The New York Times

September 23, 2004

Good Stock Advice or Online Noise?


TALK is cheap, particularly on the Internet. Stock message boards are a case in point. Every day participants post tens of thousands of tips about which way various stocks are heading. Is any of this worth reading?

Recently, two financial economists from the University of British Columbia, Werner Antweiler and Murray Z. Frank, examined the message board phenomenon in a paper entitled "Is All That Talk Just Noise? The Information Content of Internet Stock Message Boards," published in the June 2004 issue of The Journal of Finance.

They collected more than 1.5 million messages from two online boards, Yahoo Finance and Raging Bull, and analyzed them using methods of computational linguistics and econometrics.

The computational linguistics techniques allowed them to classify messages with respect to whether they advocated buying, holding or selling the stock in question. Most of the messages were short and direct, allowing the algorithms to do a pretty good job of classification.

Of course, some postings could not be easily classified. As the authors charitably remark, "A remarkable range of sometimes quite odd things are said in the messages."

Mr. Antweiler and Mr. Frank then merged the estimated buy/sell/hold signals into one "bullishness measure," which was a slightly modified version of the ratio of buy to sell recommendations.

During the period they examined in their article, January to December 2000, the Internet bubble dissipated. Bullish messages, however, continued to proliferate throughout the year, perhaps reflecting a "buy on the dips" sentiment popular at the time. Or perhaps the messages represented attempts by day traders to talk up the value of the stocks in their dwindling portfolios.

But did the message volume, timing and sentiment forecast anything useful? The three most interesting features of a stock on a given day are its return (how much it increased or decreased), its volume (how many shares were traded) and its volatility (how much the price fluctuated).

The authors found that the characteristics of messages helped predict volume and volatility. Perhaps more surprisingly, they also found that the number of messages on one day helped predict stock returns the next day. The degree of predictability, however, was weak and reversed itself the next trading day. Perhaps cheap talk can move stock prices a tiny bit, but if so, the response was only temporary.

The bullish sentiment of messages was positively associated with contemporaneous returns, but has no predictive power for future returns. Traders post bullish messages about a given stock on days when its price goes up, but it is hard to determine which way the causation runs. The fact that the postings do not predict future returns suggests that, if anything, it is the returns that cause the postings rather than the reverse.

Economists who believe in efficient markets can breathe a sigh of relief. It would indeed be a problem if publicly available information like message board postings had meaningful predictive power for stock returns.

The story was different with respect to volatility. It appeared that the more messages posted about a stock one day, the higher its volatility was the next day.

Trading volume was also correlated with messages. However, the apparent causation here was somewhat subtle. Message posting appeared to cause volume when the researchers examined daily data. But when they looked at the market at 15-minute intervals, trading volume seemed to cause messages.

One hypothesis consistent with these observations about volume and messages was that people might tend to post messages shortly after buying a stock. Even though there are two sides to every trade, the seller has little incentive to brag about the dog he just unloaded, while the buyer has a strong incentive to recommend that others buy the stock he has just purchased.

This effect may lead to the bullish bias in postings alluded to above, which persisted even in the face of declining markets.

An old adage says, "It's differences of opinion that make horse races." Just so with stock markets: the more agreement in the buy/sell messages on a given day, the less trading volume was observed. But substantial agreement of message postings on one day was correlated with a higher level of trading volume the next. Perhaps this is a momentum effect - if the crowd says it is going to buy, individual participants might decide they should buy too.

Finally, the authors asked whether the message boards simply repeat information already publicly available, or whether they add new information. It appeared from their analysis that high message volume helps predict articles in The Wall Street Journal published two or three days later, though again the causality was unclear.

It could be, for example, that the messages and the articles were stimulated by the same news, but it takes longer for the stories to be published than for the messages to be posted.

In summing up their study, the authors conclude that the talk on message boards is not just noise. Though the predictive power for returns is too small to be meaningful, message board activity does seem to help predict volatility and volume.

Furthermore, the correlations between message board postings, volume and volatility are not only statistically significant, they are quite large compared with the magnitude of correlation one typically observes in financial markets.

This is noteworthy in itself. Volatility is a critical factor in options prices, and stocks of Wall Street companies respond to changes in trading volume. So it seems there is something to be learned from the textual analysis of message board data.

More important, perhaps, the authors have shown that computational linguistics can be a useful tool in financial economics. In current work, the authors are examining the effect of news stories on stock prices using similar methods. One could also analyze Securities and Exchange Commission filings with such techniques, perhaps shedding some light on corporate accounting practices.

In the 1970's we saw the rise of Wall Street quantitative analysts. Then came program trading. Perhaps computational linguistics and textual data mining will become the new hot technologies in financial economics.

Hal R. Varian is a professor of business, economics and information management at the University of California, Berkeley.

Copyright 2004 The New York Times Company | Home | Privacy Policy | Search | Corrections | RSS | Help | Back to Top