Confirmation Bias, Overconfidence, and Investment

Performance: Evidence from Stock Message Boards

JaeHong Park
Ulsan National Institute of Science and Technology
[email protected]

Prabhudev Konana, Bin Gu

University of Texas at Austin
[email protected], [email protected]

Alok Kumar
University of Miami
[email protected]

Rajagopal Raghunathan
University of Texas at Austin
[email protected]

Confirmation Bias, Overconfidence, and Investment

Performance: Evidence from Stock Message Boards

Abstract
Using data from a new field experiment in South Korea, we study how information from virtual
communities such as stock message boards influences investors trading decisions and
investment performance. Motivated by recent studies in psychology, we conjecture that investors
would use message boards to seek information that confirms their prior beliefs. This
confirmation bias would make them more overconfident and adversely affect their investment
performance. An analysis of 502 investor responses from the largest message board operator in
South Korea supports our conjecture. We find that investors exhibit confirmation bias when
they process information from message boards. We also demonstrate that investors with
stronger confirmation bias exhibit greater overconfidence. Those investors have higher
expectations about their performance, trade more frequently, but obtain lower realized returns.
Collectively, these results suggest that participation in virtual communities increases investors
propensity to commit investment mistakes and is likely to be detrimental to their investment
performance.

Keywords: Confirmation bias, overconfidence, investment decisions, Korean individual investors,

virtual communities, stock message boards.

1. Introduction
Investors are increasingly flocking to virtual communities or message boards to seek, clarify, and
exchange information. Businesses like Seekingalpha.com and leading business magazines (e.g.,
Fortune) are evaluating, synthesizing, and reporting comments made on message boards or blogs.
Numerous online services (e.g., PredictWallStreet.com, Marketwatch.com, socialpicks.com,
Stockpkr.com) have started to aggregate stock sentiments from these message boards and make
the information available to online brokerages (e.g., TDAmeritrade) and other financial
informediaries. In May 2010, Yahoo! Finance Message Boards recorded 41.4 million unique
visitors (comScore Media Metrix 2010).
The non-trivial number of users of message boards and the increasing use of community
sentiments raise the question of how investors use information from message boards to
formulate their investment decisions. It is important to understand whether these message
boards are beneficial or detrimental to investors financial health. This research investigates how
investors process message board information and analyzes the impact of their information
gathering activities on their return expectations and investment performance.
Prior studies in accounting and finance suggest that virtual communities often provide
more accurate information than analyst forecasts (Bagnoli et al 1999, Clarkson et al. 2006). Early
evidences of Enron collapse from accounting tricks were first reported in message boards
(Felton and Kim 2002). Some studies show that the volume of posts or the level of investor
sentiments in virtual communities has a positive relation with market activity (e.g., trading
volume or volatility) (Antweiler and Frank 2004, Bagnoli et al. 1999, Das and Chen 2007,
Tumarkin and Whitelaw 2001). These findings indicate that message boards attract insiders and
informed investors and information in virtual communities can have an impact on investor
investment decisions and trading activities. However, research explicating how this information
influences their investment performance is still evolving.

The value of these message boards in information processing and decision-making can
be explained using both economic and psychological perspectives. However, these two
perspectives suggest different behaviors. From an economic perspective, one would expect that
message boards have a beneficial effect, since they provide timely information at a much lower
search cost (Birchler and Butler 2007). The economic principle suggests that individuals are
rational in obtaining information (Birchler and Butler 2007). That is, they tend to seek
information without any bias, objectively assess the relevance and veracity of each piece of
information, and integrate the information in forming an investment strategy that may yield
better investment performance. On the contrary, psychological studies indicate that virtual
communities may not necessarily make investors more informed or lead to better investment
performance. In particular, the behavioral finance literature shows that psychological biases,
particularly in an uncertain and noisy environment, influence investors information processing
behavior (Kahneman and Riepe 1998, Barber and Odean 2001).
Our study extends this behavioral literature and develops several hypotheses motivated
primarily by the psychology literature on confirmation bias. Our main goal is to examine whether
psychological factors influence how individuals process information from virtual communities
and whether such behavior in turn influences their investment expectations and actual
performance.

Our key conjecture is that investment-related virtual communities would not

necessarily benefit investors because they are likely to seek information that aligns more closely
with their prior beliefs, i.e., they would exhibit confirmation bias. Further, since the messages
posted on virtual communities are likely to reinforce their prior beliefs, investors would become
overconfident and hence more prone to making investment mistakes such as excessive trading
(Barber and Odean 2001, 2002, Graham et al. 2009). Such behavior would lead to less carefully
considered investment decisions, and ultimately, lower returns on their investments.

To test these conjectures, we conducted a field experiment on investors participating in

Naver.com, which is a popular message board community in South Korea. Our empirical results
indicate that contrary to the predictions of standard economic theory (e.g., Birchler and Butler
2007), investors process information from stock message boards inefficiently and exhibit
confirmation bias. In particular, investors with stronger prior beliefs are more likely to accept
confirming opinions from virtual communities. Further, we find that investors with stronger
confirmation bias have higher expectations about their investment performance, but they engage
in excessive trading and experience lower realized performance. A natural interpretation of this
evidence is that confirmation bias makes investors overconfident and overly optimistic, which
results in lower returns on stock investments. We also find that investors perceived knowledge
negatively moderates the confirmation bias such that investors with a higher level of perceived
knowledge have less motivational needs to confirm their own opinions from others.
These findings contribute to the finance literature that investigates the economic value
of information disseminators such as virtual communities. We demonstrate that message board
participation may not necessarily improve individuals financial well-being. Our study also
contributes to a better understanding of investors behavioral biases in financial markets. While
investors are known to demonstrate overconfidence and illusion of knowledge, our analysis takes
a step further by showing that virtual investment-related communities contribute to these biases.
Last, our paper makes significant contributions to the virtual community literature in
information systems (IS). Prior IS literature has focused mainly on the motivations to participate
(e.g., Wasko and Faraj 2005, Ma and Agarwal 2007), but has not investigated the performance
impact of participation in virtual communities.
The rest of the paper is organized as follows. In the next section we discuss the related
literature and outline the main testable hypotheses. In Section 3, we summarize our survey

methods and data. We present our main empirical results in Section 4 and conclude in Section 5
with a brief discussion.

