Social Interaction and Investing:Evidence from an Online Social Trading Network∗

Manuel Ammann†, Nic Schaub‡

June 2, 2016

Preliminary first draft

Abstract

This paper investigates the role of social interaction in investment decisions of indi-viduals. We use unique data from an online social trading network that allows tradersto implement trading strategies that followers can directly invest in. We document thattraders with good past performance are more likely to talk about their implementedstrategies. There is strong evidence that traders’ communication impacts investment de-cisions of followers. However, comments do not contain informational value, suggestingthat followers’ investment behavior is mainly driven by sentiment. Finally, we show thatit is primarily small investors that rely on social interaction when making investmentdecisions.

JEL Classification: D14, G11, G23Keywords: social interaction, investor behavior, online communities, textual analysis

∗We are grateful to Marc Gerritzen, Florian Weigert, Felix von Meyerinck, conference participants at thethe 2016 Boulder Summer Conference on Consumer Financial Decision Making, and seminar participants atthe University of St. Gallen for helpful comments. We gratefully acknowledge data from Boerse Stuttgart.†Swiss Institute of Banking and Finance, University of St. Gallen, CH-9000 St. Gallen, Switzerland,

manuel.ammann@unisg.ch‡Swiss Institute of Banking and Finance, University of St. Gallen, CH-9000 St. Gallen, Switzerland,

nic.schaub@unisg.ch

1 Introduction

Shiller (2000, 2005) proposes that investor sentiment transmitted through social interaction

plays a crucial role in the formation of asset pricing bubbles.1 A growing number of empirical

studies finds that social interaction indeed affects investment decisions of individuals and

professional money managers.2 However, existing research typically does not observe the

communication between any two individuals or managers but relies on geographical proximity

to infer variation in the level of interaction between investors. Thus, previous research is

generally not able to determine whether correlated behavior of nearby investors is driven by

investors passing along real information about fundamentals to their neighbors or ‘irrationally

exuberant’ sentiment (Hong et al., 2005).

In this study, we use unique data from an online social trading network that allows us to

disentangle the two effects.3 Online social trading platforms provide an ideal setting for such

an investigation as they let traders implement trading strategies that followers can directly

invest in. In addition, these platforms allow traders to talk about their implemented trading

strategies. Thus, we observe the shared investment ideas, how traders communicate with

followers about these ideas, as well as the investment decisions of followers. We perform four

sets of novel tests: First, we analyze what makes traders chatting about their implemented

ideas. Second, we analyze whether the posting of comments by traders and the enthusiasm

transmitted through these comments affect investment decisions of followers. Third, to un-1Shiller (2005) says the following about social interaction and asset pricing bubbles: “I define a speculative

bubble as a situation in which news of price increases spurs investor enthusiasm, which spreads by psychologicalcontagion from person to person, in the process amplifying stories that might justify the price increases andbringing in a larger and larger class of investors, who, despite doubts about the real value of an investment,are drawn to it partly through envy of others’ successes and partly through a gambler’s excitement.” (p. 2)

2Using survey data, Shiller and Pound (1989) show that interpersonal communication is important forprofessional and individual investors’ decision making. Hong et al. (2005), Cohen et al. (2008), and Pool etal. (2015) provide further evidence that fund managers’ portfolio choices are influenced by word-of-mouthcommunication. Hong et al. (2004), Brown et al. (2008), and Kaustia and Knüpfer (2012) find that socialinteraction affects stock market participation of individual investors. Moreover, Feng and Seasholes (2004),Ivković and Weisbenner (2007), Heimer and Simon (2012), and Hvide and Östberg (2015) provide evidencethat social interaction matters for individual investors’ trading behavior.

3Online social trading networks are considered to be a subcategory of traditional online social networks.Social trading has experienced tremendous growth in the recent past. Hundreds of thousands of individualsare using these networks and assets under management of social trading platforms are reported to amount toseveral billion euros (see, e.g., “Social trading takes off for the masses”, The Financial Times, November 5,2014; “UK’s financial regulator warns on copy trading”, The Financial Times, March 10, 2014; “Social tradingtargets savvy retail investors”, The Financial Times, June 22, 2013).

1

derstand whether comments contain informational value or whether investments of followers

are mainly driven by sentiment, we investigate whether the posting of comments and the

tone of comments have predictive power for the future performance of investment strategies.

Finally, we shed light on the type of follower that trades on posted commentaries.

Our data come from a large European social trading platform and cover the time pe-

riod from January 2013 to December 2014. The sample contains more than 1,000 investible

strategies shared on this network. Traders managing these strategies post about 20,000 com-

mentaries on their profile pages. In addition, follower flows into and out of these investment

ideas amount to more than EUR 230 million over our sample period.

In the first part of the study, we analyze what makes traders posting comments and what

drives the tone of commentaries. We show that traders are more likely to communicate the

better their recent performance is. To extract the tone of posted commentaries, we follow

previous research and compute the fraction of positive and negative words in the text (e.g.,

Tetlock, 2007; Li, 2008; Kothari et al., 2009; Feldman et al., 2010; Loughran and McDonald,

2011; Engelberg et al., 2012; Larcker and Zakolykina, 2012; Chen et al., 2014; Hillert et al.,

2014; Huang et al., 2014). We find strong evidence that price increases in trading strategies

spur investor enthusiasm and significantly increase the fraction of positive words in comments.

There is also evidence that price decreases lead to more pessimistic commentaries. However,

this effect is much less pronounced than the relation between past returns and the fraction

of positive words in commentaries. Thus, the communication by traders is clearly biased

towards good outcomes.

Next, we investigate how the (biased) communication of traders affects the behavior of

followers. We document robust evidence that posting comments induces followers to trade.

Specifically, if a trader posts a comment today this increases tomorrow’s net investments of

followers by about 10%. In addition, we show that a one percentage point increase in the

fraction of positive (negative) words in a comment increases (decreases) net investments by

followers by about 2% (3%) on average. Thus, our analysis provides direct empirical evidence

that enthusiasm (and pessimism) spreads from person to person and affects individuals’ in-

vestment decisions. We run our analysis separately for inflows and outflows and show that

2

comments not only lead to higher inflows but also trigger higher outflows. Moreover, the flow

reaction to comments is not only observed on the day after the positing but lasts for about

three weeks. We run all our analyses in a within-trading-strategy setting by including strategy

fixed effects. Moreover, we account for potential time trends by adding day fixed effects to all

our regression specifications. We also control for variables that might influence the posting of

comments and at the same time investments of followers such as past performance and past

flows. Taken together, our results provide strong evidence that social interaction between

traders and followers matters for the investment decisions of followers.

In the third part of the paper, we investigate why followers trade on comments posted

by traders. We show that the posting of comments and the tone of comments do not have

predictive power for the future performance of trading strategies. Thus, our results support

the view that followers’ investment decisions are driven by sentiment rather than by a rational

reaction.

Finally, we shed light on the type of followers that listens to commentaries. We rely on

previous literature and use trade size to categorize followers (e.g., Lee and Radhakrishna,

2000; Malmendier and Shanthikumar, 2007; Mikhail et al., 2007). Large investors tend to be

more sophisticated market participants while small investors often show signs of naiveté. We

find a highly significant flow reaction to comments for small investors but not reaction for

large investors. This suggests that it is mainly unsophisticated individuals that rely on social

interaction when making investment decisions, but social interaction does not help them to

improve investment quality.

Our study contributes to three stands of the literature. First, existing literature shows

that social interaction matters for individuals’ investment decisions (e.g., Feng and Seasholes,

2004; Hong et al., 2004; Ivković and Weisbenner, 2007; Brown et al., 2008; Heimer and Simon,

2012; Kaustia and Knüpfer, 2012) as well as for mutual fund managers’ stock selection (e.g.,

Hong et al., 2005; Cohen et al., 2008). However, previous literature is typically not able to

distinguish whether correlated stock purchases in a peer group are driven by transmission

of information about stock fundamentals among the peer group or by investor sentiment as

the communication between individuals is not directly observed. Hvide and Östberg (2015)

3

and Pool et al. (2015) make a first step towards disentangling these two effects. Pool et

al. (2015) investigate the behavior of mutual fund managers living in the same neighborhood

and provide some evidence that socially connected fund managers share valuable information.

Hvide and Östberg (2015) focus on co-workers and find strong influence of co-workers on

investment decisions, but the effect on performance is rather negative than positive as there

is some evidence that social interaction propagates investment mistakes. To the best of our

knowledge, we are the first to observe the actual communication between investors and how

they react to this communication. Thus, we are able to clearly disentangle whether individual

investors’ reaction to social interaction is triggered by rationality or by sentiment. Our results

support the view that it is mainly sentiment that is transmitted through social interaction.4

Second, our study also contribute to the literature on the value of investment ideas shared

on the internet. These studies report mixed results. While some papers document that there

is a statistically and economically meaningful relation between opinions posted in online

networks and future returns (e.g., Chen et al., 2014; Avery et al., 2015), others find not

effect (e.g., Tumarkin and Whitelaw, 2001; Dewally, 2003; Antweiler and Frank, 2004; Das

and Chen, 2007). However, none of these existing studies observes how followers react to

shared investment ideas. Thus, our work adds to this strand of the literature by documenting

that investors follow the comments posted in online communities, even if there is not much

evidence that they contain valuable information.

Finally, our study also contributes to the literature on textual analysis in finance. A

growing body of research shows that the tone in corporate disclosures, media articles, analyst

reports, and internet postings predicts future stock price changes (Tetlock, 2007; Feldman et

al., 2010; Loughran and McDonald, 2011; Engelberg et al., 2012; Chen et al., 2014; Huang

et al., 2014). However, existing studies typically only indirectly observe investors’ reaction to

the tone of text by investigating price changes on financial markets. An exception is Hillert et4Our study also contributes to the literature on social interaction and investing by clearly disentangling

the effects of social interaction from community effects (e.g., Manski, 1993; Moffitt, 2001). Since prior workuses geographical proximity to infer variation in the level of social interaction between investors, these studiesface the problem that their results could also be driven by community effects rather than by social interactionbetween individuals. Our analysis is not subject to this problem as we do not rely on geographical proximitybut directly observe how traders communicate with followers.

