Seokje Lim

This study proposes a large language model (LLM) framework that separates prediction from explanation for early warning of high-volatility regimes in the Korean stock market: a dedicated model issues the forecast and an LLM produces auditable, evidence-grounded explanations. The predictor's probability, SHAP attributions, a market-state summary, and a risk-factor taxonomy form a Structured Evidence Packet that is the LLM's only input. Explanation quality is assessed with a Grounded Evidence Alignment Score (GEAS), repetition stability, input ablation, and deletion/injection interventions. GEAS rose from 0.183 without SHAP evidence to 1.000/0.998 with it, and the dedicated model reached PR-AUC 0.799 against 0.615 for direct LLM prediction. The findings support using LLMs as grounded, auditable explanation layers rather than autonomous forecasters.