The weight of words: textual data versus sentiment analysis in stock returns prediction

The focus of this paper is to understand whether the words contained in a text corpus improves the explained variance of stock returns better than the use of the polarity of the same texts, obtained through a sentiment analysis using a generic ontological dictionary. The empirical analysis is based on the content of a weekly column in the most important Italian financial newspaper, which published past information and analysts’ recommendations on listed companies. The use of textual data clearly increases the explained variance of stock returns but, through comparisons between data mining techniques, we observed minor differences in terms of MSE, by adding a selection of specific terms as features. In this context, the text mining approach proved to be very useful to improve the explanatory power of forecasting models, while it emerged the limited explanatory power of an automatic sentiment analysis based on a generic lexicon.