Dependence-Informed Sparse Neural Architecture for Stock Return Prediction

Battle Count: 15

Highly relevant to quantitative trading. The paper directly addresses cross-sectional stock return prediction and ranking, which are core tasks in equity long-short strategies. The HNN achieves the highest Pearson IC (0.0642) and equal-weighted decile spread (3.95% gross, 3.34% net at 25bps) among all models tested. The 80x parameter reduction vs. dense networks is practically significant for production systems. The interpretable clique structure allows traders to understand which characteristic interactions drive predictions. However, the asset-weighted performance is weaker than NN3, and the advantage is concentrated in smaller stocks, which may limit capacity for large institutional strategies.

Moderate to high complexity. Requires implementing the MFCF algorithm (graph filtering with admissible separators and decomposability constraints), constructing the clique-to-neural-unit mapping, building sparse weight matrices following set inclusion, implementing multi-layer readout with layer normalization, and managing the walk-forward protocol with annual graph re-estimation. The sparse connectivity pattern and the requirement to compute clique subsets add implementation overhead compared to standard MLPs. However, the architecture is fixed before training, simplifying the training loop. The 10-seed ensemble and hyperparameter search over K and learning rate add computational cost.

The paper uses the well-known GKX (Gu, Kelly, Xiu) U.S. equity panel with 94 firm characteristics and monthly returns from 1957-2016. The walk-forward protocol, hyperparameter grids, and model specifications are described in detail. However, no GitHub repository or code link is provided. The MFCF and HNN implementations reference prior work (Lin et al. 2026, Wang et al. 2023) but do not include explicit code availability. The 10-seed ensemble averaging and specific training details (Adam, batch size 10000, gradient clipping) are documented.