2. Related Literature and Hypothesis Development

2.1 Related Research
Our study is related to three distinct lines of research.

First, our paper adds to a growing body

of behavioral literature on virtual communities. It is well accepted that decision makers are
often influenced by multiple psychological biases that distort their decision making and
economic outcomes (Barber and Odean 2001, 2002, Kahneman and Riepe 1998, Raghunathan
and Corfman 2006). In particular, Barber and Odean (2001) argue that the illusion of control
(e.g., people believe that they can influence the outcome of chance events), the illusion of
knowledge (e.g., when people have far more data, they believe that they are more knowledgeable
than they really are), and self-attribution bias (e.g., people tend to attribute their success to their
own abilities while attribute their failure to bad luck) drive investors to be overconfident.
Overconfident investors are known to trade more frequently and have negative abnormal returns
amongst stock market traders, especially when they are less experienced yet successful (Odean
1998, Barber and Odean 2001). These studies provide valuable insights that help investigate the
relation between overconfidence and investment performance. The current study extends this
analysis to participants in virtual communities.
Second, our study is related to the research stream that explicates homophilous behavior.
Homophily is defined as an individuals tendency to associate with others who have similar
backgrounds or beliefs (McPherson et al. 2001, Gu et al. 2010). Brown et al. (2007) suggest that
users identity information may help to form a homophily of interests with virtual communities.
They also argue that the virtual community itself may create homophily. Since virtual
communities are formed based on a topic, activity, hobby, or ideology, the design encourages
community members to form social interactions with people who share similar interests (Best

and Krueger, 2006, Brown et al, 2007).

Gu et al. (2010) suggest that homophilous behavior of

investors can be explained based on cognitive dissonance theory whereby investors are
motivated to reduce dissonances between their own opinions and opinions expressed in virtual
communities.
This research is also related to the IS literature that investigates individuals motivation
to participate in virtual communities (Bagozzi and Dholakia 2002, Butler 2001). Understanding
this motivation helps reveal information processing behavior and the motivation for users to
contribute in virtual communities.

These studies suggest that social and economic benefits are

major motivations to participate. Individuals perceive extrinsic (e.g., reciprocity, reputation) and
intrinsic benefits (e.g., self-efficacy, enjoyment of helping others) from their participation. In
particular, IS literature explains such motivations by using social capital and social exchange
theories (Wasko and Faraj, 2005, Ma and Agarwal 2007, Constant et al. 1996). However, current
IS literature has not investigated how individuals process information from virtual communities
and the impact of virtual communities on individuals decision process and resulting
performance. Our study adds to the current body of this literature.

2.2 Main Testable Hypotheses

Our main testable hypotheses are motivated by the literature on confirmation bias and the recent
evidence on individual investor behavior from the behavioral finance literature. The
confirmation bias literature from psychology suggests that decision makers tend to put too much
weight on evidence that confirms their prior views and too little weight on evidence that
contradicts or invalidates their views (Shefrin 1999, Lord et al. 1979). Decision makers are often
disposed to the confirmation bias and depreciate information that opposes their beliefs
(Festinger 1957, McMillan and White 1993, Nickerson 1998, Raghunathan and Corfman 2006,).
This indicates that, when decision makers process information, they often discount
disconfirming evidences while selecting and emphasizing confirming evidences. The bias can be

explained by the cognitive dissonance theory which suggests that individuals attempt to reduce
cognitive dissonance by distorting information in favor of the chosen alternative (e.g., Russo et al.
1996, Festinger 1957, Frey 1986). The need for individuals to reduce cognitive dissonance leads
to confirmation bias among investors in seeking information in message boards.
The behavioral finance literature has long argued that retail investors adhere excessively
to previously formed beliefs in spite of new information (Lord et al. 1979, Ko and Hansch 2008).
This behavior generates under-reaction to new information and reinforces prior beliefs. Such
adherence becomes more substantial in online environments (Barber and Odean 2001, Konana
and Balasubramanian 2005). Online investors exhibit strong sentiments about their stock
holdings (i.e., sentiments to either buy or sell) and trading decisions as they can access more data
and information over the Internet (Barber and Odean 2001). A greater volume of information is
known to increase the illusion of knowledge, whereby individuals subjective assessment of the
depth, relevance, or accuracy of their own knowledge exceeds the objective assessment of that
knowledge by a disinterested expert (Barber and Odean 2001, Konana and Balasubramanian
2005). Thus, these investors are likely to have strong beliefs about their trading abilities and/or
the markets future performance.
These above arguments indicate that online investors may be driven by their desire to
reduce cognitive dissonance between their own opinion and the widespread opinions in virtual
communities, where the magnitude of the dissonance may be determined by the strength of their
beliefs.

When investors with strong beliefs receive disconfirming messages from virtual

communities, their cognitive dissonance is likely to become more severe. Thus, they would be
more motivated to accept confirming information but exert more effort to scrutinize
disconfirming messages to reduce the dissonance. We thus expect investors confirmation bias to
become more salient when they hold a stronger belief about their stock position. Therefore, we
conjecture that:

