Market State-Aware Dynamic Weighting Model for Price-Based Investment Decision

Abstract

Financial markets exhibit significant regime-switching characteristics, and the value of information sources varies across different market states. While existing fusion models have been widely studied, the current paper focuses on the price modality to validate the proposed dynamic weighting mechanism. Most existing fusion approaches adopt static weighting or implicit dynamic interactions, lacking explicit awareness of the macro-market environment, which leads to performance degradation when market states switch. To address this issue, this paper proposes a functional dynamic attention fusion model. The model uses 20-day historical volatility as a proxy for market state and generates dynamic weights through a piecewise linear function, reducing the weight of price features as volatility increases, thereby decreasing reliance on price trends during high-volatility periods. Using daily data of the CSI 300 index from 2020 to 2024, the prediction accuracy and investment performance of the static model and the dynamic model are compared. Empirical results show that the dynamic model achieves an annualized return of 4.91%, numerically higher than the static model’s 2.67%; its Sharpe ratio (0.206) is better than that of the static model (0.046); annualized volatility is slightly lower; and the maximum drawdown remains the same as the static model. Prediction accuracy improves from 50.76% to 50.78%. The proposed model requires no training, is computationally simple, fully transparent, and can be reproduced entirely in Excel, providing a low-cost, highly interpretable new paradigm for dynamic feature weighting in quantitative investment research.