Extracting Relevance from Virtual Investing-Related Community Postings

The rapid growth of online investing and virtual investing-related communities (VICs) has a wideraging impact on research, practice and policy. In this context, this research addresses how information is generated, discussed, and diffused within and across VICs, and how such activities impact market efficiency. Regulators are particularly interested given the potential for fraud and spreading of false rumors. However, understanding information processing in VIC is a challenge given enormity of posted messages. Automated analysis of these messages is primarily complicated by three factors: (a) the amount of irrelevant messages or "noise" messages (e.g., spam, insults), (b) the highly unstructured nature of the text (e.g., abbreviations), and finally, and (c) the wide variation in what is considered relevant information for a given company. We have developed a mechanism relying on commonly occurring terms and a set of classifying criteria to identify: (1)"noisy" messages that bear no relevance to the topic at hand, (2) messages that have relevance to the topic at hand, but do not express an opinion as to the quality of the investment, and (3) messages that are both relevant and express a sentiment about the quality of the investment. To test our mechanism we have collected approximately 3 million messages related to 46 stocks over a 2-year period. Preliminary results show sufficient promise to classify messages and how participants react. Preliminary results show the classifier a classification accuracy of 54% . 1 This research is supported by NSF ITR grant number IIS-0218988