H1: Investors with strong prior beliefs about a stocks future performance are more likely to show
confirmation bias in processing information from the virtual communities.
By selectively screening out non-confirming evidence, confirmation bias is known to
increase investor overconfidence that has been widely reported in financial literature (Burger
1989, Deci and Ryan 1987, Barber and Odean 2001). When investors seek information that
confirms their prior beliefs, it is likely to enhance overconfidence (Barber and Odean 2001). The
literature suggests that overconfident investors overestimate the precision of information or
underestimate the volatility of random events in financial markets (David et al. 2007, Glaser and
Weber 2007). Such investors miscalibration is known to increase the differences of opinion
among investors, which, in turn, leads to excessive trading (Varian 1989). The mis-calibration is
estimated by asking for upper and lower bounds of 90% confidence intervals for the return on a
stock in the future (Glaser et al. 2004, Hilton 2001).
The extant literature shows that overconfident investors confidence intervals are very
tight and this estimate is negatively related to trading frequency (Glaser and Weber 2007).
Further, the literature demonstrates that overconfident investors excessive trading leads to poor
investment performance (Barber and Odean 2000, 2001, Konana and Balasubramanian 2005,
Glaser et al. 2004). Although individual investors are motivated to trade when they expect to
earn high stock returns to offset their trading costs, overconfident investors experience poor
overall performance because they trade frequently and speculatively and have greater trading
costs (Barber and Odean 2000).
In our research context, investors who confirm their beliefs using information from
virtual communities are more likely to become overconfident. These investors would reinforce
their decisions and would be disposed to overestimate the precision of their decisions or
underestimate the variance of their information signals. Therefore, confirmation bias can drive
investors to engage in excessive trading and has lower investment performance.

By screening out disconfirming evidence, confirmation bias can also lead to higher
perceived competence. Recent studies in finance suggest that perceived competence is positively
related to investor trading frequency (Graham et al. 2009). Investors who feel more competent in
their investment-related skills tend to trade with their own beliefs than investors who feel less
competent. Therefore, these investors are more likely to make investment mistakes (e.g.,
excessive trading) and have lower returns on their investments.

In the context of virtual

communities, investors can easily find the confirming opinions that reinforce their prior beliefs.
Consequently, they are likely to feel more competent, which reduces their investment
performance due to excessive trading. More formally, our second hypothesis posits that:
H2: Investors who exhibit confirmation bias would trade excessively and experience lower investment
performance.
We also posit that confirmation bias would be positively related to investor optimism.
Optimism represents a decision makers tendency to be overoptimistic about the future and the
outcome of planned actions (Hilton 2001, Weinstein 1980). People often have unrealistic
positive views about themselves or believe that their skills are better than others. For instance,
optimistic managers in firms have positive performance estimates and believe that their firms
chance of success is greater than others in the same business category (David et al. 2007, Simon
et al. 1997). The literature has reported that optimism is common in various contexts and has a
strong impact on an individuals decision making (Weinstein 1980, Kahneman and Riepe 1998).
In particular, finance literature suggests that individual investors often have optimistic views of
their stock investment performance since they perceive to have made rational choice (Hilton
2001, Glaser et al. 2004, Glaser and Weber 2007). The optimism bias makes investors
overestimate their abilities and thus have unrealistic optimistic views about the future
performance (Kahneman and Riepe 1998).

In virtual investment-related communities, investors may foster optimistic view about

their stock investment performance. As discussed earlier, online investors driven by
confirmation bias can reinforce their prior beliefs using information from virtual communities.
Those investors may believe that they are better than others in investment-related tasks and
develop unrealistic optimism about their investment abilities and performance. Therefore,
investors who show biased information seeking behavior in virtual communities would have a
higher level of expectation about their investment performance.

Our third hypothesis formally

states this conjecture:

H3: Investors who exhibit confirmation bias are likely to have higher expectations about their investment
performance.

2.3 Moderating Effects and Auxiliary Hypotheses

We investigate three moderating factors that can influence investors confirmation bias in virtual
investment-related communities. Studies in information processing suggest that decision makers
ability and motivation determine how to process information (Petty and Cacioppo 1986). In our
context, investors investment-related knowledge and trading experience are related to their
ability, and investment amounts are related to their motivation (Sussman and Siegal 2003). Thus,
investors knowledge, trading experience, and portfolio size would influence how they process
information from message boards.
2.3.1. Perceived Knowledge
Specifically, we posit that the relation between investors belief strength and confirmation bias in
virtual communities can be moderated by the level of the investors perceived knowledge.
Perceived knowledge is defined as the knowledge which people believe they hold irrespective of
what they actually know (Salmon 1986). When decision makers believe that they are more
knowledgeable than others, the dissonance induced by disconfirming opinions in virtual

communities becomes less significant as these investors can easily dismiss such opinions as
uninformed. Hence, their motivation to seek confirmation from outside sources will decrease.
On the contrary, a low level of perceived knowledge could reduce decision makers selfesteem and, thus, increase their motivation to seek confirming evidence. Self-esteem is a basic
human need that reflects an individuals assessment of his or her own worth or competence
(Branden 1969).

When investors believe that they are less knowledgeable in understanding

stock investment, it may become more difficult for them to dismiss disconfirming opinions.
Subsequently, these investors may become more motivated to resolve the dissonance by seeking
confirmatory evidence in virtual communities. Thus, we conjecture that, for the investors with a
high level of perceived knowledge, the strength of beliefs about the stock would have less effect
on confirmation bias in virtual communities. More formally, we posit that:
H4: The relation between an investors strength of beliefs and confirmation bias would be stronger for
investors with lower perceived investment-related knowledge.
2.3.2. Trading Experience
The relation between trading experience and the degree of overconfidence has been studied in
the behavioral finance literature (Gervais and Odean 1999, Nicolosi et al. 2010). This literature
suggests that investors are more likely to be overconfident when they are less experienced as they
learn about their true ability through experience (Barber and Odean 2001). This implies that
overconfidence would decrease with experience. There are some empirical findings indicating
that psychological biases of investors indeed decrease with trading experience due to the learning
effect (Coursey et al. 1987, Dhar and Zhu 2006). Inexperienced investors, therefore, are more
likely to be overconfident in their investment decisions. As overconfident investors demonstrate
higher self-esteem, they have less need to seek confirming evidence to support their prior belief.
Thus, we propose a pair of competing hypotheses:

H5: The relation between an investors strength of beliefs and confirmation bias would be stronger for less
experienced investors.
2.3.3. Investment Amount
In our last hypothesis, we examine whether confirmation bias is moderated by an investors stake
in a stock.

