s at a daily horizon and demand immediate
liquidity. The smaller retail investors also display significant behavioral biases, such as over -
confidence and gambling preference s, and they fail to predict and process earnings news. On the
contrary, the largest retail investors disp lay contrarian tradi ng patterns; they trade against the
behavioral biases of the other retail groups and are capable of predicting and processing earnings
news. In explaining order flow ’s predictive power for future returns, order persistence, daily
momentum trading, behavior al biases and information disadvantage s all contribute to the negative
predictive power of smaller retail investors, while contrarian trading, trading against behavioral
biases and information skills contribute to the positive predictive power of the large st retail
investors.
We also investigate other dimensions of the data and conduct several robustness checks. We
find that male investors across all ages negatively predict returns , especially t he older ones. These
findings are generally in line with Barbe r and Odean (2001). Results from all other robustness
checks are consistent with the main results.
Our study is closely related to the retail investor literature. Previous studies on retail investors
mostly use data from the U.S. and other markets , and the y mostly treat retail investors as one group . 5
For instance , using data from a discount broker in the U.S., Barber and Odean (2000, 2001, 2008 )
document many behavior al biases; Kaniel, Saar and Titman (2008), Barber, Odean, Zhu (2009),
Kelley and Tetlock (2013), and Boehmer et al. (2021) use different datasets from the U.S. and find
that retail trading can positively predict the cross -section of future returns; and Barber, Lin, and
Odean (2021) explain why U.S. retail investors lose money despite their pr edictive power for stock
returns. Recently, Barber et al. (2021), Eaton et al. (2021) and Welch (2021) study the trading
behavior of Robinhood retail investors in the U.S. Outside of the U.S., Grinblatt and Keloharju
(2000) , Linnainmaa (2010), Grinblatt , Keloharju , and Linnainma a (2012) focus on the Finland
data; Bach, Calvet, and Sodinish (2020) focus on the Sweden data; Dorn, Huberman, Sengmueller
(2008) study data from Germany; Barrot, Kaniel and Sraer (2016) study data from France; Fong,
Gallagher, and Lee (2014) study data from Australia; Barber, Lee, Liu and Odean (2009) examine
Taiwan data; and Balasubramaniam, Campbell, Ramadorai and Ranish (2021) make use of Indian
data. All these papers provide important results regarding retai l investors ’ trading activities .1
Our study is also related to studies on the rapidly growing Chinese stock market. Liu,
Stambaugh and Yuan (2019), and Liu, Zh ou, and Zhu (2021) establish asset pricing factors for
stock returns . For Chinese retail investo rs, An, Lou and Shi (202 2) study the wealth redistribution
role of financial bubbles and crashes over July 2014 and December 2015 , and they document a net
transfer of 2 50 billion CNY from the poor to ultra -wealthy retail investors over this period. Liu,

1 Please see the Appendix for the literature review of the studies on retail investors in different markets. 6
Peng, Xiong and Xiong (2021) and Liao, Peng and Zhu (2021) both focus on behavioral properties
of Chinese retail investors and document overconfidence, gambling preferences and extrapolative
expectations in these investors. Li et al. (2017), Titman et al. (2022), Hu et al. (2021) and Jiang et
al. (2020) and other papers2 focus on an earlier Chinese sample period and examine behavioral
biases and reactions to corporate events.
The above studies mostly rely on low frequency data, or data from brokerage which covers a
small part of the market, or investigate issues other than return predictability. As a result, there is
still no direct study on the heterogeneity of retail investors ’ trading behavior, their return predictive
power, and how they process informat ion, using high frequency trading data in one major stock
market where retail investors dominate . Therefore , our study makes two important contributions
to the literature. First, we separate retail investors into groups based on account size s and provide
unique, and direct evidence on investor heterogeneit y in terms of return predictability . Second, we
examine different hypothes es for the return predicti on patterns for different retail investor groups,
and we provide clear evidence on the sources of the negative or positive predictive power of
different retail investors . Our study, with its large coverage of the market for a recent sample
period , is one of the most thorough and comprehensive studies for of Chinese retail investors , and
it provides many important implications for regulators , practitioners, and academic researchers.
I. Data

2 These papers include Li, Geng, Subrahmanyam and Yu (2017), Chen, Gao, He, Jiang and Xiong (2019), Jiang, Liu,
Peng, and Wang (2020), Titman, Wei, and Zhao (2022), and Hu, Liu and Xu (2021). 7
A. Data on Stock Returns and Firm Characteristics
We obtain data on stock returns, volumes, and accounting information from Wind Information
Inc. (WIND), the largest financial data provider in China. To be consistent with our retail data, our
sample period runs from January 2016 to June 2019. We adopt the filters in Liu, Stambaugh and
Yuan (2019) and exclude stocks with less than 15 days of trading records during the most recent
month. Liu, Stambaugh and Yuan (2019) also eliminate stocks that have become public within the
past six months, stocks with fewer than 120 days of trading records during the past 12 months, and
the smallest 30% of total firms listed in the Chinese A -share market . We do not exclude these
stocks for the main results, because retail investors trade actively in small stocks and during the
IPO period. We present the results with all filters from Liu et al . (2019) in our robustness checks,
and our findings are almost the same with the additional filters. Starting from March 31, 2010,
margin buying and short selling are allowed on Chinese sto ck exchanges for subsets of stocks. We
include these leveraged trades in our main results, and provide additional analysis excluding
leveraged trading in our robustness checks. Our sample covers over 1.1 million stock -day
observations , and on each day, we have an average of around 1,200 firms.
We present summary statistics on our sample firms in Panel A of Table 1. Daily stock returns
are calculated using clos ing prices, which are dividend and split adjusted.3 The average daily stock

3 Previous literature using the U.S. data shows that microstructur e frictions can generate noise in daily return measures.
For instance, Blume and Stambaugh (1986) show that daily returns computed from the end -of-day closing prices can
have an upward bias due to bid -ask bounce. To assess the potential magnitude of the bi as, they measure the bias as
(𝑃𝐴−𝑃𝐵
𝑃𝐴+𝑃𝐵)2
, where PA and PB are closing ask and bid prices. Blume and Stambaugh (1986) find that the average bias 8
return , Ret, is -0.01% for Chinese stocks, while the average daily stock return is 0.04% in the U.S
stock market over the same sample period. Market capitalization, Size, is computed as the product
of the previous month ’s clos ing price and total A shares outstanding. The average Chinese firm
capitalization is 20.1 billion CNY or 3 billion USD, about half of the cross -sectional average in
the U.S stock market during the same period, which is 6.9 billion USD. The earnings to price ratio,
EP, is compu ted as the ratio of the most recently reported quarterly net profit excluding non -
recurrent gains/losses over last month -end’s market capitalization . According to Liu, Stambaugh
and Yuan (2019), the EP ratio captures the value effect. The average EP ratio is 0.0075 in China,
while the average EP ratio is 0.0272 in U.S stock market. This difference may be driven by high
valuation s in China. Finally, monthly turnover is calculated as monthly share trading volume
divided by tradable shares outstanding at the e nd of the previous month. The average monthly
turnover in Chin a is 48.32%, which is much large r than the monthly turnover of 22% in the U.S.
during the same period.
B. Data on Retail Investors
We obtain investors ’ daily trading and holding data of all A -share stocks listed on the Shanghai
Stock Exchange between January 2016 and June 2019. Out of the two stock exchanges in Ch ina,

for small stocks is 0.051%, and for large stocks, the bias is 0.001%, which are sizable magnitudes for d aily returns
averaging at less than 1%. Therefore, they recommend using closing bid -ask average prices to compute daily returns.
We compute this bias measure using the closing bid and ask prices for all A -share stocks listed on the SHSE. The
average bias m easure is generally below 0.0002% across all stocks, which is negligible compared to the b ias computed
in Blume and Stambaugh (1983). Therefore, we compute daily returns using daily close prices without the Blume and
Stambaugh (1983) adjustments. 9
the Shanghai Stock Exchange (SHSE) and the Shenzhen Stock Exchange (SZSE), the former
accounts for 60% of the total marke t capitaliz ation in China and thus is a reasonable representation
of the overall Chinese stock market.4 Our data contains roughly 53 million accounts , and based
on investor identities, they are first group ed into three major categories : retail ( RT), institutional
(INST ), and corporations ( CORP ). Retail investors are further stratified into five groups based on
their account sizes , which is the average portfolio value (including equity holdings in both SHSE
and SZSE -listed firms , plus cash) over the previous twelve months . As mentioned in the
introduction, t here are five subgroups: below 100,000 CNY ( RT1), 100,000 -500,000 CNY (RT2),
500,000 - 3 million CNY ( RT3), 3 million - 10 million CNY ( RT4), and above 10 million CNY
(RT5). Since our focus in this study is how retail trades are related to stock prices in the cross
section, we sum up individual investors ’ trading information at seven investor group level (RT1 -
RT5, INST and CORP) for each stock each day.
We merge the exchange data and WIND data by stock ticker and present account summary
statistics in Panel B of Table 1. During our sample period , the total number of active account s for
retail investor s, institutions and corporation s are 53.4 million , 40,000 and 47,000, respectively .
Within the retail investor category , there are 31.4 million, 15.3 million, 5.8 million, 0.7 million and

