The Power of Large Language Models: A ChatGPT-driven Textual Analysis of Fundamental Data

Abstract

In this study, we examine the effect of stale, factual information on stock price formation, utilizing Shikiho, the Japan Company Handbook, as a pivotal data source. The breadth and depth of information offered by Shikiho, covering all listed Japanese companies, and compiled by independent journalists, are unmatched by other market data sources. Our innovative approach involves the use of large language models, specifically ChatGPT, to interpret the business implications within Shikiho's corporate descriptions and future projections. The interpreted data serves as a basis to classify firms into positive and negative outlook groups, which are further quantified into confidence scores. These scores inform the construction of an equity portfolio. Benchmarking the portfolio's performance against relevant risk factors suggests that the market may not fully absorb the factual information encapsulated in Shikiho's text. These observations prompt a deeper inquiry into market efficiency, the role of public information in price discovery.

Keywords: Stale information, ChatGPT, LLM, Return predictability, Text Analysis, Shikiho

JEL Classification: G12, G14