A Walk-Forward XGBoost Strategy Using Cross-Sectional Stock Factors

Summary

This strategy note outlines a Chinese equity model that ranks stocks using three cross-sectional features: market capitalization, turnover, and five-day price change. Each feature is converted to a daily percentile rank. A shallow XGBoost regressor is trained on a rolling one-year sample and refreshed every 60 trading days; the strategy predicts scores and rebalances every five days into an equally weighted portfolio of the ten highest-ranked stocks. Its target is the cross-sectional rank of a short forward return, so the prediction represents relative strength rather than an absolute return forecast.

The write-up describes logging training-set error measures, information coefficient, and feature importance, then saving those records with backtest outputs. It reports no strategy performance results, and the listed model diagnostics are in-sample. The author flags a mismatch between the target horizon and five-day holding period, the narrow feature set, and a hard-coded end date. The suggested fixes and any claims of timing safety are not independently validated; the strategy still requires careful out-of-sample and execution testing.

Key ideas

• Daily percentile ranks place market capitalization, turnover, and five-day price change on a common scale.
• A shallow XGBoost regressor is retrained on a rolling one-year sample at 60-trading-day intervals.
• The strategy predicts relative forward-return ranks and rebalances every five days into ten equally weighted stocks.
• Training diagnostics include error measures, information coefficient, and feature importance, but no performance results are provided.
• The note identifies target-horizon mismatch, limited features, and a hard-coded end date as limitations.

Tags

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