[2I6-GS-10] AI application: Finance

[2I6-GS-10-01] Stock Selection Attempt using Sentiment of Japan Company Handbook

〇Masahiro Suzuki^1,2 (1. Nikko Asset Management Co., Ltd., 2. The Univesity of Tokyo)

Keywords:Sentiment, Financial documents, Large language model

We use several methods, including large language models (LLMs), to calculate sentiment from the Kaisha Shikiho, a databook that summarizes the trends of Japanese listed companies.
The use of LLMs, such as ChatGPT, has been popular in the financial domain, and various empirical analyses have been conducted using text.
This study constructs multiple sentiment calculation methods using a polarity dictionary, a model trained on an existing sentiment dataset, and ChatGPT.
The higher quartiles of the sentiment tended to have larger excess returns, while the lower quartiles tended to have smaller excess returns.
The analysis by stock size showed that this tendency was stronger for smaller issues.