4

al. (2014) who use textual analysis to show that mutual fund flows react to the tone in mutual

fund shareholder letters. Online social trading platforms also allows us to directly observe

the reaction of investors to the tone of comments. However, comments posted in an online

social trading network are fundamentally different from shareholder letters in that traders are

free in the wording used in these comments, while fund managers are legally constrained to

portray a fair and truthful picture of the current economic situation of a fund. Thus, we add

to this literature by providing direct evidence that the trading reaction of individual investors

depends on the specific opinion revealed in the information source in a setting without any

regulatory restrictions.

The remainder of the paper is organized as follows. In the next section, we explain the

concept of online social trading in greater detail and introduce our dataset from the online

social trading network. In Section 3, we first analyze what makes traders positing comments.

We then examine the role traders’ communication plays in the investment decisions of follow-

ers. Next, we shed light on the predictive power of comments for the future performance of

traders’ investment ideas. The last analysis in Section 3 looks at the type of followers that

listens to comments of traders. Section 4 concludes.

2 Data and variables

2.1 Online social trading networks

Online social trading networks are considered a subcategory of classical online social networks.

The first online social trading platforms were founded in the early 2000s. However, online

social trading only started to take off more recently. As of December 2014, there were more

than 30 networks worldwide that offered similar services. eToro claims to be the largest one

with over 3 million users as of 2014. Moreover, investments by followers in trading ideas

shared on these platforms are estimated to amount to several billion euros.

Registering with social trading platforms is usually free, both for traders and followers.

Traders either create a virtual portfolio or alternatively they can set up a real money account

on the platform. Some platforms also offer traders the possibility to link their third-party

5

brokerage accounts to the platform. The investment universe traders can choose from is pre-

defined by the provider and typically consists of a wide range of stocks, funds, and currencies.

Some platforms also allow traders to invest in bonds, derivatives, or commodities. In addi-

tion, traders set up profile pages on which they disclose their identity, describe themselves

and their trading strategy. The profile page also shows the trader’s current holdings, the

trading history, and the past portfolio performance. Moreover, one typically sees the number

of followers of a trading strategy or the amount of money invested by followers. Traders

communicate with followers by either posting comments on their profile pages or by sending

personal messages. Figure 1 provides a sample profile page of a trader.

Followers can scan traders’ profile pages and once they have discover a convincing in-

vestment idea, they can follow it and replicate investment decisions of traders in their own

accounts in real time. Thus, even if investors do not transfer capital to the traders’ accounts,

traders de facto manage the assets of followers. Accounts of followers are either located on

the platform or some platforms enable followers to link their third-party brokerage accounts

to the platform. While some social trading networks allow investors to copy single transac-

tions of traders, the predominant form of social trading is the replication of entire investment

strategies.

Platforms typically earn bid-ask spreads when both traders and followers turn over their

portfolios. Moreover, they usually charge a base fee on the amount of money invested by

followers as well as a performance fee that depends on the amount of money invested by

followers and the performance of a trading strategy above a high-water mark. The high-

water mark ensures that there is no performance fee until the trading strategy has recovered

past losses. Traders are typically remunerated for sharing their investment ideas by receiving

a fraction of either the base fee or the performance fee or both. The basis for actual returns

received by followers is the net returns after bid-ask spreads paid by traders and after fees.

6

2.2 Data and sample selection

Our data come from a leading European online social trading platform that has all the typical

features of an online social trading network. The investigation period starts in January 2013

and ends in December 2014 and is determined by the availability of the data of the platform.

Our initial dataset contains all trading strategies that were investible at some point during

our investigation period. As we are interested in the investment behavior of followers, we

focus on those strategies that attract investments from followers at least once during our

sample period. In order to not introduce any bias, we delete all entries of a strategy before

the first investment by a follower. Moreover, to ensure a sufficient number of observations

for measuring risk-adjusted returns, we require a trading strategy to have at least one month

of daily observations to be included in our sample. This leaves us with a sample of 1,040

portfolios and 247,048 daily observations. As of December 2014, 990 strategies are still alive

and 50 strategies are defunct. In January 2013, our sample consists of 113 portfolios. Thus,

the platform has experienced strong growth over our investigation period. The amount of

money invested by followers increases from EUR 6.2 million in January 2013 to EUR 52.3

million in December 2014. Panels A and B of Figure 2 graphically illustrate the growth in

the number of portfolios as well as the growth in the amount of money invested by followers

over our sample period.

2.3 Comment characteristics

On our platform, traders can communicate with investors by positing comments on their

profile pages. The only way for followers to gather information about a trading strategy

is to visit traders’ profile pages, where they immediately see the posted commentaries. To

understand the nature of the commentaries, it is helpful to look at examples. Comments

are either security-specific or they are of general nature. Figure 3 provides six fairly typical

examples of comments published in the network, of which three are security-specific comments

and the other three are general comments. The text of comments is typically either backward

looking and provides an explanation for the past performance of the trading strategy or it

7

contains a predicted price change and some explanations for the prediction. Most of the

explanations are rather short. We drop comments of less than five words as these comments

tend not to be informative. Our final sample consists of 19,851 commentaries. The average

(median) length of a commentary is 40 (22) words. This is consistent with Anweiler and

Frank (2004) who report that the number of words in messages posted on Yahoo! Finance

and Raging Bull is typically between 20 and 50.

Our main independent variables of interest are variables that capture different aspects

of the communication by traders. We create an indicator variable that equals one on days

on which traders post at least one comment on their profile pages, and zero otherwise. We

also count the number of comments posted on the days on which traders communicate. To

extract traders’ opinions from commentaries, we build on prior research, which suggests that

the frequency of negative and positive words used in a text measures the tone of the text (e.g.,

Tetlock, 2007; Li, 2008; Kothari et al., 2009; Feldman et al., 2010; Loughran and McDonald,

2011; Engelberg et al., 2012; Larcker and Zakolykina, 2012; Chen et al., 2014; Hillert et

al., 2014; Huang et al., 2014). As most of our comments are not in English, we cannot use

the widely used Loughran and McDonald (2011) word list that was specifically designed for

financial matters.5 Thus, we follow the approach in Hillert et al. (2014) and design our own

dictionary of positive and negative words. We randomly select 500 commentaries from our

sample. We then have two individuals looking at these comments independently of each other

and marking all words that sound positive and all words that sound negative to them. A

word is included in our word list if both individuals have marked a word as positive or if

both have marked it as negative. This results in a dictionary of 129 positive words and 134

negative words. Table A1 in Appendix B reports the 25 positive and negative words of our

word list that appear most frequently in commentaries. Among the most frequently used

positive words are ‘good’, ‘profit’, and ‘positive’. The most frequently used negative words

include ‘unfortunately’, ‘correction’, and ‘losses’. In the sample commentaries provided in

Figure 2, we underline words that are included in our dictionary. The tone measure is then5The Loughran and McDonald (2011) dictionary is constructed based on textual analysis of 10-K filings.

Traders on our platform certainly use a different writing style than that used in 10-K filings. Thus, it is notentirely clear whether the word list of Loughran and McDonald would be perfectly suitable in our setting.

8

constructed as the sum of the number of positive (negative) words in a comment divided by

the sum of the total number of words in the comment.6

Panel A of Table 1 provides descriptive statistics on our communication variables. Traders

post comments on 5.3% of all days. Moreover, on those days on which they communicate, they

post on average 1.5 commentaries. Using our own word list, the average fraction of positive

and negative words used in commentaries is 1.7% and 0.7%, respectively. The fact that the

fraction of positive words is much higher than the fraction of negative words provides first

suggestive evidence that the communication by traders is biased toward positive outcomes.7

Using the dictionary of Loughran and McDonald (2011), Chen et al. (2014) report that the

average fraction of negative words in comments posted on Seeking Alpha is 1.8%.8

2.4 Flow characteristics

Our platform allows followers to link their third-party brokerage account to the platform.

The amount followers allocate to a trading strategy then tracks the net performance of that

strategy. In total, our sample includes 43,275 transactions executed by followers in the time

period from January 2013 to December 2014, of which 28,473 are investments and 14,800 are

withdrawals.

Our main dependent variable of interest is a variable that captures investments of followers.

On each day, we compute the net flows in a trading strategy. To make our data more

normally distributed, we follow previous research and make use of the inverse hyperbolic sine

transformation (e.g., Burbidge et al., 1988; Kale et al., 2009; Karlan et al, 2016). Taking the

inverse hyperbolic sine is an alternative to a log-transformation when a variable takes on zero6We rerun all our analyses using two alternative tone measures. First, we make use of a word list that is

based on the Harvard IV-4 dictionary (Remus et al., 2010). This word list consists of 15,536 positive wordsand 14,800 negative words. Second, we also use a dictionary provided by the Linguistic Inquiry and WordCount (LIWC) package (Wolf et al., 2008). This LIWC dictionary is composed of 645 positive emotion wordsand 1,047 negative emotion words. Both the Harvard IV-4 dictionary and the LIWC dictionary are generallanguage dictionaries rather than dictionaries specifically designed for financial matters. However, both havebeen used in a finance context in previous research (e.g., Tetlock, 2007; Li, 2008; Kothari et al., 2009; Feldmanet al., 2010; Engelberg et al., 2012; Larcker and Zakolykina, 2012). In the main part of our paper, our resultsare similar albeit weaker when we use the two alternative dictionaries.

7Using the Harvard IV-4 (LIWC) dictionary, our comments contain 4.8% (3.0%) positive words and 1.6%(1.0%) negative words on average.

8Chen et al. (2014) do not report the fraction of positive words in commentaries.

9

or negative values.9

Panel B of Table 1 provides descriptive statistics on investments by followers. The average

(median) trade size is EUR 5,377 (EUR 2,251). With an average trade size of EUR 6,150

withdrawals tend to be larger than investments (EUR 4,974). The size of transactions by

followers suggests that social trading platforms mainly attracts individual investors. There

are a handful of very large investments of up to EUR 300,000, indicating that there are a few

institutional players or very wealthy individuals investing in social trading strategies. Net

flows are positive on average and amount to EUR 205. This is not surprising given the strong

growth in assets under management over our sample period.

