Bottom Up vs Top Down: What Does Firm 10-K Tell Us?

Abstract

In contrast to the recent increasing focus on large language models, we propose a bottom-up approach that exploits the predictive power of individual words. Using a data-driven word dictionary and lasso-regularized regressions on large panels of word counts, we estimate the cross-section of stocks' expected returns. A factor summarizing this information generates economically and statistically significant returns, largely unexplained by standard factor models. The dictionary includes risk-related words such as "currency," "oil," and "restructuring," which increase expected returns, while words like "acquisition," "derivatives," and "quality" decrease them.

Keywords: Text AnalysisAsset PricingWord DictionaryWord Count