SEARCH

SEARCH BY CITATION

Keywords:

  • internet stock message board;
  • online trading;
  • investor sentiment;
  • noise trader;
  • naïve bayesian;
  • text classifier

Abstract

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

Abstract:  This study extends the literature on the information content of stock message boards. To better understand the effect of online postings on trading activities and reduce the error due to stocks with small message board followings, we examine stocks with no fundamental news and high message posting activity. Such stocks tend to be of small firms with weak financials. For those stocks, we find a two-day pump followed by a two-day dump manipulation pattern among online traders, which suggests that an online stock message board can be used as a herding device to temporarily drive up stock prices. We also find that online traders’ credit-weighted sentiment index, but not the number of postings, is positively associated with contemporaneous return and negatively predicts the return next day and two days later. Also, absolute sentiment is negatively related with contemporaneous and next day's intraday volatility and positively related with the proportion of volume in small-sized trades. We conclude that message board sentiment is an important predictor of trading-related activities.


1. INTRODUCTION

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

The use of the Internet for stock trading and information collection and sharing has been growing rapidly.1 In addition to trading via online discount brokers, traders often seek information online from paid sources such as ValueLine and Zacks and free sources such as stock message boards and chat rooms. Among the free sources, stock message boards (e.g., Yahoo! Finance, Raging Bull and TheLion.com) are popular for learning other investors’ opinions about specific stocks or the market. There are numerous message boards where potential investors post, read and reply to messages. Not only the number of such boards but also the number of participants on these boards has exploded. For example, TheLion.com, a message board aggregator web site, tracks over 100 million message postings from more than 25,000 message boards and attracts over 250 million page views and two million monthly visitors as of June 2007. With their increasing prominence, stock message boards have been getting the attention of practitioners, policymakers, and researchers since the late 1990s.

The message board sentiments are closely watched by investors who seek inputs to enhance their trading profits. However, there is a concern that participants may post messages designed to improve their positions at the expense of other investors. Such trading behavior is known as ‘pump-and-dump’ (Antweiler and Frank, 2004; and Jiang et al., 2005).2 Many low priced stocks are subject to stock manipulation because of their thin float and lack of broad ownership (Wysocki, 1999; and Aggarwal and Wu, 2006), and regulators are concerned about this issue. The Securities and Exchange Commission (SEC) and the Federal Trade Commission (FTC) are especially interested in tracking the activities on online message boards in order to prevent internet investment scams and protect investors’ interests. Among researchers, studies such as van Bommel (2003) and Eren et al. (2009) provide theoretical models of the manipulation of stock prices by spreading rumors.

In this paper, we first argue that an online stock message board can be used to manipulate the price of a vulnerable stock with small market capitalization and weak fundamentals. Specifically, an online message board can be used as a herding mechanism to temporarily drive up the stock price; subsequently, the stock is sold and the price falls. We then examine if the empirical evidence supports this argument. Our argument, described in detail in the next section, is based on the theoretical model in van Bommel (2003).

Another issue we examine is how online message posting activities influence short-term security trading. Although previous studies document that online talk is not just noise, it remains unclear whether message board activities influence trading activities or is it the other way around. Also, previous studies do not find a relation between investor sentiment on the message boards and future stock returns.

In this paper, we revisit the issue of the causality between online posting activities and trading activities, including online investor sentiment's ability to predict subsequent returns. Further, as discussed above, we empirically examine the use of an online message board to pump and dump stocks. We also examine the characteristics of stocks likely to be targeted by online traders for their pump and dump strategy.

The approach we use for investigating the above issues has some important differences from other studies. Specifically, we examine stocks for which there is no material news and the message board activity is high. If there is material news about a stock, market activities are likely to influence message board activities instead of the other way around. Therefore, in earlier studies, which do not specifically control for news, evidence is confounded by the presence of stocks with material news. By examining stocks with no material news, we are able to better understand the effect of online postings on market trading activities. Also, consistent with Tumarkin (2002), who suggests that only stocks with high message volume should be included in the sample in order to reduce error introduced by stocks with small bulletin board followings, we only include stocks with high message posting activity. For this purpose, we use TheLion.com, a novel data source that provides a list of the ten most heavily discussed stocks on each trading day. In our sample, we include these most heavily discussed stocks, instead of subjectively selecting a research sample that could include stocks without a significant online following.

In constructing a measure of online investor sentiment, we include messages with and without a self-reported sentiment. We use a text classifier to assess the sentiment of a message in which it is not self-reported. TheLion.com has a reward system in which a poster earns reputation credit from other users for posting quality messages. This enables us to use a credit-weighted measure of investor sentiment.

While there is no sample selection bias in our study, one caveat is that our sample, which includes heavily discussed stocks without any material news, mainly has small cap stocks with weak financials.3 Therefore, while this study enhances our understanding of the influence of online message board activities on trading activities in such stocks, the findings of the study should not be generalized to larger firms with strong financials.

The main findings of this study are as follows. First, the empirical evidence is consistent with our pump and dump argument. There is a significant positive abnormal return on the day a stock is heavily discussed on the message board and also on the preceding day. There are significant negative abnormal returns on the next two days. These abnormal returns suggest a two-day pump followed by a two-day dump stock manipulation pattern among online traders. There are no long-term economic effects of the pumping as the stock price reverts to the pre-pump level within a few days. Second, online traders prefer non-financial stocks with a low price, a high trading volume, a small float, and a low price-to-book ratio.

We also find that online posting activities affect online trading activities. Online traders’ sentiment index on the day a stock is heavily discussed is positively related with the return on that day and negatively related with the return on the following day. This is consistent with our event study results suggesting a pump and dump pattern. Of the three posting activity measures examined (sentiment index, disagreement index, and the number of messages), sentiment index dominates in explaining returns. Further, absolute sentiment helps explain volatility and the proportion of volume in small-sized trades. These results show that message board sentiment is an important predictor of trading-related activities.

The rest of the paper is organized as follows. The next section discusses the main issues and hypotheses. The design-related features of this study are discussed in Section 3. Section 4 includes details of sample construction, data sources, and sample characteristics. Section 5 provides event study results. In Section 6, we look at online investors’ trading incentives and characteristics of targeted stocks. Section 7 examines if posting activities can explain contemporaneous and subsequent trading activities. Section 8 discusses the robustness of our results and Section 9 concludes.

2. ISSUES AND HYPOTHESES

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

(i) Pump and Dump

Several studies have examined the manipulation of stock markets. Most of these studies, such as van Bommel (2003), Eren et al. (2009) and Kyle and Viswanathan (2008), are theoretical. There are a few empirical studies, including Jiang et al. (2005), Khwaja and Mian (2005) and Aggarwal and Wu (2006). None of these studies empirically examine online stock manipulation through a ‘pump and dump’ scheme. In this paper, we suggest that an online stock message board can be used as a herding mechanism to temporarily drive up prices; the stock is then dumped and the price falls.

Our arguments, based on the theoretical model in van Bommel (2003), are as follows. An influential poster, with a high reputation and many followers on an online message board, builds a long position in a target stock.4 The target stock is likely to be a small firm with weak financials and small institutional ownership since such a stock is easier to manipulate via an online message board. After building a long position, the poster pumps the stock on the message board by posting messages about the stock with a bullish sentiment. As in the ‘bluffing’ variant of the model in van Bommel, the poster need not be informed and he could just be bluffing that he is trading on information. Due to the poster's reputation and the strong sentiment in his messages, his followers herd and trade based on his advice, thus moving the price up. Outside traders (non-subscribers to the message board) such as chartists and momentum traders are influenced by the price movement and also take a position. Within a day or two, the price likely overshoots the trigger point for the influential poster to reverse his position. When the poster knows the price is overshooting, he dumps the stock and his followers then do the same. Both the poster and his followers profit at the expense of the outside traders.5

Based on the above arguments, we propose the following hypothesis. Initially when a stock is heavily talked about (pumped) on a message board, there are positive abnormal returns in it. Negative abnormal returns in the stock follow as the stock is then dumped.

(ii) Trading Preferences of Online Traders

A basic issue is the attributes of stocks likely to be targeted by online traders for their pump-and-dump strategy. Wysocki (1999) examines which stocks attract message postings. He finds that message-posting volume is related to firm characteristics. While Wysocki examined number of messages and our focus is on trading preferences, the findings in Wysocki indicate that online traders have a preference for certain types of stocks. Accordingly, we suggest that a firm's characteristics affect the trader's preference for that firm's stock. Maximizing trading profits is typically the primary incentive for trading a stock (Lakonishok and Smidt, 1986). Korczak et al. (2010) discuss that the abnormal return associated with a corporate news announcement reflects for the trading profit realized (foregone) by an insider if he chose to trade (not to trade) ahead of the news. In this study, we use abnormal returns as a proxy for trading preferences.

Earlier studies such as Fama and French (1992) and Daniel and Titman (1997) show that a firm's fundamental factors significantly explain its stock returns. Other studies such as Brock et al. (1992) and George and Hwang (2004) document the significant power of technical characteristics in explaining security returns. Wysocki (1999) finds that both fundamental and technical characteristics of a stock affect the message-posting volume in that stock. In view of the above studies, we propose that trading preferences of online traders are related to the fundamental and technical characteristics of the stock.

(iii) Short-Term Effect of Online Activities on Trading Activities

An extant interesting issue is whether online message posting activities on a given day influence short-term security trading on that day and the next day or two. Though previous studies document that online talk is not just noise, the causality between message board postings and trading activities is not without controversy. While some studies find significant predictive effects of the number of messages on trading activities, the studies generally find no significant predictive effects of investor sentiment.

