Rumortrage: Can Investors Profit on Takeover Rumors on Internet Stock Message Boards?

Jenni Bettmana, Aiden Hallettb, Stephen Saulta*

a School of Finance, Actuarial Studies and Applied Statistics, Research School of Business, Australian National

University, Canberra, ACT, 2000, Australia b Goldman Sachs, Sydney, NSW, 2000, Australia

Abstract

We provide the first empirical investigation into the impact of Internet Stock Message Board

takeover rumors on the price discovery process in the United States equity market. Our investigation

involves using an innovative five-stage filtering process, that employs computational linguistics

methods, to derive a unique sample of 2,898 first post Message Board takeover rumors disseminated

during the period January 2003 to December 2008, inclusive. Overall, our analysis, utilizing intraday

Trade and Quote (TAQ) data, reveals that Message Board takeover rumors generate significant

positive abnormal returns and trading volumes. Finally, after verifying the robustness of these results

to the influence of sample selection bias and an alternative matched control firm abnormal returns

metric, we find that Message Board takeover rumors present an economically exploitable opportunity

for investors.

Keywords: Internet Stock Message Boards; Price Discovery Process; Takeover; Rumor; Computational Linguistics.

* Corresponding author: Tel. +61 2 6125 4869; fax +61 2 6125 0087; e-mail Stephen.Sault@anu.edu.au. We gratefully acknowledge the comments and suggestions offered by Vidhan Goyal, Bruce Grundy, Tom Smith, Garry Twite, participants of the Asian Finance Association Conference in Yokohama, ANU Finance Conference 2007, University of Melbourne Finance Seminar Series, and financial support provided by the Australian Prudential Regulation Authority and the Reserve Bank of Australia. All errors are our own.

2

I. Introduction With the ongoing advancement in information technology, equity market participants are

increasingly using the internet and, more specifically, Internet Stock Message Boards (Message

Boards) as a medium to disseminate and collect value-relevant information (see, for example, Le

Vallée, 1997;Tumarkin and Whitelaw, 2001; Tumarkin, 2002; and, Gruhl et al, 2005). Message

Boards are organised online forums enabling users to read and post information on specific firms

and investment-related topics (Wysocki, 2000). Das and Chen (2007) contend that Message

Boards allow all market participants, irrespective of size, to express their opinions and encourage

greater market participation by small investors. This paper examines a particular class of

information posted on Message Boards, namely takeover rumors and examines whether investors

can profit by trading on these rumors. In order to determine whether significant dollar profits are

attainable we examine the price and volume (to ensure there is sufficient liquidity in these stocks)

impact of these unofficial announcements and examine whether significant profits are obtainable

from trading on takeover rumor messages. In doing this, we employ an innovative application of

computational linguistics to obtain our rumor sample.

Essentially, Message Boards assist the price discovery process and market efficiency by

enhancing the speed and accessibility of information to a wider range of market participants.

There are many types of online discussion forums ranging from real-time chat rooms, where

historical discussion archives are unavailable, through to Message Board forums which allow

retrieval of both past and present message posts (see, for example, Tumarkin and Whitelaw,

2001). Furthermore, these Message Boards can be either public or private, with public Message

Boards providing access to the largest user base.

The most popular public Message Board websites pertaining to the US equity market include:

Yahoo! Finance; RagingBull.com; The Motley Fool; and, SiliconInvestor.com (Wysocki, 2000;

3

Tumarkin and Whitelaw, 2001; Das and Chen, 2007; and, Antweiler and Frank, 2004). Yahoo!

Finance is the most widely utilised Message Board website, with over five times the number of

posts of the other Message Boards combined (Thomas, 2003), and over 50,000,000 first posts

during our observation period of 2003 to 2008. A wide range of market participants are known to

use these specific Message Boards, including: small investors (see, for example, Das et al, 2005);

day-traders (see, for example, Wysocki, 1998; and, Koski et al, 2004); professionals/institutions

(see, for example, Bagnoli, 1999); and, corporate insiders (see, for example, Carson and Felton,

2004).

The growth in the use of internet Stock Message Boards is fuelled by the comparative

advantages this medium offers over more traditional mediums in terms of instantaneous

information dissemination, communication facilitation and anonymity. By increasing the

availability of value-relevant information, and the speed at which it is incorporated into security

prices, appropriate dissemination of information via Message Boards should enhance financial

market efficiency. Conversely, this form of communication may also be utilised to fictitiously

manipulate market activity by Message Board participants for self-serving purposes. For both

reasons, it is crucial for all financial market stakeholders and regulators to enhance their

understanding of the impact of Message Boards and the various types of information

disseminated via this medium.

Wysocki (2000) and Das et al (2005) identify various categories of information disseminated

via Message Boards, including: repetition of prior news releases and related analysis; individual

opinions and predictions; private stock analysis; fictitious information; foreshadowing of future

press releases; information leaks; and, other unsubstantiated rumour information. This latter

Message Board rumour information category can be further disaggregated by the type of rumour.

4

Amongst others, these rumour information types include: ‘whispers’;1 dividend forecasts; Chief

Executive Officer (CEO) succession hearsay; operation expansion anticipation; and, takeover

activity speculation. In this paper, we explore the impact of the release of these takeover rumors

via Message Boards on US equities.

‘Rumours’ are a class of public information that is unable to be accurately and objectively

verified (Knapp, 1944; Banerjee, 1993; Kosfeld, 2005; and, Clarkson et al, 2006). By nature, a

rumour consists either of unsubstantiated information that is fictitious, an individual’s

interpretation/perception, or inside/private information deliberately leaked to the public.2

Dispersion of rumours via Message Boards is particularly attractive given their capacity to

expedite widespread dissemination of posted information in a largely anonymous and unregulated

environment (Clarkson et al, 2006).

However, these inherent features that make Message Boards attractive sources for rumour

dissemination can also detract from the credibility of the information provided. While Bagnoli et

al (1999) criticise the transparency of Message Board information, Kosfeld (2005) suggests that

the credibility regarding the accuracy of a rumour increases in line with the frequency a rumour is

communicated and the number of dissemination sources. In combination with the highly

competitive nature of financial markets (where information is of utmost value in gaining

comparative advantage) this theory suggests that the mere existence of a rumour on a Message

Board can potentially result in a substantial impact on financial market activity. Moreover,

regardless of the direction of causation between Message Board activity and market activity, the

credibility of a Message Board rumour is enhanced by prior or subsequent market activity (see,

for example, Kosfeld, 2005; and, Clarkson et al, 2006).

1 ‘Whispers’ are defined as rumour forecasts of future corporate earnings figures that are primarily disseminated via electronic communication channels such as email and Message Boards (see, for example, Bagnoli et al, 1999; Abraham, 2005; and, Bhattacharya et al, 2006). 2 Information leakage involves the unauthorised release or signalling of private information, pertinent to the value of a firm, to a subset of public market participants (de Bodt et al, 2002; Goenka, 2003; and, Brunnermeier, 2005).

5

Furthermore, although rumours cannot be objectively verified, the mere potential that they

may contain private or leaked corporate information provides sufficient incentive to motivate

market participants to acknowledge, access and possibly act on their content. The Information

Leakage Hypothesis suggests that informed market participants have a profit-maximising

incentive to release private information in order to capitalise on the likely market price reaction

that will occur once the private information becomes publicly available (de Bodt et al, 2002). In

particular, the impact of the announcement of takeover activity on a target’s stock price has been

well documented as having a substantially positive effect (see, for example, Bradley et al, 1983;

Jensen and Ruback, 1983; Bradley et al, 1988; and, Anderson et al, 1994).

Due to this seemingly predictable share price reaction, a substantial body of literature has

documented significant stock price run-ups prior to takeover announcements for target firm

shares (see, for example, Keown and Pinkerton, 1981; Arshadi and Eyssell, 1993; and, Jabbour et

al, 2000). This widely observed phenomenon lends ex-post weight to the argument that

information leakage and market anticipation facilitated by takeover rumours exert a substantial

impact on equity markets (see, for example, Jarrell and Poulsen, 1989; Aitken and Czernkowski,

1992; Jabbour et al, 2000; and, Athanasoglou et al, 2005).

However, there are conflicting views on the longer-term impact of takeover rumours on

shareholder wealth maximisation and market efficiency. Meulbroek (1992), and Cornell and Sirri

(1992) argue that information leakage and insider trading results in the incorporation of private

information into stock prices. This ultimately facilitates a more rapid and informative price

discovery process, thereby achieving greater informational efficiency in financial markets. In

contrast, Meulbroek and Hart (1997) argue that pre-announcement market rumours relating to

target firm share price effects impede the functionality of the market for corporate control. This

occurs by increasing required bid premiums and consequently pricing out many potential

acquirers.

6

Overall, it is generally acknowledged within the literature that rumours offer socially

recognised and accepted mechanisms for disseminating leaked insider/other private information,

providing an imprecise signal regarding the foreshadowing of potential corporate events and

announcements (Brunnermeier, 2005). Traditionally, print media represents a primary channel

used to publicly disseminate rumours (Aitken and Czernkowski, 1992). More recently however,

Message Boards are becoming the leading medium through which rumours are now divulged due

to their highly accessible, anonymous, and timely characteristics (see, for example, Das and

Chen, 2007; Brunnermeier, 2005; and, Clarkson et al, 2006).

The connection between the role of Message Boards and financial market activity stimulation

relates to the Theory of Communication (see, for example, Antweiler and Frank, 2004). In

particular, Message Boards facilitate interactive communication, and subsequently, belief

formation amongst market participants. DeMarzo et al (2001) suggest that market participants

overvalue the opinions of those with whom they interactively communicate when forming their

own personal beliefs. Cao et al (2002) also advocate that many potential traders would not trade

on information signals if they were unable to reinforce the interpretation of these signals through

communication with other participants. Accordingly, market participants have an incentive to

create and access Message Board posts due to the potential impact this information may have on

shaping financial markets.

While the above discussion suggests that Message Board interaction influences market

equilibrium and the price discovery process,3 the Noise Trading Theory contends that Message

Board information may simply be ‘noise’ that does not provide any new or price predictive

information to the market (see, for example, Dewally, 2003; Antweiler and Frank, 2004; and,

Koski et al, 2004). Specifically, Koski et al (2004) suggest that Message Board information

3 Shleifer and Summers (1990) characterise ‘noise trading’ as volatility in share prices caused by trade based on beliefs or sentiment that is unjustified by fundamental information. As such, this trading activity may be the result of systemic biases (see also, Menkhoff, 1998).

7

increases stock return volatility but does not contribute to predicting stock returns in any

fundamental way. As such, while this theory acknowledges that Message Boards significantly

impact market volatility, it does not recognise that Message Board information influences the

fundamental price discovery process. Further, Bagnoli et al (1999) criticize the credibility of

information provided on Message Boards, noting that they can contain bias, and frequent lack of

justification for the rumors posted on the sites examined.

