Assessing dynamic qualities of investor sentiments for stock recommendation

Investor based social networks enable investors to share sentiments (e.g., bullish or bearish) about stock trends. Modeling and predicting the qualities of investor sentiments is a critical problem when aggregating sentiments and making investment recommendations. Most previous works relied on the overall past performance of investors to assess the quality of investor sentiments. However, we show that it is beneficial to assume that the qualities of the sentiments of a single user to vastly different stocks are not the same in this work. We propose a novel method called correlation-based robust dynamic qualities (CBRDQ) to model the qualities of investor sentiments more accurately by considering the correlations among stocks. The correlations, which are designed to reflect how helpful the qualities of user sentiments about a stock is for inferring the qualities of user sentiments about another stock, are employed as weights to estimate the quality of an investor's sentiment about a given stock. We refer to this quality measurement dynamic quality since it assigns different qualities to the sentiments from a single user to different stocks. Based on a large-scale dataset from the real-world investor platform StockTwits, we evaluate CBRDQ and several conventional methods in a unifying stock recommendation framework. The results support the use of dynamic quality rather than static quality. Moreover, the comparative results demonstrate the effectiveness of our method in making investment recommendations.