WORDS BEFORE VOLATILITY: DO TEXT-BASED RISK DISCLOSURES PREDICT FUTURE STOCK RETURN VOLATILITY?

Abstract

Textual analysis of 10-K filings has long been used to examine the relationship between corporate disclosures and market outcomes. Prior research shows that textual measures of sentiment and uncertainty are associated with stock returns, volatility, and other indicators of firm risk (Li, 2008; Loughran and McDonald, 2011; Campbell et al., 2014). However, traditional dictionary-based approaches rely on predefined word lists and may fail to capture contextual meaning in financial disclosures. This study examines whether large language models (LLMs) provide a more infor mative measure of disclosure-based risk than traditional dictionary-based methods. Using Item 1A (Risk Factors) and Item 7 (Management’s Discussion and Analy sis) disclosures from U.S. 10-K filings, we construct LLM-based volatility scores and compare their performance with dictionary-based uncertainty measures. The analysis combines pooled OLS regressions, firm and time fixed effects models, and out-of-sample forecasting tests. The results show that LLM-based measures exhibit stronger explanatory power than dictionary-based measures in pooled cross-sectional regressions and are pos itively associated with subsequent realized stock return volatility. However, these relationships largely disappear once firm and time fixed effects are introduced, in dicating that textual measures primarily capture persistent differences across firms rather than time-varying changes in risk. Consistent with this interpretation, nei ther LLM-based nor dictionary-based measures improve out-of-sample volatility forecasts beyond standard financial control variables. The findings suggest that while LLMs provide a richer representation of disclosure based risk, their value lies primarily in characterizing cross-sectional differences in firm risk rather than improving the prediction of future stock volatility.