The Nonstationarity-Complexity Tradeoff in Return Prediction

Battle Count: 73

Highly relevant to quantitative trading. The paper directly addresses the practical problem of how much historical data to use and which model complexity to choose for return prediction in non-stationary markets. The ATOMS method provides a data-driven, implementable solution that improves out-of-sample R² by 14% and Sharpe ratios from 0.30 to 0.33. The gains are concentrated during recessions when trading decisions are most critical. The portfolio construction (mean-variance and signed) is directly implementable with realistic trading costs. The adaptive window selection addresses a fundamental challenge in live trading: when to trust recent data vs. longer history.

Moderately complex implementation. The core ATOMS algorithm involves: (1) training multiple candidate models across specifications (Ridge, LASSO, E-Net, Random Forest) and window lengths (6 configurations), (2) implementing the adaptive rolling-window pairwise comparison with Goldenshluger-Lepski-style window selection, (3) running the sequential elimination tournament. The theoretical framework is sophisticated (local Rademacher complexity, total variation distance), but the algorithmic implementation is straightforward. Computational cost scales with the number of candidate models and requires re-estimation each period. Code is provided.

Code is publicly available at https://github.com/ch3702/ATOMS. All data sources are publicly available (Kenneth French data library, CRSP, Gu et al. 2020 characteristics, Chen et al. 2024 factors). Hyperparameter grids are fully specified. Implementation uses standard Python libraries (scikit-learn, statsmodels, pandas, numpy). 20 independent repetitions of stratified 80%-20% train-validation splits are reported. Ablation studies on key hyperparameters (M') are provided.