2.5 Performance characteristics

To measure the performance of social trading strategies, we compute daily raw returns and

daily alphas. As traders invest 92% of their portfolios in equities, we employ a standard

equity asset pricing model to determine abnormal returns of trading strategies. The model

contains an equity market factor, the investment style factors of Fama and French (1993)

and Carhart (1997), and the two at-the-money option factors of Agarwal and Naik (2004).

Option factors can account for the non-linear payoff profiles that result from the traders’

use of derivative instruments. We construct factors using MSCI indices as these indices are

investible for retail clients. Since two-thirds of the stock holdings in portfolios are invested

in European stocks, we use European MSCI indices. We employ the MSCI Europe Index

as proxy for the market. The size factor (SMB) is approximated by the difference in daily

returns between the MSCI Europe Small Cap Index and the MSCI Europe Index. The value

factor (HML) is approximated by the return difference between the MSCI Europe Value Index

and the MSCI Europe Growth Index. Moreover, we use the MSCI Europe Momentum Index

as a proxy for the momentum factor. The two option factors are constructed as in Agarwal

and Naik (2004) using at-the-money European call and put options on the Euro Stoxx 50.

The risk-free rate is captured by daily returns on the J.P. Morgan 3 Month Euro Cash Index.9Alternatively, we could transform all observations by adding a constant equal to the absolute value of the

minimum net flow to each observation. For this transformation, our results are qualitatively similar to thereported results.

10

Data on indices and options are obtained from Thomson Reuters Datastream. To determine

daily alphas of a trading strategy, we estimate factor exposures over 6-month rolling windows

from t-126 to t-1. Alphas are then calculated as the difference between daily raw returns of

trading strategies and returns predicted by the estimated factor loadings.10

Panel C of Table 1 provides information on the return and alpha distribution in our

sample. Gross returns are net of bid-ask spreads paid by traders but before the base fee and

the performance fee. The average (median) gross return of trading strategies is close to zero.

The average (median) net return amounts to -2.7% p.a. (-0.7% p.a.), which is the basis for

actual returns received by followers. In practice, actual returns of investors might be even

lower due to differences between bid and ask prices of strategies. The difference between gross

and net returns of 1-2% p.a. reflects the average fee earned by the platform and the traders.

The average (median) alpha is -7.7% p.a. (-2.0% p.a.) before fees and -9.7% p.a. (-3.4% p.a.)

after fees. Thus, trading strategies in our sample tend to underperform common benchmarks

and followers would have been better off investing directly in the benchmark indices.

2.6 Portfolio and trader characteristics

We include a large number of portfolio and trader characteristics as controls in our multivari-

ate analyses. In Panel D of Table 1, we present summary statistics on portfolio characteristics.

To determine assets under management of trading strategies, we sum up net flows over time.

On average, followers have invested EUR 64,500 in the strategies of traders over our inves-

tigation period. There is substantial variation in assets under management. The trader of

the most popular trading strategy manages over EUR 8 million at some point in time. We

define the age of a trading strategy as the number of calendar days since its inception. On

average, trading strategies are 431 days old in our sample, which is equivalent to about 1.2

years. The performance fee is the percentage of daily profits over some high-water mark that

goes in similar proportions to the platform and the trader. At inception, the trader chooses a

performance fee between 5% and 30%. The average performance fee in our sample amounts10In our robustness tests, we also use alphas from a simple CAPM rather than alphas from the 6-factor

model. This does not alter our findings.

11

to 12.6%. Moreover, we create a dummy variable that equals one for traders who have their

own money invested in their trading strategy, and zero otherwise. Of all trading strategies in

our sample, 24.1% are flagged as strategies with managerial ownership.

On our platform, traders set up virtual portfolios that followers (and traders themselves)

can then invest in. The investment universe traders can pick securities from consists of stocks,

funds, and derivatives. The average portfolio contains 12 different securities. More than half

of the average portfolio is invested in stocks, of which approximately two-thirds are held in

European stocks and one-third in non-European stocks. About one-fifth of the mean portfolio

is held in mutual funds and exchange-traded funds, thereof more than 70% are equity funds.

Only about 1% of portfolio holdings are allocated to derivative instruments. Moreover, traders

hold a substantial fraction of 21.6% of their portfolios in cash.11 The remaining 2.7% of the

average portfolio are held in securities we cannot identify. Figure 4 graphically illustrates

that the average asset allocation of traders is fairly constant over time. On average, traders

place about 1.4 trades per day. Daily turnover is defined as the average of the value of all

purchases and the value of all sales executed on a specific day divided by the value of the

trader’s portfolio at the beginning of the day. The mean daily turnover is 4.6%, implying that

traders on this platform tend to trade excessively. For comparison, in the discount brokerage

dataset of Barber and Odean (2000), the average household turns over 75% of its portfolio

annually.

Finally, in Panel E of Table 1 we report trader characteristics. The 1,040 trading strategies

in our sample are managed by 740 different traders. Only 1.0% of all trading strategies are

managed by professional money managers. Thus, users sharing investment ideas on this

network are mainly individuals. More than 97% of traders claim to have several years of

trading experience. Only 3.5% of all traders are women. Finally, traders manage about 2.5

portfolios on average over our investigation period. This also includes portfolios that are

dropped in our sample selection procedure. Appendix A provides detailed descriptions of all

variables use throughout the study.11Traders are generally not able to use leverage. However, the deduction of fees from traders’ accounts might

result in negative cash positions.

12

3 Empirical analysis

Our unique dataset allows us to perform four sets of novel tests. First, we analyze the

determinants of traders’ communication (Section 3.1). Second, we investigate how followers

react to comments posted by traders (Section 3.2). We then go on to examine why followers

listen to the comments of traders (Section 3.3). Finally, we analyze which followers tend to

trade on comments (Section 3.4).

3.1 Determinants of comments

We begin our empirical investigation by looking at what makes traders posting comments

and what determines the tone of comments. To do so, we conduct multivariate analyses and

use the different comment characteristics as dependent variables and relate them to flow,

performance, portfolio, and trader characteristics. All time-varying explanatory variables are

lagged by one day to address potential reverse causality concerns. We include day fixed effects

in all regressions to control for the overall market environment. Standard errors are clustered

at the strategy level.

Results on the determinants of comments are presented in Table 2. In Column 1, the

dependent variable is the indicator variable that is equal to one on days on which traders

post at least one comment, and zero otherwise. Thus, we run a logit regression and report

marginal effects. The coefficient on lagged daily net returns is positive and highly statisti-

cally significant, implying that traders are more likely to post comments when their strategies

performed well. The magnitude of the coefficient suggests that a one standard deviation in-

crease in daily net returns increases the probability of posting a commentary by 4.9%. We

also find a positive impact of investments of followers on the likelihood of observing a trader

posting a comment, even after controlling for past performance. Consistently, the positive

and significant coefficient on the variable capturing assets under management indicates that

traders of more popular strategies tend to post more comments. Age is significantly neg-

atively correlated with the comment dummy. This implies that traders are more likely to

communicate after they have set up a strategy and the communication frequency decreases

13

over time. Moreover, the positive and significant coefficient on the managerial ownership

dummy indicates that those traders who have their own money at stake tend to communicate

more actively. The performance fee and all portfolio and trader characteristics do not have a

significant effect on whether or not a trader posts a commentary.

In Column 2, we focus on days with at least one posted comment and use the number

of posted comments per day as the dependent variable. Thus, we run an OLS regression

rather than a logit regression. The coefficients on past short-term performance and recent

net flows are again positive and statistically significant at least at the 5% level. Hence,

past performance and past flows not only affect whether traders post comments but they

also impact the number of posts by traders on days with at least one comment. Consistent

with our findings in Column 1, the coefficient on age is negative and significant. Moreover,

we document that individuals who trade more excessively are more active communicators.

There is also evidence that professionals and females post fewer commentaries, while traders

who claim to be experienced post more frequently.

In Columns 3 and 4, we use the fraction of positive words and the fraction of negative

words as dependent variables, respectively. We again focus on days with at least one posted

commentary. In Column 3, the past performance of trading strategies has a strongly positive

effect on the fraction of positive words in commentaries. Thus, if trading strategies perform

well, traders tend to be enthusiastic about their investment ideas and share their enthusiasm

with followers. A one standard deviation increase in daily net returns increases the fraction of

positive words by 9.3%. When focusing on the fraction of negative words in Column 4, past

performance also matters, but the effect is both statistically and economically substantially

weaker compared to the results in Column 3. The coefficient is only statistically significant at

the 10% level and a one standard deviation increase in net returns results in a 4.9% increase

in the fraction of negative words. Thus, the effect of past returns on negativity is only about

half the size of the effect we document for positivity. This again provides evidence that traders

are more reluctant to talk about failure than too talk about success.

Looking at the other explanatory variables in Columns 3 and 4, we find that professional

asset managers use significantly more negative words in their comments. This suggests that

14

their communication is less biased. Those who claim to be experienced traders communicate

much more enthusiastically. Finally, the negative coefficients on the female dummy in both

columns indicate that female traders’ communication tends to be more neutral than the

communication of male traders.

In summary, we find strong evidence that the communication of traders is biased towards

good outcomes. This is consistent with the self-enhancing transmission bias described by Han

and Hirshleifer (2015), which states that investors like to recount to others their investment

victories more than their defeats.

