FinTransformer: A Time-Biased Attention Mechanism with Feature Fusion for Enhanced Fund Return Prediction

Fund return prediction is critical for investment decisions, but traditional methods struggle to capture the complex nonlinear relationships and long-term dependencies in financial data. In this paper, we propose FinTransformer, a deep learning model designed for fund return forecasting. Although Transformer performs well in sequence modelling, its direct application to financial forecasting faces three major challenges: inability to distinguish the importance of near- and far-term data, lack of feature type differentiation, and insufficient quantification of forecast uncertainty. To address these challenges, we design three key innovations: (1) a finance-specific attention mechanism, which enables the model to adaptively adjust the importance of data at different time points through temporal bias; (2) a feature fusion layer, which intelligently integrates features of different types and abstraction levels; and (3) a multi-task learning framework, which simultaneously forecasts returns, market states, and uncertainty estimates. Experiments on a real dataset containing 26,093 funds show that FinTransformer significantly outperforms existing methods with a 15.3% reduction in RMSE and a 12.0% improvement in $R^{2}$. The model attention weighting analysis reveals the contribution of different features to the prediction and enhances the interpretability. This study not only advances the application of deep learning in financial forecasting, but also provides a more accurate tool for investment decision-making.