AlphaQuantFormer: A Market-Entropy Time-Biased Attention Mechanism with Hierarchical Feature Fusion for Enhanced Mutual Fund Return Prediction

Fund return forecasting plays a key role in investment decision-making, yet traditional methods struggle to effectively capture the complex nonlinear nature and long-term dependencies of financial data. This study proposes AlphaQuantFormer, a deep learning architecture for fund return forecasting, to address three core challenges: differential temporal significance of financial time series, effective integration of multiple types of features, and accurate quantification of forecast uncertainty. Firstly, adaptive temporal weight allocation to market states is achieved through a multi-entropy time-biased attention mechanism; secondly, exclusive processing paths are designed for fundamental, technical and macroeconomic indicators through a feature type adaptive processing network; and finally, a hierarchical feature fusion with a multi-task learning framework is used to achieve joint forecasting of returns, market states and uncertainty. On a large-scale dataset containing 26,093 funds, AlphaQuantFormer reduces RMSE by 19.4% and improves R² by 8.3% compared to the best baseline model, FinTransformer, and demonstrates superior generalisation performance across different market environments and fund types. The experimental results validate the significant enhancement of predictive performance by financial domain-specific attention mechanism design and feature processing methods, providing a more accurate and interpretable analytical tool for fund investment decisions.