Relevance Scoring in Investment Themes: Domain Adaptation of Local LLMs for Company-Theme Relevance Scoring [in Japanese]

Tatsuto Ito, Ryota Ozaki, Kentaro Imajo, Masanori Hirano

The 36th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence, pp. 61-67, Mar. 21, 2026

Conference

The 36th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence (SIG-FIN)

Abstract

This study investigates domain adaptation of local large language models (LLMs) for Japanese financial market analysis. We formulate company-theme relevance scoring as mapping a company name and an investment theme to an integer score s in {0, ..., 10} for ranking. Training via supervised fine-tuning (SFT), SFT models substantially improve output consistency compared to instruction-tuned models, achieving up to 0.707 Top-1 accuracy on sector classification and 0.651 PR-AUC on identifying theme-related stocks. We find that base models pretrained on Japanese financial data generalize better, while SFT does not uniformly improve out-of-domain generalization. Our contributions include the first LLM-based company-theme relevance dataset and alignment framework for standardized integer scoring, and an extensive evaluation of LLM generalization under domain adaptation for practical investment theme analysis in Japan.

Keywords

Large Language Model; Domain Adaptation; Fine-tuning; Investment Theme;

doi

10.11517/jsaisigtwo.2026.FIN-036_61

bibtex

@inproceedings{Ito2026-sigfin36,
  title={{Relevance Scoring in Investment Themes: Domain Adaptation of Local LLMs for Company-Theme Relevance Scoring [in Japanese]}},
  author={Tatsuto Ito and Ryota Ozaki and Kentaro Imajo and Masanori Hirano},
  booktitle={The 36th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence},
  pages={61-67},
  doi={10.11517/jsaisigtwo.2026.FIN-036_61},
  year={2026}
}