Optimal Narratives for Predicting Stock Returns

Alex G Kim, Accounting PhD Student

Firms disclose large amount of narrative disclosures on earnings announcement days and these disclosures are known to have significant impact on the capital market. However, we still do not know which narratives are relatively more useful in predicting future stock returns than others. In this project, I use Proximal Policy Optimization algorithm to fine-tune a large language model (LLaMA-2) to produce summaries that are optimized for predicting stock returns. In a pilot test, I find that optimized summaries contain more forward-looking information and that trading strategies based on optimized summaries yield higher Sharpe ratios than strategies using raw documents.Then I introduce a measure that captures the degree of textual exposure at the firm year-level and show that high textual exposure is associated with higher market beta. I plan to provide a time-series landscape of textual contents that matter for predicting returns. Overall, this study provides first evidence on the relative importance of textual information in predicting stock returns.