Predicting Stock Price from Financial Message Boards with a Mixture of Experts Framework

Many online financial message boards exist where users can discuss the latest information about their favorite stocks. However, the sheer number of postings and the high level of noise in these postings present a significant challenge to message board users seeking to extract recommendations on whether to buy or sell a stock. Earlier attempts to extract sentiment and information from these boards and relate it to stock markets have been largely unsuccessful. This paper presents a new approach which focuses on identifying the most historically accurate posters in message boards. Specifically, we use a mixture of experts framework to identify these posters and analyze their sentiments in a completely automated fashion. We find that this approach not only extracts information from message boards, but that implementable strategies based on this extracted information can automatically be devised to achieve statistically significant returns even after adjusting for both market effects and commission rates.