Risky Words and Returns

Abstract

To discover dynamic risks that determine the expected stock returns, I develop a method to predict returns through the text of firms' risk disclosures. By cross-sectionally regressing returns on the text of risk disclosures, I find certain words in the risk discussions (defined as "risky words") that have independent predictive power for the cross-section of stock returns: an out-of-sample strategy that times "risky words" earns up to 22% annual alpha between 2005-2023. Then I group risky words into 14 orthogonal clusters that are jointly eminent determinants of expected returns. Firm characteristics, industries, sentiments, and previously discovered text features do not explain the results.

Keywords: Risk sections, stock returns, textual analysis, asset pricing, machine learning