Dynamic Factor Loadings and Portfolio Return Prediction: Evidence from Deep Sequence Models

This paper examines whether factor dynamics contain predictive information for portfolio returns and whether deep sequence models can capture such information. Using factor-based portfolio data, we compare deep sequence models with statistical, machine-learning, and asset-pricing benchmarks. The results show that factor dynamics provide incremental predictive information, while model performance varies across specifications. We find differences between serial and parallel sequence architectures in forecasting and economic value. Further analysis shows that the predictive role of factor dynamics strengthens during extreme market conditions. Overall, deep sequence models provide an effective framework for capturing nonlinear temporal patterns in factor information and improving portfolio return prediction.