Sentiment-Driven Nonlinear Relationship Between Financial Market Emotion and Returns: Dynamic Adaptive Modeling Based on LLMs

This study investigates the nonlinear relationship between investor sentiment and market returns in two Chinese financial markets, emphasizing asymmetric effects across bull and bear regimes. Via a fine-tuned Chinese RoBERTa model, we analyze over financial news texts everyday to generate sentiment scores. Two novel indices, the Momentum-Liquidity Index (MLI) and Price Stability Adjusted Equal-Weight Index (PSAEWI), capture short-term volatility and long-term trends. Empirical results, validated through VAR, quantile regression, and XGBoost, confirm sentiment's stronger predictive power in bull markets and its state-dependent interactions, and we further use them in prediction of macro-performance indicator. Also, we propose a real-time monitoring system and a dual-model framework integrates sentiment classification and emotion regression, dynamically weighted by market conditions. Finally, we offer practical applications for investors and regulators in behavioral finance.