Wysocki (1999) suggests that online posters can influence ‘public opinion’ about a firm. Tumarkin and Whitelaw (2001) study Internet service stocks. They find that on days with abnormally high posting activity, changes in online investor opinion are correlated with contemporaneous abnormal industry-adjusted returns and trading volume but do not predict future returns and volume. It is ambiguous whether activity on the message board causes or is the result of abnormal returns on the stock. Similar findings are reported in Tumarkin (2002).

Using data from Yahoo! Finance and Raging Bull message boards, Antweiler and Frank (2004) suggest that the causality for volatility is more from message boards to market. But Koski et al. (2007) find the opposite. Antweiler and Frank also find that high number of messages predicts negative subsequent returns. However, investor sentiment, though positively related with contemporaneous returns, does not predict subsequent returns. Using data from Motley Fool, Das et al. (2005) report that sentiment does not apparently predict returns, but returns drive sentiment instead. Gu et al. (2006) rebut this argument and suggest that there are informed posters on stock message boards whose information is not fully incorporated into market prices. Das and Chen (2007) find that tech sector's aggregate sentiment can predict the level of the sector's aggregate index but not of individual stocks.

Overall, prior studies conclude that online talk is not just noise. However, it remains unclear whether message board activities influence market activities or is it the other way around, and whether online investor sentiment predicts future returns. The lack of clarity is puzzling, especially for the relation between online investor sentiment and subsequent returns. In this paper, we argue that the evidence in earlier studies is confounded by the presence of stocks with material news. If there is material news about a stock, market activities are likely to influence message board activities instead of the other way around (Klibanoff et al., 1998; and Chan, 2003). However, in the absence of material news, message board activities are likely to influence market activities, especially for the small-cap stocks examined in this paper. Accordingly, we propose the following hypothesis. For stocks with no material news, online posting activities (the number of messages, sentiment, and disagreement) affect short-term trading activities (returns, volatility, volume and the bid-ask spread).

3. STUDY DESIGN

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

(i) Use of a Chat-Room-Like Message Board Aggregator

Our data source is TheLion.com, a chat-room-like message board aggregator. Unlike message boards such as Yahoo! Finance that list messages on a separate webpage for each stock, TheLion.com lists messages reverse chronologically on a single page. Also, it follows a reward system in which posters earn credits from other users for posting quality messages. This system alleviates the problem stated in Vilardo (2004) that the same poster may post under multiple IDs. It also enables the construction of a credit-weighted sentiment index.

(ii) Control for Surrounding Material News

In previous studies, evidence regarding the effect of message board activities is confounded by the presence of news. In this study, we examine only those firms for which there is no material news at the time they are included in the sample.6 Our sample allows for a clearer examination of the effect of message board activities. For identifying material news, we use Ravenpack news database, an up-to-date and reliable source affiliated with Dow Jones News.7 It collects comprehensive news from all the major news sources, including the Wall Street Journal, Reuters Newswires, Press Release Wires, and many others.

(iii) Sample of most heavily discussed stocks

The main advantage of using TheLion.com is that it provides a list of the ten most heavily discussed stocks on each trading day. We include these most heavily discussed stocks in our sample. Thus, we don't have to subjectively select a research sample that could possibly include stocks without a significant online following and suffer from the noise and error introduced by such stocks (Tumarkin, 2002). In the absence of material news, the most actively discussed stocks are likely to be small-cap stocks with weak fundamentals and low institutional holdings as those stocks are more easily influenced by online postings. This study will shed more light on the role and influence of online message boards for such stocks.

(iv) Inclusion of Messages with and without a Self-Reported Sentiment

During the late 1990s and early 2000s, online message posters were not able to directly disclose their sentiment (such as buy, hold, or sell). As a result, many previous studies, such as Antweiler and Frank (2004) and Das and Chen (2007), need to use a text classifier to assess the sentiment of each message before constructing a sentiment index. However, these days many message boards provide a sentiment indicator for posters to explicitly disclose their sentiment on a voluntary basis. For instance, Yahoo! Finance allows a poster to choose one of the following sentiments: strong buy, buy, hold, sell, strong sell, or not disclose (by default). TheLion.com offers two more sentiments: short and scalp.8

Though online posters may now disclose their sentiment, not all of them do so. Some studies such as Gu et al. (2006) and Cook and Lu (2009) use only the messages with a self-reported sentiment. However, Antweiler and Frank (2004) argue that to construct a sentiment index with a minimal bias, it is necessary to include all messages from the same message board. Accordingly, in this paper, we include messages without a self-disclosed sentiment also, and use a text classifier to assess the sentiment scores of those messages.

(v) Text Classifier

Following Antweiler and Frank (2004), we employ the Naïve Bayesian (NB) text classifier to assess the sentiment score of messages without a self-reported sentiment.9 Additional details of our procedure are included in Appendix A. For all messages, including those with and without a self-reported sentiment, we follow Tumarkin and Whitelaw (2001), Tumarkin (2002) and Zhang and Swanson (2010), and code sentiment scores as −3 for short, −2 for strong sell, −1 for sell, 0 for hold and scalp, +1 for buy, and +2 for strong buy.

(vi) Credit-Weighted Sentiment and Disagreement Indexes

TheLion.com has a credit system for posters, with the credit score being a proxy for the poster's reputation and influence. Therefore, we are able to construct a credit-weighted sentiment index, as suggested in some other studies (Gu et al., 2006; and Cook and Lu, 2009). Our formula for the sentiment index is similar to Antweiler and Frank (2004), except that our index is credit-weighted, with the log of a poster's credit being the poster's weight.10

Each message has a sentiment score from Short (−3) to Strong Buy (+2), which is either provided by its poster or assessed with our NB text classifier. We first compute the bullishness ratio inline image based on all sentiment scores for stock i on event day t as follows:

  • image(1)

where inline image is the number of messages with an optimistic sentiment (buy or strong buy) on stock i on day t; inline image is the number of messages with a pessimistic sentiment (sell, strong sell or short); x is the sentiment score; and creditscore is each poster's credit score. Following the transformation in Antweiler and Frank (2004), our sentiment index for stock i on day t is:

  • image(2)

whereinline image is a standardized bullish index bounded by −1 and 1 and inline image is the total number of messages for stock i.

Antweiler and Frank (2004, p. 1268) further propose a reduced form of standard deviation of sentiments as a disagreement proxy. In their setup, sentiment is either 1 or −1, with an equal weight allocation. Because our sentiment score ranges from −3 to +2, our disagreement index is a credit-weighted standard deviation of the sentiment scores for the stock i. We first calculate each message k's weight and then the disagreement index:

  • image(3)
  • image(4)

where inline image is the sentiment score associated with message k and inline imageis the credit- weighted mean score for the stock.11

4. SAMPLE AND DATA

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

(i) Sample Construction

We download real-time messages from TheLion.com from July 18, 2005 to July 18, 2006.12 These messages are downloaded daily after the close of trading. Tumarkin (2002) suggests that only stocks with a sufficient number of messages should be included. Thus, on each trading day we retain only those stocks that are among the top ten stocks by number of messages on that day. So, the event day for a stock is the day it is among the top ten stocks. To ensure sufficient posting activity, we also require that the sample stocks have at least 30 messages during the event day (from 4:00 p.m. Eastern time the previous day to 4:00 p.m. Eastern time on the trading day). We additionally require that the stocks not be over-the-counter bulletin board (OB) or pink sheet (PK) stocks; have a share price above $1 to avoid excess transaction costs; and have at least five different posters in order to prevent hyping by one or two dominant posters.13 Finally, to avoid any contamination from news, we exclude stocks with material news on or within two days before and after the event day. Some stocks occur more than once in our sample. We include those multiple occurrences as separate observations if there is a gap of at least two trading days between any two occurrences. Otherwise, we include only the first occurrence. Our sample consists of 80 observations, accounted for by 64 distinct stocks.

(ii) Data Sources

For controlling news, we use the Ravenpack news database described in Section 3. We obtain financial and trading data from the Center for Research in Security Prices (CRSP), Trade and Quote database (TAQ), Compustat, and Thomson Financial. We use the standard procedures to filter the TAQ data. Specifically, we exclude the following: trades and quotes outside of the regular trading hours (9:30 a.m. to 4:00 p.m. Eastern time): opening and closing trades and quotes; regional trades and quotes; trades marked as irregular; negative bid or ask quotes; and spreads that are negative or in excess of 40% of the bid-ask midpoint.

(iii) Sample Characteristics

Our sample has 64 stocks, for which we download a total of about 12,000 messages. Because we exclude stocks with material news, most of these messages are related to trading. To provide a qualitative feel for their contents, we provide examples of messages in Table 1 for each of the six self-reported message sentiment categories. These messages, like the sample messages in Wysocki (2000) and Antweiler and Frank (2004), are short and in the form of conversational discourse. Admati and Pfleiderer (2001) point out that unlike newspaper articles or financial reports, forum messages are style free, short, and in an informal dialogue-like format. The messages in our study have the same characteristics.

