Enhancing Investment Decisions with Sentiment Analysis: A Probabilistic Ranking Framework*

We develop a probabilistic framework to extract sentiment information from text by training a model to predict and rank sentiments in newly encountered documents. Our approach imposes a joint semi-parametric model on text and ordinal response variables, addressing the challenges of sparse sentiment signals and complex response distributions. Through a word screening procedure and the use of normalized ranks, our approach achieves consistent sentiment ranking without estimating the full model. Applying our method to the Dow Jones Newswires, we demonstrate its effectiveness in extracting return-predictive signals.