3.2 Comments and investments of followers

Next, we analyze whether followers base their investment decisions on the (biased) com-

mentaries made by traders or whether they ignore traders’ communication. The only way

followers can gather information about a trading strategy is by visiting profile pages, where

they immediately see traders’ commentaries. Thus, it is highly likely that followers frequently

read the comments posted by traders. To analyze whether commentaries affect followers’ in-

vestment decisions, we run panel regressions and regress today’s net investments by followers

on yesterday’s comment characteristics. We use the inverse hyperbolic sine of daily net flows

as dependent variable. To control for the impact of past performance on investment decisions

of followers, we include five variables capturing the net returns of trading strategies on each

of the past five trading days.12 To control for potential effects of past net flows on future

net flows, we add five variables measuring the net investments of followers over the past five

trading days. Furthermore, we include the natural logarithm of assets under management,

the natural logarithm of age, the managerial ownership dummy, the number of securities held

in the portfolio, and the portfolio turnover as additional control variables. We do not report

coefficient estimates of control variables for space reasons. All the regressions are estimated12We use net returns to control for past performance rather than gross returns or alphas since followers only

see net returns on traders’ profile pages. Calculating gross returns requires adding back fees to net returns,which is a non-trivial task as performance fees are only paid if a high-water mark is surpassed. It is also ratherunlikely that followers regularly compute alphas when making their investment decisions. Nevertheless, asadditional robustness tests, we rerun the analysis using past gross returns and past alphas to control for pastperformance (not reported). This does not materially change our findings.

15

with portfolio fixed effects. Portfolio fixed effects account for all portfolio and trader char-

acteristics that remain constant over time (e.g., performance fee, gender). Therefore, any

effect of traders’ communication on followers’ investments we find is purely driven by within-

trading-strategy variation. To control for the general market environment, we include day

fixed effects in all our regressions. Standard errors are clustered at the trading strategy level.

Results are reported in Table 3. In Column 1, we use the dummy variable that equals

one on days on which traders post at least one comment as our main explanatory variable.

We find that whether traders post a commentary is positively related to net investments of

followers. The effect is statistically significant (at the 5% level) and the impact of comments

on investments by followers is also economically meaningful. Posting a comment increases

investment of followers by 10.7% on average. Thus, there is strong evidence that the soft

information transmitted through comments matters for investment decisions of followers.

To dig deeper, we use the lagged number of posted comments as explanatory variable in

Column 2 and focus on days with at least one comment posted on the preceding day. The

positive and significant coefficient on the lagged number of comments suggests that it not

only matters whether traders post a commentary but the more they post, the more inflows

they attract.

Finally, in Columns 3 and 4 of Table 3, we investigate the role of comment tone. We

show that the fraction of positive words is positively correlated with net investments and the

relation between the fraction of negative words and net flows is negative. Both coefficient

estimates are statistically significant at the 5% level. However, in economic terms, the effect of

negative words is about 50% larger than the effect of positive words. Increasing the fraction of

positive words in a comment by one percentage point increases net flows by 2.1% on average,

while increasing the faction of negative words by one percentage point reduces net flows by

3.2% on average. In our descriptive statistics, we saw that traders are somewhat reluctant to

use negative words in their communication. They employ about twice as much positive words

than negative words. Most likely because of the conservative use of negative words, followers

are more sensitive to negativity than to positivity.

Our results cannot be driven by reverse causality as we use lagged comment characteris-

16

tics in all regressions. However, there might be an omitted variable problem in our analysis.

It could be the case that traders post comments following firm-specific news announcements

and followers might know the securities traders hold in their portfolios. Therefore, followers’

reaction to the posting of commentaries could be a reaction to corporate news announce-

ments rather than a reaction to the posted comments. This appears rather implausible as

traders hold on average 12 securities in their portfolios and frequently change the portfolio

composition and it is therefore rather unlikely that followers are always up-to-date about the

portfolio components. However, to still address this concern, we make use of an additional

feature of our data. When positing a commentary, traders have to classify commentaries as

either security-specific or general commentaries. Figure 2 provides examples for both types

of comments. While the former might be affected by corporate news announcements, the

latter tend to talk about the trading strategies more broadly. In our sample, 50.5% of all

comments are security-specific comments and the remaining 49.5% are general comments.

We re-estimate all the regressions from Table 3 separately for security-specific comments and

general comments.

Results of this robustness test are presented in Table A2 in Appendix B. When focusing

on security-specific comments in Columns 1 to 4, none of our comment characteristics shows

a statistically significant effect on investments by followers. Instead, our results are primarily

driven by general commentaries. In Columns 5 to 8, all coefficient estimates are statically

significant except for the coefficient on the lagged number of comments. This suggests that our

findings are not caused by firm-specific news announcements that drive investors’ attention

towards certain trading strategies.

In an additional robustness test, to address potential concerns that our own dictionary is

too subjective, we rerun the regressions from Columns 3 and 4 of Table 3 using the Harvard

IV-4 dictionary and the LIWC dictionary rather than our own word list to identify positive

and negative words in comments. Results are reported in Table A3 in Appendix B. When

relying on the word list based on the Harvard IV-4 dictionary in Columns 1 and 2, we find

a positive and significant effect for the fraction of positive words on net investments by

followers but no effect for the fraction of negative words. When using the LIWC dictionary

17

in Columns 3 and 4, the fraction of positive words has no significant effect on net flows,

while the relation between yesterday’s fraction of negative words and today’s net flows is

negative and statistically significant. It is not surprising that results are weaker with these

two alternative dictionaries as both were designed to capture tone in a general context rather

than in a financial context. This supports the argument of Loughran and McDonald (2011)

that specifically-designed dictionaries are more appropriate in a financial context.

So far, we focused on net investments of followers. The documented effect of commentaries

on net investments could be driven by commentaries increasing inflows or by traders being

able to prevent followers from withdrawing their invested capital. To disentangle these two

effects, we re-estimate all specifications from Table 3 separately for inflows and outflows.

Results are presented in Table 4. In Column 1, we find that commentaries have a positive

and highly statistically significant effect on inflows (t-statistic of 5.34). Interestingly, we also

document a positive relationship between comments and outflows in Column 2, suggesting

that comments not only attract inflows but also motivate some followers to redeem their

money. The net effect is positive as posting a commentary increases inflows by 18.8%, while

it increases outflows by only 6.0% on average. Results in Columns 3 and 4 suggest that the

number of comments on days with at least one comment only affects inflows but has not

impact on outflows. In Columns 5 and 6, we analyze the relation between positivity, inflows,

and outflows. The coefficient estimate is positive for inflows and negative for outflows, that is,

the higher the fraction of positive words in a comment, the higher the inflows and the lower

the outflows. This implies that a comment with a more optimistic tone can prevent existing

investors from redeeming their invested capital. Similarly, in Columns 7 and 8, we document

that more pessimistic comments lead to lower inflows and higher outflows. Hence, negativity

in comments provides an explanation why comments can also lead to higher outflows.

Finally, in our last analysis in this subsection, we examine how the communication of

traders affects investments of followers in the longer run. Commentaries remain on the profile

18

pages of traders after they were posted.13 Thus, it might well be that they do not only attract

investments immediately after they were published but also in the longer run. Alternatively,

it could be the case that the comment-driven flow effect reverses after some time. To shed

light on the longer-term effects of traders’ communication, we turn to a longer-term analysis

over several weeks. More specifically, for the two months after the positing of a comment, we

compute average daily net flows for each week and rerun our main regression from Column 1

of Table 3.

Results are reported in Table 5. In Columns 1 to 3, we show that there is a strong

comment-driven flow reaction over the first three weeks after the posing of a comment. Net

flows are 23.7% higher in week 1, 13.3% higher in week 2, and 14.3% higher in week 3 relative

to the average daily net flows of the same strategy. The effect is always statistically significant

at least at the 5% level. After week 3, the flow reaction turns statistically and economically

insignificant. In weeks 7 and 8, the coefficient estimates on the comment dummy turn from

positive to negative but they still lack statistically significance. Thus, there are only some

very weak sings of a reversal over the longer run, suggesting that the flow reaction is relatively

persistent.

In summary, we show that social interaction affects followers’ behavior, both investments

and withdrawals. There is not only an immediate comment-driven flow reaction but the effect

lasts for several weeks.

3.3 The predictive power of comments for future performance

In this section, we investigate why investors listen the comments of traders. The flow reaction

to comments could be rational if comments contain informational value and help followers

to identify strategies that deliver superior performance. Alternatively, if comments do not

contain information about the future price development of trading strategies, followers’ seem

to be fooled by sentiment transmitted through commentaries. To differentiate between these13While traders on this platform cannot alter published comments, they can delete comments from their

profile pages. To investigate whether traders frequently drop comments, we randomly select 50 trading strate-gies and download published comments at two points in time with a gap of approximately six months. Wedo not find a single trader that has deleted commentaries from the profile page. Thus, deletion of selectedcomments ex post should not be a concern in our analysis.

19

two alternative explanations, we examine whether commentaries posted by traders have pre-

dictive power for the future performance of their trading strategies. To do so, we regress

daily returns and daily alphas from our 6-factor model on our four comment characteristics.

Comment characteristics are lagged by one trading day. The same set of control variables as

in Table 3 is included in every regression but not reported. We again also include portfolio

fixed effects and day fixed effects in all specifications. Standard errors are clustered at the

trading strategy level.

Results are presented in Table 6. In Columns 1, 3, 5, and 7, the daily net returns of

trading strategies serve as dependent variables and in Columns 2, 4, 6, and 8, we use daily

net alphas as dependent variables. The coefficient estimates on the comment characteristics

are mostly statistically insignificant. There are only two specifications with raw returns as

dependent variables in which coefficient estimates are weakly significant. In Column 3, the

coefficient on the number of comments is negative and significant at the 10% level and in

Column 7, the coefficient on the fraction of negative words is negative and significant at the

5% level. However, once we focus on alphas and control for common risk factors statistical

significance vanishes completely.

In Table A4 in Appendix B, we replicate the analysis from Table 6 using CAPM alphas

rather than 6-factor model alphas. We find our results to be virtually unchanged. This

suggests that the weak predictive power of some comment characteristics for future raw returns

is entirely driven by these characteristics being correlated with the market risk exposure.

Finally, we assess the robustness of our results by looking at a longer investment horizon.