There is also considerable debate within the literature regarding the direction of causation

between Message Board activity and related market activity (contrast, for example, Zivney et al,

1996; Tumarkin and Whitelaw, 2001; Antweiler and Frank, 2004; and, Koski et al, 2004). In the

event that Message Board activity proactively causes subsequent market activity, it follows that

information disseminated on Message Boards can be used to predict future stock returns and

volume. Conversely, if Message Board activity is simply reactive to prior market activity, then

according to concepts of market efficiency, these postings provide information that is

insignificant to the market and irrelevant in the price discovery process. While theories regarding

the equity market impact of Message Boards are mixed, there is general consensus within the

literature that this medium does provide a legitimate means of information dissemination which is

becoming increasingly important on a domestic and global scale (Ahmed et al, 2003).

Empirical evidence pertaining to the Message Board and equity market activity relationship is

also mixed. Wysocki (1998) finds that Message Board posting activity is significantly related to

prior news and earnings announcements disseminated via non-Message Board communication

mediums. However, in later research Wysocki (2000) reveals that Message Board activity

substantially impacts market activity, and consequently can be used to predict market trading

volume. Furthermore, Wysocki (2000) finds that information contained in Message Board posts

is rapidly reflected in trading volumes and concludes that a two-way causational relationship

8

exists. Specifically, Message Board activity rapidly reflects market activity and conversely,

market activity tends to rapidly reflect Message Board activity.4

In contrast, Tumarkin and Whitelaw (2001) investigate the impact of Message Board

information on stock returns and volume for technology industry firms between April 1999 and

February 2000. Their findings indicate that, while Message Board activity does not influence

technology firm returns or trading volume, it appears that stock returns tend to influence Message

Board activity. Irrespective of direction of causation, these observations provide evidence that

Message Boards are used by market participants to assist in the formation of value beliefs and

efficient price discovery (Tumarkin and Whitelaw, 2001). Indeed, in testing for the information

relevance of Message Board posts, Antweiler and Frank (2004) reject the Noise Theory

Hypothesis by revealing that the level of disagreement among message postings and volume of

overall Message Board activity helps to predict subsequent stock trading behaviour.5 Antweiler

and Frank (2004) also indicate that Message Board posts provide new information to the market

not found in other information mediums, such as the Wall Street Journal (WSJ).

The motives driving Message Board participation represent another key area of interest

within the empirical literature. In this regard, Das et al (2005) obtain survey findings that uncover

a primary reason motivating participants to access Message Boards is to compare interpretations

and learn from other participants’ opinions. This finding concurs with Baker (1984) who argues

that people engage in interactive communication to minimise the costs of bounded rationality

and/or as a result of self-serving, opportunistic motivations. Consistent with Antweiler and Frank

(2004), Das et al (2005) also find that, in many instances, value-relevant news is frequently

disseminated on Message Boards before the news is available via more traditional

4 Wysocki’s (2000) findings are based on message post observations for the top 50 most active stock forums on the Yahoo! Finance Message Board website between January 1998 and August 1998. 5 Antweiler and Frank (2004) examine the Yahoo! Finance and RagingBull.com Message Board websites (using computational linguistics methods), in 2000, in order to analyse sentiment expressed in 1.5 million message posts relating to the 45 stocks listed in the Dow Jones Industrial Average and Internet Index.

9

communication channels. This information often foreshadows price-relevant corporate

announcements and press releases, and is frequently the result of posters’ unintentional exposure

to non-public information that they subsequently post on relevant Message Boards. Furthermore,

Jones (2006) finds significant increases in a firm’s daily trading volume directly following the

creation of the specific firm’s Yahoo! Finance Message Board forum. This evidence provides

support for the theory that Message Boards induce new participants into the equity market and/or

result in increased activity amongst existing participants.

Gu et al (2006) find that the average sentiment contained in posts on Message Boards,

weighted by the message author’s reputation, also has both statistically and economically

significant predictive influences over future stock prices. This result indicates that studies which

find that Message Board information entails insignificant predictive power over future stock

prices may be due to a methodological flaw that equally weights the sentiment expressed in

message posts, irrespective of the poster’s credibility (see, for example, Wysocki, 1998; and,

Tumarkin and Whitelaw, 2001). By further testing a series of simple trading strategies based on

weighted average Message Board sentiment, Gu et al (2006) reveal their ability to generate

abnormal returns, providing further evidence that new value relevant information is generated via

Message Boards.

Overall, the aforementioned empirical findings emphasise that Message Boards appear to

successfully expedite the dissemination of financial market information and facilitate

communication amongst market participants (see, for example, Wysocki, 2000; Tumarkin and

Whitelaw, 2001; Antweiler and Frank, 2004; Das et al, 2005; Jones, 2006; and, Gu et al, 2006).

While the evidence generally indicates Message Board activity has some form of impact on

equity markets, inconsistent findings regarding the ability of Message Board information to

predict future price movements is potentially the result of studying aggregated message post

information, combined with inadequate methodological procedures (see, for example, Gu et al,

10

2006). Indeed, the extant research focuses on messages posts in aggregate to either ascertain the

impact of high frequency posting on a firm’s returns and volatility (see, for example, Antweiler

and Frank, 2002) or attempt to extract emotive content from message posts to identify buy or sell

signals (see, for example, Das and Chen, 2007; and, Tetlock et al, 2008). As such, disaggregating

Message Board information by type would allow for more accurate and detailed analysis. In light

of this, we present the first study to empirically test the impact of Message Board takeover

rumors on the price discovery process in the US. By identifying posts that explicitly discuss

companies that are takeover targets, we depart from the current extant Message Board literature

that concentrates on classifying the sentiment of posts as containing either ‘bad’ or ‘good’ news

(see, for example, Das and Chen, 2007; and, Tetlock et al, 2008), and are able to identify whether

the release of this unofficial information leads to positive abnormal returns consistent with the

extant takeover literature (see, for example, Bradley et al, 1987; Jensen and Ruback, 1983,

Bradley et al, 1988; Anderson et al, 1994, Keown and Pinkerton, 1981; and, Jabbour et al, 2000).

In addition, we offer several innovations to the extant literature. Specifically, we compile our

data set from a unique and contemporaneous sample of Message Board takeover rumors by

utilizing specially designed collection and filtering software that employs computational

linguistics methods. Furthermore, we employ multivariate analysis to gauge how specific

explanatory variables identified within both the Message Board and takeover literature impact

upon Message Board takeover rumors. We also verify the robustness of our results to sample

selection bias, and additionally, utilize an alternate abnormal return metric, which is calculated

using an approach influenced by Barber and Lyon’s (1997) matched sample methodology.

Finally, we explicitly examine whether an investor trading on the release of these rumors can

earn economically significant dollar profits.

Our results from the aforementioned analysis identifies that Message Board takeover rumors

substantially impact the US equity market during the observation period January 2003 to

11

December 2008, inclusive. Most notably, significant positive shareholder wealth and volume

effects are identified in post and surrounding rumor periods. Furthermore, multivariate analysis

reveals significant negative firm size and significant positive technology industry effects;

inferring small firms and technology sector firms are most affected by Message Board takeover

rumors. In addition we find that there are significantly positive wealth and volume effects

attributable to rumors with the highest star rating6. Our results are also robust to sample selection

bias and the return metric employed.

Despite this analysis, a question still remains as to whether an investor could obtain

economically significant profits from trading on these takeover rumors. This is especially

pertinent given that the SEC has continually expressed its concern that Message Boards may be

used to facilitate ‘pump and dump’ schemes where market manipulators (especially small size

stocks) deliberately ‘talk up’ a stock to create investor excitement leading to sharp price increases

to the benefit of the manipulator. One of the easiest ways to create excitement and increases a

company’s share price is to post a rumor regarding a potential takeover of the firm. Indeed, the

financial press has recently referred to this activity of trading on internet rumors as ‘rumortrage’.

Our inspection of the economic exploitability of Message Board takeover rumors concludes that

significant economic profits can be obtained by investors, with an average profit of $1,340 on a

$50,000 investment over a twenty-four-hour period.

The remainder of the paper proceeds on to discuss the sample identification process and our

methodological approach in Section II, and the data utilized in Section III. Section IV presents

the primary results of our investigation, and Section V concludes.

II. Takeover Rumor Acquisition and Methodology A. Takeover Rumor Acquisition and Filtering Process

6 Each first post rumor has an associated star rating. The star rating is based on the history of the person posting the rumor, as well as ratings provided by readers when the post is first made.

12

To investigate the impact of Message Board takeover rumors on the price discovery process

in the US equity market, we construct a unique sample of takeover rumors disseminated on the

public Yahoo! Finance Message Board between January 2003 and December 2008, inclusive.7 In

order to isolate the initial sample of takeover rumors from these sites, we employ a five-stage

process, which is graphically depicted in Figure 1.

[Insert Figure 1 About Here]

Consistent with Antweiler and Frank (2004), given the extremely large number of first-post

(over 50,000,000) and reply-post (over 480,000,000) messages archived on the public Yahoo!

Finance Message Board, we design a web crawler software program to automate the Message

Board first-post and reply-post downloading stage. A first-post message instigates a string of

related subsequent message posts, thereby representing the most accurate point-in-time when a

new takeover rumor is publicly disseminated on a Message Board (see, for example, Zhang and

Swanson, 2007).

While the aforementioned messages are downloaded in Hypertext Markup Language

(HTML), we convert all message posts into a plain text database to facilitate filtering and

assessment procedures. This process results in a total of 51,574,894 first-post messages and

481,037,394 reply-post messages. The plain text database contains key information relating to

each message post including: company ticker; forum name; date; posting time (accurate to the

minute); author; message post title; star rating; and, message post body text. As the first-post

message instigates a string of these related subsequent reply-post messages, we are able to keep

each string together to ensure we have a correct match between the reply-post messages and first-

post message.

7 Yahoo! Finance is widely considered to be among the largest and most prominent US Message Boards (see, for example, Antweiler and Frank, 2004; Das and Chen, 2007; and, Clarkson et al, 2006).

13

Following the plain text conversion, we initially commence the process of isolating

appropriate takeover rumors by electronically filtering the plain text database using key words

associated with takeover rumors, and date/time criteria. Specifically, we develop a program that

filters these message posts based on key words such as ‘takeover’, ‘acquisition’, ‘buyout’,

‘merger’, ‘target’, ‘bid’ and/or ‘rumor’. In addition, we only consider message posts between

January 2003 and December 2008 that are disseminated during the trading hours 10.30am to

3.00pm on trading days.8 An analysis of the data reveals that 80% of all first post takeover

rumors occur during these hours, allaying concerns regarding how representative our sample is of

the entire population. This process ultimately results in the identification of 173,798 first-post

messages and 731,795 reply-post messages. The large reduction in message posts is unsurprising

as a large number of the Yahoo! Finance Message Board forums relate to employment,

macroeconomic discussions, currencies, exchange traded funds and mutual funds.

