Evidence-Traced and Rules-Guided LLM Pipeline for Predicting China’s Stock-Market Trend in Response to FOMC Press Conferences

Abstract

This paper studies whether language in Federal Open Market Committee (FOMC) press conferences helps predict the next trading-day trend of China’s stock market. We propose an evidence-traced and rules-guided LLM pipeline for event-level prediction without model fine-tuning. In our setting, evidence-traced means that each forecast is grounded in a compact set of retrieved transcript spans, and the output includes span-level citations that make the decision auditable. The pipeline separates each press conference into an opening-statement view and a Q&A view, applies evidence retrieval, and produces a structured label in {up, down, unknown}. We use self-consistency voting and a selective threshold τ to allow abstention when evidence is weak. We also distill a lightweight rulebook from the development period and reuse it as soft guidance at test time. On a time-based split, the rules-guided opening view achieves 71.4% accuracy at 71.8% coverage on the held-out test set (AUC = 0.655), outperforming dictionary sentiment, FinBERT, TF–IDF logistic regression, and zero-shot LLM baselines. Q&A-only results stay near chance, suggesting that the prepared opening statement carries more reliable short-horizon cross-market signal.