Even though portfolio turnover is high, which suggests that traders tend to have a rather

short investment horizon, it could still be the case that they take up a longer-term stance in

their commentaries. Hence, we use 1-month cumulative raw returns and 1-month cumulative

abnormal returns (CARs) as performance measures rather than the 1-day returns and the 1-

day alphas and re-estimate the specifications from Table 6. Results are reported in Table A5

in Appendix B. We document that comment characteristics also do not have any predictive

power for the performance over the following four weeks.

Overall, these findings show that comments influence investment behavior of followers

20

despite the fact that they do not contain any information about the future price development

of the trading strategies. Thus, followers listening to comments suffer from the rather poor

performance delivered by the trading strategies on this platform. This suggests that invest-

ment decisions of followers are primarily driven by sentiment transmitted through traders’

communication rather than by rationality.

3.4 Who listens to comments?

In this last section, we investigate which investors (foolishly) follow the sentiment-driven com-

ments of traders. As we do not observe any follower characteristics directly, we use trade size

to categorize investors. Previous research shows that large trades tend to be executed by

more sophisticated (professional) investors, while small trades tend to come from unsophis-

ticated individuals (e.g., Lee and Radhakrishna, 2000; Malmendier and Shanthikumar, 2007;

Mikhail et al., 2007). Our results so far indicate that followers are not guided by rationality

but rather by sentiment. Thus, we hypothesis that our findings are mainly driven by unso-

phisticated individuals executing small transactions. In our descriptive statistics, we saw that

the average trade size in our sample is EUR 5,377. Thus, we use EUR 5,000 as the cutoff

point and classify transactions below this threshold as small trades and transactions above

the threshold as large trades. Small trades account for 74.0% of all transactions but only

for 25.7% of the overall transaction volume. In contrast, large trades make up only 26.0% of

all transactions but 74.3% of the entire trading volume in our sample period. To investigate

whether results differ for small and large investors, we rerun our main analysis from Column

1 of Table 3 separately for small and large trades.

Results of this analysis are presented in Table 7. In Column 1, we focus on small trades

and in Column 2 on large trades. Even though small trades only account for about one

quarter of the overall flows into and out of trading strategies, we find a very strong comment-

driven flow effect in Column 1. In contrast, the coefficient estimate on the comment dummy

in Column 2 is neither statistically nor economically meaningful. Thus, consistent with our

findings in the previous section that comments do not contain valuable information, we show

that it is mainly unsophisticated investors that listen to these comments.

21

4 Conclusion

This paper investigates the role of social interaction in investment decisions of individuals. We

use unique data from an online social trading platform that provides an ideal setting for such

an investigation. In this data, we observe the investment strategies implemented by traders,

how traders talk about these ideas, as well as followers’ investments in this trading strategies.

Thereby, we overcome shortcomings of previous studies on the impact of social interaction on

investment behavior that typically rely on geographic proximity to infer variation in the level

of social interaction.

In the first part of our analysis, we document that traders with good short-term per-

formance are more likely to post comments on their profile pages and they communicate

more enthusiastically. Thus, traders’ communication tends to be biased towards good out-

comes. Even though comments tend to be biased, we find strong evidence that the posting

of comments and comment tone impact investment decisions of followers. On average, net

investments of followers increase by about 10% on the day after the posting of a commentary.

However, positing commentaries is not a universal remedy as we show that some commentaries

can also motivate followers to withdraw their funds. We document that there is not only an

immediate comment-driven flow reaction but the effect lasts for about three weeks after the

posting of a commentary. We then show that comments do not have predicative power for

the future performance of trading strategies. Thus, in our setting, social interaction does not

trigger a rational reaction but rather influences sentiment among followers. Consistent with

this finding, we show that our results are mainly driven by small investors that are typically

considered to be unsophisticated individuals. Taken together, this paper suggests that it is

primarily unsophisticated individuals that rely on social interaction when making investment

decisions, but there is not much evidence that social interaction helps these unsophisticated

market participants to improve their investment quality.

22

References

Agarwal, V., and N.Y. Naik, 2004, Risk and portfolio decisions involving hedge funds, Review

of Financial Studies 17, 63-98.

Antweiler, W., and M.Z. Frank, 2004, Is all that talk just noise? The information content of

Internet stock message boards, Journal of Finance 59, 1259-1294.

Avery, C.N., J.A. Chevalier, and R.J. Zeckhauser, 2015, The "CAPS" prediction system and

stock market returns, Review of Finance, forthcoming.

Barber, B.M., and T. Odean, 2000, Trading is hazardous to your wealth: The common stock

investment performance of individual investors, Journal of Finance 55, 773-806.

Brown, J., Z. Ivković, P. Smith, and S. Weisbenner, 2008, Neighbors matter: Causal commu-

nity effects and stock market participation, Journal of Finance 63, 1509-1531.

Burbidge, J.B., L. Magee, and A.L. Robb, 1988, Alternative transformations to handle ex-

treme values of the dependent variable, Journal of the American Statistical Association 83,

123-127.

Carhart, M.M., 1997, On persistence in mutual fund performance, Journal of Finance 52,

57-82.

Chen, H., D. Prabuddha, Y.J. Hu, and B.-H. Hwang, 2014, Wisdom of crowds: The value of

stock opinions transmitted through social media, Review of Financial Studies 27, 1367-1403.

Cohen, L., A. Frazzini, and C. Malloy, 2008, The small world of investing: Board connections

and mutual fund returns, Journal of Political Economy 116, 951-979.

Das, S.R., and M.Y. Chen, 2007, Yahoo! for Amazon: Sentiment extraction from small talk

on the web, Management Science 53, 1375-1388.

Dewally, M., 2003, Internet investment advice: Investing with a rock of salt, Financial Analyst

Journal 59, 65-77.

Engelberg, J., A.V. Reed, M.C. Ringgenber, 2012, How are shorts informed? Short sellers,

news, and information processing, Journal of Financial Economics 105, 260-278.

Fama, E.F., and K.R. French, 1993, Common risk factors in the returns on stocks and bonds,

23

Journal of Financial Economics 33, 3–56.

Feldman, R., S. Govindaraj, J. Livnat, and B. Segal, 2010, Management’s tone change, post

earnings announcement drift and accruals, Review of Accounting Studies 15, 915-953.

Feng, L., and M. Seasholes, 2004, Correlated trading and location, Journal of Finance 59,

2117-2144.

Han, B., and D. Hirshleifer, 2015, Self-enhancing transmission bias and active investing,

Working Paper, Rotman School of Management.

Heimer, R.Z., and D. Simon, 2012, Facebook finance: How social interaction propagates active

investing, Working Paper, Brandeis University.

Hillert, A., A. Niessen-Ruenzi, and S. Ruenzi, 2014, Mutual fund shareholder letter tone –

Do investors listen?, Working Paper, University of Mannheim.

Hong, H., J.D. Kubik, and J.C. Stein, 2004, Social interaction and stock-market participation,

Journal of Finance 59, 137-163.

Hong, H., J.D. Kubik, and J.C. Stein, 2005, Thy neighbor’s portfolio: Word-of-mouth effects

in the holdings and trades of money managers, Journal of Finance 60, 137-163.

Huang, A.H., A.Y. Zang, and R. Zheng, 2014, Evidence on the information content of text

in analyst reports, Accounting Review 89, 2151-2180.

Huber, P.J., 1967, The behavior of maximum likelihood estimates under nonstandard condi-

tions, Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Proba-

bility 1, 221-233.

Hvide, H.K., and P. Östberg, 2015, Social interaction at work, Journal of Financial Economics

117, 628-652.

Ivokvić, Z., and S. Weisbenner, 2007, Information diffusion effects in individual investors’ com-

mon stock purchases: Covet thy neighbors’ investment choices, Review of Financial Studies

20, 1327-1357.

Kale, J.R.K., E. Reis, and A. Venkateswaran, 2009, Rank-order tournaments and incentive

alignment: The effect on firm performance, Journal of Finance 64, 1479-1512.

Karlan, D., M. McConnell, S. Mullainathan, and J. Zinman, 2016, Getting to the top of mind:

24

How reminders increase saving, Management Science, forthcoming.

Kaustia, M., and S. Knüpfer, 2012, Peer performance and stock market entry, Journal of

Financial Economics 104, 321-338.

Kothari, S.P., X. Li, J.E. Short, 2009, The effect of disclosures by management, analysts,

and business press on cost of capital, return volatility, and analyst forecasts: A study using

content analysis, Accounting Review 84, 1639-1670.

Larcker, D.F., and A.A. Zakolyukina, 2011, Detecting deceptive discussions in conference

calls, Journal of Accounting Research 50, 495-540.

Lee, C.M.C., and B. Radhakrishna, 2000, Inferring investor behavior: Evidence from TORQ

data, Journal of Financial Markets 3, 83-111.

Li, F., 2008, Annual report readability, current earnings, and earnings persistence, Journal

of Accounting and Economics 45, 221-247.

Loughran, T., and B. McDonald, 2011, When is a liability not a liability? Textual analysis,

dictionaries, and 10-Ks, Journal of Finance 66, 35-65.

Manski, C., 1993, Identification of endogenous social effects: The reflection problem, Review

of Economic Studies 60, 531-542.

Mikhail, M.B., B.R. Walther, and R.H. Willis, 2007, When security analysts talk, who listens?,

Accounting Review 82, 1227-1253.

Malmendier, U., and D. Shanthikumar, 2007, Are small investors naïve about incentives?,

Journal of Financial Economics 85, 457-489.

Moffitt, R.A.,2001, Policy interventions, low-level equilibria, and social interactions, in: Durlauf,

S.N., and H.P. Young,H.P. (eds.), Social Dynamics, MIT Press, Cambridge.

Pool, V.K., N. Stoffman, and S.E. Yonker, 2015, The people in your neighborhood: Social

interactions and mutual fund portfolios, Journal of Finance 70, 2679-2732.

Remus, R., U. Quasthoff, and G. Heyer, 2010, SentiWS – a publicly available German-

language resource for sentiment analysis, Proceedings of the 7th International Language

Ressources and Evaluation, 1168-1171.

Shiller, R. J., and J. Pound, 1989, Survey evidence on diffusion of interest and information

25

among investors, Journal of Economic Behavior and Organization 12, 47-66.

