[1Yin-B] Poster 2

[1Yin-B-02] A Fundamental Study on Stock Financial Information Evaluation Using Language Model Agents

〇Hitomi KUMON¹, Kenta OFUJI¹ (1. University of Aizu)

Keywords:

language models,financial information,investment

In this paper, we conducted a basic experimental study to examine how language models assess qualitative financial information for stocks and how this can provide a consistent/complementary perspective with quantitative machine learning models. To this end, we selected Japanese stocks that mimicked U.S. stocks in a "Buffett-like" manner, and quantified their similarity as a Buffett score (quantitative financial score). Next, we constructed a language model agent that performs structural evaluation and had it evaluate the qualitative information supporting Buffett's concept of structural advantage (Economic Moat), and calculated a comprehensive index (qualitative structural score). The results showed that the majority of stocks, naturally, had at least one low score. Cases where one score was high but the other was low suggested a complementarity between quantitative and qualitative scores. This will enable a complementary evaluation. For three stocks with high quantitative financial scores, a consistent relationship was observed between the quantitative and qualitative scores. On the other hand, for the stocks that were excluded, the reason for exclusion was connected with negative output text from the agent.