The theory of cognitive dissonance suggests that the magnitude of dissonance

increases when the issue under consideration has serious personal consequences (Festinger 1957).
In our setting, this theory implies that investors who are more involved in a stock could show a
higher level of confirmation bias. Those investors may be more motivated to seek confirmatory
evidence to reduce dissonance.
Specifically, investors are likely to be more involved in a stock when they invest a
substantial amount of money in the stock. When these investors have a strong belief and
encounter disconfirming messages in virtual communities, they are more motivated to reject
such disconfirming opinions and more motivated to seek and accept confirming opinions.

We

thus posit that:

H6: The relation between an investors strength of beliefs and confirmation bias would be stronger for
investors with higher investment amount in a stock.
To summarize, our fundamental conjecture is that investment performance is adversely
affected by participation in stock message boards. Investment performance is influenced by
trading frequency, both trading frequency and investment expectations are influenced by
confirmation bias, and the severity of confirmation bias is influenced by the strength of an
investors beliefs about a stock. We further conjecture that the relation between strength of
beliefs and confirmation bias would be moderated by perceived knowledge, online investing
experience and investment amount. The constructs and theoretical arguments of these linkages
are illustrated in Figure 1.

3. Data and Summary Statistics

3.1 Design of the Field Experiment
Our main data set comes from a field experiment on the participants of Naver stock message
boards, which are operated by Naver.com (http://stock.naver.com). Naver is the largest online
portal website in South Korea and it provides a wide range of services including financial news
service, stock quotes, stock exchange rates, corporate press releases, reports and
recommendations. Naver reports that its 1,992 stock message boards attract about 100,000
unique visitors per day. Overall, on any given day, Naver attracts about 600,000 visitors and 46
million page views. With the help of Naver.com, we posted the link to our research experiment
at the top of each Naver stock message board during the October 7, 2009 to October 21, 2009
period. We also assured participants that the results would be reported only in aggregate and that
their anonymity would be secured.
At the beginning of the experiment, participants were asked to complete an online
questionnaire. The questionnaire included investors demographic information, information
search behavior, investment patterns, perceived knowledge and perceived competence about
stock investments. Participants were asked to answer questions specific to the stock related to
the stock message board.

For instance, if they participated in our experiment on the

SAMSUNG Electronics message board, the questions would be related to SAMSUNG

Electronics. The participants were asked to report their current opinion about the stock (e.g.,
strong sell, sell, hold, buy, and strong buy), their current investment amount in the stock, and
their expected return of the stock one month following the date of participation in the
experiment. Participants were allowed to join our experiment multiple times at different stock
message boards, but they could only participate once on each stock message board.
The questionnaire was pre-tested with 24 employees of Naver who were also actual retail
investors. They were asked to comment on the questions, to raise concerns related to the

questionnaire, and to describe any ambiguities. To address the concerns of content validity, the
questionnaire was also pretested with three academic researchers familiar with virtual
investment-related communities. The survey questions were then modified according to the
comments from researchers and Naver employees.
To measure investors confirmation bias on message boards, we asked them to read five
new messages posted on stock message boards.

With the help of Naver, we created five new

messages at the top of each message board.1 The titles of two messages represented positive
opinion about the stock, two messages represented negative opinion about the stock, and the last
message represented a neutral opinion or no opinion about the stock. The messages were
randomly selected from the database of Naver stock messages and the titles of messages were
modified for clarity. These five new messages were randomly ordered at the top of the stock
message boards and marked as newly posted messages. Since we used actual Naver stock
message boards for our field experiment, we controlled for the number of replies, sentiments,
rating values, posting date, etc.

A screenshot of the experimental setup is provided in Figure 2.

For each of the five new messages, we asked participants to answer the following
questions: (1) which opinion appears to have the most widespread support, (2) which opinion
appears to be most strongly backed by news about the stock, and (3) which opinion appears to
have the most convincing argument. Participants were asked to click on one of the five new
messages for each of the above questions.
Investors driven by confirmation bias are more likely to click on messages consistent
with their prior beliefs. For instance, if an investor with a strong buy opinion about a stock
exhibits confirmation bias, he would be more disposed to accept positive messages of the stock
while rejecting negative messages. These investors may believe that positive messages among the
new five messages have the most widespread support, are most strongly backed by news, and
1

We only provided the titles of new messages since it is difficult to control for participants subjective judgment of
message content. Investors could see the titles, but not the content of new messages.

offer the most convincing argument. Overall, they are more likely to click on positive messages.
Thus, using the message clicking behavior of investors, we are able to measure their
confirmation bias.

3.2 Experimental Data

During the two-week experimental period, 651 participations clicked on the link to our
experiment. Of these responses, we dropped 149 responses because of scattered and
substantially incomplete data. We were left with 502 data points that are used in our empirical
analysis. Table 1, Panel A provides the demographic attributes for these 502 participants.

3.3 Main Variables

To test our hypotheses, we obtain measures of investors strength of beliefs about a stock and
their degree of confirmation bias:

Strength of Beliefs About a Stock: This variable takes a value of 2 if an investor has a
strong sell or strong buy opinion about a stock, 1 if he has a sell or buy
opinion, and 0 if he has a hold opinion.

Confirmation Bias: This variable takes a value of 1 if an investor clicks on positive

(negative) messages when he holds a strong buy or buy (a strong sell or sell
opinion), -1 if an investor clicks on negative (positive) messages when he holds a
strong buy or buy (a strong sell or sell opinion), and 0 otherwise.
The first variable represents an investors strength of beliefs about a stock and is a

categorical variable. The second variable is also a categorical variable and measures the
confirmation bias of an investor. Since participants were asked to click on the message titles for
each of the three questions, the maximum value of confirmation bias is 3 while the minimum
value is -3. A higher score indicates that an investor is more likely to exhibit the confirmation
bias.