4 In June 2019, there are 1,471 A -share stocks listed on the SHSE, with a total market capitalization of $ 4.6 trillion.
In comparison, 2,157 A -share stocks are listed on the SZSE, with a total market capitalization of $ 3 trillion. The
Science and Technology Innov ation Board (or STAR Market) was launched on the SHSE on July 22, 2019, and thus
is not included in our study. Given the data accessibility, our results only cover Shanghai stock exchange , but not
Shenzhen stock exchange (SZSE). Chen et al. (2019) use Shen zhen stock exchange data to examine retail investors
trading around price limits events , and find consistent patterns to those in our study, which suggest that the retail
investors at SZSE likely behave similar ly to those at SHSE. 10
0.2 million accounts for RT1 to RT5. Clearly, most of the retail investors have accounts less than
500,000 CNY . The overall tra ding volume on the SHSE averages 201 b illion CNY per day, and
retail investor s, institutions and corporation s account for 81%, 17% and 2% of the total trading
volume , respectively . Within the retail investor sector, trading volume s for RT1 to RT5 are 5%,
17%, 27%, 13% and 19% of the total trading volume , which is more evenly distributed than the
numbers of account s. For stock holding s, retail investors ’ holdings account for 22%, institutions
17% and corporation s 62% . Within the retail investor sector, the account value s for RT1 to RT5
are 1%, 4%, 6% , 3% and 7% of the total tradable market cap .
To understand the relative importance of different investment groups ’ trading over time, we
plot the time series of cross sectional means of various investors ’ trading a ctivity in Figure I. Panel
A presents each group ’s trading volume as a percentage of total trading volume. The RT3 group
has the highest trading volume, accounting for about 30% of total trading. Interestingly,
institutional trading gradually incre ases over time , from 10% in 2016 to over 20% in 2019. The
corporations barely trade and account for a negligible amount of trading volume. Panel B displays
the shares held percentage by each group, and the time -series patterns in holdings are quite stable.
Overall , in the Chinese stock market retail investors dominate in terms of trading, while
corporations dominate in holdings . The retail investors ’ dominance in trading of Chinese stock
market is likely the joint result of market development, regulations a nd investor preference. This
pattern is not particularly rare for emerging markets, but it is quite different from that of developed 11
markets. This particular dominance in trading by retail investors also renders more relevance and
significance of our study .
Finally, t o have a rough idea about holding horizons , we make a simpl e assumption that
shareholders within same investor gro up have identical holding horizon s. Then we compute the
holding period for stock i, type G investors as 1/𝑇𝑂 𝑖,𝐺 , where 𝑇𝑂 𝑖,𝐺 is the turnover (shares
traded /shares held by this type of investor) of stock i for type G investors . For example , if 1% of
the shares trade each day, then it takes 100 days for the entire stock of tradable shares held by this
group to turn over, and the average holding period would be 100 days. In the last row of Panel B,
the average holding period for the five groups of retail investors range s from 35 days to 50 days,
reflecting their active trading and short holding horizons. Institutional holdi ng periods in our
sample are much longer at 109 trading days. Corporations barely trade in our sample period, and
their estimated holding period is 6,319 trading days. In co mparison, the market overall month ly
turnover in the U.S. over the same period is 2 2%, indicating a holding period of 1/0.22 = 4.5
months, which is about 90 days .5 These dramatic differences in holding horizon s suggest that
different types of Chinese investors might have quite different trading patterns and trading
preferences .6

5 We would also like to mention that i n the U.S., there is a large amount of high frequency trading, including
establishing and closing position s on the same day. China adopts the “T+1 trading rule ”, which requires that if stocks
are bought on day T, they cannot be sold on the sa me day. The reverse trade has to be executed on day T+1 or later.
That is, there is essentially a minimum holding period of one day.
6 We present additional summary statistics on retail trading volumes and holdings in Appendix Table 1. The results
show tha t small retail investors prefer to trade and hold small, low earning/price ratio, and high turnover firms, while
the largest retail investors trading and holding are tilted towards larger, high EP firms. In terms of sectors, the small 12
We measure order flows from different group s of investors ’ using order imbalance measure s,
as in Chordia and Subra hmanyam ( 2004 ). For stock i, day d, and investor group G, we compute
𝑂𝑖𝑏(𝑖,𝑑,𝐺)=∑ 𝐵𝑢𝑦𝑉𝑜𝑙 (𝑖,𝑑,𝑗)𝑗∈𝐺 −∑ 𝑆𝑒𝑙𝑙𝑉𝑜𝑙 (𝑖,𝑑,𝑗)𝑗∈𝐺
∑ 𝐵𝑢𝑦𝑉𝑜𝑙 (𝑖,𝑑,𝑗)𝑗∈𝐺 +∑ 𝑆𝑒𝑙𝑙𝑉𝑜𝑙 (𝑖,𝑑,𝑗)𝑗∈𝐺, (1)
where the numerator is the differen ce between buy and sell volumes summed up over all individual
j’s within group G, and the denominator is the sum of buy and sell volumes of all individuals in
group G. The order imbalance measure is an order flow measure , and we directly observe each
trade ’s direction from the data . When a set of investors buy s more than they sell, the order
imbalance is positive, and vice versa. We compute the order imbalance mea sure for each investor
group as OibRT 1 to OibRT 5, OibINST and OibCORP . The overall retail or der imbalance measure,
OibRT , is calculated by summing up all trades within the five retail groups.
Table I Panel C reports summary statistics for the order imbalance measures . The average order
imbalance for RT1 to RT5 , institutions and corporations are -0.021, -0.011, -0.006, 0.002 , 0.019 ,
-0.003, -0.011 and -0.004, respectively.7 The small magnitude of the se average order imbalance
measures indicates that most buys and sells within each investo r group cancel out each other . The
standard deviation s of order imbalance s are larger for large retail investor s and institutions
compared to small and medi um retail investors , indicating that there is more cross -stock variation

retail investors pref er to trade and hold the alternative energy sector, and prefer not to trade and hold banks and life
insurance firms, while the institutions and corporates behave in opposite ways. Finally, small retail investors tend to
use small order sizes, large retail investors tend to use large orders sizes, while institutions use all order sizes.
7 We plot the time -series of the cross -sectional mean, median and 25 th and 75 th percentiles of different types of
investors ’ order imbalance in Appendix Figure I. There are no obvious time trends or structural breaks in the time
series observations. 13
in large retail investor and institution al trading activity . The one-day autocorrelation coefficient ,
AR1, for these Oib measures are 0.243, 0.259, 0.216, 0.059 and 0.102 for RT1 to RT5 , suggesting
that small and medium retail order imbalance s are generally more persistent than large retail
imbalances .
In terms of order flow correlations across the seven groups , order flows from smaller retail
investors, OibRT 1, OibRT 2 and OibRT 3, are highly correlated , with correlation coefficient s mostly
higher than 0 .60. OibRT 4 is still positively correlated with OibRT 1-OibRT 3, but with a much lower
correlation of 0.20. The large st retail investors ’ order imbalance , OibRT 5, is negatively correlated
with all four other groups, with correlation s around -0.15, indicating that this group of retail
investors might have different trading patterns from the others .8 Institution al order imbalance s are
negatively correlated with all five retail groups, with correlation s ranging from -0.380 to -0.188,
again implying different trading pat terns from retail investors, even the largest retail investors. As
we saw earlier, corporat ions barely trade and their correlations with the rest of the investor
categories are all lower than 10%.9
In this data section, we include OibINST and OibCORP for the completeness of the summary
statistics. Corporations are long -term inve stors and rarely trade, while our study focus es on trading
behavior, so we also drop corporations from the remaining empir ical results. In terms of

