The News in Earnings Announcement Disclosures: Capturing Word Context Using LLM Methods
研究用大语言模型分析盈利公告文本，发现其解释股价变动的能力远超传统方法，且电话会议文本能进一步提升解释力。
This study examines the information content of textual disclosures in firms’ earnings announcements. Using a large language model (LLM) to capture information in both words and word context, I show that the news in earnings press releases (i) explains three times more variation in short-window stock returns than a host of textual measures based on dictionary and non-LLM machine learning methods; (ii) doubles the R 2 of an array of financial statement surprises, modeled with conventional regression or machine learning approaches; and (iii) accounts for a large fraction of immediate price revisions within just five minutes of release. LLM-modeled conference calls further enhance R 2 by one fourth compared with press releases and financial surprises. Textual disclosures are more informative when earnings are less persistent and during periods of aggregate uncertainty. Most news arises from text describing numbers, at the beginning of the disclosure, and including novel contents. These findings highlight the role of firms’ textual disclosures in moving stock prices and advance our understanding of how investors utilize corporate disclosures. This paper was accepted by Suraj Srinivasan, accounting. Funding: The author gratefully acknowledges financial support from the Naveen Jindal School of Management. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2024.05417 .