Given the overwhelming volume of electronically filtered message posts, additional and more

precise automation of the takeover rumor classification process is desirable. We need only

determine if the first-post is a takeover rumor, as all subsequent reply-posts must relate to the

original post, and hence will continue the string discussion relating to the takeover rumor9. To

address this requirement, we employ a well established method of classifying Message Board

posts, namely, computational linguistics. Computational linguistics techniques employ natural

language algorithms to statistically classify text and its use in facilitating wide scale Message

Board studies is increasing in popularity in the extant literature (see, for example, Antweiler and

Frank, 2004; and, Zelikovitz and Hirsh, 2005). While a large body of literature has developed

computational linguistics techniques for text classification over the past 40 years (see, for

8 Restricting the observation time periods of takeover rumours ensures no intraday event windows are split over more than a single trading day, thereby minimizing problems associated with both compounding confounding events accruing during market close periods and abnormal market behaviour typically witnessed at market open and close. 9 Reply-posts that do not follow the string are moved to the correct string or form a new post by the forum master.

14

example, Vapnik and Chervonenkis, 1964; and, Chakrabarti et al, 1998), it is Chen et al (1999)

who are first to formally apply computational linguistics techniques to ascertain the sentiment of

messages posted on Message Boards. Subsequently, Das and Chen (2007) extensively develop a

number of computational linguistics techniques for direct application to Message Board text

classification. It is important to note that the use of computational linguistics within the Message

Board literature has focussed on classifying the sentiment of the posts themselves (see, for

example, Antweiler and Frank, 2004; Das and Chen, 2007; and, Tetlock et al, 2008). Specifically,

Antweiler and Frank (2004) classify posts to determine the bullishness of each stock message

board, while Das and Chen (2007) extract small investor sentiment from stock message boards.

In this paper we are not attempting to extract sentiment or classify a string of words as either

negative or positive. We use computational linguistics to filter our sample based on probabilities

in identifying strings of words which are likely to indicate a takeover rumor. As we are not trying

to classify the sentiment of the post, the use of computational linguistics to identify takeover

rumors is likely to result in a higher success rate than the 70-80% positive identification noted in

Antweiler and Frank (2004) and Das and Chen (2007).

To identify takeover rumor posts, we utilize the Naïve Bayes algorithm coding10 (see, for

example, Lewis, 1998; McCallum and Nigan, 1998; and, Antweiler and Frank, 2004). While a

thorough review of this method is beyond the scope of this paper, we do discuss the premise

behind the technique. As Antweiler and Frank (2004) discuss, the Naïve Bayes algorithm uses a

“bag of words” approach which assumes that occurrences of words are independent of each other,

hence the “naïve” algorithm. A training set is developed where the occurrence of particular words

of interest are identified in messages of a particular type and then in the messages of the anti-

10 For robustness and comparative purposes we also utilize Support Vector Machine coding (see, for example, Vapnik and Chervonenkins, 1964; Vapnik, 1995; and, Joachims, 1999). We utilize the software developed Joachims (1998) which can be found at http://svmlight.joachims.org/. Consistent with Antweiler and Frank (2004), results using this form of computational linguistics are consistent with the Naïve Bayes approach.

15

type. From this training set, conditional probabilities are then identified with the addition of logs

of odds-ratios used to classify messages by processing the message word by word.

For our Naïve Bayes coding we employ the Rainbow package developed by McCallum

(1996).11 The first step involves creating a training set by manually reading and classifying the

first 1,000 first-post messages lodged in 2003 as either positive (indicating the message post is a

valid takeover rumor) or negative (containing non-takeover related information). To address any

issues of look-ahead bias, this training set is then used to train the algorithm to classify all

subsequent postings. While the size of the training set may appear small, this is consistent with

the Message Board computational linguistics literature (see, for example, Antweiler and Frank,

2004; Das and Chen, 2007) which stresses the need to prevent the overfitting of data which leads

to poor out of sample performance. Of the 1,000 first-post messages, 19 are classified as positive

(indicating a valid takeover rumor), and 981 are classified as negative (containing non-takeover

related information).

The second step involves running the Rainbow package to analyse the messages in the

training set using the Naïve Bayes method with the number of words in the vocabulary restricted

to the top 1,000 words as ranked by the average mutual information with the class variable

(Antweiler and Frank, 2004). The final step sees all subsequent messages being evaluated by the

Rainbow package, with each message receiving a probability of being either a positive takeover

rumor or a negative (containing non-takeover related information) takeover rumor. We choose

the classification with the highest probability, resulting in 3,526 first-post messages being

identified as containing takeover rumors. We then manually read all 3,526 first-post messages

and identify 507 that do not contain takeover rumors. The 3,019 first post-messages which are

confirmed as takeover rumors represents an 85.62% accuracy in correctly identifying positive

11 The Rainbow package can be found at http://www.cs.cmu.edu/~mccallum/bow/ and is freely available for academic purposes.

16

takeover rumor messages. While this is a large number compared to the computational linguistics

literature (see, for example, Antweiler and Frank, 2004, Das and Chen, 2007), it is unsurprising

given that we are not performing an emotive classification. Of the remaining 170,272 messages

which the Rainbow software package classified as containing non-takeover related information,

manually reading the sample identified only 103 which were falsely classified and were actually

positive takeover rumors. This represents a 99.93% accuracy in terms of correctly identifying

messages with non-takeover related information. Again this is unsurprising as we are not

classifying the sentiment of the posts, and it is highly unlikely for a person to post a takeover

rumor without using key words such as: takeovers; mergers; acquisition, etc which the Rainbow

package would have identified. Together, these results demonstrate the value of using

computational linguistics on large text databases such as Message Boards to accurately identify

posts of a subject driven nature such as takeovers.

Subsequent to the selection of the final sample, further filtering is necessary to ensure several

additional requirements are satisfied. To facilitate data collection, we only include takeover

rumors pertaining to publicly listed target firms in the final sample. Moreover, in accordance with

the matched firm methodology we employ, the final sample only contains rumored target firms

for which a suitably matched control firm exists. Finally, for each rumored target and control firm

we require a complete data set to allow calculation of our abnormal return and abnormal trading

volume measures.

Based on the sample identification and filtering procedures outlined above, of the 3,122 first-

post takeover rumors identified, we obtain a final sample of 2,898 first-post takeover rumors

pertaining to 734 unique US publicly listed firms during the period January 2003 to December

2008, inclusive. Figure 2 depicts the distribution of takeover rumor dissemination by half hour

intraday time periods. Most notably, Figure 2 indicates the highest takeover rumor Message

Board activity between 10.30am to 3.00pm occurs in the 10.30am to 10.59am period and the

17

2.00pm to 2.29pm period. These intraday intervals correspond to the periods close to market open

and following the standard lunch period interval (ie. 12.00 pm to 2.00 pm). Furthermore, the

decline we observe in takeover rumor Message Board activity during the lunch period is

consistent with the general pattern extant studies find with respect to trading volume (see, for

example, Antweiler and Frank, 2004). Cumulatively, these 2,898 first-post messages have 19,863

reply-posts.

[Insert Figure 2 About Here]

We utilize the Event Study Methodology (Fama et al, 1969) to examine the impact of

Message Board takeover rumors on target firm returns and trading volume over a range of

intraday event windows detailed in Table I. Investigation of intraday intervals minimizes

distortions induced by confounding events, and provides a more precise insight into pre, post and

surrounding-rumor equity market impacts.

Furthermore, to facilitate the control firm benchmark methodologies utilized, we ascertain

matched sample control firms by initially revealing all firms with the same North American

Industry Classification System (NAICS) two-digit economic sector code. Subsequently, we apply

the Barber and Lyon (1997) matching criteria by further identifying all remaining firms with a

market value of equity between 70 to 130 percent of each rumored takeover target sample firm.

From this subset, the firms with the closest book-to-market ratio to the rumored target sample

firms are ultimately selected as the control firms.

[Insert Table I About Here]

B. Abnormal Price Return Calculations

18

To examine target shareholder wealth effects, we consider buy-and-hold average abnormal

returns using the traditional 0/1 Market Model approach to evaluate how the sample stock returns

perform relative to the market benchmark. Using this approach, we determine abnormal returns

as the average excess return of the rumored target firms over the market portfolio. Specifically,

we calculate the 0/1 Market Model Buy-and-Hold Average Abnormal Returns (BHAARs) as

follows:

[ , ] ,[ , ] ,[ , ] ,[ , ]1 1

1 1n n

q r i q r i q r m q ri i

BHAAR BHAR R Rn n

Where:

[ , ]q rBHAAR = The Buy-and-Hold Average Abnormal Return of all sample firms

over the event window [ , ]q r ;

,[ , ]i q rBHAR = The Buy-and-Hold Abnormal Return for sample firm i over the

event window [ , ]q r ;

,[ , ]i q rR = The buy-and-hold return for sample firm i over the event window

[ , ]q r ;

,[ , ]m q rR = The buy-and-hold return for the market portfolio over the event

window [ , ]q r ; and,

n = The number of firms in the sample.

C. Abnormal Trading Volume Calculations

To evaluate the impact of Message Board takeover rumors on target firms’ trading volumes, we

calculate a ratio of each target firm’s volume over the event window divided by the average

historical volume of the firm in a window of the same length. Historic trading volumes are

measured over the ten to thirty trading days prior to the release of the rumor. For example, if the

rumor is released at 11am on a Tuesday, for the [0,10] minute event window, the actual volume

(1)

19

within this window is divided by the average volume every Tuesday from 11am to 11:10am in

the ten to thirty trading days before the rumor is disseminated. In an effort to assess the presence

of abnormal trading activity, this data is applied to generate firm-specific volume benchmarks

that are compared to the realised trading volume over the event window of interest.

Therefore, abnormal trading for a company over the event window of interest is

calculated using the Cumulative Average Abnormal Trading Volume (CAATV) measure, which

is expressed as follows:

11

,

,

],[,1],[

ti

b

atti

bai

n

iba AV

VATV

nCAATV (2)

Where:

],[ baCAATV = The cumulative average abnormal trading volume over the event window

[a,b];

],[, baiATV = The abnormal trading volume for sample firm i over the event window

[a,b];

b

attiV , = The trading volume of sample firm i over the event window [a,b];

tiAV , = The average historical trading volume obtained by averaging the volume

that occurred at the same time and day of the week as the actual rumor

event in the ten to thirty days preceding the rumor; and,

n = The number of firms in the sample.