8 In addition to the cross correlation analysis, we also estimate a VAR specification for the oib ’s from different group
of investors. Results are similar and are available on re quest .
9 Appendix Figure 1 present the time -series plot of the order imbalance measures for each investor group , and we find
no evidence of time trends or breaks. 14
institutional investors, given that retail investors are commonly assumed to be less sophisticated
than institutional investors, we keep institutions in our main empirical results for comparison
purposes.
II. Can Retail Order Flows Predict Future Stock Returns?
Can retail investors ’ activity predict future stock returns in China? If they can, it is possible
that these retail investors trading may contain information about future stock price movements.
We s tart by investigating the whether retail investors could predict future short term and the long
term returns with Fama -MacBeth regressions in Section II.A and II.B, respectively. In Section II.C,
we examine the predictive patterns for different subsets of stocks based on firm and stock
characteristics .
A. Predicting Next Day Stock Returns Using Retail Order Flows
To investigate the role s different retail investors play in the price discovery process, we first
examine the predictive power of various order flow variables for next -day returns using the two -
stage Fama -MacBeth regression. For the first stage, we estimate the following cross -sectional
regression for each day d,
𝑅𝑒𝑡(𝑖,𝑑)=𝑎0(𝑑)+𝑎1(𝑑)𝑂𝑖𝑏(𝑖,𝑑−1)+𝑎2(𝑑)′𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑑−1)+𝑢1(𝑖,𝑑), (2)
where the dependent variable 𝑅𝑒𝑡(𝑖,𝑑) is the stock return for firm i on day d, and the independent
variable s include order imbalance measure s from the previous day, 𝑂𝑖𝑏(𝑖,𝑑−1) , and control
variables , 𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑑−1). We follow previous literature for the choices of control variables. 15
To control for potential momentum/reversal from past returns, we include returns from the
previous day, Ret(-1), returns from the previous week, Ret(-6,-2), and returns from the previous
month, Ret(-27, -7). For size, value and liquidity effect s, we include log market size ( Size),
earnings -to-price ratio ( EP), and turnover, all computed from the previous month -end.
From the first stage estimation, we obtain a daily time-series of coeffic ients,
{𝑎0(𝑑),𝑎1(𝑑),𝑎2(𝑑)′}. For the second stage estimation, we conduct statistical inf erence based on
the mean and s tandard errors of the first stage coefficients, and we compute Newey -West standard
errors with 5 lags, which is the optimal lag number using a Bayesian Information Criterion (BIC) .
If the order flow variable from a specific investor group predict s future returns in the right direction,
in the sense that more past purchases are associated with higher future returns, and more past sales
are associated with lower future returns, we expect the coefficient a1 to be significantly positive,
and vice versa . The set -up of other Fama -MacBeth regressions in this study are similar to this
benchmark case. Therefore, we omit the similar details when int roducing the other specifications.
The estimation results for equation (2) are reported in Panel A of Table 2 , which display s
distinctive predictive patterns across different groups of retail investors. For the smallest retail
investor group , RT1, t he coe fficient on retail order flow variable is -0.0093 , with a significant t-
statistic of -24.98 . The negative coefficient shows that if retail investors RT1 buy more than they
sell o n a given day, the next day return on that stock is significantly negative. To understand the
economic magnitude of the coefficients , we report the inter -quartile range for OibRT 1 at the bottom. 16
Multiplying the interquartile range, 0.2222, by the regression coefficient of -0.0093 generates an
interquartile daily return difference of -21 basis points (more than 50% annualized !). For retail
investors in groups RT2 to RT4, the predictive patterns are qualitatively similar. All coefficients
are negative and statistically significant, and the daily interquartile return differences are -17, -11,
and -2 basis point s for RT2, RT3 and RT4 , respectively. That is to say, the first four groups of
investors all trade in the wrong direction vs. future price movements. Interestingly, when we move
from the smaller account size s to the larger ones, th e negative coefficients become smaller,
indicating that larger retail investors trade less incorrectly than smaller retail investors.
Indeed, for t he largest retail investors , RT5, the coefficient on past day order imbalance is
0.0012 , which is positive a nd significant with a t-statistic of 12.26 . The interquartile daily return
difference now is 5 basis points per day (over 12% per year) . It seems that the largest retail
investors ’ trading predict s the cross -section of future stock price movements in the correct
direction.
As a comparison, the coefficient on the previous day order imbalance is 0. 0016 for institution s,
with a t-statistic of 20.34 . That is to say, institutional order flows predict future stock price
movements in the right direction, and the interquartile return difference is 10 basis points per day ,
about twice the magnitude of the RT5 estimate . This finding is consistent with many previous
studies that institutional investors are more informed than retail investor in general.10

10 Table 1 Panel C shows a negative correlation of -0.188 between OibRT5 and OibINST . Readers might find it
confusing that both OibRT5 and OibINST positively predict returns while they have a negative correlation coefficient. 17
For the control variables, the coefficients on previous day return have mixed signs, while the
coefficients on previous week and previous month return s are all negative and significant,
indicating strong reversals over weekly and monthly horizons. Size is mostly insignificant, while
the earnings -to-price variable s are always positive and significant , indicating a strong value effect.
The coefficients on turnover are always negative and significant, suggesting that higher turnover
leads to lower returns in t he future. The above findings are mostly consistent with previous studies
of the Chinese stock market , such as Liu, Stambaugh and Yuan (2019) . These result s also confirm
that the predictive power of various order flow variables for future stock returns is not a
manifestation of siz e, value , liquidity or momentum/reversal effect .
B. Predicting Long Term Stock Returns Using Retail Order Flows
The exercise in Section II.A focuses on next-day return prediction. It is natural to ask whether
the predictive patterns carry on for longer term s. If the predictive pattern quickly vanishes or
reverses, what we observe migh t be driven by short -term noise . If the predictive pattern persists
over longer horizons , it is more likely the return predictability is linked to firm fundamentals or
persistent biases . Therefore, we extend the Fama -MacBeth specification in equation ( 2) to longer
horizons up to 12 weeks :
𝑅𝑒𝑡(𝑖,𝑤)=𝑏0(𝑤)+𝑏1(𝑤)𝑂𝑖𝑏(𝑖,𝑑−1)+𝑏2(𝑤)′𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑑−1)+𝑢2(𝑖,𝑤). (3)

In our opinion, both OibRT5 and OibINST contain positive information for future returns, which leads to the positi ve
predictive coefficients. However, the information contained in OibRT5 and OibINST are probably different, and the
trading of RT5 and INST might be quite distinctive, which leads to the negative correlation between the two. 18
That is, we use previous day order imbalance, 𝑂𝑖𝑏(𝑖,𝑑−1), to predi ct the cumulative returns
over the next w-week s. To be more specific, 𝑅𝑒𝑡(𝑖,𝑤) is calculated as a cumulative return from
day d+1 to the end of week w, where w=1, …, 12. For instance, when w=1, 𝑅𝑒𝑡(𝑖,𝑤) is the
cumulat ive return over day d+1 to d+5; when w=12, 𝑅𝑒𝑡(𝑖,𝑤) is the cumulative return over day
d+1 to d+60. If order imbalances have only short -lived predictive power for future returns, we
should observe the coefficient b1 decrease to zero quickly when w increases or even reverses .
Alternatively, if the specified retail order imbalance has longer predictive power, the coefficient
b1 should remain statistically significant for a longer period.
We present the estimat es of coefficient b1 in equation (3) in Table II Panel B. To save space,
we only report the coefficients and the statistical significance level by asterisks, with ***, ** and