Shiller, R.J., 2000, Irrational exuberance, Princeton University Press, New Jersey.

Shiller, R.J., 2005, Irrational exuberance, Princeton University Press, New Jersey.

Tetlock, P.C., 2007, Giving content to investor sentiment: The role of media in the stock

market, Journal of Finance 62, 1139-1168.

Tumarkin, R., and R.F. Whitelaw, 2001, News or noise? Internet postings and stock prices,

Financial Analyst Journal 57, 41-51.

White, H., 1982, Maximum likelihood estimation of misspecified models, Econometrica 50,

1-26.

Wolf, M., A. Horn, A., M. Mehl, M., S. Haug, J.W., Pennebaker, and H. Kordy, 2008,

Computer-aided quantitative text analysis: Equivalence and robustness of the German adap-

tion of the Linguistic Inquiry and Word Count, Diagnostica 54, 85-98.

26

Tables

Table 1: Descriptive statisticsThis table presents descriptive statistics on comment characteristics (Panel A), flow characteristics (Panel B),performance characteristics (Panel C), portfolio characteristics (Panel D), and trader characteristics of tradingstrategies (Panel E). Appendix A provides detailed descriptions of all variables used throughout the study.

Mean Min Median Max Std.dev.

N

Panel A: Comment characteristicsComment (d) 0.053 0.000 0.000 1.000 0.223 247,048# comments 1.52 1.00 1.00 26.00 1.20 13,014% positive words 1.73 0.00 0.00 40.00 3.13 13,014% negative words 0.73 0.00 0.00 25.00 1.85 13,014

Panel B: Flow characteristicsTrade size (EUR) 5,377 0 2,251 324,587 11,285 43,275Net flows (EUR) 205 -1239722 0 1,402,852 9,346 247,048

Panel C: Performance characteristicsGross return (%) -0.002 -99.694 0.000 304.911 2.793 247,048Net return (%) -0.011 -99.694 -0.003 304.911 2.784 247,048Gross alpha (%) -0.032 -108.135 -0.008 342.803 2.424 221,910Net alpha (%) -0.040 -108.135 -0.014 342.803 2.419 221,910

Panel D: Portfolio characteristicsAssets under management (EUR) 64,500 0 4,724 8,299,194 369,396 247,048Age (days) 431 22 414 1,184 201 247,048Performance fee (%) 12.57 5.00 10.00 30.00 7.51 247,048Managerial investment (d) 0.241 0.000 0.000 1.000 0.428 247,048# securities 12.30 0.00 8.00 194.00 14.74 247,048% stocks 57.22 0.00 67.97 134.60 37.94 218,465% funds 17.50 0.00 0.00 109.45 30.00 218,465% derivatives 1.02 0.00 0.00 102.25 7.34 218,465% cash 21.56 -34.60 8.05 100.00 27.80 218,465# trades 1.41 0.00 0.00 278.00 5.40 247,048Turnover (%) 4.57 0.00 0.00 4,693.56 32.61 247,048

Panel E: Trader characteristicsProfessional (d) 0.010 0.000 0.000 1.000 0.100 247,048Experienced (d) 0.971 0.000 1.000 1.000 0.169 247,048Female (d) 0.035 0.000 0.000 1.000 0.183 247,048# portfolios 2.50 1.00 2.00 11.00 1.81 247,048

27

Table 2: Determinants of commentsThis table presents the results from a logit regression (Column 1) and the results from OLS regressions(Columns 2 to 4). The dependent variable is either a dummy variable that equals one on days on whichtraders post at least one comment, and zero otherwise (Column 1), the number of posted comments (Column2), the fraction of positive words per comment (Column 3), or the fraction of negative words per comment(Column 4). In Columns 2 and 4, we focus on days with at least one posted comment. Appendix A providesdetailed descriptions of all variables used throughout the study. In Column 1, we report marginal effects. Thet-values (in parentheses) are based on the cluster-robust variant of the Huber-White (Huber, 1967; White,1982) sandwich estimator which accounts for the dependence of observations within clusters (different dailyobservations for one trading strategy). ***, **, * denote statistical significance at the 1%, 5%, 10% level.

Comment (d) # comments % positivewords

% negativewords

(1) (2) (3) (4)

Net return (%)t-1 0.002*** 0.024*** 0.130*** -0.029*(6.26) (2.69) (4.77) (-1.85)

Log(net flows)t-1 0.001*** 0.007** -0.001 -0.005(2.99) (2.34) (-0.09) (-0.90)

Log(AuM)t-1 0.006*** 0.002 -0.002 0.005(6.26) (0.28) (-0.09) (0.47)

Log(age) -0.016*** -0.104*** -0.001 -0.042*(-12.77) (-5.26) (-0.01) (-1.88)

Performance fee (%) -0.000 -0.003 0.009 0.001(-0.40) (-0.89) (0.86) (0.14)

Managerial investmentt-1 (d) 0.016** -0.004 -0.159 -0.096(1.96) (-0.05) (-0.81) (-1.06)

# securitiest-1 0.000 0.002 0.004 -0.004**(0.61) (0.97) (0.65) (-2.29)

Turnover (%)t-1 0.000 0.002** -0.002** 0.000(1.17) (2.54) (-2.26) (0.68)

Professional (d) 0.030 -0.298*** -0.047 1.245***(1.24) (-4.77) (-0.22) (3.98)

Experienced (d) -0.005 0.186** 0.666** 0.159(-0.26) (2.26) (2.08) (1.13)

Female (d) 0.018 -0.222*** -0.424** -0.181**(1.16) (-3.50) (-2.41) (-2.09)

# portfoliost-1 0.001 -0.002 -0.022 -0.003(0.63) (-0.10) (-0.62) (-0.18)

Constant 2.498*** 0.815 0.279(3.26) (0.83) (0.97)

Day fixed effects Yes Yes Yes YesPseudo R2 0.058Adj. R2 0.027 0.019 0.025N 247,048 13,014 13,014 13,014

28

Table 3: Comments and investments of followersThis table presents the results from panel regressions with portfolio and day fixed effects. The dependentvariable is the inverse hyperbolic sine of daily net flows in trading strategies. Taking the inverse hyperbolicsine is an alternative to a log-transformation when a variable takes on zero or negative values. In Columns2 to 4, we focus on days with at least one comment posted on the preceding day. The variables Net return(%)t-1, Net return (%)t-2, Net return (%)t-3, Net return (%)t-4, Net return (%)t-5, Log(net flows)t-1, Log(netflows)t-2, Log(net flows)t-3, Log(net flows)t-4, Log(net flows)t-5, Log(AuM)t-1, Log(age), Managerial investment(d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 are included as controls in every regressions butnot reported. Appendix A provides detailed descriptions of all variables used throughout the study. Thet-values (in parentheses) are based on the cluster-robust variant of the Huber-White (Huber, 1967; White,1982) sandwich estimator which accounts for the dependence of observations within clusters (different dailyobservations for one trading strategy). ***, **, * denote statistical significance at the 1%, 5%, 10% level.

Log(net flows)

(1) (2) (3) (4)

Comment (d)t-1 0.107**(2.37)

# commentst-1 0.053*(1.95)

% positive wordst-1 0.021**(2.01)

% negative wordst-1 -0.032**(-2.04)

Constant 3.085*** 8.578*** 8.687*** 8.680***(6.41) (4.87) (4.86) (4.87)

Controls Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes YesDay fixed effects Yes Yes Yes YesAdj. R2 0.137 0.179 0.179 0.179N 247,048 13,087 13,087 13,087

29

Table 4: Comments, inflows, and outflowsThis table presents the results from panel regressions with portfolio and day fixed effects. The dependent variable is the inverse hyperbolic sine of either dailyinflows (Columns 1, 3, 5, and 7) or daily outflows in trading strategies (Columns 2, 4, 6, and 8). Taking the inverse hyperbolic sine is an alternative to alog-transformation when a variable takes on zero or negative values. In Columns 3 to 8, we focus on days with at least one comment posted on the precedingday. The variables Net return (%)t-1, Net return (%)t-2, Net return (%)t-3, Net return (%)t-4, Net return (%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(netflows)t-3, Log(net flows)t-4, Log(net flows)t-5, Log(AuM)t-1, Log(age), Managerial investment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 areincluded as controls in every regressions but not reported. Appendix A provides detailed descriptions of all variables used throughout the study. The t-values(in parentheses) are based on the cluster-robust variant of the Huber-White (Huber, 1967; White, 1982) sandwich estimator which accounts for the dependenceof observations within clusters (different daily observations for one trading strategy). ***, **, * denote statistical significance at the 1%, 5%, 10% level.

Log(inflows) Log(outflows) Log(inflows) Log(outflows) Log(inflows) Log(outflows) Log(inflows) Log(outflows)

(1) (2) (3) (4) (5) (6) (7) (8)

Comment (d)t-1 0.188*** 0.060**(5.34) (2.39)

# commentst-1 0.040* -0.005(1.93) (-0.26)

% positive wordst-1 0.009 -0.010*(1.19) (-1.81)

% negative wordst-1 -0.018* 0.008(-1.69) (1.05)

Constant 3.009*** 0.266 8.704*** 2.764 8.785*** 2.753 8.781*** 2.755(7.96) (0.73) (4.00) (1.11) (4.01) (1.11) (4.02) (1.11)

Controls Yes Yes Yes Yes Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes Yes Yes Yes Yes YesDay fixed effects Yes Yes Yes Yes Yes Yes Yes YesAdj. R2 0.388 0.319 0.475 0.405 0.475 0.405 0.475 0.405N 247,048 247,048 13,087 13,087 13,087 13,087 13,087 13,087

30

Table 5: Comments and investments of followers in the longer runThis table presents the results from panel regressions with portfolio and day fixed effects. The dependent variable is the inverse hyperbolic sine of average dailynet flows in trading strategies. Taking the inverse hyperbolic sine is an alternative to a log-transformation when a variable takes on zero or negative values.We compute average daily net flows for each week after the posting of a comment. The variables Net return (%)t-1, Net return (%)t-2, Net return (%)t-3, Netreturn (%)t-4, Net return (%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(net flows)t-3, Log(net flows)t-4, Log(net flows)t-5, Log(AuM)t-1, Log(age), Managerialinvestment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 are included as controls in every regressions but not reported. Appendix A providesdetailed descriptions of all variables used throughout the study. The t-values (in parentheses) are based on the cluster-robust variant of the Huber-White(Huber, 1967; White, 1982) sandwich estimator which accounts for the dependence of observations within clusters (different daily observations for one tradingstrategy). ***, **, * denote statistical significance at the 1%, 5%, 10% level.

