eInformation: a clinical study of investor discussion and sentiment

Financial Management (Financial Management Association), Autumn, 2005 by Sanjiv Das, Asis Martinez-Jerez, Peter Tufano

There is a less obvious conclusion from this analysis. Although the boards are places where many rumors are suggested, we did not see evidence that they were "rumor mills," where these rumors themselves were the source of sustained discussion. In 47 of 54 instances, each of the rumors generated fewer than four subsequent posts, and unsubstantiated rumors generated less discussion.

If we were to construct a "wheat and chaff" measure for the boards, they would probably perform poorly. On the positive side, of the 16 actual news events in Table III, half were foreshadowed on the board discussions. This average sounds good, but this is conditional on knowing that something actually happened. In contrast, an avid reader of the boards continually scanning for merger announcements would have had useful information 2% of the time, with the remainder of the stories being unsubstantiated rumors. Although it is impossible to compare these percentages without knowing the gains and losses of trading on this information, it seems that the gains from being right 2% of the time would be more than offset by being wrong 98% of the time.

IV. The Concept and Measurement of eInformation

In the remainder of the study, we use computer algorithms to classify the 170,953 messages and to relate our measures of sentiment and disagreement to both information sources and stock prices.

A. Definition and Motivation of eInformation

The simplest characterizations of the flow of information are activity measures: simple counts of the numbers of news stories or postings, or the length of news story or posting. These metrics are used by Mitchell and Mulherin (1994; number of news stories) and Wysocki (1999; number of postings). These activity measures indicate the level of interest, excitement, puzzlement, or "buzz" about the information set, similar to the measure of the decibels of noise in trading pits used by Coval and Shumway (2001). Activity measures are based on the notion that discussion (whether in person, by electronic posting, or news stories) is correlated with the salience and newness of information releases.

Capturing the content of the information is a more complicated matter. Although we did some of this by hand, this method is infeasible for a large data sample. Therefore, we extract a subjective measure of the meaning of the information by using computer algorithms that read and categorize the content of each individual message. The algorithms parse the degree to which the message conveys a buy, sell, or neutral sentiment about a stock. By aggregating these messages over some time period, we can gauge the average sentiment as well as the distribution of "posting sentiment" manifested by the stock message board information flow. We also use this method to classify the "news sentiment" of press stories. We call the combination of activity measures and content measures (distribution of sentiment indices) "eInformation." (3)

Sentiment is an intangible quality that is critical to many models used in financial economics. In behavioral finance, investor sentiment (or noise trader sentiment) is used to explain deviations in prices from "rational" levels (see Delong, Shleifer, Summers, and Waldman, 1990). To measure sentiment, academics have used the closed-end fund discount (Lee, Shleifer, and Thaler, 1991), flows into mutual funds (Goetzman, Massa, and Rouwenhorst, 2004), and subjective determinations based on reading various news stories (Gay, Kale, Kolb, and Noe, 1994). We create new measures of sentiment by examining posting or news stories.