This paper conducts a comprehensive comparative analysis of state-of-the-art deep learning methods for asset pricing, focusing on predicting individual stock excess returns. The study uses a large dataset of approximately 30,000 US stocks over 40 years (1976–2016), with predictors including 81 firm characteristics, 6 macroeconomic variables, and 5 Fama-French factors. The models evaluated range from simple linear regression to deep feedforward networks (DNNs, Residual DNNs), convolutional networks (CNNs, Residual CNNs), recurrent networks (RNNs, LSTM, GRU), and the Transformer. All models are trained using mean squared error loss and the Adam optimizer, with regularization via dropout and normalization.

The empirical results show that all deep learning models significantly outperform the linear baseline, with RNNs incorporating memory mechanisms (LSTM, GRU) and the Transformer achieving the highest out-of-sample R² values. This demonstrates the predictive power of historical information and sequential dependencies in stock returns. In contrast, CNNs perform relatively poorly, suggesting that local patterns are less important than temporal dynamics. The study also finds that models trained on shorter time spans (5 years) outperform those trained on longer spans (40 years), indicating that return distributions are non-stationary and that recent data is more informative. Additionally, skip connections provide only marginal benefits, implying that moderate network depth is sufficient for this task.

The paper emphasizes the importance of incorporating financial theory into model design, as domain-specific architectures (e.g., RNNs for time series) outperform generic ones. It also highlights the challenge of distribution shift in financial time series and suggests that explainable AI methods could help uncover underlying economic mechanisms. Overall, the findings justify the value of deep learning in fintech and demonstrate its advantages over traditional machine learning for asset pricing.