In addition to these two key variables, we obtain a measure of trading frequency of each
investor, which is used as a mediating variable. Participants were asked to report their average
weekly trading frequency. This self-reported measure is used to capture an investors trading
behavior.
Our set of moderating variables includes perceived knowledge, investment experience,
and amount invested in a stock. An investors perceived knowledge is measured based on
responses to three items in the questionnaire: (1) I am well informed about the stock market, (2)
I am familiar with the stocks I trade (e.g., their business, financial status), and (3) I am well
informed about the major economic news that impacts the stock market. A seven-point Likert
scale is used to collect these responses. To ensure the reliability and convergent validity of this
scale, we use the Cronbach alpha measure and confirmatory factor analysis (CFA), respectively.
We find that the scale has a Cronbach alpha of 0.81 for the sample while the CFA indicates that
a single factor solution can extract 72.79% of the item variance. Both Cronbach and CFA values
meet the recommended threshold values.
Investors online investing experience is measured as 1 if an investors experience is less
than 1 year, 2 if the experience is between 1 and 2 years, 3 if the experience is between 3 and 6
years, and 4 if the experience is greater than 6 years. Investors were also asked to report their
average amount invested in each stock (in Korean Won).

We use the standardized value (i.e.,

the mean is set to zero and the standard deviation is one) of this variable in our empirical analysis.
One of our key dependent variables is an investors realized investment performance.
We compute the realized stock return during the one month period following the participation
date:
R , 

P, P,
P,

In the above equation, R , represents the actual return of an investor i who holds
the stock from time t to t+1. P, is the stock price at the end of time period t+1 and P, is

the stock price at the end of time period t. The time period t is the date on which investor i
participated in our experiment and the time period t+1 refers to the one month period following
the participation date. The prices of all stocks are obtained from the Naver stock price database.
Our second key dependent variable R
, is the expected stock return for the one month
period following the participation date, as reported by the investor in the questionnaire.

3.4 Summary Statistics

Table 1, Panels B and C present the summary statistics for all dependent and independent
variables.

There is significant heterogeneity in the investors beliefs about the future stock

performance. The statistics presented in Panel B indicate that about 50% of the respondents
have a hold opinion about their stock position, 32.90% of them have directional opinions (e.g.,
sell or buy), and 16.70% of them have strong directional opinions (e.g., strong sell or strong buy).
Further, the evidence in Panel C indicates that the mean value of their expected one-month
return is 28.43%, which indicates that most investors expect significant positive returns from
their stock investment. These expectations are very optimistic in comparison to the performance
of the overall market (Korea Composite Stock Price Index), which yielded only +3.49% during
the study period. The mean value of investors realized return, however, is -4.9%, which indicates
that many participants experienced negative returns from their stock investments.

4. Empirical Results
4.1 Estimation Framework: Systems of Equations
We estimate several systems of equations to test our key hypotheses.

To begin, we examine

whether investors prior beliefs influence their degree of confirmation bias, which subsequently
influences their trading frequency and realized stock performance. The first system of equations
is specified as follows:
Realized Performance System of Equations:

ConfirmationBias,   StrengthOfBelief,  Knowledge, $

(1a)
StrengthOfBelief,

% Experience, $ StrengthOfBelief,  * Amount ,

$ StrengthOfBelief,  - Knowledge,  . Experience,
 / Amount ,  , ;
(1b)
TradingFrequency,   ConfirmationBias,  , ;
(1c)
R ,   TradingFrequency,  Age,  % Sex,  * Eduation, 
- Income,  , .
In this set of equations, i = 1, 2, , N are the investors who participated in our
experiment, t denotes each investors participation date, and t+1 represents the period one
month following the participation date. In equation (1a), ConfirmationBiasi,t captures the
confirmation bias of investor i at time t and StrengthOfBeliefi,t denotes the strength of beliefs of
investor i at time t. We also include three interaction terms with StrengthOfBeliefi,t to estimate
the moderating effects of perceived knowledge (Knoweldgei,t), trading experience (Experiencei,t),
and investment amount (Amounti,t).
In equation (1b), TradingFrequencyi,t captures the trading frequency of investor i at time
t.

In equation (1c), Ri,t+1 represents the actual realized stock return of investor i during the time

period t+1. Drawing on the previous literature (e.g., Graham et al. 2009), we consider four
control variables that may influence investors performance, including investors age (Agei,t),
gender (Sexi,t), level of education (Eduationi,t), and income (Incomei,t).
Next, we investigate whether confirmation bias affects performance through the
overconfidence channel. One of our key conjectures is that investors reinforce their prior beliefs
in virtual communities and they become more overconfident or feel more competent. Increased

overconfidence subsequently induces investors to trade excessively and lowers their performance
levels.
To test this conjecture, we include both overconfidence and perceived competence
measures in our empirical model. We estimate overconfidence using a mis-calibration measure
that is based on an investors estimates of the upper and lower bounds of 90% confidence
intervals for the return and value of a stock in the future (Glaser et al. 2004, Hilton 2001). And
we measure investors perceived competence based on their responses to three items in the
questionnaire: (1) I am comfortable with my ability to understand investment products,
alternatives and opportunities, (2) I am competent in my ability to invest successfully, and (3) I
am skillful in investment activities (e.g., stock picking, position, volume) 2.
To examine the relation between confirmation bias and overconfidence/perceived
competence (see Figure 3), we estimate the following system of equations:
Trading Frequency System of Equations:
(2a)
Miscalibration,   ConfirmationBias,  ,
(2b)
Competence,   ConfirmationBias,  ,
(2c)
TradingFrequency,   Miscalibration,  Competence,  Age,  % Sex, 
* Eduation,  - Income, 

,

The Miscalibrationi,t variable in equation (2a) is the difference between an investors

upper and lower bounds of the 90% confidence interval for the stock return during the one
month period following the participation date. Motivated by Glaser and Weber (2007), we use

A seven-point Likert scale is used to collect the responses.

this measure as a proxy for the degree of investor overconfidence. The Competencei,t variable in
equation (2b) is motivated by Graham et al. (2009).

