Abstract
A large body of literature has tried to identify the relative information contributions of different characteristics - jointly or in isolation - to the cross-section of stock returns. These characteristics generally cover three data sources: market, fundamentals, and analyst recommendations. In this paper, we complement these traditional data sources by processing a large corpus of 1,309,335 sell-side analyst reports, covering 2,515 stocks over the last three decades. Our analysis leverages on machine learning (ML) to emulate a sophisticated agent who utilizes all data sources to form forward-looking portfolios. We find that analyst textual content on top of conventional data sources has a higher-order complex effect, which feeds into the portfolio selection process in a nonlinear form. Differently from a simple linear technology, a nonlinear machine learning model yields positive and strongly significant alphas. Finally, the ML long-short portfolio strategy exhibits a higher correlation with common characteristics related to R&D, leverage, and cash holdings, stressing the intertwined and complex nature of analyst content.

This publication has 52 references indexed in Scilit: