Grouped Time-Series Cross-Validation to Reduce Financial Model Overfitting

Summary

This study compares six cross-validation approaches for tuning machine-learning models on financial time series. It argues that ordinary shuffled K-fold validation can leak temporal structure and select overly complex models. Expanding-window time-series validation trains on earlier observations and validates on later ones; the grouped variant additionally splits at month boundaries so samples from one month cannot be divided between training and validation. Additional control methods vary sample size or shuffle temporal order to help identify why the time-aware methods perform better.

Using rolling equity-selection experiments with logistic regression and XGBoost, the report evaluates model metrics, factor tests, and portfolio backtests. It finds that grouped time-series validation generally performs best on held-out model and single-factor measures, while both time-aware methods outperform K-fold; portfolio results are less consistent. The controls suggest preserving chronology accounts for most of the improvement, with reduced training sample size contributing some. The evidence comes from a specific historical Chinese equity universe and period, and the authors note possible underfitting and future regime change. The method reduces one source of validation bias; it does not guarantee future profitability or eliminate research overfitting.

Key ideas

• Randomized K-fold validation can overstate performance when observations have temporal dependence.
• Expanding-window validation trains on earlier periods and evaluates on later periods.
• Grouped time-series splits keep all observations from a calendar month in the same fold.
• Control experiments attribute most of the reported improvement to preserving chronology, with a smaller contribution from using fewer samples.
• Results favor time-aware validation for held-out model and factor tests, while portfolio findings are mixed and historically specific.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.