It captures the degree of perceived

competence of an investor in his abilities or skill related to stock investing. In equation (2c), we
incorporate these two measures to investigate how overconfidence and perceived competence
influence investor trading behavior.
Last, using a similar structure, we use the following system of equations to examine how
confirmation bias affects an investors expectations about future stock return:
Expected Return System of Equations:
(3a)
Confirmation bias equation: same as equation (1a);
(3b)
R
,   ConfirmationBias,  Age,  % Sex,  * Eduation,  - Income,
In equation (3b), R
, is the expected stock return of investor i for time period t+1.
Since we only replace the actual return with expected return as the new final dependent variable,
other variables in the specification remain the same as in the actual return system of equations
(see system of equations (1)).

4.2 System of Equation Estimation Results: Tests of the Main Hypotheses

We begin our formal empirical analysis by estimating the first system of equations that examines
the link between confirmation bias and realized performance. The estimation results are
presented in Table 2. When we estimate the model without any moderating or control variables,
we find that an investors strength of belief has a significant impact on the degree of
confirmation bias (see Column (1)). Further, confirmation bias positively affects trading
frequency, while trading frequency adversely affects realized performance (see Columns (2) and
(3)). Both these effects are statistically significant and are consistent with our first two
hypotheses (H1 and H2).

Next, we consider the impact of various moderating variables, including perceived

knowledge, online investing experience, and investment amount in a stock. We include the
moderating variables in equation (1a) and also add various control variables such as gender, age,
education level, and income to equation (1c). The estimates reported in Columns (4)-(6) indicate
that the coefficient estimates of our key independent variables remain qualitatively similar as we
add the moderating and control variables into the system of equations.
Specifically, we find that the impact of the strength of beliefs on confirmation bias is
positive at the 10% significance level (coefficient estimate = 0.260). This evidence is consistent
with our first hypothesis, which posits that investors with stronger beliefs exhibit stronger biases
in processing information from virtual communities. These investors are more likely to seek and
accept confirming evidences from stock message boards and ignore opinions that contradict
with their beliefs.
We also find that the impact of confirmation bias on trading frequency is positive at the
1% significance level (coefficient = 0.185), while the impact of trading frequency on realized
return is negative and statistically significant (coefficient = -0.079). This evidence supports our
second hypothesis (H2) and indicates that investors with stronger confirmation bias are likely to
trade more frequently, which has a detrimental effect on performance. The result is consistent
with the view that confirming opinions reinforce investors prior beliefs, which makes them
more overconfident and induces them to trade excessively. This behavior in turn leads to lower
investment performance.
The effect of confirmation bias on realized performance is economically significant. We
observe a considerable shift in realized return as an investors confirmation bias changes. When
an investors confirmation bias increases by one standard deviation, the realized stock return
decreases from its mean value of -4.92% to -5.42%. This evidence indicates that the average
investment loss of an investor increases from 494,108.37 ($407.04) to 544,322.63 ($448.41)

when his confirmation bias increases by one standard deviation. Thus, participation in virtual
communities can make investors considerably worse-off because it amplifies their information
processing biases.
The model fit statistics in Table 2 indicate that our system of equations model is wellspecified. Our model explains 7.54% of variance in confirmation bias, 3.40% of variance in
trading frequency, and 7.70% of variance in realized performance. Further, the 2/df (Marsh and
Hocevar 1985) of the model is 3.24, where a small value represents better model fit.

Also, the

root mean square error of approximation (RMSEA) is less than 0.08, suggesting a reasonable fit
of the model in relation to the degrees of freedom (Browne and Cudeck 1993). Other model fit
indices also suggest a reasonable fit (>0.90).
To investigate whether the decline in performance can be attributed to an increase in
investor overconfidence and/or perceived competence, we estimate the second system of
equations. The results reported in Table 3 are consistent with our conjecture that confirmation
bias increases both investor overconfidence as well as perceived competence, which leads to
increased levels of trading. The evidence reported in Columns (1) and (4) of Table 4 indicates
that the impact of confirmation bias on mis-calibration is negative and statistically significant at
the 10% level (coefficient = -0.109). This finding indicates that confirmation bias makes
investors overconfident and consequently their confidence intervals become tighter. We also find
that confirmation bias has a positive and significant impact on perceived competence (coefficient
= 0.263), which indicates that investors who confirm their prior beliefs using information from
virtual communities have higher levels of perceived competence in their investment abilities.
The estimates from the trading frequency equation reported in Column (6) show that the
impact of mis-calibration on trading frequency is significantly negative (coefficient = -0.237),
while the impact of perceived competence on trading frequency is significantly positive
(coefficient = 0.254).

This finding indicates that both overconfidence and perceived

competence affects trading frequency, although the effect of perceived competence is marginally
higher. The evidence is consistent with recent studies which demonstrate that trading frequency
is explained more by perceived competence effect than due to mis-calibration (e.g., Graham et al.
2009).
Examining the goodness of fits for the models with mis-calibration and perceived
competence measures, we observe that the model explains 1.20% of variance in mis-calibration,
6.90% of variance in perceived competence, and 14.40% of variance in trading frequency. The
2/df of the model is 3.16 and the root mean square error of approximation (RMSEA) is 0.067.
All model fit statistics point to a reasonable model fit.
In the next set of tests, we gather support for the third hypothesis. Specifically, we
estimate the expected return system of equations (see equation set (3)) and report the results in
Table 4. The estimates in Column (2) of this table indicate that the impact of confirmation bias
on expected return is positive and significant at the 1% level (coefficient estimate = 0.115). Thus,
consistent with our third hypothesis (H3), confirmation bias increases investors expectation of
investment performance. Investors who exhibit stronger biases in processing information from
virtual communities have more distorted expectations about their investment performance.
Again, the model fit statistics suggest a reasonable overall model fit. Specifically, we find that
8.53% of variance in expected returns can be explained by the model. The 2/df of the model is
3.41 and the root mean square error of approximation (RMSEA) is 0.069.
The impact of confirmation bias on the investment expectation is economically large.
The coefficient estimates reported in Table 4 indicate that when an investors confirmation bias
increases by one standard deviation, their stock return expectation increases from its mean of
28.43% to 32.01%. In monetary terms, this evidence indicates that when an investors
confirmation bias increases by one standard deviation, the investor expects that his investment
gain on average will increase from 2,852,170.25 (= $2,349.59) to 3,214,717.24 (= $2,648.26).