• indicating significance at 1%, 5% and 10% level, respectively. For the smallest retail investors
  RT1, the coefficient o n OibRT1 monotonically increases from -0.0226 at week one to -0.0458 for
  a 12-week horizon, and all coefficients are statistically significant. Same patterns are also observed
  for OibRT 2, OibRT 3, and OibRT4. The positive predictive power of OibRT 5 and OibINST also
  persist significantly for at least 12 weeks , and there are no obvious reversals. The persistence of
  cross -sectional predictability indicates that the predictive power is likely rooted in information
  related to fundamentals or from persistent noise trading or behavioral biases .
  C. Predicting Patterns Across Firms with Different Characteristics 19
  Previous studies show that stock returns can be significantly affected by firm and stock
  characteristics, such as size, EP ratio and liquidity. Do predictive pattern s of retail order flows
  differ across firms with different characteristics? To answer this question, we modify the Fama -
  MacBeth specification in equation (2) and allow different coefficients for firms with different
  characteristics , by i ncluding interactions with characteristics dummies as follows,
  𝑅𝑒𝑡(𝑖,𝑑)=𝑐0(𝑑)+[𝑐1(𝑑)𝐷𝑢𝑚𝑚𝑦 1(𝑖,𝑑−1)+𝑐2(𝑑)𝐷𝑢𝑚𝑚𝑦 2(𝑖,𝑑−1)+
  𝑐3(𝑑)𝐷𝑢𝑚𝑚𝑦 3(𝑖,𝑑−1)]𝑂𝑖𝑏(𝑖,𝑑−1)+𝑐4(𝑑)′𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑑−1)+𝑢3(𝑖,𝑑). (4)
  Take size as an example. We first separate all firms on day d into three groups, based on previous
  month -end firm market capitalization . The dummy variable, 𝐷𝑢𝑚𝑚𝑦 1(𝑖,𝑑−1), takes value 1 if
  firm i belongs to the smallest 1/3 of firms, zero otherwise; 𝐷𝑢𝑚𝑚𝑦 2(𝑖,𝑑−1) takes value 1 if
  firm i belongs to the medium 1/3 of firms, zero otherwise; and 𝐷𝑢𝑚𝑚𝑦 3(𝑖,𝑑−1) takes value 1
  if firm i belongs to the largest 1/3 of firms, zero otherwise. The coefficients 𝑐1,𝑐2 𝑎𝑛𝑑 𝑐3
  provide information on whether the predictive pattern changes for firms with different sizes.
  Estimation results for equation (4) are reported in Table III. In the first three rows, we separate
  firms by their market capitalization. The negative predictive pattern of order flow from RT1 -RT4
  for next day return, as observed in Table II, is quite robust for firms with different sizes. But it is
  interesting to notice that the magnitudes generally decrease from the smallest firms to the largest
  firms, indicating that the negative predictive pattern is the strongest for smaller firms. For the large
  retail investors, RT5, the positive predi ctive pattern remains for the small and medi um-sized firms, 20
  but not for large firms, indicating that their information advantage, if any, might be concentrated
  in smaller firms. As a comparison, order flows from institutions significantly predict next day
  returns in all three rows, and more so for the large firms , suggesting that their information
  advantage , if any, might be m ore prominent for larger firms.
  When we separate firms by EP, turnover and stock price, we observe similar interesting patterns.
  That is, the predictive patterns in Table II are generally robust across firms with different
  characteristics, and the negative (positive) predictive power of smaller (large r) retail investors is
  stronger for small, low EP, and higher turnover firms, while th e positive predictive power of
  institutional investors is stronger for large and high EP firms. 11 12
  III. What Drives the Order Imbalance Predictive Power for Future Returns ?
  Given the large differences in predictive power for future returns of different in vestor groups’
  order flows, it is important to understand the driving forces for these differences. Previous
  literature provides several hypotheses for explaining investor order flows in gen eral, and th ese
  might help to explain the heterogeneous predictive patterns from different investor group s for

11 The Appendix Table 1 Panel A and Panel B show that small retail investors trade and hold more of smaller stocks,
firms with lower EP and higher turnovers. Combining with the predictive patterns in Table 3, small retail investors
likely have the largest information disadvantage or be havioral biases in these firms. In contrast, institutions have more
trading and holding in larger stocks, firms with lower EP and higher turnovers. Those are also the firms that institutions
have the highest predictive power, indicating they might have mor e information advantage over these firms.
12 Given the positive and significant coefficients on OibRT5, one might wonder who are those larger retail investors.
With a subsample between January 2019 to March 2019 with demographic information, we find RT5 are mostly male,
with age above 45. Previous literature also provide s additional information on the retail investors with larger accounts.
For instance, An, Lou, Shi (2022) find that a transfer of 250 billion CNY from the poor to wealthy retail investors
during bubbles and crashes period. Titman, Wei, and Zhao (2022) find larger retail investors tend to accumulate
positions before suspicious split announcements and sell in the post -split period, indicating that they might have
information advantage. 21
future returns. In Section III.A, we introduce a two-stage decomposition for the order flow ’s
predictive power for future returns. We present the empirical resu lts for the decomposition in
Section III.B. We take a closer look at the information channel using event days in Sect ion III.C .
A. A Two-Stage Decomposition to Explain Order Imbalance ’s Predictive Power
We consider four hypotheses for explaining the order flow dynamics and their predictive power
for future stock returns. First, Chordia and Subrahmanyam (2004) state that order flows tend to be
persistent, and persistent buying/selling pressure could lead dir ectly to the predictability of future
returns. Second, Kaniel, Saar, and Titman (2008) argue that retail traders in the U.S. are mostly
contrarian , which provides liquidity to the market, and investors receive future positive returns .
Following this logic, if the retail trades are momentum, which demand liquidity, then it is possible
that the momentum trades might negatively predict future returns . Third , Liu et al. (2021) connect
retail trading motives to behavioral biases, and they find that over -confidence about information
advantage and gambling preference s are the two dominant behavioral biases that affect trades of
Chinese retail investors .13. Finally, Kelley and Tetlock (2013) find that retail investors, especially
the aggressive ones, may have valuable information about fundamental firm news, and thus their
trading could correctly predict t he direction of future returns. The above hypotheses are not
mutually exclusive.

13 There are many other interesting behavior biases of retail investors, such as disposition effects, extrapolation . Due
to space limit of this study, here we choose the two most behavioral biases that affect trades of Chinese retail investors:
over-confiden ce about information advantage and gambling preferences. W e leave the other s to future research. 22
To find out whether the above hypotheses help to explain the t rading behavior of different
retail investor groups, and their predictive power for future stock returns, we follow the two -stage
decomposition method as in Boehmer et al. (2021). For the first stage, we use the above hypothese s
to explain the retail flow measures to find out which ones are important drivers for the order flows.
This step also helps to decompose the retail order flows into hypothesis -implied components for
each hypothesis . For the second stage, we investigate which of the hypothesis -implied component s
contributes to the predictive pattern of different investor order flow measures.
To estimate the two -stage decomposition, we first identify proxies f or each hypothesi s. The
proxies for the first two hypotheses are relatively easy to construct. For the order -persistence
hypothesis, we adopt the previous day order imbalance measure, Oib(i,d-1), as the proxy. For the
liquidity provision hypothesis, since it is directly linked to previous contrarian/momentum trading,
we use returns from the previous day, week and month as proxies . For the overconfidence measure ,
we follow Barber et al . (2008) and Liu et al. (2021) and proxy it with corresponding investor
group ’s turnover on that stock , which is the investor group ’s average of daily buy volume plus the
daily sell volume divided by the investor group ’s holding shares at the end of previous day . Then
we computed the final over confidence variable, 𝑂𝑣𝑒𝑟𝑐𝑜𝑛𝑓 (𝑖,𝑑), as an average of daily group
turnover from the previous 20 days .14 For gambling preference s at stock level, 𝐺𝑎𝑚𝑏𝑙𝑒 (𝑖,𝑑), we

14 For overconfidence proxy, Barber and Odean (2000) use each household ’s portfolio ’s turnover to proxy their
overconfidence, which is aggregated at household level. Here we focus on investor groups rather than households, so
we carry the spirit from th e previous literature and use the stock -level turnover from each investor group. We also
acknowledge that group turnover can potentially contain information other than over -confidence. 23
follow Bali et al. (2011) and compute the maximum daily returns from the previous 20 days as the
proxy.15
For the information hypothesis, the most influential information at the firm level is earnings
news , hence we follow Kelley and Tetlock (2013) and measure firm -level information by the
cumulative abnormal returns (CAR) over the earnings announcement period. However, unlike t he
proxies for order persistence, liquidity provision and behavioral biases, whi ch can be computed
for each stock on each day, the news prox ies are only available on earnings news days, which
account for 1.58% of stock -days, and would render our two -stage estimation imprecise. To cope
with this missing data issue for the news hypothes is, in this section we only consider the order
persistence, liquidity provision and behavioral bias hypotheses, and we focus on the news
hypothesis using an event -day approach in Section III.C.
After we collect all the proxies, we estimate the first stage for the two -stage decomposition.
For each day d, we estimate a cross -sectional specification,
𝑂𝑖𝑏(𝑖,𝑑)=𝑑0(𝑑)+𝑑1(𝑑)𝑂𝑖𝑏(𝑖,𝑑−1)+𝑑2(𝑑)′𝑅𝑒𝑡(𝑖,𝑑−1)+𝑑3(𝑑)𝑂𝑣𝑒𝑟𝑐𝑜𝑛𝑓 (𝑖,𝑑−
1)+𝑑4(𝑑)𝐺𝑎𝑚𝑏𝑙𝑒 (𝑖,𝑑−1)+𝑢4(𝑖,𝑑). (5)
After we obtain the time-series of coefficients, {𝑑0̂(𝑑),𝑑1̂(𝑑),𝑑2̂(𝑑)′,𝑑3̂(𝑑),𝑑4̂(𝑑)} , we

