eInformation: a clinical study of investor discussion and sentiment

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

We examine the information flow for four stocks over seven months to trace the relationship between on-line discussion, news activity, and stock price movements. On-line discussions support numerous unsubstantiated rumors, substantial on-point exchanges, and quick dissemination of imminent and recently released information. Applying language-processing routines to message board postings and news, we create sentiment and disagreement measures or "eInformation." We analyze the determinants of sentiment and disagreement, and trace links between news, eInformation, and stock returns. This intensive clinical study of on-line discussions suggests mechanisms individual investors and groups can use to analyze and digest company information.

**********

In light of the large body of research on informationally efficient markets, there seems little left to learn from the continued empirical examination of information and markets. It would seem similarly pointless for individual investors to try to compete with professional analysts. However, understanding the individuals' investing decisions has been one of the most vibrant research streams in recent years (see Roll, 1986, Odean, 1998, Gervais and Odean, 2001, and Barber and Odean, 2001).

Technology, in the form of stock chat message boards, now provides a new real-time window into discussions by individual investors. It is instructive to peek through this window to observe how information is digested, how sentiment evolves, and how perceptions are related to prices.

The method we adopt in this article is to use a clinical, i.e., small sample, approach to understanding investor behavior. Before framing hypotheses or constructing tests, it is important to establish a base level of understanding in an area. Thus, our article is decidedly descriptive, part of a long inductive tradition in economics (Blaug, 1992). We do not attempt to either affirm or reject theory. Rather, we suggest a series of working conjectures (or hypotheses) that can be developed through subsequent model building and large-scale empirical study.

We have three goals in this article. First, we closely analyze the people who share their opinions (posters) and their discussions surrounding a few stocks. Given the anonymous nature of this activity, we instead choose to study an outlier by interviewing an extensive poster. Doing so enables us to understand why someone would spend substantial amounts of time posting messages to one of the boards we study.

As part of our analysis, we also focus attention on the discussions themselves. Although there is a perception that postings are "garbage," to the contrary, discussions sustain on-point exchanges, generate possibly non-public information, quickly disseminate public information from news stories, and serve as forums where investors can extract meaning from information. Chat rooms and postings are also sources of numerous unsubstantiated rumors, adding noise to the information flow. Nevertheless, the fact that even some nonpublic information may be released on the boards--and the observation that posters use the boards to test their own analyses and obtain those of others--may explain why posters and surfers continue to frequent these chat board sites.

Second, using language-processing algorithms, we measure the intensity and dispersion of sentiment (which we dub eInformation) for over 170,000 messages posted about four stocks. We analyze the determinants of the level of sentiment and disagreement among posters, and find that there is a close relationship between sentiment levels, stock prices, and trading volume. We also find that disagreement is related to the intensity of discussion.

Finally, we explore the usefulness of expressed investor sentiment (eInformation) to predict stock returns. Our clinical study confirms other studies that fail to find predictive power forecasting returns (Antweiler and Frank, 2002, 2004, and Das and Chen, 2003).

The article proceeds as follows. Section I deals with our clinical design. In Section II, we discuss the demographics of posters, detailing our interview with an especially active investor-discussant. Section III reports on our clinical examination of the nature of the discussions and the quality of information in those discussions. Section IV describes our computer-generated measures of sentiment and disagreement (eInformation) that are extracted using language-processing algorithms. In Section V, we analyze the determinants of our eInformation measures. In Section VI, we examine the relationship of eInformation to the price formation process. Finally, in Section VIII, we summarize the hypotheses that emerge from this clinical investigation.

I. Sample Design

We study four firms over a period of seven months. We use these four firms as archetypes for different information environments where traditional and new eInformation flows vary. As befits a clinical study, we attempt to dig deeply into these four firms, using our observations to derive hypotheses for large-scale studies. We have deliberately not selected pathological examples where posters have used stock message boards to explicitly manipulate prices (Leinweber and Madhavan, 2001).