Table 1.  Examples of Messages
SentimentMessage NumberStock SymbolPoster's IDPosting TimeMessage Content
Strong Buy821429IPIIdreadknought402005–8-24 10:02:27People still trying to short stock here, HEDGE funds will punish that soon, IMO
875337ODMOOTCBBKing2005–12-16 11:10:24HUGE volume spike New HOD of $1.93
876264FORDlooktothefuture2005–12-19 15:35:07Its about to run, get in today.
942162OLABdreadknought402006–5-10 15:34:31Adding a few more dont let the shorts win here.
Buy804159CAFEGOPHERBROK2005–7-21 8:08:40Buying opportunity thia AM after yesterday selloff
821338BIDUmtmover6902005–8-24 9:02:03Down a little but gaining ground,still looks good for a run, imo.
925474NXXIGoingin602006–4-4 12:49:45This week 2.50 2.70 $3 ?
930660ARTWdreadknought402006–4-17 12:06:11Looks cheap here esp. if BEETS could be a new energy tie in 980k floater
Hold859475SYNXhy_hawarya2005–11-11 13:38:45Acting peculiarly …
820154SMTXtoughmarket2005–8-22 10:06:59Still undervalued.
916763CHCIGOPHERBROK2006–3-16 9:56:43Nice article by Street.com $12 and $16 target
927150OLABLionmaster2006–4-7 10:02:59Tons of shorting OLAB here next CHNR?
Sell836582COOLwindy2005–9-20 10:41:45Out all COOL 1.34 from 1.25 nice. Thanks
848671HEBTheLaker2005–10-17 11:47:46Out Heb rest of heb here Congrats all
849770ENWVglamba2005–10-19 9:39:13Shorted ENWV at 14 this AM … will cover at 11
965273BIDUKATN2006–7-13 14:12:43Tech thats still defying reality, needs to fall hard imo.
Strong Sell818079CAFETraderspot1232005–8-16 08:18:35Seems like CAFE failed to appeal and not even pinks taking them in. Very bad.
823721SMTXEPR531672005–8-29 10:42:45Out at 3.5 from 2.85.
871389PPTVdanshadow2005–12-8 15:40:03VOLUME IS DEAD
877407NVAXmumic2005–12-21 15:58:27Looks like back to 2.80 from here …
Short809773JRJCjjkool_012005–8-1 12:45:14Short into earnings 6.08
833759CTIBNapolion2005–9-14 12:44:35Re: Short @$7.25 … I Plan to cover 1/2 @$3.80 and the rest @$3.20 …
836954HOKUtjtrader2005–9-20 18:40:27Negative Motley Fool article on HOKU just out afterhours. Looks like a good short …
924971IPIILionmaster2006–4-3 13:36:56Short 1k shr here $20.32, way too fast here, target $16.

The details of our 64 distinct sample stocks are included in Appendix B. Among all the stocks in our sample, 31.25% (20 out of 64) are in the technology sector, only 18.75% (12 out of 64) have options listed, and most are listed on NASDAQ (58 out of 64).

Table 2 provides summary statistics of the sample. The table discloses some interesting characteristics of stocks targeted by online investors. Panel A includes a summary of fundamental characteristics. The average market equity is above three times the book equity, suggesting that these stocks are more likely to be characterized as growth stocks. The mean ROE is 24.92% but the median is −2.80%. The profit margins (mean of −261.06%, median of 0.03%) and quarterly revenue growth rates (mean of −187.61%, median of 6.80%) are weak. The mean operating cash flow is $9,420. Overall, these summary statistics suggest that stocks targeted by online investors have weak fundamentals and may be vulnerable to pump-and-dump trading.

Table 2.  Summary Statistics
VariableMeanStd. Dev.MinimumMedianMaximum
  1. Notes:  Our sample includes a total of 80 observations, accounted for by 64 distinct stocks. Event day is the day on which a stock appears in the list of ten most actively discussed stocks on that day. We include two observations of the same stock if they are at least two trading days apart. The summaries are based on the statistics on the event day.

Panel A: Summary of Fundamental Aspects of 64 Stocks
Price-to-book ratio3.393.450.292.0517.13
ROE (%)24.92405.91−243.37−2.802945.21
Profit margin (%)−261.461129.74−6754.020.0370.71
Quarterly revenue growth (%)−187.612292.07−17072.706.802353.50
Debt/Equity0.521.06−2.750.194.53
Operating cash flow (‘000 $)9.4248.00−104.700.88188.90
Panel B: Summary of Technical Aspects of 64 Stocks
Prior three-month volume (in millions)0.901.690.040.298.65
Floating shares (in thousands)14.4318.140.587.2085.67
Stock price ($)8.4413.481.434.1797.90
Held by institutions (%)20.3623.680.1011.5099.60
Short-sell ratio2.323.830.000.7016.70
Options listed (binary)0.190.390.000.001.00
Technology sector (binary)0.310.470.000.001.00
Panel C: Summary of Trading and Posting Activities for 80 Observations
Number of messages44.4114.8230.0039.0098.00
Number of posters11.554.345.0011.0023.00
Bullishness index [−1,1]0.650.38−0.790.751.00
Sentiment index [−∞, +∞]2.441.40−2.712.733.99
Disagreement index [0, +∞]1.030.290.460.991.84
Raw return on posting day (%)13.6520.40−67.849.87105.88
CRSP equally-weighted return (%)−0.020.06−1.65−0.012.45
PowerShare QQQ return (%)−0.200.09−2.17−0.312.71
Volume-weighted volatility0.320.360.020.212.32
Trading volume (in millions)4.6011.300.111.2481.20
Number of trades (in thousands)7.8016.150.202.4194.83
Small trades (<=$3,750) proportion0.780.170.000.820.96
Medium trades proportion0.210.150.040.180.89
Large trades (>=$30,000) proportion0.010.020.000.000.15
Normalized spread0.020.010.000.010.06

The summary of technical aspects in Panel B shows a low trading volume in the three months prior to the event day (mean of 0.90 million shares, median of 0.29 million shares). The range of average daily trading volume is from a minimum of merely 40,000 shares to a maximum of 8.65 million shares. Similar to the low trading volume, average float is small with a mean of 14,430 shares and a median of 7,200 shares. The stock price ranges from $1.43 to $97.90, with the mean being fairly low at $8.44. Wysocki (1999) reports that online investors chase stocks with low institutional holdings in order to generate a bigger impact on the stock price. Our sample also reflects this attribute with the average institutional holding being only 20.36% and the median being even lower at 11.50%. The short-sell ratio indicates that it takes investors an average of 2.32 days to cover the current short interest.

Panel C includes a summary of trading and posting activities. Consistent with previous literature (Tumarkin and Whitelaw, 2001; and Tumarkin, 2002), we find that on average, online posters are bullish with a mean bullishness ratio of 0.65 (zero implies a neutral opinion) and a mean credit-weighted sentiment index of 2.44. The credit-weighted disagreement index (mean of 1.03) shows that investors express diverse opinions (disagreement index > 0) although a majority of investors are optimistic. The average event-day return on our sample stocks is 13.65%. In contrast, the CRSP equally-weighted average return on these days is about zero and the average return on PowerShare QQQ fund, a proxy for the NASDAQ composite index, is −0.20%. The average trading volume in a stock on the event day is 4.60 million shares compared with the average prior three-month volume of just 0.9 million shares per day. Thus, the trading volume is much larger when heavy online talk occurs. We also report the percentage of small-, medium-, and large-sized trades on the event day. We follow Lee and Radhakrishna (2000) to classify trades. A trade is classified as small if it is worth $3,750 or less and large if it is worth more than $30,000.14 Small trades account for 78% of all trades, suggesting that the trading volume surge is likely induced by small-sized trades.

5. EVENT STUDY

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

We have argued in Section 2 that if an online message board is used as a herding mechanism to temporarily drive up prices, initially there would be significant positive abnormal returns in a stock when it is heavily talked about (pumped) on a message board. Subsequently, there would be negative abnormal returns when the stock is dumped. To test this argument, we measure abnormal returns using the standard event study methodology.15 We include the methodology details in Appendix C.

In Table 3, we report the results of daily event study for abnormal returns and volumes from day t−5 to day t+5, with day t as the event day when the stock is heavily discussed among posters. For a side-by-side comparison, the table also includes online posting statistics. Cumulative abnormal return and volume are also computed for two wider windows: (t−30, t−1) and (t+1, t+30). The event study results and posting statistics are shown pictorially in Figures 1 and 2, respectively.

Table 3.  Return and Volume Event Study and Posting Statistics
Panel A: Daily
DayAAR (%)Patell ZCowan ZCAAR (%)AAV (%)Patell ZCowan ZCAAV (%)NMessagesNPostersSentimentDisagreement
  1. Notes: We use a 255-day estimation window starting on t-300, wherever possible, and ending on t−46. We require three days as the minimum length of the estimation window. AAR is the average abnormal return, CAAR is the cumulative average abnormal return, AAV is the average abnormal volume, and CAAV is the cumulative average abnormal volume. Patell Z-statistic is based on Patell (1976) test statistic and Cowan Z-statistic is based on Cowan (1992) nonparametric generalized sign test. NMessages represents the average number of messages and NPosters represents the average number of posters. Sentiment represents the average credit-weighted sentiment index and Disagreement represents the average credit-weighted disagreement index. ***, ** and * denote two-tailed significance at the 1%, 5% and 10% levels, respectively.

−50.000.340.590.0317.3244.22***5.88***317.33.603.000.191.69
−42.093.96***0.822.1486.3472.62***6.14***803.73.892.12−0.131.75
−33.604.24***0.135.7714.4280.14***5.62***1518.14.312.570.032.20
−20.641.00−0.346.3645.7060.30***6.14***2163.83.383.28−0.081.98
−14.919.76***2.68**11.2804.4193.87***7.98***2968.25.043.790.252.01
 013.9321.80***7.08***25.2878.55196.15***12.98***3846.744.4111.552.441.03
 1−1.65−2.22**−2.19**23.5425.22138.28***11.93***4272.057.8722.72−1.723.47
 2−2.06−3.84***−1.73*21.5933.8297.67***10.62***5205.823.0418.87−0.382.37
 3−1.19−1.92*−1.2620.3652.8971.68***9.56***5858.719.718.33−0.041.69
 4−0.46−1.13−1.96*19.8439.0439.86***9.30***6297.79.877.970.121.55
 5−0.040.660.2419.8422.2147.52***8.37***6719.910.135.320.390.99
Panel B: Longer Windows     
WindowCAAR (%)Patell ZCowan ZWindowCAAV (%)Patell ZCowan Z     
(−30, −1)20.305.20***2.68***(−30, −1)>999.99161.43***8.51***     
(+1, +30)−15.33−4.12***−3.34***(+1, +30)>999.99195.82***9.30***     
image

Figure 1. Abnormal Return and Volume

Notes:  We use a 255-day estimation window starting on t-300, wherever possible, and ending on t-46. We require three days as the minimum length of the estimation window. AAR is the average abnormal return, CAAR is the cumulative average abnormal return, AAV is the average abnormal volume, and CAAV is the cumulative average abnormal volume.