Log(net flows)

t,t+4 t+5,t+9 t+10,t+14 t+15,t+19 t+20,t+24 t+25,t+29 t+30,t+34 t+35,t+39

(1) (2) (3) (4) (5) (6) (7) (8)

Comment (d)t-1 0.237*** 0.133** 0.143** 0.071 0.017 0.040 -0.067 -0.047(4.23) (2.04) (2.02) (0.91) (0.25) (0.58) (-0.90) (-0.76)

Constant 4.007*** 3.556*** 2.901*** 2.437*** 2.317*** 2.510*** 2.006*** 1.427**(7.03) (6.04) (4.52) (3.93) (3.78) (4.11) (3.64) (2.21)

Controls Yes Yes Yes Yes Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes Yes Yes Yes Yes YesDay fixed effects Yes Yes Yes Yes Yes Yes Yes YesAdj. R2 0.185 0.137 0.113 0.096 0.088 0.083 0.080 0.078N 242,888 237,688 232,488 227,288 222,122 217,044 212,051 207,143

31

Table 6: The predictive power of comments for future performanceThis table presents the results from panel regressions with portfolio and day fixed effects. The dependent variable is either the daily net return (Columns 1, 3,5, and 7) or the daily net alpha of trading strategies (Columns 2, 4, 6, and 8). In Columns 3 to 8, we focus on days with at least one comment posted on thepreceding day. Daily alphas are calculated as the difference between daily net returns and predicted returns using a 6-factor model. The factor exposures usedto predict returns are estimated over 6-month rolling windows from t-126 to t-1. The 6-factor model includes the MSCI Europe Index as proxy for the market,an SMB factor (return difference between the MSCI Europe Small Cap Index and the MSCI Europe Index), an HML factor (return difference between theMSCI Europe Value Index and the MSCI Europe Growth Index), a momentum factor (MSCI Europe Momentum Index), and two at-the-money option factorsconstructed as in Agarwal and Naik (2004) using options on the Euro Stoxx 50. The variables Net return (%)t-1, Net return (%)t-2, Net return (%)t-3, Netreturn (%)t-4, Net return (%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(net flows)t-3, Log(net flows)t-4, Log(net flows)t-5, Log(AuM)t-1, Log(age), Managerialinvestment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 are included as controls in every regressions but not reported. Appendix A providesdetailed descriptions of all variables used throughout the study. The t-values (in parentheses) are based on the cluster-robust variant of the Huber-White(Huber, 1967; White, 1982) sandwich estimator which accounts for the dependence of observations within clusters (different daily observations for one tradingstrategy). ***, **, * denote statistical significance at the 1%, 5%, 10% level.

Return (%) Alpha (%) Return (%) Alpha (%) Return (%) Alpha (%) Return (%) Alpha (%)

(1) (2) (3) (4) (5) (6) (7) (8)

Comment (d)t-1 0.012 -0.010(0.56) (-0.39)

# commentst-1 -0.036* -0.020(-1.89) (-1.42)

% positive wordst-1 0.001 -0.001(0.29) (-0.29)

% negative wordst-1 -0.019** -0.013(-2.11) (-1.28)

Constant 2.033*** 0.676* 2.385* -0.551* 2.312* -0.583** 2.309* -0.589**(10.23) (1.78) (1.84) (-1.93) (1.75) (-2.03) (1.75) (-2.05)

Controls Yes Yes Yes Yes Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes Yes Yes Yes Yes YesDay fixed effects Yes Yes Yes Yes Yes Yes Yes YesAdj. R2 0.064 0.047 0.088 0.125 0.088 0.125 0.088 0.125N 247,048 221,910 13,087 10,861 13,087 10,861 13,087 10,861

32

Table 7: Who listens to comments?This table presents the results from panel regressions with portfolio and day fixed effects. The dependentvariable is the inverse hyperbolic sine of daily net flows in trading strategies. Taking the inverse hyperbolicsine is an alternative to a log-transformation when a variable takes on zero or negative values. In Column 1, wecompute net flows only taking into account small trades and in Column 2, we focus on large trades. We classifytrades as small trades if they are smaller than or equal to EUR 5,000 and as large trades if they are largerthan EUR 5,000. The variables Net return (%)t-1, Net return (%)t-2, Net return (%)t-3, Net return (%)t-4,Net return (%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(net flows)t-3, Log(net flows)t-4, Log(net flows)t-5,Log(AuM)t-1, Log(age), Managerial investment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 areincluded as controls in every regressions but not reported. Appendix A provides detailed descriptions of allvariables used throughout the study. The t-values (in parentheses) are based on the cluster-robust variantof the Huber-White (Huber, 1967; White, 1982) sandwich estimator which accounts for the dependence ofobservations within clusters (different daily observations for one trading strategy). ***, **, * denote statisticalsignificance at the 1%, 5%, 10% level.

Log(net flows)

Small trades Large trades

(1) (2)

Comment (d)t-1 0.104*** 0.007(2.79) (0.20)

Constant 2.719*** 0.877**(6.90) (2.38)

Controls Yes YesPortfolio fixed effects Yes YesDay fixed effects Yes YesAdj. R2 0.141 0.085N 247,048 247,048

33

Figures

Figure 1: Sample profile pageThis figure shows a sample profile page of a trader on an online social trading platform.

34

Figure 2: Number of portfolios and assets under managementThis figure shows the number of trading strategies (Panel A) as well as the total assets under management(Panel B) in our sample in the time period from January 2013 to December 2014.Panel A: Number of portfolios

Panel B: Assets under management

35

Figure 3: Sample commentariesThis figure shows six sample comments published on our online social trading network, of which three aresecurity-specific comments (Panel A) and the other three are general comments (Panel B). Positive and negativewords that are included in our dictionary are underlined.

Panel A: Security-specific commentsApril 24, 2014, regarding GlaxoSmithKline (GB0009252882):I sell GlaxoSmithKline with a gain of 2%. In the last couple of days, prices ofpharma stocks increased substantially in the course of merger fantasies. I usethis opportunity and exit.

June 7, 2013, regarding Lyxor ETF Turkey (FR0010326256):Due to the protests in Turkey the Lyxor ETF TURKEY performed rather poorly.However, as the overall economic environment has not changed, I remain investedwith 5% of my portfolio.

March 19, 2013, regarding Apple (US0378331005):Apple is expected to release good quarterly results.

Panel B: General commentsFebruary 20, 2014:This is the perfect point in time to invest in my strategy. I am waiting for newopportunities and currently hold 100% of the portfolio in cash.

October 27, 2014:Today, I lost money. The way back will be cumbersome. I was too greedy and didnot implement the strategy properly. I should trade on facts and should not fallin love with certain securities. The market is always right.

August 22, 2014:Investing in the stock market is currently a tough business. However, theportfolio is still on track to generate a return of 23% p.a. To date, theannualized return is about 25.7%. Just to remind you, an annual return of 23%implies that the value of the portfolio doubles every three years. Yet, theperformance of the portfolio is heavily dependent on only a few companies.Recently, Norma and SHW have performed rather poorly. However, as fundamentalshave not changed I leave my portfolio as it is.

36

Figure 4: Asset allocationThis figure shows the average asset allocation of portfolios in our sample in the time period from January 2013to December 2014.

37

Appendix

Appendix A: Variable descriptionsThis table defines the variables used throughout the study.

Variable Description

Comment characteristics

Comment (d) Dummy variable that equals one on days on which traders post at least onecomment, and zero otherwise

# comments Number of comments posted by a trader on a certain day

% positive words Number of positive words per comment / Number of words per comment.If there are multiple comments per day, we take the average across allcomments

% negative words Number of negative words per comment / Number of words per comment.If there are multiple comments per day, we take the average across allcomments

Flow characteristics

Trade size (EUR) Trade size (in EUR)

Net flows (EUR) Daily net investments by followers in a trading strategy (in EUR)

Log(net flows) Inverse hyperbolic sine of daily net flows

Performance characteristics

Gross return (%) Daily raw return of a trading strategy net of bid-ask spreads paid by thetrader but before fees

Net return (%) Daily raw return of a trading strategy net of bid-ask spreads paid by thetrader and after fees

Alpha Daily alphas are calculated as the difference between daily (net or gross)returns and returns predicted using a 6-factor model. The factor exposuresused to predict returns are estimated over 6-month rolling windows fromt-126 to t-1. The 6-factor model includes the MSCI World Index as proxyfor the market, an SMB factor (return difference between the MSCI WorldSmall Cap Index and the MSCI World Index), an HML factor (returndifference between the MSCI World Value Index and the MSCI WorldGrowth Index), a momentum factor (MSCI World Momentum Index), andtwo at-the-money option factors constructed as in Agarwal and Naik (2004)using options on the Euro Stoxx 50

Portfolio characteristics

Assets under management(AuM; EUR)

Amount of money followers have invested in a trading strategy (in EUR)

Log(AuM) Natural logarithm of assets under management

Age (years) Number of calendar days since inception of the trading strategy

Log(age) Natural logarithm of age

Performance fee (%) Fraction of profits earned by the trader and the platform. The performancefee is paid only if the returns surpass some high-water mark, that is, thereis no performance fee until the trading strategy has recovered past losses

Managerial investment (d) Dummy variable that equals one if the trader invests own money in thetrading strategy, and zero otherwise

# securities Number of securities in a trader’s portfolio

38

% stocks Value of stocks in a trader’s portfolio / Portfolio size

% funds Value of mutual funds and exchange-traded funds in a trader’s portfolio /Portfolio size

% derivatives Value of derivative instruments in a trader’s portfolio / Portfolio size

% cash Cash position in a trader’s portfolio / Portfolio size

# trades Number of trades in a trader’s portfolio on a given day

Turnover (%) 12 value of all transactions executed in a trader’s portfolio on a certain day/ Portfolio size at the beginning of the day

Trader characteristics

Professional (d) Dummy variable that equals one if the trader is a professional assetmanagement firm, and zero otherwise

Experienced (d) Dummy variable that equals one for traders that claim to have several yearsof investment experience, and zero otherwise

Female (d) Dummy variable that equals one for female traders, and zero otherwise

# portfolios Number of portfolios managed by a trader. This also contains portfolios wedropped in the sample selection process

39

Appendix B: Results from robustness tests

Table A1: Most frequently used positive and negative wordsThis table shows the 25 most frequently mentioned positive and negative words in traders’ comments basedon our own word list. We provide one possible translation of these words.