4.3 System of Equations Estimation Results: Tests of Auxiliary Hypotheses

Next, we test our hypotheses related to the effects of moderating variables. Specifically, we
assess the influence of moderating variables on confirmation bias. The interaction term estimates
reported in Column (4) of Table 2 show that perceived knowledge has a significantly negative
moderating impact on the strength of belief-confirmation bias relation (coefficient estimate = 0.158). This evidence supports our fourth hypothesis (H4) and indicates that while investors with
stronger beliefs about a stock are more disposed to seek confirming evidences from message
boards, the extent of this behavior depends on the level of their perceived knowledge. The
strong belief about a stock has a greater impact on confirmation bias among investors with a low
level of perceived knowledge. In contrast, investors with a higher level of perceived knowledge
have less motivational need to confirm their own opinions using the opinions of other investors.
We also find that online investing experience and investment amount do not
significantly affect the strength of belief-confirmation bias relation. The impact of online
investing experience on confirmation bias is negative, while the impact of investment amount on
confirmation bias is positive, but both interaction term estimates are statistically insignificant.
Thus, our results do not support Hypotheses 5 and 6.

4.4 Discussion
Taken together, our empirical results provide strong support to our main hypotheses. The
strength of investors beliefs influences their degree of overconfidence and perceived
competence, which subsequently affects investors trading frequency and realized performance.
Investors with strong beliefs exhibit stronger confirmation bias and experience a larger adverse
impact on investment performance. These empirical findings provide interesting new insights
into the behavior of investors on stock message boards.
Our study also has several limitations. For example, we do not observe the antecedents
of investor beliefs. Such beliefs are potentially influenced by environmental factors such as stock

market movements, firms announcements, news reports and other information sources. Future
studies may develop a better understanding on how individuals form their beliefs on individual
stocks.

Further, our experiment provides new messages on top of each message board and

focus on how investors evaluate these messages.

Prior studies on confirmation bias however

reveal that confirmation bias not only influences individuals interpretation and evaluation of
information but also their information search process. Future studies are needed to develop a
better understanding on the influence of confirmation bias on both the information search
process and information processing behavior.
Third, our study is a cross-sectional experiment by analyzing investors with different
degrees of confirmation bias.

We do not observe how investors confirmation bias changes

over time and how such changes influence their investment expectation and actual performance.
Obtaining longitudinal data on individual behavior in virtual communities and their investment
performance will provide a more comprehensive understanding on the impact of confirmation
bias in virtual communities.

Finally, we conjecture that confirmation bias in virtual

communities leads to illusion of knowledge, which in turn leads to greater overconfidence. As

people reinforce their prior beliefs from virtual communities, they believe that they are more
knowledgeable than they really are. This illusion of knowledge ultimately makes them
overconfident (Barber and Odean 2001).

However, the studies in psychology and finance

literature posit that confirmation bias could directly lead to overconfidence (Daniel et al. 1998,
Barber and Odean 2001). Future research could identify other psychological biases among
message board participants and provide a more comprehensive understanding of decision
makers information processing behavior.

5. Summary and Conclusions

Virtual communities have attracted millions of retail investors and have aroused interests among
financial executives and academic researchers. Economic theory suggests that virtual

communities could be beneficial to investors as they provide more information with much lower
costs (Antweiler and Frank 2004, Tumarkin and Whitelaw 2001). Our results, however, suggest
that participation in virtual communities could adversely affect retail investors because they are
likely to exhibit biases in processing information gathered from these settings. In particular,
investors use the gathered information to confirm their prior beliefs. This confirmation bias
strengthens their prior beliefs and makes them optimistic as well as overconfident.
Subsequently, confirmation bias induces investors to form higher expectations about their stock
investments, but those investors experience lower actual performance.
This analysis takes a first step to understand how investors information processing
behavior influences their trading frequency, investment performance, and expectations.
Investors are known to demonstrate behavior biases in making investment decisions.

However,

little is known what contributes to such behavior biases. In this analysis, we show that an
important contributing factor is the manner in which investors process information from
external sources.
While our study focuses on investors information processing behavior on message
boards our finding may be generalized to other contexts where decision makers with prior beliefs
process new information before making decisions. For example, when a consumer who has a
strong preference about a particular brand visits online Word-Of-Mouth networks (e.g.,
consumers review websites), she is more likely to accept the opinions that are consistent with
her prior belief (Duan et al. 2009).

Thus, she may make purchase decisions based on her biased

beliefs, and overlook or underweight different opinions that are related to her purchase decisions.
Also, a project manager within organizations may participate in internal knowledge management
networks or project management networks where her team members discuss a projects
directions or decisions. However, if a project manager has a strong belief toward a particular way
of the project, she will be disposed to seek and accept opinions that are aligned with her own

opinions. Such behavior may lead to a projects failure. Our model thus provides a foundation
that can be extended to other contexts.
literature.

This study also informs the behavioral finance

While investor behavior biases such as overconfidence have been widely observed in

financial markets, less is known about the fundamental factors that generate those biases. Our
analysis reveals that confirmation bias in investors interpretation of information could be a key
factor that drives investor overconfidence.
In addition, our results could provide guidance to online investors and financial
infomediaries. Confirmation bias suggests that individuals become selective in information
acquisitions according to their prior beliefs. To be unbiased in information acquisition and
processing, investors should expose themselves to both confirming and disconfirming
information, objectively assess the relevance and veracity of new information, and integrate all
the information before making a decision. Message board operators can help investors by
presenting both confirming and disconfirming opinions and warning participants that the
selective information seeking may yield undesirable results.