Download figure to PowerPoint

image

Figure 2. Online Posting Statistics

Notes:  Average posting statistics are based on 80 observations. NMessages represents the average number of messages and NPosters represents the average number of posters. Sentiment is the average credit-weighted sentiment index and Disagreement is the average credit-weighted disagreement index. NMessagesinline image [30, +∞] on day t, the event day, and NMessagesinline image [0, +∞] on other days. NPostersinline image [5, +∞] on day t and NPostersinline image [0, +∞] on other days; Sentimentinline image[−∞, +∞]; and Disagreementinline image[0, +∞].

Download figure to PowerPoint

On the event day t, both average abnormal return (AAR) and average abnormal volume (AAV) are significantly higher (13.93% and 878.55%, respectively). The average number of messages (44.41) and the average number of posters (11.55) also proliferate on that day. Meanwhile, average sentiment jumps from 0.25 on day t−1 to 2.44 on event day t and disagreement drops from 2.01 on day t−1 to 1.03 on day t. Both AAR and AAV are also significantly higher on day t−1 (4.91% and 804.41%, respectively). DeMarzo et al. (2003) argue that online investors are likely to build their positions prior to posting messages. Our finding of significant positive abnormal return on day t−1 supports this argument. The two-day period of days t−1 and t can be characterized as a pumping period with overly optimistic and herding behavior, as documented in Shiller (1995) and Hirschey et al. (2000).

The heavy trading volume continues on days t+1 (AAV of 425.22%) and t+2 (AAV of 933.82%) but the price drops significantly (AARt+1 of −1.65%, and AARt+2 of −2.06%). The post-event period of days t+1 and t+2 can be described as a two-day dumping period. The average sentiment changes from positive or bullish on day t (2.44) to negative or bearish on day t+1 (−1.72). The trading on the days after t+2 is associated with a mild price drop and smaller volume and is possibly due to some outside traders (non-subscribers to the message board) liquidating their positions. In almost 30 days, the price reverts back to the pre-event level. Thus, there are no long-term economic effects of the message board activity.

Overall, the event study findings support our hypothesis that that initially when the stock is heavily talked about (pumped) on a message board, there are positive abnormal returns in the stock, followed by negative abnormal returns when the stock is dumped.

6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

We now test our second hypothesis that the trading preferences of online traders, as proxied by a stock's abnormal returns, are related to the fundamental and technical characteristics of the stock. We choose fundamental and technical characteristics as per the suggestions in Wysocki (1999) and test them in separate models.

Earlier studies such as Fama and French (1992) and Daniel and Titman (1997) show that a firm's fundamental factors significantly explain its stock returns. Based on these studies, our cross-sectional regression model for fundamental variables is as follows.

  • image(5)

inline image is the cumulative abnormal return on stock i from t1to t2based on the Scholes and Williams (1977) market model, PTB is the price-to-book ratio, ROE is the return on equity, PRM is the profit margin percentage, QRG is the quarterly revenue growth rate, DTE is the debt-to-equity ratio, and OCF is the operating cash flow. We control for industry effects using six industry dummy variables, D1to D6, corresponding to consumer goods, financial, healthcare, industrial goods, services, and technology industries. Basic materials is the reference industry. We estimate the model separately for three CAR windows (0, 0), (0, 10) and (0, 30), where 0 is the event day. The t-statistics in all our regressions are adjusted based on the heteroskedasticity-consistent covariance matrix developed by White (1980).

In Panel A of Table 4, we find that the fundamental variables lack a cross-sectional explanatory power for CAR except for the price-to-book ratio being significant in the (0, 30) window (β =−0.022, t-stat =−2.17). We also find that online investors have some preference for consumer goods stocks (β(0, 0)= 0.237, t-stat = 1.89) and technology stocks (β(0, 30)= 0.278, t-stat = 1.74).

Table 4.  Trading Incentives and Target Stock Characteristics
Dependent VariableCAR0, 0CAR0, 10CAR0, 30
Coeff.t-statCoeff.t-statCoeff.t-stat
  1. Notes:This table reports the results of two sets of cross-sectional regressions. The dependent variable inline image is the cumulative average abnormal return between time t1 and t2. In the model based on fundamental variables, PTB is the price-to-book ratio, ROE is the return on equity, PRM is the profit margin percentage, QRG is the quarterly revenue growth rate, DTE is the debt-to-equity ratio, and OCF is the operating cash flow. In the model based on technical variables, AVO is the log of average prior 3-month volume, FLS is the log of the number of floating shares, PRC is the log of stock price, INS is percentage institutional holdings, SSI is the short-sell interest, and OPT is a dummy variable that assumes a value of one if the stock has options listed and zero otherwise. Six industry dummy variables are used. An industry dummy variable assumes a value of one if the firm belongs to that industry and zero otherwise. The t-statistics are adjusted based on the heteroskedasticity-consistent covariance matrix developed by White (1980). ***, ** and * denote two-tailed significance at the 1%, 5% and 10% levels, respectively.

Panel A: Fundamental Aspects
Intercept0.1441.350.0660.54−0.050−0.33
PTB−0.009−1.25−0.014−1.65−0.022**−2.17
ROE−0.000−0.85−0.000−0.580.0000.18
PRM0.0000.04−0.000−0.17−0.000−0.61
QRG0.0000.19−0.000−0.09−0.000−0.62
DTE0.0190.920.0351.520.0511.56
OCF0.000−0.38−0.000−0.560.000−0.24
DConsumer Goods0.237*1.890.1861.300.1801.00
DFinancial−0.336−1.56−0.323−1.31−0.174−0.56
DHealthcare0.1130.940.1320.960.2361.37
DIndustrial Goods0.0740.540.1390.890.1680.86
DService0.0160.140.0050.040.0750.46
DTechnology0.0890.800.1901.500.278*1.74
R-squared0.110.120.12
Panel B: Technical Aspects
Intercept0.3420.55−0.332−0.53−0.058−0.09
AVO0.066*1.730.070*1.810.0491.17
FLS−0.088*−1.68−0.021−0.40−0.012−0.21
PRI−0.096*−1.89−0.124**−2.43−0.203***−3.67
INS−0.000−0.220.0010.330.0020.72
SSI−0.026−0.92−0.046−1.61−0.016−0.52
OPT0.0350.260.0390.29−0.021−0.14
DConsumer Goods0.1260.540.1380.590.3141.23
DFinancial−0.954**−2.28−1.128**−2.69−0.906**−1.99
DHealthcare0.3491.570.4081.630.4571.58
DIndustrial Goods−0.238−0.93−0.096−0.37−0.121−0.43
DService−0.129−0.60−0.205−0.95−0.220−0.94
DTechnology−0.091−0.41−0.040−0.18−0.015−0.06
R-squared0.320.360.40

Previous studies such as Brock et al. (1992) and George and Hwang (2004) also document the significant power of technical characteristics in explaining returns. Our cross-sectional regression model for technical variables is as follows.

  • image(6)

AVO is the average prior 3-month trading volume, FLS is the log of the number of floating shares, PRI is the log of the stock price, INS is the institutional holding percentage, SSI is the short-sell interest, OPTis a binary variable with a value of one if the stock has options listed and zero otherwise, and D1to D6 are six industry dummy variables as discussed above.

Panel B of Table 4 shows that the abnormal return on event day 0 is significantly positively related to the average prior 3-month trading volume (β = 0.066, t-stat = 1.73) and significantly negatively related to the log of the number of floating shares (β =−0.088, t-stat =−1.68) and the log of the stock price (β =−0.096, t-stat =−1.89). Thus, stocks with a high trading volume, a small float, and a low price are more likely to be targeted by online traders. These findings support Kumar and Lee (2003), who document that individual investors have a special interest in low-priced small cap stocks, and Wysocki (1999), who finds that online traders have a greater trading incentive for stocks with a higher trading volume. We also find that financial stocks are less preferred by online investors (β =−0.954, t-stat =−2.28). We further examine CAR0, 10 and CAR0, 30 as the dependent variables and find similar results. Overall, we find partial support for our second hypothesis. While the fundamental aspects other than the price-to-book ratio are irrelevant to online investors’ trading incentives, the technical aspects are not.

7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

We now examine whether message posting activity, as reflected in the number of messages, aggregate sentiment, and aggregate disagreement among investors, can (a) explain contemporaneous return, volatility, volume, and spread, and (b) predict these variables. Previous studies find a good explanatory and predictive power of the number of messages, but not of sentiment and disagreement indices. However, in contrast to previous studies that use non-weighted indices, we use credit-weighted indices. The higher a poster's reputation, the greater is the weight given to the poster's sentiment. Also, we include messages with and without a self-reported sentiment, with the sentiment of messages without a self-reported sentiment assessed using a text classifier.

We follow Antweiler and Franks' (2004) contemporaneous and time sequencing lead-lag regression models. We first discuss the contemporaneous models followed by the lead-lag predictive models. For both these sets of regression models, we first calculate the various endogenous variables, including the daily raw return, volume-weighted volatility, proportion of volume induced by small trades, and bid-ask spread. We then test each of them as an endogenous variable in separate regression models.

