Risk Exposures from Risk Disclosures: What They Said and How They Said It

Abstract

We extract information from 10-K risk disclosures: topic models measure what is discussed, and context models measure how it is discussed. We find that both contain significant predictive information about future aggregate risk exposures, even controlling for (structured) firm characteristics, and that this information is economically valuable. For market exposure, only management's choice of context is useful; for other factors, tradable and nontradable, both context and topic information is useful. Further, we show that topics and contexts are statistically distinct. We present evidence that management's discussion helps to predict some future corporate actions, which in turn can alter the firm's exposure to aggregate risk.

Keywords: Text Analysis, Asset Pricing, Factor Model, Conditional Betas, Word embedding

JEL Classification: C23, C53, G11, G12, G17