Decoding mutual fund performance: Dynamic return patterns via deep learning

Abstract

This paper applies the Temporal Fusion Transformer (TFT) model to learn dynamic time-series patterns in mutual fund performance and to assess whether these patterns predict future alpha. I summarize the model’s cross-sectional ranking power using diagnostic portfolio spreads: a top-minus-bottom decile exhibits an annualized Carhart four-factor alpha spread of 2.8%, with dispersion persisting for up to four years. In panel regressions controlling for standard predictors and fund and time fixed effects, TFT forecasts improve explanatory power by more than 25% in adjusted R 2 . Leveraging TFT’s interpretable outputs, I show that historical fund returns receive the largest weight (about 29%), their importance displays earnings-cycle seasonality, and attention to past observations rises by 46% during crisis periods. Using fund-by-month variable-importance weights, I define fund-specific informativeness states and construct conditional skill measures that predict and persist precisely when the same signal becomes informative again, beyond coarse macro conditioning. Together, these results provide an alternative explanation for why unconditional performance persistence appears weak: skill is episodic and becomes visible when a manager’s key signals regain relevance. • TFT deep learning model predicts mutual fund alpha from time-series patterns. • Top-minus-bottom decile spread yields 2.8% annualized four-factor alpha. • Model attention to past observations rises 46% during crisis periods. • Conditional skill measures persist when key signals regain informativeness. • Episodic skill explains why unconditional persistence appears weak.