D. Multivariate Analysis

We argue that the impact of Message Board takeover rumors on a firm’s stock varies

depending on the interaction of industry-specific, firm-specific and rumor-specific variables. We

20

utilize multivariate regressions to analyze the combined influence of these variables on abnormal

returns and abnormal trading volumes. Specifically, we employ the following three-model

framework12 to investigate the cross-sectional variation of abnormal returns, trading volumes, and

bid-ask spreads in our takeover rumor sample for each event window and calculation

methodology:

Abnormal Return Multivariate Model

iiiirqi FIRSTTIMESTARTECHSIZELNAR 43210],[, )(

iiii TIMESPECCONFLICT 765

Abnormal Trading Volume Multivariate Model

iiiirqi FIRSTTIMESTARTECHSIZELNATV 43210],[, )(

iiii TIMESPECCONFLICT 765

Where: ,[ , ]i q rAR = The Abnormal Return for sample firm i over the event window

,q r ;

,[ , ]i q rATV = The Abnormal Trading Volume for sample firm i over the event

window ,q r ;

iLN SIZE = The natural logarithm of the market capitalization of firm i ;

iTECH = Dummy variable equal to 1 if sample firm i belongs to the

technology industry sector, and 0 otherwise;

iSTAR = Dummy variable equal to 1 if the rumor pertaining to the sample firm

i has a star rating of 4 or 5, and 0 otherwise;

iFIRSTTIME = Dummy variable equal to 1 if this is the first time the sample firm i

has been subject to a takeover rumor on the public Yahoo! Finance

Message Board, and 0 otherwise;

12 To address potential heteroskedasticity in the data, we estimate the regressions using the heteroskedasticity-consistent standard errors method prescribed by White (1980).

(3)

(4)

21

iCONFLICT = Dummy variable equal to 1 if more than 15% of the reply posts to the

initial takeover rumor on sample firm i are contradictory within the

sample window, and 0 otherwise;

iSPEC = Dummy variable equal to 1 if sample firm i has been subject to media

speculation regarding a takeover in the previous six months, and 0

otherwise;

iTIME = Dummy variable equal to 1 if rumor is disseminated during the

period 2006 to 2008, and 0 otherwise; and,

i and i = The error terms.

E. Robustness Measures

Barber and Lyon (1997) criticize the standard 0/1 Market Model event study methodology by

highlighting that this method inherently mis-specifies test statistics and therefore produces biased

results. To address this shortcoming, we employ an alternative abnormal returns event study

approach influenced by the Barber and Lyon (1997) Matched Sample Control Firm methodology.

We advocate that this Matched Sample Control Firm methodology more accurately accounts for

intervening factors, firm characteristics and risk by examining affected sample firms’

comparative performance against a control sample.

Furthermore, analysis of our rumored target firm sample may be subject to sample selection

bias in the event that the non-random characteristics of these firms, relative to the general

population of firms, increase their predisposition to takeover rumor dissemination (Heckman,

1979). This sample selection issue potentially hinders the transferability of inferences deducted

from our findings to the broader US equity market. Consequently, we employ the Heckman

(1979) two-stage method to test and correct for potential sample selection bias in our univariate

and multivariate analysis. In the first stage, we estimate a probit model over the full sample of

22

both event and control firms to measure the conditional probability of a company being included

in our sample (Inverse Mills ratio). The probit model is formally defined as:

Heckman Selection Equation (Probit Model)

Di = f (γ0+ γ1LN(SIZE)i + γ2TECHi + γ3TIMEi + γ4DEBTi + γ5ROAi

+ γ6BHRi + γ7PREVi + μi) (5)

Where:

Di = Dummy variable equal to 1 if firm i is a sample firm, 0 otherwise;

f = The normal link of the function;

iLN SIZE = The natural logarithm of the market capitalization of firm i ;

iTECH = Dummy variable equal to 1 if sample firm i belongs to the

technology industry sector, and 0 otherwise;

iTIME = Dummy variable equal to 1 if rumor is disseminated during the

period 2006 to 2008, and 0 otherwise;

DEBTi = Debt to Asset ratio of firm i prior to rumor dissemination;

ROAi = Return on Assets ratio of firm i prior to rumor dissemination;

BHRi = Pre-announcement abnormal returns of firm i, for a one week period

prior to rumor dissemination;

PREVi = Pre-announcement abnormal trading volume of firm i, for a one week

period prior to rumor dissemination; and,

μi = The error term with 0u and 1u .

The query year, defined as a dummy variable equal to 1 if the rumor is disseminated in the

2006 to 2008 the calendar years (and 0 otherwise), is selected to account for potential differences

in the use Message Boards over the time period. Pre-dissemination abnormal returns and trading

23

volume, the likely causes of takeover rumor speculation, are measured over a one-week interval

immediately prior to the rumor dissemination. In addition, the presence of firm leverage and

company profitability ratios in the selection equation account for firm-specific characteristics that

may influence the occurrence of a Message Board rumor.

Finally, we investigate whether trading on the dissemination of Message Board takeover

rumors yields economically significant profits after incorporating transaction costs and liquidity

constraints in the form of bid-ask spreads and the volume of shares traded. To this end, we extend

upon Bradley et al’s (1988) examination of the wealth effects from announced takeovers to the

market, by calculating the actual dollar profit attainable to investors. As the major participants of

Message Boards are personal investors or day traders we examine whether these smaller traders

could execute profitable trades on the dissemination of Message Board takeover rumors. This

strategy entails purchasing $50,000 of the rumored target sample company’s stock commencing

5-minutes after the dissemination of the rumor, and selling the stock commencing 24-hours after

the release of the rumor. The purchase of stock 5-minutes after the dissemination of the rumor

allows time for participants to read the rumor and then act on it if they choose to do so. In

purchasing the stock in the rumored company, we utilise TAQ data, allowing us to replicate the

trading behavior of investors by purchasing the exact volume of shares traded in each parcel at

the ask quote immediately preceding the trade, until we have purchased the required $50,000.

Then, 24-hours after the release of the rumor, we commence selling all shares originally

purchased at their relevant bid quote. Specifically, using TAQ data, we sell parcels of shares at

their bid price immediately preceding the trade, until all stock has been sold. Furthermore, by

purchasing and selling all shares at their relevant ask-and-bid-quotes immediately prior to each

trade, we largely incorporate the impact of transaction costs.

24

III. Data A. Data Collection

Once we have identified our sample we need to collect firm specific and market related data.

Specifically, we utilize TAQ data in investigating the impact of Message Board takeover rumors

on the US equity market. We obtain the reference date and time (to the minute) of takeover

rumor message posts from the original HTML message posts incorporated in our plain text

database. A summary of the abnormal return and trading volume data we analyze in this paper is

presented in Table II as follows:

[Insert Table II About Here]

To construct the matched control firm sample, we collect NAICS, market capitalization

adjusted for option dilution, and book-to-market values for all publicly listed US firms covered

by the Centre for Research in Security Prices (CRSP)/COMPUSTAT Merged Industrial Annual

database. Furthermore, the Standard and Poor’s 500 (S&P 500) market index is selected as the

appropriate market proxy to compare sample firm returns. Consistent with Antweiler and Frank

(2004), we proxy trade-by-trade returns on the market portfolio using the exchange-traded fund

that mimics the S&P 500 US market index that is available on the New York Stock Exchange

Trade-And-Quote (NYSE TAQ) database (ticker: SPY). To construct our returns on assets and

debt to asset ratios for our Heckman selection equation, we collect all necessary data from the

CRSP/COMPUSTAT Merged Industrial Annual database.

Finally, Table III outlines the data we use to construct the explanatory variables we employ in

our multivariate analysis methodology:

[Insert Table III About Here]

25

B. Descriptive Statistics

Characteristics pertaining to the 2,898 unique first-post rumored takeover target firms and

734 unique control firms are reported in Table IV. Sample firms are subject to 1.3458 first-post

Message Board takeover rumors on average and 1.0000 on median during the period January

2003 to December 2008, inclusive. With over three-quarters of the sample firms subject to 2.0000

or less first-post takeover rumors, it appears to be rare for a particular firm to be consistently the

subject of new Message Board takeover rumors. Furthermore, with rumored takeover firms’

market capitalizations ranging between US$3.10 million to US$250.3 billion, our sample covers

the entire size spectrum of rumored takeover targets.

[Insert Table IV About Here]

Table V reports summary statistics for the dummy and continuous variables we employ in our

multivariate analysis. Notably, there are more takeover rumors in the period 2006 to 2008 than

2003 to 2005. This suggests the use of Message Boards has increased over time. We observe a

substantial proportion of takeover rumor Message Board activity relates to the technology

NAICS industry sectors. Further, 30 percent of the rumors have star ratings in excess of 4, and

over 2/3 of companies are only subject to one takeover rumor. Finally, approximately 35 percent

of all first-post rumors have more than 15 percent of reply-posts disagreeing with the takeover

rumor, and over 20 percent of first-post rumors have been subject to media speculation regarding

a takeover in the previous six months.

[Insert Table V About Here]

26

IV. Results A. The Aggregate Shareholder Wealth and Trading Volume Effects

We commence our analysis by documenting the shareholder wealth effects of US Message

Board takeover rumors. Overall, the abnormal return results documented in Table VI provide

strong evidence that Message Board takeover rumors generate significant positive abnormal

returns in comparison to the market portfolio. Explicitly, abnormal returns relative to the market

portfolio are significant in all post and surrounding rumor event intervals observed; with the

exception of the one-minute rumor intervals. In particular, abnormal returns relative to the market

portfolio of 1.86 percent are observed in the twenty-four-hour period surrounding the initial

dissemination of Message Board takeover rumors, with 1.58 percent after the dissemination of

the rumor. These results indicate that, while there is no significant share price movement in the

period leading up to the rumor dissemination, substantially positive responses occur in the five-

minute to twenty-four-hour period post dissemination. This provides clear evidence that the

release of takeover rumors on Message Boards can have a significant impact on shareholder

wealth.

Table VI also details our findings relating to the trading volume impact of US Message Board

takeover rumors. The results are largely consistent with those found for the shareholder wealth

effects, with substantial increases in the volume of shares traded for takeover rumored firms than

their historical average in the post-dissemination and surrounding windows of five-minutes to

twenty-four-hours. An average firm has a significant increase in the volume of shares traded

ranging from 11 percent, five-minutes after the rumor is released, to 26 percent one-hour after the

rumor is released. The twenty-four-hour period after the rumor is disseminated sees an increase

in volume of 16.5 percent. Consistent with our abnormal return results, there is no significant

change in firm volume in any of the event windows leading up to the release of the takeover

rumor.