15 An alternative measure for gambling preference proxy is introduced in Liu et al. (2021), which rely on events when
the stock return hits 10% price limit . However, the 10% price limit hit only accounts for 0.07% of our total sample,
and isn ’t suitable for our purpose on daily*stock frequency. Thus we choose the maximum daily return as our main
gambling measure. We also consider other alternative proxies fo r gambling preferences, such as idiosyncratic volatility
and skewness. These proxies deliver similar results to those using maximum daily return s, and are available on request. 24
conduct statistical inference using the time -series means and standard errors, which are adjusted
using Ne wey-West with five lags, in order to understand how each of the four hypothese s
contributes to retail order flows. Meanwhile, t he first stage estimation allows us to decompose
Oib(i,d) into five components:
𝑂𝑖𝑏(𝑖,𝑑)=𝑂𝑖𝑏̂𝑖,𝑑𝑝𝑒𝑟𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑒+𝑂𝑖𝑏̂𝑖,𝑑𝑙𝑖𝑞𝑢𝑖𝑑𝑖𝑡𝑦+𝑂𝑖𝑏̂𝑖,𝑑𝑜𝑣𝑒𝑟𝑐𝑜𝑛𝑓+𝑂𝑖𝑏̂𝑖,𝑑𝑔𝑎𝑚𝑏𝑙𝑒+𝑂𝑖𝑏̂𝑖,𝑑𝑜𝑡ℎ𝑒𝑟, (6)
with 𝑂𝑖𝑏̂𝑖,𝑑𝑝𝑒𝑟𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑒=𝑑1̂(𝑑)𝑂𝑖𝑏(𝑖,𝑑−1),𝑂𝑖𝑏̂𝑖,𝑑𝑙𝑖𝑞𝑢𝑖𝑑𝑖𝑡𝑦=𝑑2̂(𝑑)′𝑅𝑒𝑡(𝑖,𝑑−1),𝑂𝑖𝑏̂𝑖,𝑑𝑜𝑣𝑒𝑟𝑐𝑜𝑛𝑓=
𝑑3̂(𝑑)𝑂𝑣𝑒𝑟𝑐𝑜𝑛𝑓 (𝑖,𝑑−1),𝑂𝑖𝑏̂𝑖,𝑑𝑔𝑎𝑚𝑏𝑙𝑒=𝑑4̂(𝑑)𝐺𝑎𝑚𝑏𝑙𝑒 (𝑖,𝑑−1) 𝑎𝑛𝑑 𝑂𝑖𝑏̂𝑖,𝑑𝑜𝑡ℎ𝑒𝑟=𝑢4̂(𝑖,𝑑−
1)+𝑑0̂(𝑑−1). That is, the “persistence ” part is related to the order persistence hypothesis, the
“liquidity ” part is related to the liquidity pr ovision hypothesis, the “overconf ” and “gamble ” are
both related to behavioral biases, and the “other ” component is the residual component, which
potentially contains other relevant information about future returns.
For the second stage of the decomposition, we relate future returns to each individual
component of order flow by estimating the foll owing specification using the Fama -MacBeth
methodology:
𝑅𝑒𝑡(𝑖,𝑑+1)=𝑒0(𝑑+1)+𝑒1(𝑑+1)𝑂𝑖𝑏̂𝑖,𝑑𝑝𝑒𝑟𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑒+𝑒2(𝑑+1)𝑂𝑖𝑏̂𝑖,𝑑𝑙𝑖𝑞𝑢𝑖𝑑𝑖𝑡𝑦+𝑒3(𝑑+
1)𝑂𝑖𝑏̂𝑖,𝑑𝑜𝑣𝑒𝑟𝑐𝑜𝑛𝑓+𝑒4(𝑑+1)𝑂𝑖𝑏̂𝑖,𝑑𝑔𝑎𝑚𝑏𝑙𝑒+𝑒5(𝑑+1)𝑂𝑖𝑏̂𝑖,𝑑𝑜𝑡ℎ𝑒𝑟+𝑒6(𝑑+1)′𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑑)+
𝑢5(𝑖,𝑑+1). (7)
With the decomposition in equation (6), the coefficient estimates in equation (7) show how each
component of various order flows helps to predict future stock returns. 25
According to Boehmer et al. (2021) , the advantage of the two -stage decomposition approach
is that it includes various components of 𝑂𝑖𝑏(𝑖,𝑑) from alternative hypotheses in a unified and
internally consistent empirical framework. The caveat of this approach is that we need to make
empiri cal assumptions when choosing proxies for different hypothes es. Even though t hese
assumpti ons seem to us to be reasonable, we still need to be cautions that the results depend on the
validity of our empirical assumptions.
B. Estimation Results for the Two -Stage Decomposition
We report first -stage estimation results in Table IV Panel A. In the first row, the coefficients on
lagged order flow variables are always positive and significant, indicating that order persistence is
an important driver for order flows. For the next three rows, we connect order flows with returns
from previous day, week and month, and the patterns are quite interesting. The order imbalances
of RT1, RT2, and RT3 load positively and significantly on the previous day return, indicating that
these investors buy more if the previous day return is positive, and sell more if the previous day
return is negative . This corresponds to a daily momentum trading strategy , which demand s
immediate liquidity. For larger r etail investors in RT4 and RT5 , order imbalances load negatively
and significantly on returns from the previous day, indicating that they are contrarian investors,
buying low and selling high, and possibly provid ing immediate liquidity. If we extend the horizon
to previous one week or one month, then the coefficients on all returns are negative and significant, 26
indicating that all retail investors follow contrarian strategies, buying losers and selling winners
over the longer term .16
The next two rows present results on how behavioral biases are relate d to order flows . The
coefficients on the overconfidence proxy are all positive and significant for RT1 -RT4, indicating
that overconfidence , proxied by group turnover, might be a strong driver for these retail investors ’
trading . Intriguingly, the magnitude of the coefficients gradually decrease s from 0. 0894 for RT1
to 0.0418 for RT4, implying a decreasing impact of overconfidence for retail investors as their
account si zes increase. For the largest retail group, RT5, the coefficient becomes -0.0881 with a
significant t-stat of -8.11. That is, the largest retail investors ’ trade s are in the opposite direction
of the overconfidence proxy . In terms of the gambling preference, for RT1 -RT4, the coefficients
are always positive and significant, indicating these retail investors like to buy stocks with lottery
features. Interestingly, the coefficients gradually increase from 0.0330 for RT1 to 0.2423 for RT4,
suggesti ng that larger retail investors trades have higher association with gambling preferences .17