References
References are omitted due to the page limit of WISE 2010.
Table 1: Summary Statistics
Panel A: Demographic Characteristics
Characteristic
Gender
Age

Education

Income

Category
Female
Male
< 25 years
25 to 35 years
36 to 50 years
51 to 65 years
> 65 years
High school
2-year college
4-year college
Masters degree
PhD degree
< 15 million won
15 to 50 million won
50 to 100 million won
100 to 200 million won

Sample (%)
12.97%
87.03%
14.88%
61.22%
23.69%
0.21%
0.00%
14.01%
16.77%
62.85%
5.73%
0.64%
29.30%
25.90%
32.70%
10.40%

>200 million won

1.80%

Panel B: Investment Behavior

Characteristic
Time Spent on Message
Boards Per Day

Category
< 1 hour
1 to 2 hours
3 to 6 hours
7 to 10 hours
> 10 hours
0
1 to5
6 to10
11 to20
>20
Strong Sell
Sell
Hold
Buy
Strong Buy
1
2
3
-3
-2
-1
0
1
2
3
< 5 million won
5~10 million won
10 ~ 50 million won
50~100 million won
> 100 million won
<1
1~2
3~6
7~10
> 10

Number of Postings

Current Opinion

Strength of Beliefs

Confirmation Bias

Investment Amount

Trading Frequency

Sample (%)
44.49%
32.06%
17.64%
4.81%
1.00%
48.06%
38.24%
6.75%
4.50%
2.45%
3.00%
5.20%
50.40%
27.70%
13.70%
50.40%
32.90%
16.70%
5.20%
4.20%
10.80%
17.90%
24.10%
13.30%
21.50%
36.53%
22.94%
33.53%
3.59%
2.99%
43.49%
25.45%
14.43%
8.42%
8.22%

Panel C: Other Variables

Mean

S.D

Min

Max

25th Pctl

Median

75th Pctl

Expected Return

28.43%

47.16

-100%

500%

10%

15%

30%

Actual Return

-4.92%

0.29

-82.81%

82.31%

-13.15%

-5.67%

5.33%

Miscalibration

4.96%

4.57

26.41%

1.57%

3.77%

6.41%

Min. Expected Return

8.91%

30.32

-120%

400%

0%

5%

10%

Max. Expected Return

49.66%

94.51

-50%

1000%

10%

25%

50%

Perceived Knowledge

4.10

1.09

3.50

4.50

Perceived Competence

4.12

1.17

3.33

4.67

Variables

Table 2: Realized Performance System of Equations Estimates

Independent Variables
Strength of Beliefsi,t

CBias
(1)

TrdFreq
(2)

Dependent Variable:
RealPerf
Cbias
(3)
(4)

0.234***
(5.332)

Strength of Beliefsi,t * Experience

Strength of Beliefsi,t * Inv Amount
Knowledge
Experience
Investment Amount
0.186***
(4.348)

Trading Frequencyi,t

0.185***
(4.309)
-0.075*
(-1.663)

Age
Gender
Education
Income
Adjusted R2
Number of Obs.
Model Fits
Chi-square/df
RMSEA
Normed Fit Index
Relative Fit Index
Tucker-Lewis Coefficient
Comparative Fit Index

5.54%
502

RealPerf
(6)

0.260*
(1.832)
-0.158*
(-1.663)
-0.050
(-0.378)
-0.028
(-0.156)
0.251**
(2.647)
0.052
(0.502)
0.130
(1.170)

Strength of Beliefsi,t * Knowledge

Confirmation Biasi,t

TrdFreq
(5)

3.40%
502

6.20%
502

4.97
0.089
0.762
0.207
0.246
0.774

7.54%
502

3.40%
502

-0.079*
(-1.750)
0.030
(0.587)
-0.021
(-0.433)
0.014
(0.302)
-0.050
(-0.977)
7.70%
502

3.24
0.067
0.937
0.906
0.933
0.955

Table 3: Trading Frequency System of Equations Estimates

Dependent Variable:
Independent Variables
Confirmation Biasi,t

MisCal

Comp

TrdFreq

MisCal

Comp

TrdFreq

(1)

(2)

(3)

(4)

(5)

(6)

-0.109*

0.263***

-0.109*

0.263***

(-1.926)

(5.900)

(-1.927)

(5.895)

Miscalibrationi,t

-0.237***

-0.237***

(-5.046)

(-4.721)

Competencei,t

0.254***

0.254***

(5.839)

(5.501)

Age

0.044
(0.926)

Gender

0.063
(1.432)

Education

-0.078*

Income

-0.108**

(-1.762)
(-.2.251)
Adjusted R

Number of Obs.

1.20%

6.90%

13.70%

1.20%

6.90%

14.40%

502

502

502

502

502

502

Model Fits
Chi-square/df

4.38

3.16

RMSEA

0.082

0.067

Normed Fit Index

0.919

0.804

Relative Fit Index

0.594

0.607

Tucker-Lewis Coefficient

0.829

0.812

Comparative Fit Index

0.931

0.847

Table 4:: Expected Return System of Equations Estimates

Dependent Variable:
Independent Variables
Confirmation Biasi,t

ExpRet

ExpRet

(1)

(2)

0.116***

0.115**

(2.516)

(2.526)

Age

0.024
(0.316)

Gender

-0.028
(0.601)

Education

0.016
(-1.349)

Income

-0.033
(-1.199)

Adjusted R

7.92%

8.53%

502

502

Chi-square/df
square/df

9.65

3.41

Number of Obs.
Model Fits
RMSEA

0.131

0.069

Normed Fit Index

0.781

0.945

Relative Fit Index

0.315

0.913

Tucker-Lewis
Lewis Coefficient

0.364

0.937

Comparative Fit Index

0.733

0.960

Figure 1: The Main Conceptual Model

Figure 2:: A Snapshot of Experiment on Naver Stock Message Boards

Figure 3: Conceptual Model for Overconfidence and Perceived Competence as

Mediating Variables