27

[Insert Table VI About Here]

B. Cross-Sectional Shareholder Wealth and Trading Volume Analysis

While the aforementioned results pertain to the overall impacts of Message Board takeover

rumors, we now consider the combined influence of specific variables that potentially explain

cross-sectional variations evident in these aggregated results.13 For brevity, we present cross-

sectional variation results for only the five-minute, thirty-minute and twenty-four-hour event

windows with regard to the various methodologies we employ in calculating abnormal returns

and trading volumes.14

Table VII presents cross-sectional variation results for the 0/1 Market Model abnormal return

approach we employ. Abnormal returns multivariate analysis reveals significant firm size15 and

star rating effects. A significant negative firm size effect is apparent in all event windows that

follow and surround rumor dissemination. This suggests the equity market response to Message

Board takeover rumor dissemination is substantially stronger for smaller firms. A negative firm

size effect is consistent with general observations apparent in broader extant US Message Board

studies (Bagnoli et al, 1999), thus highlighting that Message Boards provide an avenue of

information dissemination for small companies that suffer an inherent analysis coverage bias.

Furthermore, we find that those takeover rumors which have a higher star rating (specifically

a rating of four or higher), have significantly positive higher abnormal returns in all windows that

follow and surround the release of the rumor. This indicates that participants on Message Boards

13 For brevity, we do not report the results of our univariate stratification analysis. However, the qualitative findings are consistent with the multivariate analysis results. 14 It is important to note that, for brevity, we do not present tables for all event windows. However, we do provide a discussion of the main findings, including results pertaining to event windows not reported. 15 We considered the inclusion of analyst coverage as an additional explanatory variable. However, size and analyst coverage are highly correlated and as such we choose size as an explanatory variable to avoid potential multicollinearity issues. Nonetheless, we find testing using analyst coverage rather than size, results in no significant differences.

28

are more likely to act on takeover rumors that have been posted by other participants with a

reliable track record in terms of the quality of their posting.

[Insert Table VII About Here]

We also find that those firms with prior media speculation regarding a takeover have higher

abnormal returns in the twenty-four-hour window after the release of a rumor. This suggests a

larger price impact for those firms where there has been some corroborating media speculation

regarding the takeover.

Table VIII presents cross-sectional variation results for the five-minute, thirty-minute and

twenty-four-hour abnormal volume event windows. Across all post and surrounding event

windows there is a clear negative firm size effect, positive technology sector effect, and a positive

star rating effect. The firm size and star rating effects for our trading volume analysis are

consistent with the findings for our shareholder wealth analysis. In comparison to larger firms,

small firms subject to a takeover rumor have a greater percentage of their shares traded than their

historical average. In addition, those firms in the technology sector who have takeover rumors

disseminated also have a higher share turnover. Those firms with takeover rumors which have a

higher star rating have a significantly greater volume of shares traded than those with lower star

ratings.

Collectively, the findings from Tables VII and VIII suggest that those firms which are:

smaller; subject to takeover rumors with higher star ratings; in the technology sector; and, have

some corroborating prior media speculation regarding a takeover are likely to be the most

positively impacted in terms of shareholder wealth and volume of shares traded.

[Insert Table VIII About Here]

29

D. Robustness Analysis

We perform three forms of robustness analysis. Namely, we: evaluate an alternative Matched

Sample Control Firm abnormal returns methodology; investigate the robustness of our results to

sample selection bias; and, explicitly see if a profitable trading strategy can be formed based on

these takeover rumors. For brevity, we do not present all of our robustness results; however, we

do discuss all significant and overall findings.

We advocate that our Barber and Lyon (1997) influenced matched sample control firm

abnormal returns framework provides a more robust methodological approach than the 0/1

Market Model method. However, the results derived from this alternate abnormal return metric

are largely consistent with the 0/1 Market Model findings, providing strong evidence that

Message Board takeover rumors generate significant positive abnormal returns in comparison to

matched sample control firms. Specifically, all post and surrounding rumor intervals are

significant with the exception of the one-minute intervals. Finally, both the 0/1 Market Model

and our Matched Sample Control Firm abnormal return multivariate regressions provide largely

consistent cross-sectional findings.

Furthermore, we ensure the robustness of our aggregate and multivariate analysis results to

sample selection bias by utilizing the Heckman (1979) two-stage sample selection test and

adjustment procedure. Employing this procedure identifies the Inverse Mill’s Ratio is statistically

insignificant in all univariate and multivariate regressions capturing abnormal return and trading

volume impacts during the eighteen event windows we examine. As such, we can conclude that

Message Board takeover rumors are applied equally to all firms, and there is no sample selection

bias. Furthermore, subsequent to inclusion of the Inverse Mill’s Ratio, the statistical significance

of all original multivariate explanatory variables are consistent with the aforementioned findings.

Consequently, we conclude all aggregate and cross-sectional results presented in this paper are

robust to sample selection bias.

30

Finally, despite our findings which indicate that Message Board takeover rumors lead to

significant abnormal returns and trading volumes, a question still remains regarding whether an

investor could extract positive dollar profits from investing in these rumored companies. In an

attempt to answer this question, we formulate a trading strategy of purchasing $50,000 worth of

stock in these rumored companies, commencing five-minutes after the release of the rumor, and

selling these shares after a twenty-four-hour period. All shares are purchased at their relevant ask

quote immediately prior to the trade, and all shares are sold at their relevant bid quote

immediately prior to the trade. This allows us to consider transaction costs which may impact on

the profitability of the strategy.

Further, given our results that that those firms which are: smaller; subject to takeover rumors

with higher star ratings; and, are in the technology sector have greater positive impacts in regards

to wealth and volume effects, we also re-perform our trading strategy using different criteria.

Rather than purchasing shares in all rumored companies, we invest in only those companies with:

a market capitalization of less than $500 million; the associated rumor has a star rating of four or

five; and, the company is restricted to the technology sector.

Table IX reveals that a Message Board participant who trades on the release of a takeover

rumor can make significant profits over a twenty-four-hour period. Specifically, a participant

who invests $50,000 in any company which is subject to a Message Board takeover rumor can on

average earn a profit of $920, or a return of 1.84 percent over twenty-four-hours. Those investors

who limit their strategy to small technology companies who are associated with a high quality

rumor can on average earn superior profits of $1,340, or a return on investment of 2.68 percent

over twenty-four-hours. This indicates that there is sufficient liquidity in these smaller

technology companies to execute a profitable trading strategy.

[Insert Table IX About Here]

31

V. Conclusion The impact of Message Board takeover rumors on the price discovery process presents a

relatively new and highly dynamic facet to the publicly available information set. While the

extant literature provides convincing evidence that aggregated Message Board information exerts

a substantial influence on equity markets (see, for example, Antweiler and Frank, 2004), we

provide a disaggregated analysis of this Message Board activity by investigating takeover rumor

information to gain a more focused insight into the impact of this communication medium on

equity market activity. Furthermore, we consider a range of intraday event windows that allows

us to more accurately isolate the takeover rumor impact by minimizing the potential for

confounding events distorting equity market responses.

Overall, our findings suggest that Message Board takeover rumors impact on the US equity

market in the period January 2003 to December 2008, inclusive. We identify significant positive

abnormal returns and trading volumes in post and surrounding rumor dissemination periods. In

addition, we conduct multivariate analysis to examine the cross-sectional variations of the impact

of Message Board takeover rumors on the US equity market. Most notably, we document

evidence of significant: firm size; rumor rating; technology industry; and, prior media speculation

effects. In addition, these aggregate and cross-sectional findings are robust to sample selection

bias and an alternative Matched Sample Control Firm abnormal returns methodology.

Most notably, we find that a participant who trades on Message Board takeover rumors can

make significant profits in the magnitude of $1,340 on a $50,000 investment over a twenty-four-

hour period. Taken as a whole, our findings support the view that Message Board takeover

rumors provide an interesting and useful insight into the impact of an increasingly influential

channel of value-relevant information dissemination within contemporary financial markets. An

understanding of the shareholder wealth and trading volume effects of this type of takeover

rumor highlights substantial implications that should be given due consideration by financial

32

market participants, corporations and regulators when assessing and/or protecting the integrity

and efficacy of equity markets.

Further, the results of this study may have implications in the field of empirical finance research. If

events such as takeovers are anticipated, then announcement returns around actual events may be

understated. Taking into account this anticipation would be important for future event studies. Also,

Schwert (1996) documents substantial mark-up pricing in takeovers, which is usually a premium of the

previous months trading prices. As such, if rumors do lead to an increase in the share price prior to the

actual announcement of a takeover, do the eventual acquirers end up paying more for targets that were

subject to these rumors?

33

References Abraham, R., (2005), “An Exploration of Earnings Whispers Forecasts as Predictors of Stock Returns",

Journal of Economic Studies, vol. 32, pp. 524-539.

Ahmed, A., Schneible, R., & Stevens, D., (2003), “An Empirical Analysis of the Effects of Online

Trading on Stock Price and Trading Volume Reactions to Earnings Announcements”, Contemporary

Accounting Research, vol. 20, pp. 413-439.

Aitken, M. & Czernkowski, R. (1992), “Information Leakage Prior to Takeover Announcements: The

Effect of Media Reports”, Journal of Accounting and Business Research, vol. 23, pp. 3-20.

Anderson, D., Haynes, A. & Heaney, R., (1994), “Company Takeovers and Equity Returns: The Target

Size Effect”, Australian Journal of Management, vol.19, pp.1-30.

Antweiler, W. & Frank, M., (2004), “Is All That Talk Just Noise? The Information Content of Internet

Stock Message Boards”, Journal of Finance, vol. 59, pp. 1259-1294.

Arshadi, N. & Eyssell, T., (1993), “Insiders, Outsiders, or Trend Chasers?: An Investigation of Pre-

takeover Transactions in the Shares of Target Firms”, Journal of Financial Research, vol. 16, pp. 49-

59.

Athanasoglou, P., Asimakopoulos, I. & Georgiou, E., (2005), “The Effect of Merger and Acquisition

Announcement on Greek Bank Stock Returns”, Bank of Greece Economic Bulletin, vol. 0, pp. 27-44.

Bagnoli, M., Beneish, M. & Watts, S., (1999), “Whisper Forecasts of Quarterly Earnings Per Share”,

Journal of Accounting and Economics, vol. 28, pp. 27-50.

Baker, W., (1984), “The Social Structure of a National Securities Market”, American Journal of

Sociology, vol. 89, pp. 775-811.

Banerjee, A., (1993), “The Economics of Rumors”, Review of Economic Studies, vol. 60, pp. 309-327.

Barber, B. & Lyon, J., (1997), “Detecting Long-Run Abnormal Stock Returns: The Empirical Power and

Specification of Test Statistics”, Journal of Financial Economics, vol. 43, pp. 341-372.

