Interpretable Deep Learning for Stock Returns: A Consensus-Bottleneck Asset Pricing Model

Battle Count: 103

Highly relevant to quantitative trading. The paper directly addresses cross-sectional stock return prediction with economically interpretable signals. Key findings include: annual return R² of 10.46% (vs 7.63% benchmark), H-L portfolio spreads of 2.3% per month, Sharpe ratio of 1.44 for long-short portfolios, and robust performance across macroeconomic regimes. The consensus-bottleneck provides actionable trading signals while maintaining interpretability. Transaction cost analysis shows performance remains economically meaningful at 75 bps costs. However, high turnover (~60%) and the focus on annual horizons limit applicability to high-frequency strategies.

Moderate to high complexity. Requires: (1) training a feedforward neural network consensus module with GELU activations, (2) training an autoencoder for macroeconomic embedding (32-dim latent), (3) implementing joint optimization with weighted loss function, (4) ensemble of 10 models, (5) expanding-window evaluation with proper data lagging and preprocessing, (6) portfolio construction and GRS testing. The architecture itself is relatively simple (2 hidden layers, 64/32 nodes), but the full pipeline including data preprocessing, macro encoding, and evaluation is substantial. Uses standard PyTorch components (Adam, ReduceLROnPlateau, gradient clipping, dropout, layer normalization).

The paper provides detailed hyperparameter settings (Table C.1), complete data preprocessing pipeline, architecture specifications, and uses publicly available datasets (CRSP, FRED-MD, A.Y. Chen and Zimmermann 2022 open-source asset pricing panel). Expanding-window evaluation scheme is clearly described. However, no GitHub repository is mentioned, and some implementation details are deferred to Internet Appendices. The ensemble of 10 models and specific training procedures are documented.