Residual Learning in Empirical Asset Pricing

Battle Count: 53

Highly relevant for quantitative researchers and practitioners. It provides a concrete method to improve out-of-sample Sharpe ratios (2.07 vs 0.89 for deep feedforward) by addressing the depth-performance trade-off in neural networks. The robustness to transaction costs and market capitalization screens makes it practically viable for long-short equity strategies.

Moderate to High. Requires implementing custom residual blocks tailored for asset pricing (handling width changes via projections), managing expanding training windows, and running ensembles. However, standard deep learning frameworks (PyTorch/TensorFlow) support ResNet architectures, and the authors provide code.

The paper provides a GitHub repository with code, detailed data preprocessing steps (Jensen et al. 2023 screens, Welch and Goyal 2008 macro predictors), specific model architectures (NN+, ResNet+), and training protocols (Adam/AdamW, expanding windows, ensemble seeds).