Bhattacharya, N., Sheikh, A. & Thiagarajan, S., (2006), “Does the Market Listen to Whispers”, Journal of

Investing, vol. 15, pp. 16-26.

Bradley, M., Desai, A. & Kim, E., (1983), “The Rationale Behind Interfirm Tender Offers: Information or

Synergy?”, Journal of Financial Economics, vol.11, pp.183-206.

Bradley, M., Desai, A. & Kim, E., (1988), “Synergistic Gains from Corporate Acquisitions and their

Division between the Stockholders of Target and Acquiring Firms”, Journal of Financial Economics,

vol. 21, pp. 3-40.

Brunnermeier, M., (2005), “Information Leakage and Market Efficiency”, Review of Financial Studies,

vol. 18, pp. 417-457.

Cao, H., Coval, J. & Hirshleifer, D., (2002), “Sidelined investors, trade-generated news, and security

returns”, Review of Financial Studies, vol. 15, pp. 615-648.

Carson, J. & Felton, J., (2004), “New Technology Issues for Corporate Governance: Internet Message

Boards”, in: International Corporate Governance, 2nd Ed., pp. 184-194, Joseph C. F. Lufkin, ed.,

Euromoney Books.

34

Chakrabarti, S., Dom, B., Agrawal, R. & Raghavan, P., (1998), “Scalable Feature Selection, Classification

and Signature Generation for Organising Large Text Databases into Hierarchical Topic Taxonomies”,

The VLDB Journal, vol. 7, pp. 163-178.

Chen, M., Das, S., Tom, D. & Waddle, J., (1999), “On Small Investor Information: Feature Selection and

Categorization of Stock Message Board Postings”, Working Paper, UC Berkeley.

Clarkson, P., Joyce, D. & Tutticci, I., (2006), “Market Reaction to Takeover Rumour in Internet

Discussion Sites”, Journal of Accounting and Finance, vol. 46, pp. 31-52.

Cornell, B. & Sirri, E., (1992), “The Reaction of Investors and Stock Prices to Insider Trading”, Journal

of Finance, vol. 47, pp. 1031-1059.

Das, S. & Chen, M., (2007), “Yahoo! for Amazon: Sentiment Parsing from Small Talk on the Web”,

Management Science, vol. 53, pp.1375-1388.

Das, S., Martinez-Jerez, A. & Tufano, P., (2005), “eInformation: A Clinical Study of Investor Discussion

and Sentiment”, Financial Management, vol. 34, pp. 103-137.

de Bodt, E., Aktas, N. and Declerck, F., (2002), “Is there Information Leakage Around Business

Combinations on the French Market?”, Working Paper, EFMA 2002 London Meetings.

DeMarzo, P., Vayanos, D. & Zwiebel, J., (2001), “Persuasion Bias and Social Influence’, Working paper,

Stanford University.

Dewally, M., (2003), “Internet Investment Advice: Investing with a Rock of Salt”, Financial Analysts

Journal.03.23.0l, vol. 59, pp. 65-77.

Fama, E., Fisher, L., Jensen, M. & Roll, R., (1969), “The Adjustment of Stock Prices to New

Information”, International Economic Review, vol. 10, pp. 1-21.

Goenka, A., (2003), “Informed Trading and the 'Leakage' of Information”, Journal of Economic Theory,

vol. 109, pp. 360-377.

Gruhl, D., Guha, R., Kumar, R., Novak, J. & Romkins, A., (2005), “The Predictive Power of Online

Chatter”, Proceeding of the eleventh ACM SIGKDD international conference on Knowledge

Discovery in Data Mining, ACM Press: New York, NY, USA, pp. 78-87.

Gu, B., Konana, P., Liu, A., Rajagopalan, B. & Ghosh, J., (2006), “Predictive Value of Stock Message

Board Sentiments”, Working Paper, McCombs School of Business, The University of Texas at

Austin.

Heckman, J., (1979), “Sample Selection Bias as a Specification Error”, Econometrica, vol. 47, pp. 153-

162.

Jabbour, A., Jalilvand, A. & Switzer, J., (2000), “Pre-bid Price Run-Ups and Insider Trading Activity:

Evidence from Canadian Acquisitions”, International Review of Financial Analysis, vol. 9, pp. 21-43.

Jarrell, G, & Poulsen, A., (1989), “Stock Trading Before the Announcement of Tender Offers: Insider

Trading or Market Anticipation?”, Journal of Law, Economics and Organisations, vol. 5, pp. 225-248.

Jensen, M. & Ruback, R., (1983), “The Market for Corporate Control: The Scientific Evidence”, Journal

of Financial Economics, vol. 11, pp. 5-50.

Joachims, T., (1999), “Making Large-Scale SVM Learning Practical”, B. Scholkopf, C. Burges, A. Smola,

eds. Advances in Kernel Methods—Support Vector Learning. MIT Press, Cambridge, MA, pp. 243–

253.

Jones, A., (2006), “Have Internet Message Boards Changed Market Behavior?”, Info, vol. 8, pp. 67-76.

35

Keown, A. & Pinkerton, J., (1981), “Merger Announcements and Insider Trading Activity: An Empirical

Investigation”, Journal of Finance, vol. 36, pp. 855-869.

Knapp, R., (1944), “A Psychology of Rumour”, Public Opinion Quarterly, vol. 8, pp. 22-27. Kosfeld, M., (2005), “Rumours and Markets”, Journal of Mathematical Economics, vol. 41, pp. 646-664.

Koski, J., Rice, E. & Tarhouni, A., (2004), “Noise Trading and Volatility: Evidence from Day Trading and

Message Boards”, Working Paper, University of Washington.

Le Vallée, J., (1997), “Market Information Sources Available Through the Internet: Daily to Yearly

Market and Outlook Reports, Prices, Commodities and Quotes”, Working Paper, Michigan State

University.

Lewis, D.D., (1998), “Naïve (Bayes) at Forty: The Independence Assumption in Information Retrieval”,

in Proceedings of the 10th European Conference on Machine Learning (ECML).

McCallum, A., (1996), “Bow: A Toolkit for Statistical Language Modeling, Text Retrieval, Classification,

and Clustering”, School of Computer Science, Carnegie Mellon University.

McCallum, A. & Nigam., K., (1998), “A Comparison of Event Models for Naïve Bayes Text

Classification”, in AAAI-98 Workshop on Learning for Text Categorization, 1998.

Menkhoff, L., (1998), “The Noise Trading Approach--Questionnaire Evidence from Foreign Exchange”,

Journal of International Money and Finance, vol. 17, pp. 547-564.

Meulbroek, L., (1992), “An Empirical Analysis of Illegal Insider Trading”, Journal of Finance, vol. 47,

pp. 1661-1699.

Meulbroek, L. & Hart, C., (1997), “The Effect of Illegal Insider Trading on Takeover Premia”, European

Finance Review, vol. 1, pp. 51-80.

Tetlock, P.C., Saar-Tsechansky, M & Macskassy, S,. (2008), “More than Words: Quantifying Language to

Measure Firms’ Fundamentals”, Journal of Finance, vol. 63, pp. 1437-1467.

Thomas, D. J., (2003), “News and Trading Rules”, PhD Dissertation, Carnegie Mellon University.

Tumarkin, R., (2002), “Internet Message Board Activity and Market Efficiency: A Case Study of the

Internet Service Sector Using RagingBull.Com”, Financial Markets, Institutions and Instruments, vol.

11, pp. 313-335.

Tumarkin, R. & Whitelaw, R., (2001), “News or Noise? Internet Postings and Stock Prices”, Financial

Analysts Journal, vol. 57, pp. 41-51.

Schwert, G. William, (1996), "Markup pricing in mergers and acquisitions", Journal of Financial

Economics, vol. 41, pp. 153-192.

Shleifer, A. & Summers, L., (1990), “The Noise Trader Approach to Finance”, Journal of Economic

Perspectives, vol. 4, pp. 19-33.

Vapnik, V., (1995), “The Nature of Statistical Learning Theory”, Springer-Verlag, New York.

Vapnik, V. & Chervonenkis, A., (1964), “On the Uniform Convergence of Relative Frequencies of Events

to their Probabilities”, Theory of Probability and its Applications, vol. 16, pp. 264-280.

White, H., (1980), “A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for

Heteroskedasticity”, Econometrica, vol. 48, pp. 817-838.

Wysocki, P., (1998), “Cheap Talk on the Web: the Determinants of Postings on Stock Message Boards”,

Working Paper, University of Michigan.

Wysocki, P., (2000), “Message Boards Speak Volumes – and Volatility”, Futures, vol. 29, pp. 42-44.

36

Zelikovitz, S. & Hirsh, H., (2005), “Improving Text Classification Using EM with Background Text”,

Working Paper, Rutgers University.

Zhang, Y. & Swanson, P., (2007), “Issues in Creating a Sentiment Measurement Based on Internet Stock

Message Boards”, Working Paper, The University of Texas at Arlington.

Zivney, T., Bertin, W. & Torabzadeh, K., (1996), “Overreaction to Takeover Speculation”, Quarterly

Review of Economics and Finance, vol. 36, pp. 89-115.

37

FIGURE 1: Five-Stage Initial Event Sample Selection Process

Figure 1 diagrammatically depicts the five-stage process we employ to acquire the Message Board takeover rumor sample.

FIGURE 2: Intraday Distribution of Takeover Rumors by Time

Figure 2 depicts the distribution of the final sample takeover rumor dissemination by half hour intraday time periods.

Furthermore, Figure 2 also illustrates both the number of takeover rumors and the proportion of the total takeover rumor sample

occurring in each half hour period.

16.67%

11.80% 12.01%

7.18%6.73% 6.49%

8.52%

16.46%

14.15%

0.00%

2.00%

4.00%

6.00%

8.00%

10.00%

12.00%

14.00%

16.00%

18.00%

10.30am to 10.59am 11.00am to 11.29am 11.30am to 11.59am 12.00pm to 12.29pm 12.30pm to 12.59pm 1.00pm to 1.29pm 1.30pm to 1.59pm 2.00pm to 2.29pm 2.30pm to 3.00pm

Intraday Time Period

Proportion of Sample

483 rumors

342rumors

348 rumors

208rumors

195rumors

188rumors

247rumors

477rumors

410 rumors

Downloading

(Web crawler program)

Plain Text Database

(HTML conversion program)

Key Word Filtering

(Filtering program)

Computational Linguistics

(Naïve Bayesian)

Manual check and selection

38

TABLE I: Description of Event Windows

Table I outlines the intraday observation periods examined in this paper. The intraday windows are presented in minutes relative to the time the intraday Message Board rumor is disseminated (time 0), unless otherwise stated.16 Correspondingly, the various windows we utilize are classified into Pre-Rumor Windows (Panel A), Post-Rumor Windows (Panel B), or Surrounding-Rumor Windows (Panel C).

