Presentation Information
[A-11-07]LLM-Based Financial Performance Prediction Using Corporate Disclosure Documents
◎Mahisa Hoshino1, Takahiro Baba1 (1. Kurume Institute of Technology)
Keywords:
Large Language Model(LLM),Retrieval-Augmented Generation (RAG),Financial Performance Prediction
This study proposes a method for corporate performance prediction using large language models (LLMs) based on quantitative and qualitative information in annual securities reports. A Retrieval-Augmented Generation (RAG) framework was constructed using reports from the EDINET disclosure system of Japan’s Financial Services Agency, and it was applied to predict sales revenue and net profit in semiannual financial statements released six months later. The dataset consists of 105 listed, consolidated companies with fiscal year ending on March 31, selected from 11 major industries. GPT-5 was used as the base model, and Dify was employed to build the RAG system. To prevent information leakage, only each company’s own annual securities report was used for prediction. Evaluation metrics included MAE, RMSE, MedAE, Bias, and R². Results showed strong performance in sales revenue prediction with R² = 0.9979, while net profit prediction achieved R² = -0.1434, indicating poor performance overall; however, the MedAE was 613,000 yen, suggesting relatively small errors for many firms. These results demonstrate the effectiveness of LLM-based sales forecasting using securities reports. Future work will focus on outlier handling and improving evaluation of generated explanations and prediction rationale quality.