Abstract: Predicting stock behavior via social media has attracted a great deal of attentions in the finance and knowledge management disciplines. Due to the diversity of social media participants and discussion topics, it is difficult to improve the accuracy of stock behavior prediction by applying traditional methods which based on whole level of social media information. In this paper, text feature extraction technology, principal components analysis and EM clustering are used to identify stakeholders and topics related to a special firm by social media messages' similar writing style and content feature. Furthermore, four types of social media variable are extracted from information activity intensity and sentiment inclinations to build stock behavior regression models for each stakeholder and topic. Finally, Bank of America Company's message board on Yahoo! Finance forum is chosen as our experimental platform. The validity and practicability of our proposed method are tested in experimental result.

Key words: social media, stock behavior, feature extraction, sentiment analysis