Publications
カンファレンス (国際) Faithful and Plausible Explanations for Market Prediction from Noisy Forum Posts via XAI–LLM Hybrid Rationale Selection
Kota Kotegawa (Nara Institute of Science and Technology), Kentaro Ueda (Nara Institute of Science and Technology), Ren Hosokawa (Nara Institute of Science and Technology), Yuki Ogawa (Ritsumeikan University), Eiichi Umehara (Niigata University of International and Information Studies), Kota Tsubouchi, Keiichi Yasumoto (Nara Institute of Science and Technology), Hirohiko Suwa (Nara Institute of Science and Technology)
21st International Conference on Persuasive Technology 2026 (PERSUASIVE 2026)
2026.3.13
Large language models (LLMs) enable fluent natural-language explanations for prediction models, but their outputs are often not faithful to the underlying model behavior, especially when the input text is noisy. This issue is particularly critical in financial market prediction based on forum posts, where user-generated content is highly fragmented, emotional, and contextually inconsistent. In this paper, we propose a faithful and plausible explanation framework for market prediction models using noisy financial forum posts. Our framework first extracts model-faithful information by ranking posts according to their contribution scores computed by explainable AI (XAI) methods. It then applies an LLM-based rationale selection process to identify posts that are not only important to the model but also suitable as human-understandable explanation grounds, followed by natural-language explanation generation. We conduct user studies comparing the proposed hybrid approach with baseline methods, including random selection, XAI-only selection, and LLM-only selection. The results show that the proposed method consistently improves users' overall understanding and perceived explanation quality. Furthermore, our analysis reveals that directly generating explanations from XAI-ranked posts can negatively affect user comprehension in noisy text settings, highlighting a gap between model faithfulness and human interpretability. Our findings suggest that hybrid rationale selection offers a promising direction for bridging this gap in explainable financial AI systems.
Paper :
Faithful and Plausible Explanations for Market Prediction from Noisy Forum Posts via XAI–LLM Hybrid Rationale Selection
（外部サイト）