From words to returns: sentiment analysis of Japanese 10-K reports using advanced large language models

The advent of advanced natural language processing techniques and large language models (LLMs) has revolutionized the analysis of qualitative financial data. This research harnesses the capabilities of LLMs—specifically ChatGPT, Claude, and Gemini—to extract sentiment from Japanese 10-K reports, aiming to predict future stock returns. By analyzing an extensive dataset encompassing all companies listed on the Tokyo Stock Exchange from 2014 to 2023—a total of 11,135 firm-years and over 70 million words—we conduct the first comprehensive study of its kind in Japan. Comparative analyses are performed using traditional dictionary-based methods and a DeBERTaV2-based model to evaluate efficacy in information extraction. Our findings reveal substantial differences in the models’ abilities to predict stock performance. Notably, while dictionary-based methods show no significant relationship between sentiment and subsequent stock returns, LLM-derived sentiments exhibit a significant negative correlation with future returns. These results challenge the efficient market hypothesis by demonstrating that sentiment extracted from publicly available reports can predict stock performance. This study reveals the transformative potential of advanced Natural Language Processing (NLP) technologies in financial analysis, highlighting how sophisticated language models can uncover predictive signals previously undetected by traditional methods. The article details the methodologies employed, the challenges encountered, and the implications for integrating advanced sentiment analysis into financial forecasting.