16 It is important to note that we examine 24-hour intraday event windows rather than one-day closing windows. This ensures that the observation periods are equally distributed on both sides of the intraday rumour release.

Event Window Description PANEL A: Pre-Rumor Windows

[-24 hrs, 0] The pre-24-hour period: 24 hours before the takeover rumor [-60, 0] The pre-60-minute period: 60 minutes before the takeover rumor [-30, 0] The pre-30-minute period: 30 minutes before the takeover rumor [-10, 0] The pre-10-minute period: 10 minutes before the takeover rumor [-5, 0] The pre-5-minute period: 5 minutes before the takeover rumor [-1, 0] The pre-1-minute period: 1 minute before the takeover rumor

PANEL B: Post-Rumor Windows [0, +1] The post-1-minute period: 1 minute after the takeover rumor [0, +5] The post-5-minute period: 5 minutes after the takeover rumor [0, +10] The post-10-minute period: 10 minutes after the takeover rumor [0, +30] The post-30-minute period: 30 minutes after the takeover rumor [0, +60] The post-60-minute period: 60 minutes after the takeover rumor

[0, +24 hrs] The post-24-hour period: 24 hours after the takeover rumor PANEL C: Surrounding-Rumor Windows

[-1, +1] The 2-minute period: 1 minute before to 1 minute after the takeover rumor [-5, +5] The 10-minute period: 5 minutes before to 5 minutes after the takeover rumor

[-10, +10] The 20-minute period: 10 minutes before to 10 minutes after the takeover rumor [-30, +30] The 60-minute period: 30 minutes before to 30 minutes after the takeover rumor [-60, +60] The 120-minute period: 60 minutes before to 60 minutes after the takeover rumor

[-24 hrs, +24 hrs] The 48-hour period: 24 hours before to 24 hours after the takeover rumor

39

TABLE II: Message Board Takeover Rumor Analysis Data Sources

Table II reports the data sources utilized in the calculation of abnormal return, trading volume and bid-ask spread impacts of Message Board takeover rumors.

Calculation Description Data Source

Abnormal Returns

We obtain intraday trade-by-trade share price data for affected sample firms, matched sample control firms and the S&P 500 market index proxy (ticker: SPY). We ensure the data we collect covers the 48-hour trading period surrounding each corresponding takeover rumor between January 2003 and December 2008, inclusive.

NYSE TAQ database

Abnormal Trading Volumes

We obtain intraday trade-by-trade volume data for affected sample firms. We ensure the data we collect cover the 48-hour trading period surrounding each corresponding takeover rumor between January 2003 and December 2008, inclusive, as well as the ten to thirty days prior to the dissemination of the rumor.

NYSE TAQ database

40

TABLE III: Cross-Sectional Explanatory Variables Data Sources

Table III reports the data sources utilized in the calculation and identification of cross-sectional explanatory variables employed in the multivariate analysis of the impact of Message Board takeover rumors.

Variable Description Data Source

LN SIZE To facilitate construction of the natural logarithm of firm size series, we measure ‘firm size’ as the firm’s market capitalization (data25, data199) adjusted for capitalization changes (data27).

CRSP/COMPUSTAT Merged Industrial Annual database

TECH

We classify sample firms into economic sector industry groups based on their first two-digit NAICS codes. There are 20 economic industry sectors under the NAICS. In order to construct the Technology (TECH) industry group dummy variable, we identify all firms in the final sample that belong to the Information (technology) NAICS economic sector industry group.

CRSP/COMPUSTAT Merged Industrial Annual database

STAR

The star rating of each first-post takeover rumor is recorded directly from the takeover rumor message post in our plain text database. We construct the STAR rating dummy variable using this plain text database to indentify rumors with a rating of 1 to 3 stars, and those with a rating of 4 or 5 stars. The star rating is based on the history of the person posting the rumor, as well as ratings provided by readers when the post is first made.

Message Board Takeover Rumor

Plain Text database

FIRSTTIME We classify rumors as to whether this is the first time the company has been subjected to a takeover rumor on the public Yahoo! Finance Message Boards using our plain text database.

Message Board Takeover Rumor

Plain Text database

iCONFLICT

To determine whether there is disagreement over a takeover rumor for a sample firm we utilise our plain text database. Specifically, we examine the reply-posts to ascertain the level of disagreement regarding a takeover rumor.

Message Board Takeover Rumor

Plain Text database

iSPEC

We determine whether a sample firm has been subject to takeover speculation in the financial press by identifying press articles and firm announcements occurring in the six months prior to rumor dissemination. Using Factiva, we search titles and lead paragraphs of key financial information channels (e.g. ‘PR Newswire (U.S.)’, ‘Reuters News’ and ‘The Wall Street Journal’) for the rumored target company’s name and/or ticker listings.

Factiva

iTIME

The date and time of each sample takeover rumor is recorded directly from the source takeover rumor message post in our plain text database. We construct the time period dummy variable using this plain text database to identify sample takeover rumors that are disseminated during the periods 2003 to 2005 and 2006 to 2008.

Message Board Takeover Rumor

Plain Text database

41

TABLE IV: Rumored Takeover Target and Control Firm Descriptive Statistics

Table IV summarizes firm characteristics pertaining to the rumored takeover target sample firms and the matched sample control firms. Specifically, Panel A and Panel B present descriptive statistics associated with the rumored takeover target sample and matched sample control firms pertaining to: the number of Message Board takeover rumors associated with each unique rumored target sample firm; market capitalization; and, book-to-market ratios. We calculate Market Capitalization as each firm’s number of shares outstanding multiplied by the closing share price (adjusted for dilution) at fiscal year-end immediately prior to takeover rumor dissemination. We calculate Book-to-Market ratios as the accounting book value of equity over the market value of equity, on an adjusted basis.

Measure Mean Standard Deviation

Minimum 25th Percentile

Median 75th Percentile

Maximum

PANEL A: Rumored Takeover Target Sample Firms

Number of Rumors

1.3458 1.1456 1.0000 1.0000 1.0000 2.0000 9.0000

Market Capitalization

(US$ mil) 4,945.1765 18,264.9755 3.1039 184.4579 596.4423 1,955.9751 250,318.3045

Book-to-Market 0.4456 0.8143 -6.1452 0.3159 0.5499 0.8903 3.3596

PANEL B: Matched Sample Control Firms

Market Capitalization

(US$ mil) 4,541.9654 17,744.1863 3.2369 158.4766 534.9547 1,989.2298 280,456.7842

Book-to-Market 0.4399 0.7122 -5.9955 0.2842 0.4977 0.7403 3.9855

42

TABLE V: Descriptive Statistics for Cross-Sectional Explanatory Variables

Table V presents the descriptive statistics for the dummy and discrete cross-sectional explanatory variables we employ in our multivariate analysis of the impact of Message Board takeover rumors. Panel A reports the number of takeover rumor observations pertaining to each dummy variable value and the corresponding proportion of the total sample this represents. Panel B outlines descriptive statistics pertaining to the continuous variable we employ in the multivariate analysis. With regard to reported explanatory variables: LN(SIZE) is the natural logarithm of the rumored takeover target firm’s market capitalization; TECH is a dummy variable equal to 1 if the rumored takeover target firm operates within the technology industry sector, and 0 otherwise; STAR is a dummy variable equal to 1 if the message post has a star rating of 4 or 5, and 0 otherwise; FIRSTTIME is a dummy variable equal to 1 if the sample firm has not been subject to a takeover rumor on the public Yahoo! Finance Message Board before, and 0 otherwise; CONFLICT is a dummy variable equal to 1 if more than 15 percent of the reply posts to the initial takeover rumor on the sample firm are contradictory within the sample window, and 0 otherwise; SPEC is a dummy variable equal to 1 if the sample firm has been subject to media speculation regarding a takeover in the previous six months, and 0 otherwise; and, TIME is a dummy variable equal to 1 if the takeover rumor is disseminated in the period 2006 to 2008, and 0 otherwise.

PANEL A: Dummy Variables

Variable Variable Value Number of Observations Proportion of Sample

TECH 1

0

1159

1739

39.99%

60.01%

STAR 1

0

874

2024

30.15%

69.85%

FIRSTTIME 1

0

1967

931

67.87%

32.13%

CONFLICT 1

0

1004

1894

34.64%

65.36%

SPEC 1

0

618

2280

21.33%

78.67%

TIME 1

0

1798

1100

62.04%

37.96%

PANEL B: Continuous Variable

Variable Sample

Size Mean

Standard Deviation

25th Percentile

Median 75th

Percentile

LN(SIZE) 2,898 6.4598 3.1245 5.2174 6.3909 7.5786

43

TABLE VI: The Aggregate Effects of Message Board Takeover Rumor Dissemination

Table VI reports the shareholder wealth and trading volume effects of Message Board takeover rumors for the period January 2003 to December 2008, inclusive. Panels A, B and C present pre, post and surrounding-rumor event windows (in minutes), respectively. BHAARs are constructed as the average of the aggregate sample (2,898 observations) individual rumored takeover target Buy-and-Hold Abnormal Returns (BHARs) over the specified event window (one-tailed t-statistics are presented in

parentheses), where: ,[ , ] ,[ , ] ,[ , ]i q r i q r m q r

BHAR R R . CAATVs are calculated as the average of the total sample individual rumored

target firm Abnormal Trading Volumes (ATVs) over the specified event window (one-tailed t-statistics are presented in

parentheses), where: 11

,

,

],[,1],[

ti

b

atti

bai

n

iba AV

VATV

nCAATV .

Window

Abnormal Returns

Abnormal

Volume

BHAAR CAATV

PANEL A: Pre-Rumor Windows

[-24 hrs, 0] 0.0027

(1.5074)

0.0094(0.9874)

[-60, 0] 0.0014

(1.0236)

0.0116(1.2368)

[-30, 0] 0.0003

(0.7457)

0.0050(0.5423)

[-10, 0] 0.0001

(0.5712)

0.0063(0.7456)

[-5, 0] 0.0001

(0.3657)

-0.0019(-0.2395)

[-1, 0] -0.0001

(-0.7849)

-0.0023(-0.3845)

PANEL B: Post-Rumor Windows

[0, +1] 0.0002

(1.5436)

0.0134 (1.4685)

[0, +5] 0.0012

(2.7988)**

0.1111 (2.9928)**

[0, +10] 0.0012

(3.1866)**

0.1422 (3.2492)**

[0, +30] 0.0020

(3.9864)**

0.2277 (3.9611)**

[0, +60] 0.0028

(4.3568)**

0.2677 (4.3319)**

[0, +24 hrs] 0.0158

(5.9864)**

0.1654 (4.6789)**

PANEL C: Surrounding-Rumor Windows

[-1, +1] 0.0001

(1.6549) 0.0054

(1.2168)

[-5, +5] 0.0012

(2.5469)** 0.0546

(2.9553)**

[-10, +10] 0.0014

(3.0896)** 0.0742

(3.2492)**

[-30, +30] 0.0023

(4.1056)** 0.1163

(4.0218)**

[-60, +60] 0.0042

(4.6587)** 0.1397

(4.4699)**

[-24 hrs, +24 hrs] 0.0186

(6.0123)** 0.0874

(4.7896)**

*Significant at the 5% level; **Significant at the 1% level.