(i) Contemporaneous Regressions

In the contemporaneous models, the endogenous variables mentioned above and the message board activity variables are on the same-day basis. The contemporaneous relation between posting activities and the stock's raw return is tested using the following model:16

  • image(7)

where Rt is the stock raw return, Sentiment is the credit-weighted sentiment index as shown in formula (2), Disagreement is the credit-weighted disagreement index as per (4), NMessages is the log of the number of messages posted for the stock, Size is the log of the firm's market capital, MktRet is the return on the PowerShares QQQ fund, and Rt-1 is the one-day lagged stock raw return in order to control for autocorrelation.17

The contemporaneous effect on volatility is tested using the following model:

  • image(8)

We use intraday standardized volume-weighted intraday price volatility. In Model (8) and the next two models, we use the absolute value of the sentiment index (AbsSentiment) because both optimistic and pessimistic sentiments may induce more volatility, volume, and spread. Since trading volume is positively correlated with volatility (Jones et al., 1994), we add NTrades, the log of the number of trades of the stock, as a control variable. Similar to Antweiler and Frank (2004), we include Mkt, the log of PowerShares QQQ fund's price level, to control for market movements. inline imageis included to control for autocorrelation.

We next examine the effect of message board activities on the proportion of small-sized trading volume. The idea is that retail investors, who trade in smaller sizes, are the ones likely to be influenced by message board activities. Accordingly, we use the ratio of dollar trading volume in small-sized trades (trades of $3,750 or less) and total dollar trading volume as the dependent variable. The model is as follows:

  • image(9)

Because more small trades occur on NASDAQ, we include Listing, a dummy variable for NASDAQ-listed stocks. One-day lagged volume is included to control for autocorrelation.

The contemporaneous relation between posting activities and liquidity, with the bid-ask spread as a proxy for liquidity, is tested using the following model:

  • image(10)

Spread is the ratio of the stock's average bid-ask spread and mean price on day t. Because spread and trading volume are negatively related, we include Trade, the log of the number of trades of the stock, as a control variable. inline image is included to control for autocorrelation.

Table 5 reports the results of the contemporaneous regressions. First, returns are significantly positively related with Sentiment (β= 0.006, t-stat = 2.13) and one-day lagged value of returns (β= 0.305, t-stat = 2.64), and significantly negatively with Size (β=−0.045, t-stat =−2.66). Thus, the same-day stock returns are higher for stocks with higher sentiment, smaller size, and high previous day raw return. Significant one-day lagged return suggests a two-day price momentum, which is consistent with a pre-event two-day pumping hypothesis. Different from previous research, the number of messages does not significantly explain contemporaneous returns when there is no material news about a stock. The disagreement index is also insignificant.

Table 5.  Contemporaneous Effects of Posting Activities
Dependent VariableRaw ReturntVolatilitytSmall-Trade Vol.tSpreadt
Coeff.t-statCoeff.t-statCoeff.t-statCoeff.t-stat
  1. Notes: The message board variables and control variables are for the same day as the dependent variables. Sentiment is the credit-weighted sentiment index, Abs. Sentiment is the absolute value of the sentiment index, Disagreement is the credit-weighted disagreement index, NMessages is the log of the number of messages posted for the stock, Size is the log of the firm's market capital, MktRet is the return on the PowerShares QQQ fund, Mkt is log price of the PowerShares QQQ fund, NTrades is the log of the number of trades in the stock, and Listing is a dummy variable that assumes a value of one for stocks listed on NASDAQ and zero otherwise. Lagged values of the dependent variables are included as independent variables to control for autocorrelation. t-statistics reported are adjusted based on the heteroskedasticity-consistent covariance matrix developed by White (1980). ***, ** and * denote two-tailed significance at the 1%, 5% and 10% levels, respectively.

Intercept0.674**2.190.0140.010.1940.890.038*1.88
Msg. board variables        
Sentimentt0.006**2.13      
Abs. sentimentt  -0.015*-1.700.003**2.49-0.001-0.92
Disagreementt0.0030.02-0.017-0.140.0150.41-0.002-0.71
NMessagest−0.004−0.05−0.005−0.060.013*1.68−0.001−0.52
Control variables        
Sizet−0.048**−2.660.0020.12−0.020***−2.71−0.003***−3.25
MktRett0.8970.35      
Mktt  −0.002−0.140.0010.280.0000.66
NTradest  0.043***2.85  −0.001**−2.18
Listing    0.051**2.02  
Lagged variables        
Returnt-10.305**2.64      
Volatilityt-1  0.521***9.82    
Volumet-1    0.846***17.37  
Spreadt-1      0.435***7.10
R-squared0.110.680.890.84

Second, the contemporaneous volatility test shows that intraday volatility is significantly negatively related with AbsSentiment (β=−0.015, t-stat =−1.70). This suggests possible herding among online traders when there is a strong sentiment, either optimistic or pessimistic, which is consistent with Hirschey et al. (2000). The other two facets of message board activity, the number of messages and disagreement index, are insignificant. As expected, the number of trades is significantly positively correlated with volatility (β= 0.043, t-stat = 2.85). One day lagged volatility is also significantly positive (β= 0.521, t-stat = 9.82).

Third, the contemporaneous test of the proportion of trading volume in small-sized trades yields several interesting findings. The greater the absolute sentiment, the greater the proportion of trading volume in small-sized trades (β= 0.003, t-stat = 2.49). Further, the proportion of trading volume in small-sized trades is greater for stocks with a higher number of messages (β= 0.013, t-stat = 1.68); smaller stocks (β=−0.020, t-stat =−2.71), and NASDAQ-listed stocks (β= 0.051, t-stat = 2.02). Finally, one-day lagged value of the dependent variable is significantly positive (β= 0.846, t-stat = 17.37).

Last, the contemporaneous bid-ask spread regression shows that AbsSentiment is not significantly related to the bid-ask spread. However, the signs of absolute sentiment, disagreement, and the number of messages are all negative, suggesting that higher levels of these variables tend to reduce the spread (improve liquidity). As expected, the bid-ask spread is negatively related to the control variables firm size (β=−0.003, t-stat =−3.25) and the number of trades (β=−0.001, t-stat =−2.18) and positively related to the one-day lagged spread (β= 0.435, t-stat = 7.10).

To conclude our discussion of contemporaneous tests, investors’ credit-weighted sentiment dominates the other two message posting parameters (number of messages and disagreement) in terms of the relation with return, intraday volatility and small-sized trading volume. Higher sentiment is associated with higher returns while higher absolute sentiment is associated with less volatility and more small-sized trades.

(ii) Predictive Power

To most investors, the ability to predict subsequent changes is more important than contemporaneous relations. So, we now examine if sentiment, disagreement, and the number of messages on a given day can predict market activities on the next two days. Our time sequencing lead-lag models follow Antweiler and Frank (2004). We test the one-day (two-day) predictive power of posting activities by leading the endogenous variables and the control variables by one day (two days) relative to the explanatory variables in the contemporaneous models examined earlier.

The one-day predictive results are included in Table 6. Similar to our contemporaneous results, we find that credit-weighted sentiment, not the number of messages, predicts the return on the following day. Consistent with our event study results that suggest a two-day dumping pattern, a higher sentiment predicts a lower return on the following day (β=−0.022, t-stat =−1.77). The next day significant drop could be due to dumping by the influential posters who had originally pumped up the stock and profit taking by them and their followers on the message board. Overall, based on our contemporaneous and predictive regression results, greater bullish opinion about a stock on event day t is associated with a price jump on day t but a price drop on day t+1. Theoretical foundation of such a pattern (overreaction on private information followed by a correction) is presented in Daniel et al. (1998) and Aggarwal and Wu (2006). Our findings are also consistent with prior studies such as Neumann and Kenny (2007) and Keasler and McNeil (2010) that document a significant stock price drop after media hype. We also find that a higher disagreement among posters on day t predicts a price drop on the next day (β=−0.155, t-stat =−2.37).

Table 6.  One-Day Predictive Power of Posting Activities
Dependent VariableRaw Returnt+1Volatilityt+1Small-Trade Vol.t+1Spreadt+1
Coeff.t-statCoeff.t-statCoeff.t-statCoeff.t-stat
  1. Notes:This table includes regressions to test the one-day predictive power of posting activities by leading the dependent variables and the control variables by one day relative to the message board variables. Lagged values of the dependent variables are included as independent variables to control for autocorrelation. Sentiment is the credit-weighted sentiment index, Abs. Sentiment is the absolute value of the sentiment index, Disagreement is the credit-weighted disagreement index, NMessages is the log of the number of messages posted for the stock, Size is the log of the firm's market capital, MktRet is the return on the PowerShares QQQ fund, Mkt is log price of the PowerShares QQQ fund, NTrades is the log of the number of trades in the stock, and Listing is a dummy variable that assumes a value of one for stocks listed on NASDAQ and zero otherwise. t-statistics reported are adjusted based on the heteroskedasticity-consistent covariance matrix developed by White (1980). ***, ** and * denote two-tailed significance at the 1%, 5% and 10% levels, respectively.

Intercept0.1180.531.399**2.250.0170.090.0321.32
Lagged msg. board variables        
Sentimentt−0.022*−1.77      
Abs. sentimentt  −0.010**−2.470.004*1.75−0.001−1.12
Disagreementt−0.155**−2.37−0.011−0.120.0240.85−0.003−0.89
NMessagest−0.005−0.10−0.094−1.360.0140.80−0.001−0.38
Control variables        
Sizet+10.0100.820.0140.68−0.003−0.48−0.001−0.77
MktRett+10.1950.13      
Mktt+1  −0.026**−2.17−0.000−0.08−0.000−0.72
NTradest+1  0.0140.80  −0.001**−2.33
Listing    0.022*1.70  
Lagged dependent variables        
Returnt−0.004−0.06      
Volatilityt  0.658***11.57    
Volumet    0.986***28.86  
Spreadt      0.842***8.32
R-squared0.090.720.930.83

Similar to our contemporaneous tests, we find that absolute sentiment predicts next day's intraday volatility and volume. A higher absolute sentiment is associated with a lower next-day volatility (β=−0.010, t-stat =−2.47) and a greater proportion of small-sized trading volume (β= 0.004, t-stat = 1.75).