Rank Positive words Negative words

1 good unfortunately2 profit correction3 positive losses4 profits loss5 plus bad6 high risk7 recovery negative8 rise minus9 profit-taking weakness10 opportunities declines11 all-time-high weak12 increases volatility13 increase risks14 interesting declined15 increasing declining16 uptrend decline17 growth downtrend18 strength lost19 top uncertainty20 opportunity loses21 potential danger22 increased fluctuations23 outperformance downturn24 successful speculation25 gains crisis

40

Table A2: Comments and investments of followers – security-specific vs. general commentsThis table presents the results from panel regressions with portfolio and day fixed effects. The dependent variable is the inverse hyperbolic sine of daily netflows in trading strategies. Taking the inverse hyperbolic sine is an alternative to a log-transformation when a variable takes on zero or negative values. InColumns 1 to 4, we only use security-specific comments to determine comment characteristics and in Columns 5 to 8, we rely on general comments. Moreover,in Columns 2 to 4 and Columns 6 to 8, we focus on days with at least one comment posted on the preceding day. The variables Net return (%)t-1, Netreturn (%)t-2, Net return (%)t-3, Net return (%)t-4, Net return (%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(net flows)t-3, Log(net flows)t-4, Log(net flows)t-5,Log(AuM)t-1, Log(age), Managerial investment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 are included as controls in every regressions but notreported. Appendix A provides detailed descriptions of all variables used throughout the study. The t-values (in parentheses) are based on the cluster-robustvariant of the Huber-White (Huber, 1967; White, 1982) sandwich estimator which accounts for the dependence of observations within clusters (different dailyobservations for one trading strategy). ***, **, * denote statistical significance at the 1%, 5%, 10% level.

Log(net flows)

Security-specific comments General comments

(1) (2) (3) (4) (5) (6) (7) (8)

Comment (d)t-1 0.032 0.124**(0.46) (2.54)

# commentst-1 0.007 -0.009(0.17) (-0.13)

% positive wordst-1 -0.010 0.050***(-0.65) (2.84)

% negative wordst-1 -0.016 -0.048*(-0.75) (-1.75)

Constant 3.030*** 9.250*** 9.270*** 9.254*** 2.875*** 4.162*** 4.185*** 4.159***(6.22) (4.39) (4.41) (4.38) (6.00) (4.28) (4.32) (4.30)

Controls Yes Yes Yes Yes Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes Yes Yes Yes Yes YesDay fixed effects Yes Yes Yes Yes Yes Yes Yes YesAdj. R2 0.132 0.197 0.197 0.197 0.132 0.155 0.157 0.156N 239,456 5,495 5,495 5,495 240,552 6,591 6,591 6,591

41

Table A3: Comments and investments of followers – alternative word listsThis table presents the results from panel regressions with portfolio and day fixed effects. The dependentvariable is the inverse hyperbolic sine of daily net flows in trading strategies. Taking the inverse hyperbolicsine is an alternative to a log-transformation when a variable takes on zero or negative values. We focus ondays with at least one comment posted on the preceding day. In Columns 1 and 2, we rely on a word list basedon the Harvard IV-4 dictionary (Remus et al, 2010) and in Columns 3 and 4, we use the LIWC dictionary(Wolf et al, 2008). The variables Net return (%)t-1, Net return (%)t-2, Net return (%)t-3, Net return (%)t-4,Net return (%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(net flows)t-3, Log(net flows)t-4, Log(net flows)t-5,Log(AuM)t-1, Log(age), Managerial investment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 areincluded as controls in every regressions but not reported. Appendix A provides detailed descriptions of allvariables used throughout the study. The t-values (in parentheses) are based on the cluster-robust variantof the Huber-White (Huber, 1967; White, 1982) sandwich estimator which accounts for the dependence ofobservations within clusters (different daily observations for one trading strategy). ***, **, * denote statisticalsignificance at the 1%, 5%, 10% level.

Log(net flows)

Harvard IV-4 word list LIWC word list

(1) (2) (3) (4)

% positive wordst-1 0.015** -0.001(2.34) (-0.14)

% negative wordst-1 -0.010 -0.025*(-0.82) (-1.69)

Constant 8.619*** 8.698*** 8.688*** 8.713***(4.80) (4.85) (4.86) (4.86)

Controls Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes YesDay fixed effects Yes Yes Yes YesAdj. R2 0.179 0.179 0.179 0.179N 13,087 13,087 13,087 13,087

42

Table A4: The predictive power of comments for future performance – CAPMalphasThis table presents the results from panel regressions with portfolio and day fixed effects. The dependentvariable is the daily net alpha of trading strategies. In Columns 2 to 4, we focus on days with at least onecomment posted on the preceding day. Daily alphas are calculated as the difference between daily net returnsand predicted returns using the CAPM. The factor exposures used to predict returns are estimated over 6-month rolling windows from t = -126 to t = -1. The CAPM uses the MSCI Europe Index as proxy for themarket. The variables Net return (%)t-1, Net return (%)t-2, Net return (%)t-3, Net return (%)t-4, Net return(%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(net flows)t-3, Log(net flows)t-4, Log(net flows)t-5, Log(AuM)t-1,Log(age), Managerial investment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 are included ascontrols in every regressions but not reported. Appendix A provides detailed descriptions of all variables usedthroughout the study. The t-values (in parentheses) are based on the cluster-robust variant of the Huber-White (Huber, 1967; White, 1982) sandwich estimator which accounts for the dependence of observationswithin clusters (different daily observations for one trading strategy). ***, **, * denote statistical significanceat the 1%, 5%, 10% level.

Alpha (%)

(1) (2) (3) (4)

Comment (d)t-1 -0.004(-0.17)

# commentst-1 -0.015(-1.09)

% positive wordst-1 0.003(0.73)

% negative wordst-1 -0.014(-1.54)

Constant 0.660* -0.488* -0.508* -0.519*(1.91) (-1.77) (-1.82) (-1.87)

Controls Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes YesDay fixed effects Yes Yes Yes YesAdj. R2 0.051 0.112 0.112 0.112N 221,910 10,861 10,861 10,861

43

Table A5: The predictive power of comments for future performance – 1-month performanceThis table presents the results from panel regressions with portfolio and day fixed effects. The dependent variable is either the 1-month net return (Columns1, 3, 5, and 7) or the 1-month cumulative abnormal return (CAR) of trading strategies (Columns 2, 4, 6, and 8) . In Columns 3 to 8, we focus on days withat least one comment posted on the preceding day. Daily abnormal returns are calculated as the difference between daily net returns and predicted returnsusing a 6-factor model. The factor exposures used to predict returns are estimated over 6-month rolling windows from t = -126 to t = -1. The 6-factor modelincludes the MSCI Europe Index as proxy for the market, an SMB factor (return difference between the MSCI Europe Small Cap Index and the MSCI EuropeIndex), an HML factor (return difference between the MSCI Europe Value Index and the MSCI Europe Growth Index), a momentum factor (MSCI EuropeMomentum Index), and two at-the-money option factors constructed as in Agarwal and Naik (2004) using options on the Euro Stoxx 50. The variables Netreturn (%)t-1, Net return (%)t-2, Net return (%)t-3, Net return (%)t-4, Net return (%)t-5, Log(net flows)t-1, Log(net flows)t-2, Log(net flows)t-3, Log(net flows)t-4,Log(net flows)t-5, Log(AuM)t-1, Log(age), Managerial investment (d)t-1, # securitiest-1, Turnover (%)t-1, and # portfoliost-1 are included as controls in everyregressions but not reported. Appendix A provides detailed descriptions of all variables used throughout the study. The t-values (in parentheses) are basedon the cluster-robust variant of the Huber-White (Huber, 1967; White, 1982) sandwich estimator which accounts for the dependence of observations withinclusters (different daily observations for one trading strategy). ***, **, * denote statistical significance at the 1%, 5%, 10% level.

1-monthreturn (%)

1-monthCAR (%)

1-monthreturn (%)

1-monthCAR (%)

1-monthreturn (%)

1-monthCAR (%)

1-monthreturn (%)

1-monthCAR (%)

(1) (2) (3) (4) (5) (6) (7) (8)

Comment (d)t-1 -0.220* -0.073(-1.89) (-0.31)

# commentst-1 0.034 0.011(0.52) (0.25)

% positive wordst-1 0.022 -0.001(1.00) (-0.06)

% negative wordst-1 -0.011 0.000(-0.36) (0.01)

Constant 2.772*** 1.498 1.321 1.869 1.393 1.891 1.388 1.891(2.98) (1.39) (0.65) (0.70) (0.68) (0.71) (0.68) (0.71)

Controls Yes Yes Yes Yes Yes Yes Yes YesPortfolio fixed effects Yes Yes Yes Yes Yes Yes Yes YesDay fixed effects Yes Yes Yes Yes Yes Yes Yes YesAdj. R2 0.243 0.239 0.549 0.409 0.549 0.409 0.549 0.409N 226,248 226,248 12,076 12,076 12,076 12,076 12,076 12,076

44