Beyond Sentiment: LLM-Extracted Linguistic Features from Earnings Calls and the Cross-Section of Stock Returns

Quantitative analysis of corporate disclosures has largely relied on bag-of-words sentiment dictionaries that reduce a one-hour earnings call to a single net-positive word count. This paper tests whether a structural linguistic feature, deviation from a firm's own prior-year earnings-call language, carries incremental return predictability after controlling for standard risk factors. Deviation is measured on verbatim earnings-call transcripts filed as Exhibit 99 attachments to Form 8-K on SEC EDGAR, using TF-IDF cosine similarity between a firm's current and year-earlier call. Across 24 firms and 358 firm-quarters filing verbatim transcripts on EDGAR between 2018 and 2024, a long-short portfolio sorted on this deviation measure earns 1.27% per month (t = 0.88) after controlling for the Fama-French five factors plus momentum. The point estimate is economically large but not statistically significant.