Alternative Labels and Model Ensembles for Chinese Equity Selection

Summary

This research summary compares machine-learning stock-selection models trained on conventional return labels with models trained on alternative labels based on information ratio and Calmar ratio. The rationale is that these labels capture more of the path within a forecast window, including volatility and drawdown, while the endpoint return label uses only prices at the window’s boundaries. Labels also shape factor weights, so alternative choices may produce different exposures and reduce reliance on a crowded modeling convention.

The study tests 67 training-window lengths from six to 72 months, then combines predictions using equal weights, historical information coefficients, or historical factor returns. It reports stronger excess return and information ratio for alternative labels across the broad A-share universe, with smaller and less consistent advantages among CSI 500 and CSI 800 constituents. Alternative-label models had worse excess-return maximum drawdown in all reported tests. Ensembles improved average excess return and information ratio and are presented as a way to diversify model risk. These are backtest findings; the summary gives no detail on transaction costs or live performance.

Key ideas

• Alternative labels based on information ratio and Calmar ratio incorporate return-path characteristics such as volatility and drawdown.
• The comparison spans 67 training-window lengths from six to 72 months.
• Alternative-label models performed more strongly across the broad A-share universe than in the named index constituent universes.
• The alternative-label models had worse excess-return maximum drawdown in all reported tests.
• Equal-weight and historical-performance-based ensembles improved average excess return and information ratio.

Tags

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