We now examine the two-day predictive results included in Table 7. It comes as no surprise that the longer the window, the harder it is to predict. Neither the message board variables nor firm size and market return are effective in predicting the return two days later. However, the absolute sentiment is still effective in predicting intraday volatility (β=−0.001, t-stat =−2.06) and the proportion of small-trade volume (β= 0.007, t-stat = 1.87). The number of messages continues to be insignificant in predicting trading-related activities.

Table 7.  Two-Day Predictive Power of Posting Activities
Dependent VariableRaw Returnt+2Volatilityt+2Small-Trade Vol.t+2Spreadt+2
Coeff.t-statCoeff.t-statCoeff.t-statCoeff.t-stat
  1. Notes: This table includes regressions to test the two-day predictive power of posting activities by leading the dependent variables and the control variables by two days relative to the message board variables. One-day lagged values of the dependent variables are included as independent variables to control for autocorrelation. Sentiment is the credit-weighted sentiment index, Abs. Sentiment is the absolute value of the sentiment index, Disagreement is the credit-weighted disagreement index, NMessages is the log of the number of messages posted for the stock, Size is the log of the firm's market capital, MktRet is the return on the PowerShares QQQ fund, Mkt is log price of the PowerShares QQQ fund, NTrades is the log of the number of trades in the stock, and Listing is a dummy variable that assumes a value of one for stocks listed on NASDAQ and zero otherwise. t-statistics reported are adjusted based on the heteroskedasticity-consistent covariance matrix developed by White (1980). ***, ** and * denote two-tailed significance at the 1%, 5% and 10% levels, respectively.

Intercept0.0490.34−0.517−1.04−0.018−0.10−0.024−0.90
Lagged msg. board variables        
Sentimentt−0.002−0.13      
Abs. sentimentt  −0.001**−2.060.007*1.87−0.001−0.88
Disagreementt−0.022−0.40−0.006−1.06−0.031−0.84−0.004−1.07
NMessagest−0.037−1.40−0.004−0.100.0381.01−0.004−1.60
Control variables        
Sizet+20.0091.310.032*1.76−0.003−0.58−0.002**−2.57
MktRett+20.4510.63      
Mktt+2  −0.015−1.37−0.002−0.420.0011.15
NTradest+2  0.039**2.79  −0.001−1.52
Listing    0.048***3.08  
Lagged dependent variables        
Returnt+10.1080.93      
Volatilityt+1  0.985***7.69    
Volumet+1    0.922***27.78  
Spreadt+1      0.935***9.95
R-squared0.070.850.920.82

To summarize, we find that when there is no material news about a stock, credit weighted sentiment index significantly negatively predicts the next day's return. Our findings are consistent with a pump-and-dump trading scheme.18 The significant price drop during the dumping period suggests possible trading opportunities for certain traders such as arbitrageurs and contrarian traders who are willing to take opposite positions based on stock message board information. We also find that the absolute sentiment significantly negatively predicts the volatility on the next day as well as two days later and significantly positively predicts the small-trade volume on those days. We conclude that message board sentiment is an important predictor of trading-related activities.

Overall, the results in this section support our hypothesis that online posting activities affect online intraday activities as message board sentiment is significantly related with same day's and next two days’ trading-related activities.

8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

We now discuss the robustness of our regression results.

(i) Multicollinearity

For each regression model, we test for multicollinearity by calculating the uncentered variance inflation factors (VIFs) for the independent variables. We find that the VIFs do not indicate the presence of multicollinearity. Further, although sentiment index is not functionally correlated with disagreement index, we orthogonalize the two indexes and include only the residual as a proxy for disagreement. We find no change in results.

(ii) Alternative Measure of Volatility

The posting of a message does not imply an actual trade at the same time. Also, some online traders might use limit orders rather than market orders. Some of these limit orders may be reflected only in bid-ask prices but not in the actual trade price. Therefore, as an alternative to volume-weighted volatility of actual trade price, we use the standard deviation of midpoints of bid-ask quotes as a measure of volatility. We find consistent results.

(iii) Control for Monday, After holidays, Friday, and the Third Friday Effects

Antweiler and Frank (2004) control for the first trading day after a weekend or a holiday. Also, previous studies find that people post more messages on Fridays. In addition, every third Friday is an option expiry day, which may induce more online talk. We include dummies for these days as control variables. Finally, since some investors prefer trading options over stocks, we include a dummy variable to control whether a stock has options. We find that the inclusion of these dummy variables does not qualitatively alter our results.

(iv) Control for Industries

We control for industry effects in all the regressions in Section 7 using industry dummies included in the regressions in Section 6. We find that our results are robust to the inclusion of these industry dummy variables.

9. CONCLUSIONS

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

In this study, we examine pump and dump behavior via an online message board when no material news is present. Consequently, our sample is not contaminated by news and allows for a clearer examination of the effect of message board activities. Also, to reduce error introduced by stocks with small bulletin board followings, we only include stocks with high message posting activity. Our data source is TheLion.com, a message board aggregator that provides a list of the ten most heavily discussed stocks on each trading day. Our sample consists of these stocks. All messages, including those with and without a self-reported sentiment, are used in measuring online investor sentiment. We use a text classifier to assess the sentiment of a message in which it is not self-reported. TheLion.com has a reward system in which posters earn reputation credits from other users for posting quality messages. This enables us to use a credit-weighted measure of investor sentiment.

We find that stocks without any material news, but still heavily discussed, are generally small cap stocks with weak financials and low institutional holdings – stocks that are more easily influenced by online postings. The main contribution of our study is that it sheds more light on the role and influence of online message boards for such stocks. In view of the special characteristics of the sample firms in this study, the findings summarized below should not be generalized to larger firms with strong financials and high institutional holdings.

Our main findings are as follows. First, using standard event study methodology, we find that there are significant positive abnormal returns on days t−1 and t and significant negative abnormal returns on days t+1 and t+2, with day t being the event day when the stock is heavily discussed among posters. This suggests a two-day pump followed by a two-day dump stock manipulation pattern among online traders. There are no long-term economic effects of the pumping as the stock price reverts to the pre-pump level within a few days.

Second, we find that online traders prefer low-priced stocks with a thin float, a high trading volume, and a low price-to-book ratio. We also find that online traders prefer non- financial stocks.

Third, in our contemporaneous tests, investors’ credit-weighted sentiment dominates the other two message posting parameters (number of messages and disagreement) in explaining return, volatility, and small-sized trading volume. Higher sentiment is associated with higher returns while higher absolute sentiment is associated with less volatility and more small-sized trades. In our predictive tests, sentiment index significantly negatively predicts the next day's return. This is consistent with a pump-and-dump trading behavior. The significant price drop during the dumping period also suggests possible trading opportunities for contrarian traders who are willing to take opposite positions based on stock message board information. We also find that absolute sentiment significantly negatively predicts intraday volatility on the next day as well as two days later and significantly positively predicts the small-trade volume on those days.

Overall, this study shows that for stocks of small firms with weak financials and low institutional holdings, online talk is not trivial. More importantly, this study provides clear evidence that message board sentiment in such stocks is an important predictor of short-term trading-related activities. Online sentiment about a stock on a given day has significant impacts on the stock's return, intraday volatility, and volume on that day and the following day. The findings of this study also suggest the possible use of online message boards for manipulation of small-cap stocks with weak fundamentals. With stock message boards becoming increasingly popular, influential posters on these boards could indulge in manipulation to maximize their trading profits through pump and dump activities. Therefore, regulators need to closely monitor stock message boards.

Appendices

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES

APPENDIX A

Details Related to the Use of the Text Classifier

We use software called RAINBOW (McCallum, 1996) to build our NB text classifier based on training datasets (messages with a self-reported sentiment) and then use it to evaluate messages without a self-reported sentiment. When building the NB text classifier, we restrict each of our training datasets to exactly 1,000 messages per class in order to avoid the class imbalance problem.19 Also, following Das and Chen (2007) who suggest using a smaller sample of 350–500 messages per class to control for over-fitting of data, we repeat all our tests using a standard of 500 messages per class. Since smaller sample results do not alter any of our conclusions appreciably, we allocate 1,000 messages as training data set per sentiment class. When evaluating messages without a self-reported sentiment, NB classifier returns a probability p for each of the predetermined class θj (θjinline image {Short, Strong Sell, Sell, Hold, Scalp, Buy, Strong Buy}). We assign the sentiment class with the highest p to the evaluated message.

APPENDIX B

Details of Sample Stocks
CompanyTickerSectorOpExCompanyTickerSectorOpEx
  1. Notes:

  2. The sectors are: BM-Basic Materials, CG-Consumer Goods, FI-Financial, HC-Healthcare, IG-Industrial Goods, SE-Services, and TE-Technology. Op is one for stocks with options and zero for other stocks. Ex represents the exchange on which the stock is listed and has a value of one for stocks listed on the NYSE, two for stocks listed on AMEX, and three for stocks listed on NASDAQ.