16 Our finding that large retail investors are contrarian and smaller ones are momentum traders over daily horizon is
quite interesting and different from some previous studies. For instance, contrarian patterns have been documented in
Kaniel, Saar and Titman (2008) using monthly horizons in the U.S., and Barrot, Kaniel and Sraer (20 16) using daily
and weekly horizons in France. Using U.S. data, Kelley and Tetlock (2013) and Boehmer et al . (2021) both find that
retail trades follow momentum over daily horizons, but are contrarian at weekly horizons. In our setting, we find the
trading patterns from investors with smaller account sizes are similar to those in Kelley and Tetlock (2013) and
Boehmer et al . (2021), while the investors with the largest account sizes behave similarly to the patterns documented
in Kaniel et al . (2008) and Barr ot et al . (2016).
17 The coefficients of gambling preferences monotonically increase from RT1 to RT4 are not necessary inconsistent
with the finding that RT1 has the most negative return predictive power than other retail investor groups . There might
be many rational and behavior al drivers for investors ’ trading behaviors, which potentially affect the overall predictive
power of order flow for future returns . Here we only include the two most important behavior biases , as suggested by
Liu et al. (2021), and the rest would be in the “other ” component of order imbalance measures. 27
When we move on to RT5, the coe fficient is -0.0671 with a significant t-stat of -3.09, which
indicate s that the largest retail investors trade s are in the opposite direction of the gambling motive .
We report the second stage of the decomposition results in Panel B of Table IV. We take
the first retail group, RT1, as an example. The coefficient estimate on Oib(Persistence) is -0.0333,
with a t-statistic of -15.84, wh ich implies that order persistence significantly and negatively
contributes to the predictive power of RT1 trading flow. The coefficient estimate on Oib(Liquidity)
is -0.0088, with a t-statistic of -2.61, which probably implies that daily momentum trading
probably significantly and negatively contributes to the predictive power of RT1 trading flow. The
coefficient of Oib(Overconf) is -0.1024, with a t-statistic of -2.84, and the coefficient for
Oib(Gamble ) is insignificant. For the Oib(Other) component, the coefficient is -0.008 5, with a
significant t-statistic of -27.11, indicating that there is other information, other than those
incorporated in the three hypotheses, that significantly contribute s to RT1’s negative predictive
pattern for fut ure returns. In terms of economic magnitude, we compute the interquartile range of
all five components of the order imbalance measure. For the smallest retail group RT1, if we move
from the 25th percentile to the 75th percentile in the distribution, the interquartile differences in
future one -day stock return, for the Oib(Persistence) , Oib(Liquidity ), Oib(Overconf ),
Oib(Gamble ) and Oib(Other) , are -0.1177%, -0.0290%, -0.0337%, -0.0314 %, -0.1778 %,
respectively. That is to say, order persistence, liquidity demand , overconfidence, and gambl ing
preference s all contribute to the negative predictive po wer of RT1 for next day returns , while the 28
first term has the largest magnitude . Similar patterns are observed for other smaller retail investor
groups RT2-RT4.
If we turn our attention to the largest retail investors, RT5, the patterns are quite different. In
terms of coefficient estimates, we find the order persistence and other are both positive and
significant. In terms of economic magnitude, if we move from the 25th percentile to the 75th
percentile in the distribution, the interquartile differences in future one -day stock return, for the
Oib(Persistence) , Oib(Liquidity) , Oib(Overconf ), Oib(Gamble) and Oib(Other) , are 0.0281 %,
0.0097 %, 0.0257 %, 0.0425 %, 0.0490 %, respectively . This indicates all three hypotheses
contribute to RT5’s positive predictive pattern for future returns , while only the first is significant .
Overall, our decomposition exercise shows that a substantial part of the negative predictive
power of the retail investors with smaller account sizes c omes from order persiste nce, liquidity
demand , and behavioral biases , while the positive predictive power of the retail investors wi th
larger account balances comes from order persistence and trading against overconfidence and
gambl ing preferences . Across all investor groups, t he significance and the large magnitude of the
“other ” component indicates that existing hypothes es can not fully explain the trading behaviors
and their predictive power fo r returns. So what does “other ” stand for? One possibility is
information, which we take a close look at in the next subsection.
C. A Close Look at the Information Channel 29
It is important to unde rstand how various retail investors participate in the information
discove ry process. As mentioned earlier, the most influential information at the firm level is
earnings news , hence we follow Kelley and Tetlock (2013) and measure firm -level information by
the cumulative abnormal returns (CAR) over the earnings announcement period. Notice that
earnings news only happens quarterly rather than daily, so the daily Fama -MacBeth estimation we
adopt for the two stage estimation might not be pr oper for understanding how Chinese retail
investors process information. As an alternative , in this section, we focus on event days to study
this issue. To capture each retail investor groups ’ participation in the information discovery proc ess,
we proceed in three steps.
In the first step, we examine whether different retail investors can predict earnings news the
next day. A positive answer indicates that these investors anticipate the information before the
information becomes public, either because they have access to private information or they have
better skills . In the second step, we check whether they can proces s contemporaneous earnings
news to find out whether they have skills to gather information from available public news. In the
third step, we investigate whether retail order flow ’s predictive power for future returns improves
or deteriorates on earnings ev ent days, to find out whether information is a significant contributor
for the overall prediction pattern documented in previous sections.
For this first step, t o find out whether retail order flows can predict earnings news, we estimate
the following cross -sectional specification for each quarter q: 30
𝐶𝐴𝑅 (𝑖,𝑑−1,𝑑)=𝑓0(𝑞)+𝑓1(𝑞)𝑂𝑖𝑏(𝑖,𝑑−2)+𝑓2(𝑞)′𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑞−1)+𝑢6(𝑖,𝑞). (8)
Assuming the earnings announcement day is day d, we compute the cumulative returns over day
d-1 and day d, and subtract the market returns over the same period to obtain cumulative abnormal
returns for each stock, 𝐶𝐴𝑅 (𝑖,𝑑−1,𝑑).18 The main predictive vari able on the right hand of the
equation is order imbalance measure from day d-2. Notice that each firm only has one earnings
day each quarter, and equation (8) is estimated for each quarter in the cross section to make sure
we cover all firms each quarter . As in a standard Fama -MacBeth setting, the statistical inferences
are based on the quarterly time -series of the estimated coefficients , and standard errors are
computed using Newey -West with 4 lags . If retail order flows can predict earnings surprises in the
right direction, the coefficient f1 should be significantly positive, and vice versa .
We present the estimation results in Panel A of Table V . For retail investors RT1-RT3, the
coefficients f1 are -0.0251, -0.0234, and -0.0166, respectively, all with highly significant t-statistics.
These negative and significant coefficients indicate that these investors incorrectly predict earnings
surprises. The coefficient f1 for RT4 is close to zero and insignificant . In contrast, the coefficients
f1 for RT5 is 0.0023, positive and statistically significant, implying that these investors are able to
correctly predict future earnings surprises.
For the second step, we examine whether different retail groups can proces s contemporaneous
public news. Here the depend ent variable is retail order flow, Oib(i,d), and we connect it to