44

TABLE VII: Abnormal Returns Cross-Sectional Variation Results

Table VII presents the results of our multivariate regression analysis of the abnormal returns associated with Message Board takeover rumors during the selected event windows examined. Specifically, we present the multivariate regression results for all five-minute, thirty-minute and twenty-four-hour event windows. In particular, we present multivariate regression estimates with respect to abnormal returns we calculate using the 0/1 Market Model Buy-and-Hold Abnormal Returns (BHAR) methodology. With regard to reported explanatory variables: LN(SIZE) is the natural logarithm of the rumored takeover target firm’s market capitalization; TECH is a dummy variable equal to 1 if the rumored takeover target firm operates within the technology industry sector, and 0 otherwise; STAR is a dummy variable equal to 1 if the message post has a star rating of 4 or 5, and 0 otherwise; FIRSTTIME is a dummy variable equal to 1 if the sample firm has not been subject to a takeover rumor on the public Yahoo! Finance Message Board before, and 0 otherwise; CONFLICT is a dummy variable equal to 1 if more than 15 percent of the reply posts to the initial takeover rumor on the sample firm are contradictory within the sample window, and 0 otherwise; SPEC is a dummy variable equal to 1 if the sample firm has been subject to media speculation regarding a takeover in the previous six months, and 0 otherwise; and, TIME is a dummy variable equal to 1 if the takeover rumor is disseminated in the period 2006 to 2008, and 0 otherwise. Two-tailed t-statistics are reported in parentheses below the associated explanatory variable regression coefficient values, calculated using the heteroskedasticity-consistent standard errors method prescribed by White (1980).

Window Variable

5 Minutes 30 Minutes 24 Hours [-5, +5] [-5, 0] [0, +5] [-30, +30] [-30, 0] [0, +30] [-24, +24] [-24, 0] [0, +24]

0/1 Market Model Abnormal Returns (BHAR)

Constant 0.0055

(4.4505)** -0.0002

(-0.3700) 0.0056

(5.1553)**

0.0058 (3.4208)**

-0.0001 (-0.0893)

0.0060 (5.6795)**

0.0124 (1.7074)

-0.0028 (-0.4607)

0.0147 (3.9852)**

LN(SIZE) -0.0005

(-4.1011)** 0.0001

(1.2241) -0.0006

(-5.1975)**

-0.0005 (-2.5951)**

0.0002 (1.1887)

-0.0007 (-6.0069)**

-0.0015 (2.5467)**

0.0007 (0.4391)

-0.0021 (4.2158)**

TECH 0.0006

(0.7028) 0.0002

(0.4144) 0.0004

(0.4781)

0.0002 (0.1410)

0.0001 (0.0908)

0.0001 (0.0589)

0.0024 (0.4447)

0.0063 (1.4127)

-0.0039 (-1.4411)

STAR 0.0028

(3.4052)** -0.0006

(-1.2737) 0.0035

(4.7056)**

0.0042 (3.6808)**

-0.0017 (-0.8001)

0.0060 (5.4213)**

0.0190 (3.8590)**

-0.0019 (-0.4614)

0.0205 (5.2396)**

FIRSTTIME -0.0005

(-0.6494) -0.0001

(-0.3013) -0.0004

(-0.5206)

-0.0001 (-0.0533)

0.0001 (0.1020)

-0.0002 (-0.2845)

-0.0082 (-1.4433)

-0.0060 (-1.5478)

-0.0020 (-0.8565)

CONFLICT -0.0007

(-0.9059) -0.0002

(-0.5676) -0.0005

(-0.6671)

-0.0011 (-1.0652)

-0.0007 (-0.7149)

-0.0004 (-0.5428)

-0.0026 (-0.5714)

0.0009 (0.2280)

-0.0035 (-1.4954)

SPEC -0.0009

(-0.8903) 0.0004

(0.7331) 0.0014

(1.5011)

-0.0020 (-1.3819)

-0.0009 (-0.6662)

-0.0011 (-1.2630)

-0.0014 (-1.2174)

-0.0017 (-0.3260)

0.0004 (3.1423)**

TIME -0.0019

(-1.0095) -0.0004

(-1.0529) -0.0016

(-0.9411)

-0.0013 (-1.2892)

-0.0001 (-0.1200)

-0.0012 (-0.8855)

-0.0127 (-1.5163)

-0.0088 (-0.6411)

-0.0034 (-1.5141)

Adjusted R2 0.0529 0.0083 0.0804 0.0367 0.0080 0.1464 0.0489 0.0029 0.1034 F-Statistic 5.291** 0.7893 8.285** 3.612** 0.7643 16.25** 4.865** 0.9864 10.92**

*Significant at the 5% level; **Significant at the 1% level.

45

TABLE VIII: Abnormal Trading Volumes Cross-Sectional Variation Results for Various Observation Periods

Table VIII presents the results of our multivariate regression analysis of the abnormal trading volume associated with Message Board takeover rumors during the event windows selected for reporting. Specifically, we present the multivariate regression results for all five-minute, thirty-minute and twenty-four-hour event windows. With regard to reported explanatory variables: LN(SIZE) is the natural logarithm of the rumored takeover target firm’s market capitalization; TECH is a dummy variable equal to 1 if the rumored takeover target firm operates within the technology industry sector, and 0 otherwise; STAR is a dummy variable equal to 1 if the message post has a star rating of 4 or 5, and 0 otherwise; FIRSTTIME is a dummy variable equal to 1 if the sample firm has not been subject to a takeover rumor on the public Yahoo! Finance Message Board before, and 0 otherwise; CONFLICT is a dummy variable equal to 1 if more than 15 percent of the reply posts to the initial takeover rumor on the sample firm are contradictory within the sample window, and 0 otherwise; SPEC is a dummy variable equal to 1 if the sample firm has been subject to media speculation regarding a takeover in the previous six months, and 0 otherwise; and, TIME is a dummy variable equal to 1 if the takeover rumor is disseminated in the period 2006 to 2008, and 0 otherwise. Two-tailed t-statistics are reported in parentheses below the associated explanatory variable regression coefficient values, calculated using the heteroskedasticity-consistent standard errors method prescribed by White (1980).

Window Variable

5 Minutes 30 Minutes 24 Hours [-5, +5] [-5, 0] [0, +5] [-30, +30] [-30, 0] [0, +30] [-24, +24] [-24, 0] [0, +24]

Cumulative Abnormal Trading Volumes (CATV)

Constant 0.0328

(0.6895) -0.0026

(-0.5139) 0.0682

(0.7136)

0.0945 (1.2769)

0.0067 (0.9013)

0.1823 (1.2391)

0.1080 (3.1610)**

0.0119 (0.8736)

0.2041 (3.1692)**

LN(SIZE) -0.0122

(-2.7211)** 0.0003

(0.2797) -0.0227

(-2.7406)** -0.0207

(-3.2375)**

-0.0007 (-0.6323)

-0.0408 (-3.2022)**

-0.0135 (-4.5623)**

-0.0014 (-0.5800)

-0.0256 (-4.5907)**

TECH 0.2120

(4.8295)** -0.0011

(-1.2227) 0.4252

(4.8215)**

0.3223 (4.7147)**

0.0029 (1.2628)

0.6416 (4.7233)**

0.0579 (1.8374)

0.0034 (0.8968)

0.1124 (2.2319)*

STAR 0.1351

(3.2204)** -0.0027

(-0.4061) 0.2729

(3.2380)**

0.2463 (3.7707)**

0.0065 (0.7973)

0.4861 (3.7442)**

0.1334 (4.4262)**

0.0122 (1.2745)

0.2546 (4.4819)**

FIRSTTIME -0.0398

(-1.0316) -0.0004

(-0.6697) -0.0792

(-1.0220)

-0.0512 (-0.8516)

0.0008 (0.7210)

-0.1031 (-0.8638)

0.0263 (0.9477)

0.0049 (1.4694)

0.0476 (0.9111)

CONFLICT -0.0316

(-0.8301) 0.0003

(0.4295) -0.0635

(-0.8299)

-0.0417 (-0.7038)

0.0011 (0.9869)

-0.0846 (-0.7175)

0.0338 (1.2361)

0.0055 (1.4477)

0.0621 (1.2051)

SPEC -0.0630

(-1.2369) 0.0001

(0.1549) -0.1261

(-1.2327)

-0.0941 (-1.1875)

-0.0014 (-0.9557)

-0.1869 (-1.1861)

-0.0298 (-0.8161)

-0.0032 (-0.7286)

-0.0565 (-0.8190)

TIME 0.1010

(1.2887) -0.0004

(-0.6719) 0.2024

(1.1781)

0.1524 (1.5826)

0.0010 (1.1582)

0.3037 (1.4388)

0.0252 (0.9697)

0.0012 (0.3789)

0.0493 (1.0045)

Adjusted R2 0.0827 0.0073 0.0829 0.0923 0.0061 0.0974 0.0835 0.0057 0.0851 F-Statistic 8.536** 0.403 8.562** 9.629** 0.678 9.867** 8.634** 0.585 8.813**

*Significant at the 5% level; **Significant at the 1% level.

46

TABLE IX: Message Board Rumor Trading Strategy Profits

Table IX reports the economic significance of Message Board takeover rumors trading opportunities for our 24 hour trading strategy. We report the average dollar profit from a $50,000 investment in each rumor company. Specifically, using TAQ data, we buy (sell) parcels of shares at their ask (bid) price immediately preceding the trade. Panel A reports the dollar profits from investing in all rumor companies over the twenty-four-hour period, while Panel B reports the dollar profits from investing only in rumor companies with a market capitalization under $500 million, are associated with a rumor star rating of 4 or 5, and are restricted to the technology sector. Two-tailed t-statistics, are presented in parentheses.

Panel A: Takeover Rumor Trading Strategy Dollar Profits for all Companies

Average Dollar Profits (in thousands) 0.9212

(3.2091)**

Panel B: Takeover Rumor Trading Strategy Dollar Profits for Small High Star Rumor Companies

Average Dollar Profits (in thousands) 1.3404

(3.6842)***Significant at the 5% level; and, **Significant at the 1% level