Incremental Value of Conditional Volatility in Extreme Gradient Boosting Stock Selection: Evidence from Dynamic China Securities Index 300 Constituents

Abstract

This paper tests whether stock-level GARCH conditional volatility adds predictive and economic value to an XGBoost stock-selection model in China’s A-share market. Using a point-in-time panel of 553 historical CSI 300 constituents from 2018 to 2025, the conventional, GARCH, and GARCH-plus-market models attain ten-day AUC values of 0.531, 0.534, and 0.534. With next-open execution and a 20-basis-point one-way cost, their top-decile annualized net returns are 10.93%, 12.14%, and 11.32%. The GARCH signal is visible across several checks, but increments are statistically weak and sensitive to implementation. The evidence supports a marginal, conditional, short-horizon role for conditional volatility rather than a robust standalone trading advantage.