Learning Fundamentals from Text

Abstract

We introduce a novel approach to learning the information that investors react to when processing textual information. We use the attention mechanism that learns to identify content that triggers market reactions to disclosed information. The explanatory power of the attention-based model significantly exceeds that of attention-free models. We then develop and analyze a comprehensive set of topics discussed in companies' annual reports. Segment information, goodwill and intangibles, revenues, and operating income are the topics that receive the most attention from investors. Despite their prominence in the public discourse, sustainability and governance are consistently among the least important topics judging by the market reactions. Building on our approach, we show that regulatory interventions can successfully enhance the relevance of textual communication. We also show that firms strategically position information within MD&A to influence investor focus. Our findings underscore the value of attention-based analysis of corporate communications and open new avenues for future work.

Keywords: Attention mechanism, information relevance, fundamentals, information processing, corporate disclosure, investor attention, large language models, LLM

JEL Classification: C45, C55, G12, G17