18 We also examine wider window such as CAR( -1,1) and CAR( -3,3), the results are similar and available on request. 31
contemporaneous earnings news, CAR(i,d -1,d). The specification is similar to equation (8), except
the timeline is different :
𝑂𝑖𝑏(𝑖,𝑑)=𝑔0(𝑞)+𝑔1(𝑞)𝐶𝐴𝑅 (𝑖,𝑑−1,𝑑)+𝑔2(𝑞)′𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑑−2)+𝑢7(𝑖,𝑞). (9)
If a particular type of retail order imbalance can process contemporaneous and public earnings
news in the right direction, we expect the associated coefficient g1 to be significantly positive , and
vice versa .
Panel B of Table V report the estimat ion results. For retail investors RT1-RT4, the coefficients
g1 are -1.9225, -1.8291, -1.4349 , -0.8781 respectively, all with highly significant t-statistics. These
negative and significant coefficients indicate that these retail investor groups process the
contemporaneous public earnings news in the wrong direction. In contrast, the coefficient g1 for
RT5 is 0.1583 , implying that RT5 might be able to correctly process contemporaneous public
earnings news . However, the coefficie nt is statistically insignificant .
For the third step, we examine whether retail order flows ’ predictive power for future returns
improves or deteriorates on event days to understand how much the information hypothesis help s
to explain the return predictive patterns we observe in Section II . We estimate a modified version
of equation (2), by adding the event da y dummy and an interaction term:
𝑅𝑒𝑡(𝑖,𝑑)=ℎ0(𝑑)+[ℎ1(𝑑)+ℎ2(𝑑)𝐸𝑣𝑒𝑛𝑡 (𝑖,𝑑−1)]𝑂𝑖𝑏(𝑖,𝑑−1)+
ℎ3(𝑑)𝐸𝑣𝑒 𝑛𝑡(𝑖,𝑑−1)+ℎ4′(𝑑)𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 (𝑖,𝑑−1)+𝑢8(𝑖,𝑑) . (10) 32
Here the event dummy Event( i, d-1), is equal to one if the firm i has earnings news on day d-1,
and zero otherwise . For non -news days , the predictive power of retail tra des is measured by
coefficient h1; for news days , the p redictive power is measured by (h1+h2). If coefficient h2 is
significantly different from zero, that group of retail investors anticipates future stock returns
differently on these news days.
In the U.S., firm earnings announcements are chosen by firms and scattered throughout the
year. In China, all firms are required to report their financial statements to regulators before four
preset deadline date s each year. As a result, firms mostly announce their ear nings within a short
period before these deadline dates, and there would be zero announcements outside of these short
periods . To make sure that we have enough observations to estimate the Fama -Macbeth
coefficients in equation (10) , we only include days with at least 5% of total number of firms with
earnings announcement s, which gives us 6 8 days, or 8% of the total days in our sample.
The results ar e presented in Table V Panel C . Here we take the smallest retail investors, RT1,
as an example. The coefficients on order imbalan ce, h1, is -0.0079 and is statistically significant,
indicating that on average the trades from RT1 negatively predict future returns. When there is
earnings announcement news, the coefficient on the interaction of event dummy and th e order
imbalance is -0.0080 , with a significant t-statistic of -3.20, implying that the negative prediction
of RT1 for future stock returns doubles on earnings news days . This is consistent with our earlier
finding that the smaller retail investors fail to predict and process the earnings news, which leads 33
to more negative prediction for returns on event days. We observe similar patterns for RT2, RT3
and RT4. For the largest retail investors, RT5, the coefficients h1 and h2 are 0.0005 and 0.0014 ,
both statistically significant. That is to say, the large retail investors ’ predictive power for future
returns quadruples on earnings news days, possibly because these retail investors can correctly
predict and process the earnings news, which enhance s their ability to predict future stock returns.
Overall, our results reveal interesting heterogeneous patterns of how retail investors predict
and process public information . On one hand, smaller retail investors are unable to predict future
news and lack skills to correctly process public news, while the largest retail investors and
institutions are able to correctly anticipate future earnings news and incorporate the
contemporaneous news into their trading . The differences in information -processing abili ties of
different retail investors clearly contribute to the differences in their predictive powers for future
returns.
IV. Further Discussions and Robustness
A. Ages and Genders
In this section, we examine heterogeneity through demographic differences, such as gender
and age, of retail investors. According to Barber and Odean (2001), male investors could be more
susceptible to behavioral biases, such as overconfidence and lack of attention. Due to the limited
access to data, we only have a three -month sample period from January 2019 to March 2019 on
investor gender and age. We first present summary statistics on age and gender in Table VI Panel 34
A. Male investors contribute 67% of t rading volume on average, and females account for 33%.
Within the male group, the trading volume (%) across age groups below 45, and above 45 is 29%
and 38% (summing to the 67% male total), while the trading volume (%) for the same age groups
for females i s 13% and 20% (summing to the 33% of volume traded by females). That is, across
all gender -age groups, older male investors trade the most.
Next, we examine the determinants of return prediction for each gender -age group specified in
equation (2). The resu lts are reported in Table VI Panel B. For return predictions, we find male
investors across all ages significantly and negatively predict returns, especially for older males .
The predictive coefficients are insignificantly different from zero for female re tail investors. These
interesting patterns across age and gender are mostly consistent with previous findings in Barber
and Odean (2001) provide further evidence regarding heterogeneity of retail investors.
B. Applying stricter filters from Liu, Stambaugh and Yuan (2019)
In this study, we apply a filter from Liu et al . (2019) and discard stocks with less than 15 days
of trading during the most recent month. In addition, Liu et al. (2019) also eliminate stocks that
have become public within the past six mon ths, stocks with less than 120 days of trading during
the past 12 months, and the smallest 30% of firms listed in SHSE and SZSE. We add all these
additional filters and check the robustness of our results.
In Table VI Panel C, the order imbalance predicti on direction s are similar to the results in Table
II. The first four groups of retail investors tend to trade in the wrong direction for future price 35
movements , while the largest retail investor group RT5 and institutions trade in the same direction
as the cross -section of future stock return s. The economic magnitudes for the first four type of
retail investors are quantitatively similar, while RT5’s economic magnitude is only half as large
when add ing these additional f ilters, perhaps because RT5 ’s positive return mainly comes from
small stocks. The economic magnitude for institutions is still large . In conclusion, our main results
are robust to the stricter filters from Liu, Stambaugh and Yuan (2019).
C. Leveraged positions
Our trade level data also identify i nvestor s’ margin buys, short s ales and collateral trades .
Leverage d trading may be different from non -leverage trading. On each day, margin buy s account
for 10% of the trading vol ume, short s ales account for 0.2 % and col lateral trading accounts for 15%
during our sample period. We exclude the leverage trades and re-estimate equation (2) .
Results are report ed in Table VI Panel E. The order imbalance prediction direction s are similar
to the results in Table I I. The first four groups of retail investors trade in the wrong direction of
future price movements, while the largest retail investor group RT5 and institutions trade in a way
that positively predict s the cross -section of future stock return s. The economic ma gnitud es are
quantitatively simila r. In conclusio n, our results are robust to whether or not we include these
leverage trades.
D. Price limits 36
One institutional feature of the Chinese market is the price limit restrictions . That is, investors
can buy and sell stocks freely when the stock ’s price is within ±10% from previous day close price;
if the price move out of the ±10% range, trading stops till the next day open. Chen et al. (2019)
focus on the price limit days and fin d that large investors tend to trade differently on these days.
Here we examine whether our results still hold when we exclude the price limit day s.
We re -estimate equation (2) without the price limit days and present the r esults in Table VI Panel
E. The o rder imbalance prediction directions are similar to the results in Table II. The first four
groups of retail investors trade in the wrong direction of future price movements, while the largest
retail investor group RT5 and institutions trade in a way that positively predicts the cross -section
of future stock returns. The economic magnitudes are quantitatively similar. That is , our results are
robust to whether o r not we include these price limits days .
E. Forming Portfolios Using Retail Order Flows
Our main results in previous sections are based on Fama -MacBeth regressions, which assumes
linear relations between the future returns and order flow variables. In this section, we adopt an
alternative portfolio approach and examine whether our results still hold. To be more specific, we
sort firms into five groups each day , based on previous day ’s order imbalance from a particular
investor group , buying and selling the 20% of stocks with the highest and lowest order imbalance
measures for that particular investor g roup. We report the risk adjusted returns (alphas) on this 37
long-short strategy for one to 60 days , where we conduct risk adjustment using the Liu, Stambaugh
and Yuan (2019) three factor model .
From Table V I Panel F, the one -day long -short portfolio alpha, using the previous day order
imbalance from RT1, is -0.0042, and highly significant. From one week to 12 weeks, the
cumulative alphas for the long-short portfolio decrease from -0.008 9 to -0.0183, and they are all
highly significant. That is, the cumulativ e alphas using OibRT1 is consistently negative and
significant, and there are no signs of reversal within 12 weeks, which echoes our earlier results in
Table 2. Similar patterns exist for RT2 and RT3. For RT4, the one -day alpha is negative at -0.0007,
but it quickly becomes insignificant when we extend the holding horizon to one week, indicating
that RT4 for horizons longer than one day. For the largest retail investors in RT5, the one -day alpha
is 0.0017, which is positive and significant. The 12 -week cumu lative alpha is 0.0057, still positive
and significant, confirming the results in Table 2 that RT5 has both short and long term predictive
power for future returns. Results on OibINST are similar to those for OibRT5 .
F. News from CFNDS