Blue Dolphin Energy CoBDCOBM03Escala Group IncESCLSE03
Empire Resources IncERSBM02EVCI Career Colleges IncEVCISE03
Fieldpoint Petroleum CorpFPPBM02Jewett Cameron Trading LtdJCTCFSE03
TOR Minerals Intl IncTORMBM03China Fin Online Co LtdJRJCSE03
USEC IncUSUBM11Magal Security Systems LtdMAGSSE03
CTI Industries CorpCTIBCG03Mediabay IncMBAYSE03
Forward Industries IncFORDCG03Movie Gallery IncMOVISE13
Leading Brands IncLBIXCG03MPW Industrial Services GP IncMPWGSE03
Oralabs Holdings CorpOLABCG03NYER Medical Group IncNYERSE03
Reunion Industries IncRUNCG02Odimo IncODMOSE03
International Assets HLDG CorpIAACFI03Ruths Chris Steak House IncRUTHSE13
Biocryst Pharmaceuticals IncBCRXHC03Top Tankers IncTOPTSE13
Generex Biotechnology CorpGNBTHC03Phazar CorpANTPTE03
Hemispherx Biopharma IncHEBHC02Altigen Communications IncATGNTE03
Memory Pharmaceuticals CorpMEMYHC03Baidu Com IncBIDUTE03
NABI BiopharamaceuticalsNABIHC13BOS Better Online Solns LtdBOSCTE03
Neurocrine Biosciences IncNBIXHC13Majesco Entertainment CoCOOLTE03
Novavax IncNVAXHC13Endwave CorpENWVTE13
Nutrition 21 IncNXXIHC03Eon Communications CorpEONCTE03
Occulogix IncRHEOHC03IPIX CorpIPIXTE03
Spherix IncSPEXHC03Lightpath Technologies IncLPTHTE03
Threshold Pharamaceuticals IncTHLDHC13Millennium Cell IncMCELTE03
ACS Tech 80 LtdACSEFIG03Makemusic IncMMUSTE03
Arts Way Manufacturing IncARTWIG03Pacific Internet LtdPCNTFTE03
Comstock Homebuilding Cos IncCHCIIG13PPT Vision IncPPTVTE03
China Yuchai International LtdCYDIG11Pixelplus Co LtdPXPLTE03
Highway Holdings LtdHIHOIG03Simclar IncSIMCTE03
HOKU Scientific IncHOKUIG03SMTC CorpSMTXTE03
Imperial Industries IncIPIIIG03Sento CorpSNTOTE03
Host America CorpCAFESE03Synergy Systems IncSYNXTE03
Acacia Research CorpCBMXSE03Tramford International LtdTRFDFTE03
Dryships IncDRYSSE13Vocaltec Communications LtdVOCLTE03

APPENDIX C

Event Study Methodology

For estimating the market model, we use the standard event study methodology, with a 255-day pre-event estimation window starting on t-300, if possible, and ending on t-46. We require three days as the minimum length of the estimation window. The abnormal return inline image for stock j on day t is computed as:

  • image

 where the coefficients inline image and inline image are Scholes-Williams (1977) estimates of inline image and inline image, inline image is the return on stock j on day t and inline image is the return on an equally-weighted CRSP index, with dividend reinvested, on day t.20 The average abnormal return (AAR) on day t is the sample mean:

  • image

 where N is the number of sample stocks on event day t. Furthermore, the cumulative average abnormal return (CAAR) on a portfolio starting from day t1and ending on day t2 is as follows:

  • image

 We use the Patell (1976)Z test to examine the null hypothesis that the average abnormal stock return equals zero assuming asymptotically unit normal.21 To avoid the dependence on normality of return distributions as in the Patell parametric test, we also use the Cowan (1992)Z test, a nonparametric generalized sign test. We also perform an event study for trading volume using a similar methodology as above.

To check that our event study results are not affected by the choice of our methodology, we perform some robustness tests. On average, our sample firms are fundamentally poor and their stock prices might follow a negative drift over time. To capture any drift that might occur after the event day, we use a post-event 255-day estimation window instead of the pre-event window.22 To also capture any drift that might occur before the event day, we conduct the event study using a more comprehensive pooled estimation window, including both the pre-event 255 days and the post-event 255 days, a total of 510 days. Finally, instead of the Scholes-Williams market model, we use a market-adjusted model, with the abnormal return being the difference between the stock's raw return and the return on the market index. The market is proxied by the CRSP equally-weighted index. We run this model over pre, post and pooled estimation windows.

The above additional tests do not change the results appreciably. Therefore, we only report the results for the pre-event estimation window with the Scholes-William market model. Additional test results are available on request from the authors.

Footnotes
  • 1

    Rubin and Rubin (2010) discuss the characteristics of the Internet that distinguish it from traditional information sources and have contributed to the rapid increase in its use. Orens et al. (2010) focus on the use of Internet by firms as a separate source for disclosing non-financial information and find evidence of economic relevance of such disclosure.

  • 2

    Following links provide information about SEC enforcement actions against internet fraud and manipulation: http://www.sec.gov/investor/pubs/pump.htm; http://www.sec.gov/divisions/enforce/internetenforce.htm

  • 3

    This is expected as such stocks are the ones most vulnerable to online manipulation.

  • 4

    An influential poster may work alone. Or, similar to Khwaja and Mian (2005) in which brokers in an emerging market collude in a ‘pump and dump’ stock price manipulation scheme, two or more influential posters may collaborate. The collaboration could occur through private communication channels such as email and instant messenger.

  • 5

    As the ‘pump and dump’ works for the message board subscribers, it reinforces the original poster's reputation.

  • 6

    Material news is news that is likely to affect the value of the firm's securities or influence investors’ decisions. Material news includes information such as unusual corporate events (for example, mergers, divestitures, significant changes in management, downsizing, FDA approvals, and new product introductions), earnings announcements, analyst recommendations, stock splits, etc. Technical news such as 200-day moving average alerts and abnormal volume reports are not considered material news.

  • 7

    RavenPack is a leading provider of news analytics and machine-readable content. The company specializes in linguistic analysis of high volume, real-time news from high-end newswire services. All the news for any firm at any specific time can be identified using their database. For more details, please see http://www.ravenpack.com.

  • 8

    Scalping implies the intent to buy and sell quickly to make a day trade profit without any specific sentiment.

  • 9

    For details of text classifiers, in general, please see Sebastiani (2002) and Das and Chen (2007). For details of the NB algorithm, please see McCallum and Nigam (1998) and Lewis (1998).

  • 10

    The relation between credit score and its impact on the message board is likely to be non-linear. Therefore, we use a log transformation to capture the decreasing marginal effect of the poster's credit.

  • 11

    Note that inline image and is not correlated with Sentiment or the standardized bullishness index. Therefore, we do not expect multicollinearity to be an issue in a model with both Sentiment and Disagreement.

  • 12

    We download each message's detailed information such as the associated stock symbol, time posted, poster's name, message length, and message content. The content of each message is also saved in a separate text file. These text files are used for building the text classifier, sample training, and classification evaluation.

  • 13

    Technically, a poster can register multiple screen names. However, the reputation system of Thelion.com reduces the probability of forum participants doing so. Posts written by authors with more aggregate credits are more likely to be read and paid attention to. Registering multiple accounts would reduce the accumulated reputation for any particular account. Thus, posters have little incentive to register multiple usernames.

  • 14

    Lee and Radhakrishna (2000) define a small trade as <=$2,500 and a large trade as >$20,000 based on 1990–91 data. From 1990–91 to 2007, the Consumer Price Index has increased by 48.7%, or about 50%. We take this inflation into account and accordingly our cutoffs are 50% greater than Lee and Radhakrishna's. Further, Lee and Radhakrishna find that dollar-value based proxies are generally less noisy than trade-size based proxies in separating small and large investor transactions. Therefore, we use dollar-value based proxies in this paper.

  • 15

    Please see Strong (1992) for a guide to event study methodology and procedures for modeling abnormal returns.

  • 16

    In order to generate comparable results, we follow Antweiler and Frank (2004) and use raw return instead of market-adjusted return. This market return is included as a control variable. We also run tests using market-adjusted return, with market return not included as a control variable in the model. The results are consistent.

  • 17

    Antweiler and Frank (2004) use S&P 500 as the market index. Because our sample stocks are generally small and listed on NASDAQ, we use the NASDAQ-100 index to proxy for the market.

  • 18

    Because of data limitations, it is not possible to establish ‘pump and dump’ definitively. We are mainly examining if our findings are consistent with ‘pump and dump’.

  • 19

    The training data for building the text classifier under each class are set to be equal in order to avoid an information imbalance. For instance, the classifier will have an optimistic bias if there are more training data in Buy and Strong Buy (buy-side) than in Sell and Strong Sell (sell-side). For further discussion of this issue, please see Japkowicz and Stephen (2002) and Estabrooks et al. (2004).

  • 20

    Using an equally-weighted index is appropriate in this study since most sample stocks are small-cap.

  • 21

    Event study with serial dependence (CDA) and potential variance change tests yield similar results. These results are not included in the paper but are available on request from the authors.

  • 22

    Please see MacKinlay (1997) and Cowan et al. (1990) for a justification of using a post-event window.