Our earlier results sh ow that the smaller retail investors lack skills to predict or process public
earnings news, while the largest retail investors are able to correctly predict and process future
earnings news and incorporate the contemporaneous news into their trading. In this section, we
use an alternative public news dataset to investigate whether the results from earnings news can
be extended to other news. We obtain news data from the Financial News Database of Chinese 38
Listed Companies (CFND), which includes news on all A-share stocks from over 400 internet
media and over 600 newspapers. In comparison with earnings news, the data coverage is more
substantial, but the news content is more diverse.
We estimate equation (10) and report the results in Table VI Pa nel G . For the smallest retail
investors, RT1, the coefficient on order flow is -0.0075 and highly significant, confirming that
their order flows predict returns negatively. The coefficient on the interaction between order flow
and the event dummy is -0.0045 , again h ighly significant, suggesting the negative predictive power
is significantly stronger on news days, which is consistent with our results in Section III.C. Similar
patterns are observed for RT2 -RT4. For the largest retail investors, RT5, the coefficient on order
flow is 0.0008, and on the interaction is 0.0011, both highly significant, indicating that the RT5
order flow on average predict s future returns in the correct direction, and their prediction becomes
much stronger on news days. To summarize, we confirm with an alternative news dataset that
smaller retail investors lack skills to process public earnings news, and their negative predictions
for future returns are worse on news days, while the largest retail investors and institutions are able
to co rrectly process future earnings news and enhance their predictive power for future returns.
V . Conclusion
Using comprehensive retail trading and holding data from 2016 to 2019 , we separate tens of
millions of retail investors into five groups by their ac count sizes, and examine heterogeneity in
retail investors ’ return predictabilities , and sources of the return predictabilities . 39
We provide strong and direct evidence on retail investors ’ heterogeneity. Retail investors with
account sizes less than 3mil CN Y buy and sell stocks in the wrong directions. The prices of stocks
they buy experience negative returns the next day, while the ones they sell experience positive
returns. For retail investors with large account balances, their trading predict s returns in the correct
direction. In tracing their differences in predicting future returns, we provide evidence that the
negative predictive power of the retail investors with smaller account sizes are mostly related to
their order persistence, daily mom entum trading, behavioral biases and failures in processing
earnings news. In contrast, the positive predictive power of the large retail investors is mostly
associated with order persistence, contrarian trading, trading against behavioral biases and
advan tages in processing earnings news.
Our results on the heterogeneity of retail investors help to understand the conflicting empirical
results in the previous literature regarding retail investors. In addition, the exchange itself
acknowledges the heterogen eity in retail investors and is focused on adopt ing policies on investor
education and suitability that restrict some kinds of trading for the smallest accounts. For example,
the Shanghai S tock Exchange require s a retail investor to have at least 500k CNY holding s of
stocks for at least 20 trading days to open a leverage trading account or to trade on the riskier
Science and Technology Innovation Board (or STAR Market). Th ese policies effectively exclude
the smallest retail investor s from leverage trading and trading on riskier start -ups, which could
help protect the se small retail investors from even worse losses . 40
Our study clearly leaves many interesting questions unsolved. For example, why retail
investors dominate trading in the Chinese stock market? Is it the T+1 trad ing rule that discourage
participation of institutional investors in trading? How much do retail investors gain or loss with
their investments in stock market? We leave these interesting and important questions to future
research . 41
References
An, Li , Dong Lou, and Donghui Shi , 2022, Wealth redistribution in bubbles and crashes. Journal
of Monetary Economics , 126(3), 134-153.
Anagol, Santosh, Vimal Balasubramaniam, and Tarun Ramadorai , 2021 , Learning from noise:
Evidence from India ’s IPO lotteries. Journal of Financial Economics , 140(3), 965-986.
Bach, Laurent, Laurent E. Calvet, and Paolo Sodini , 2020, Rich pickings? Ri sk, return, and
skill in household wealth. American Economic Review , 110(9), 2703 -47.
Balasubramaniam, V imal, John Y. Campbell , Tarun Ramadorai , and Benjamin Ranish , 2021 ,
Who o wns w hat? A f actor m odel for d irect s tockholding. NBER Working paper.
Bali, Turan G., Nusret Cakici, and Robert F. Whitelaw, 2011, Maxing out: Stocks as lotteries and
the cross -section of expected returns , Journal of Financial Economics , 99(2), 427 -446.
Barber, Brad M., Xing Huang , Terrance Odean , and Christopher Schwarz , 2021 , Attention induced
trading and returns: Evidence from robinhood users. Journal of Finance , Forthcoming.
Barber, Brad M. , Yi-Tsung Lee , Yu-Jane Liu , and Terrance Odean , 2009, Just how much do
individual investors lose by trading?. The Review of Financial Studies , 22(2), 609 -632.
Barber, Brad M., Yi-Tsung Lee, Yu -Jane Liu, and Terrance Odean , 2014 , The cross -section of
speculator skill: Evidence from day trading. Journal of Financial Markets , 18, 1 -24.
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 Terrance Odean, 2000, Trading is hazardous to your wealth: The common
stock investment performance of individual investors, Journal of Finance 55, 773−806.
Barber, Brad M., and Terrance Odean, 2001, Boys will be boys: Gender, overconfidence, and
common stock investment, The Quarterly Journal of E conomics , 116(1), 261 -292.
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. 42
Barber, Brad M., Terrance Odean, and Ning Zhu, 2009, Do retail trades move markets? Review of
Financial Studies , 22, 151−186.
Barrot, Jean -Noel, Ron Kaniel , and Da vid Alexandre Sraer, 2016, Are retail traders
compensated for providing l iquidity? Journal of Financial Economics , 120, 146 -168.
Black, Fischer , 1986 , Noise, Journal of Finance , 41(3), 528-543.
Blume, Marshall E. and Robert F. Stambaugh, 1983, Biases in computed returns: An application
to the size effect, Journal of Financial Economics, 12, 387 -404.
Boehmer, Ekkehart, Charles M. Jones, Xiaoyan Zhang , and Xinran Zhang, 2021, Tracking retail
investor activity , Journal of Finance , 76(5), 2249 -2305 .
Carpenter, Jennifer N., Fangzhou Lu, and Robert F. Whitelaw, 2021, The real value of China’s
stock market. Journal of Financial Economics, 139(3), 679 -696.
Chiang, Y ao-Min, David Hirshleifer , Yiming Qian , and Ann E. Sherman , 2011, Do investors
learn from experience? Evidence from frequent IPO investors. The Review of Financial Studies ,
24(5), 1560 -1589.
Chen, Ting, Zhenyu Gao, Jibao He, Wenxi Jiang, and Wei Xiong, 2019, Daily price limits and
destructive market behavior , Journal of Econometrics , 208(1), 249 -264.
Chordia, Tarun, and Avanidhar Subrahmanyam, 2004, Order imbalance and stock returns: Theory
and evidence, Journal of Financial Economics , 72, 485−518.
Dorn, Daniel, Gur Huberman, and Paul Sengmueller, 2008, Correlated trading and returns. The
Journal of Finance , 63(2), 885 -920.
Eaton, Gregory W., T. Clifton Green, Brian Roseman, and Yanbin Wu, 2021, Zero -commission
individual investors, high frequency tr aders, and stock market quality . Working paper .
Fama, Eugene F., and James D. MacBeth, 1973, Risk, return, and equilibrium: Empirical tests,
Journal of Political Economy , 81, 607−636.
Fong, Kingsley YL, David R. Gallagher, and Adrian D. Lee, 2014, Individual investors and broker
types. Journal of Financial and Quantitative Analysis , 49(2), 431 -451. 43
Gao, Xiaohui, and Tse -Chun Lin , 2015, Do individual investors treat trading as a fun and
exciting gambling activity? Evidence from repeated natural experiments. The Review of Financial
Studies , 28(7), 2128 -2166.
Grinblatt, Mark, and Matti Keloharju , 2000, The investment behavior and performance of various
investor types: a study of Finland 's unique data set. Journal of F inancial E conomics , 55(1), 43 -67.
Grinblatt, Mark, Matti Keloharju, and Juhani T. Linnainmaa, 2012, IQ, trading behavior, and
performance. Journal of Financial Economics , 104(2), 339 -362.
Hu, Conghui, Yu -Jane Liu , and Xin Xu , 2021, The valuation effect of stock dividends or splits:
Evide nce from a catering perspective, Journal of Empirical Finance , 61, 163 -179.
Jiang, Lei, Jinyu Liu , Lin Peng, and Baolian Wang, 2019, Investor a ttention and a sset p ricing
anomalies , Working paper, Tsinghua University .
Kaniel, Ron, Saar Gideon, and Titman, Sheridan, 2008, Individual investor sentiment and stock
returns, Journal of Finance , 63, 273−310.
Kaniel, Ron, Liu, Shuming , Saar, Gideon, and Titman, Sheridan, 2012, Individual investor trading
and return patterns around earnings announcements, Journal of Finance , 67, 639 -680.
Kelley, Eric K. and Paul C. Tetlock, 2013, How w ise are crowds? Insights from r etail o rders and
stock r eturns , Journal of Finance , 68, 1229 -1265.
Leippold, M arkus, Qian Wang, and Wenyu Zho u, 2022 , Machine learning in the Chinese stock
market. Journal of Financial Economics , forthcoming.
Li, Xindan, Ziyang Geng, Avanidhar Subrahmanyam , Honghai Yu , 2017, Do wealthy investors
have an informational advantage? Evidence based on account classifications of individual investors ,
Journal of Empirical Finance 44, 1 -18.
Liao, J ingchi, Cameron Peng, and Ning Zhu, 2022 , Extrapolative bubbles and trading volume ,
Review of Financial Studies , , 35(4), 1682 -1722.
Linnainmaa, Juhani T ., 2010, Do limit orders alter inferences about investor performance and
behavior?. The Journal of Finance , 65(4), 1473 -1506. 44
Liu, Jianan, Robert F. Stambaugh, and Yu Yuan , 2019 , Size and value in China, Journal of
Financial Economics , 134(1), 48 -69.
Liu, Hongqi , Cameron Peng , Wei A. Xiong, and Wei Xiong , 2021, Taming the bias zoo , Journal
of Financial Economics , 143(2), 716 -741.
Newey, Whitney K., and Kenneth D. West, 1987, A simple, positive semi -definite,
heteroskedasticity and autocorrelation consistent covaria nce matrix, Econometrica , 55, 703−708.
Ozik, Gideon , Ronnie Sadka, and Siyi Shen , 2021, Flattening the illiquidity curve: Retail
trading during the CO VID-19 lockdown , Journal of Financial and Quantitative Analysis , 56(7),
2356 -2388.
Stoffman, Noah , 2014 , Who trades with whom? Individuals, institutions, and returns. Journal
of Financial Markets , 21, 50 -75.
Titman, Sherid an, Chishen Wei, and Bin Zhao, 2022, Corporate actions and the manipulation of
retail investors in Chin a: An analysis of stock splits. Journal of Financial Economics , forthcoming.
Welch, I vo, 2020, The wisdom of the Robinhood crowd , Journal of Finance , forthcoming.
45