REFERENCES

  1. Top of page
  2. Abstract
  3. 1. INTRODUCTION
  4. 2. ISSUES AND HYPOTHESES
  5. 3. STUDY DESIGN
  6. 4. SAMPLE AND DATA
  7. 5. EVENT STUDY
  8. 6. TRADING INCENTIVES AND TARGET STOCK CHARACTERISTICS
  9. 7. SHORT-TERM EFFECTS OF ONLINE POSTING ACTIVITIES ON TRADING ACTIVITIES
  10. 8. ROBUSTNESS TESTS FOR CONTEMPORANEOUS AND LEAD-LAG REGRESSIONS
  11. 9. CONCLUSIONS
  12. Appendices
  13. REFERENCES
  • Admati, A.R. and P.C. Pfleiderer (2001), ‘Disclosing Information on the Internet: Is It Noise or Is It News?’, Technical Report (Graduate School of Business, Stanford University).
  • Aggarwal, R.K. and G. Wu (2006), ‘Stock Market Manipulations’, Journal of Business, Vol. 79, No. 4, pp. 191553.
  • Antweiler, W. and M.Z. Frank (2004), ‘Is all that Talk Just Noise? The Information Content of Internet Stock Message Boards’,Journal of Finance, Vol. 59, No. 3, pp. 125995.
  • Brock, W., J. Lakonishok and B. LeBaron (1992), ‘Simple Technical Trading Rules and the Stochastic Properties of Stock Returns’, Journal of Finance, Vol. 47, No. 5, pp. 173164.
  • Chan, W.S. (2003), ‘Stock Price Reaction to News and No-news: Drift and Reversal after Headlines’, Journal of Financial Economics, Vol. 70, No. 2, pp. 22360.
  • Cook, D.O. and X. Lu (2009), ‘Information Matters, Noises Don’t: Internet Message Boards Affect Stock Returns’, Working Paper (University of Alabama).
  • Cowan, A.R. (1992), ‘Nonparametric Event Study Tests’, Review of Quantitative Finance and Accounting, Vol. 2, No. 4, pp. 34358.
  • Cowan, A.R., N. Nayar and A.K. Singh (1990), ‘Stock Returns Before and After Calls of Convertible Bonds’, Journal of Financial and Quantitative Analysis, Vol. 25, No. 4, pp. 54954.
  • Daniel, K. and S. Titman (1997), ‘Evidence on the Characteristics of Cross Sectional Variation in Stock Returns’, Journal of Finance, Vol. 52, No. 1, pp. 133.
  • Daniel, K., D. Hirshleifer and A. Subrahmanyam (1998), ‘Investor Psychology and Security Market Under- and Over-Reactions’, Journal of Finance, Vol. 53, No. 6, pp. 183985.
  • Das, S.R. and M.Y. Chen (2007), ‘Yahoo! For Amazon: Sentiment Extraction from Small Talk on the Web’, Management Science, Vol. 53, No. 9, pp. 137588.
  • Das, S.R., A. Fa. De. Martinez-Jerez and P. Tufano (2005), ‘eInformation: A Clinical Study of Investor Discussion and Sentiment’, Financial Management, Vol. 34, No. 3, pp. 10337.
  • DeMarzo, P., D. Vayanos and J. Zwiebel (2003), ‘Persuasion Bias, Social Influence, and Uni-dimensional Opinions’, Quarterly Journal of Economics, Vol. 118, No. 3, pp. 90967.
  • Eren, N., H.N. Ozsoylev and P.E. Street (2009), ‘Hype and Dump Manipulation’, Working Paper (University of Minnesota).
  • Estabrooks, A., T. Jo and N. Japkowicz (2004), ‘A Multiple Resampling Method for Learning from Imbalances Data Sets’, Computational Intelligence, Vol. 20, No. 1, pp. 1836.
  • Fama, E.F. and K.R. French (1992), ‘The Cross-Section of Expected Stock Returns’, Journal of Finance, Vol. 47, No. 2, pp. 42765.
  • George, T.J. and C.Y. Hwang (2004), ‘The 52-week High and Momentum Investing’, Journal of Finance, Vol. 59, No. 5, pp. 214575.
  • Gu, B., P. Konana, A. Liu, B. Rajagopalan and J. Ghosh (2006), ‘Predictive Value of Stock Message Board Sentiments’, Working Paper (University of Texas at Austin).
  • Hirschey, M., V.J. Richardson and S. Scholz (2000), ‘Stock-Price Effects of Internet Buy-Sell Recommendations: The Motley Fool Case’, Financial Review, Vol. 35, No. 2, pp. 14774.
  • Japkowicz, N. and S. Stephen (2002), ‘The Class Imbalance Problem: A Systematic Study’, Intelligent Data Analysis Journal, Vol. 6, No. 5, pp. 42950.
  • Jiang, G., P.G. Mahoney and J. Mei (2005), ‘Market Manipulation: A Comprehensive Study of Stock Pools’, Journal of Financial Economics, Vol. 77, No. 1, pp. 14770.
  • Jones, C.M, G. Kaul and M.L. Lipson (1994), ‘Transactions, Volume and Volatility’, Review of Financial Studies, Vol. 7, No. 4, pp. 63151.
  • Keasler, T.R. and C.R. McNeil (2010), ‘Mad Money Stock Recommendations: Market Reaction and Performance’, Journal of Economics and Finance, Vol. 34. No. 1, pp. 122.
  • Khwaja, A.I. and A. Mian (2005), ‘Unchecked Intermediaries: Price Manipulation in an Emerging Stock Market’, Journal of Financial Economics, Vol. 78, No. 1, pp. 20341.
  • Klibanoff, P., O. Lamont and T.A. Wizman (1998), ‘Investor Reaction to Salient News in Closed-End Country Funds’, Journal of Finance, Vol. 53, No. 2, pp. 67399.
  • Korczak, A., P. Korczak and M. Lasfer (2010), ‘To Trade or Not to Trade: The Strategic Trading of Insiders Around News Announcements’, Journal of Business Finance & Accounting, Vol. 37, Nos. 3 & 4, pp. 36907.
  • Koski, J.L, E.M. Rice and A. Tarhouni (2007), ‘Noise Trading and Volatility: Evidence from Day Trading and Message Boards’, Working Paper (University of Washington).
  • Kumar A. and C.M.C. Lee (2003), ‘Individual Investor Sentiment and Comovement in Small Stock Returns’, Working Paper (Cornell University).
  • Kyle, A.S. and S. Viswanathan (2008), ‘How to Define Illegal Price Manipulation’, American Economic Review, Vol. 98, No. 2, pp. 27479.
  • Lakonishok, J. and S. Smidt (1986), ‘Volume for Winners and Losers: Taxation and Other Motives for Stock Trading’, Journal of Finance, Vol. 41, No. 4, pp. 95174.
  • Lee, C.M.C. and B. Radhakrishna (2000), ‘Inferring Investor Behavior: Evidence from TORQ Data’, Journal of Financial Markets, Vol. 3, No. 2, pp. 83111.
  • Lewis. D.D. (1998), ‘Naive Bayes at Forty: The Independence Assumption in Information Retrieval’, In Proceedings of the European Conference on Machine Learning, pp. 415.
  • MacKinlay, A. C. (1997), ‘Event Studies in Economics and Finance’, Journal of Economic Literature, Vol. 35, No. 1, pp. 1339.
  • McCallum, A.K. (1996), ‘Bow: A Toolkit for Statistical Language Modeling, Text Retrieval, Classification and Clustering’ (http://www.cs.cmu.edu/~mccallum/bow).
  • McCallum, A.K. and K. Nigam (1998), ‘A Comparison of Event Models for Naive Bayes Text Classification’, in Proceedings of the AAAI-98 Workshop on Learning for Text Categorization, pp. 4148.
  • Orens, R., W. Aerts and D. Cormier (2010), ‘Web-Based Non-Financial Disclosure and Cost of Finance’, Journal of Business Finance & Accounting, Vol. 37, Nos. 9 & 10, pp. 105793.
  • Neumann, J.J. and P.M. Kenny (2007), ‘Does Mad Money Make the Market go Mad? Quarterly Review of Economics and Finance, Vol. 47, No. 5, pp. 60215.
  • Patell, J.M. (1976), ‘Corporate Forecasts of Earnings per Share and Stock Price Behavior: Empirical Test’, Journal of Accounting Research, Vol. 14, No. 2, pp. 24676.
  • Rubin, A. and E. Rubin (2010), ‘Informed Investors and The Internet’, Journal of Business Finance & Accounting, Vol. 37, Nos. 7 & 8, pp. 84165.
  • Scholes, M. and J. Williams (1977), ‘Estimating Betas from Nonsynchronous Data’, Journal of Financial Economics, Vol. 5, No. 3, pp. 30927.
  • Sebastiani, F. (2002), ‘Machine Learning in Automated Text Categorization’, ACM Computing Surveys, Vol. 34, No. 1, pp. 147.
  • Shiller, R.J. (1995), ‘Conversation, Information, and Herd Behavior’, American Economic Review, Vol. 85, No. 2, pp. 18185.
  • Strong, N. (1992), ‘Modelling Abnormal Returns: A Review Article’, Journal of Business Finance & Accounting, Vol. 19, No. 4, pp. 53353.
  • 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, No. 4, pp. 31335.
  • Tumarkin, R. and R.F. Whitelaw (2001), ‘News or Noise? Internet Message Board Activity and Stock Prices’, Financial Analyst Journal, Vol. 57, No. 3, pp. 4151.
  • van Bommel, J. (2003), ‘Rumors’, Journal of Finance, Vol. 58, No. 4, pp. 1499520.
  • Vilardo, M.F. (2004), ‘Online Impersonation in Securities Scams’, IEEE Security and Privacy, Vol. 2, No. 3, pp. 8285.
  • White, H.L. (1980), ‘A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity’, Econometrica, Vol. 48, No. 4, pp. 81738.
  • Wysocki, P.D. (1999), ‘Cheap Talk on the Web: The Determinants of Postings on Stock Message Boards’, Working Paper (University of Michigan).
  • Wysocki, P.D. (2000), ‘Private Information, Earnings Announcements and Trading Volume’, Working Paper (MIT Sloan School of Management).
  • Zhang, Y. and P.E. Swanson (2010), ‘Are Day Traders Bias Free? Evidence from Internet Stock Message Boards’, Journal of Economics and Finance, Vol. 34, No. 1, pp. 96112.