Alternative Labels and Ensembles for AI Equity Selection

Summary

This report compares machine-learning stock-selection models trained with conventional return labels against models using information ratio or Calmar ratio labels. The alternative labels incorporate aspects of the path within the evaluation period, including returns, volatility, and drawdowns, while a simple return label mainly reflects prices at the interval’s endpoints. The comparison spans 67 training-window lengths from 6 to 72 months and covers the broad A-share universe and index constituent pools.

In the broad A-share tests, alternative-label models more often improved annualized excess return, information ratio, and Calmar ratio; their excess-return win rate was about 90%. Results were weaker among CSI 500 and CSI 800 constituents, where the more consistent edge was in excess return. Across the tests, alternative-label models had worse maximum excess drawdowns. An ensemble that combined model predictions using equal weights, historical IC, or historical factor returns showed more stable excess returns and information ratios, with model diversification as an additional rationale. These are reported backtest findings; the summary does not establish live performance.

Key ideas

• Information-ratio and Calmar-ratio labels encode more of the price path than endpoint returns alone.
• The report compares label choices across 67 training-window lengths.
• Alternative labels show stronger reported benefits in the broad A-share universe than in index constituent pools.
• Alternative-label models have worse maximum excess drawdowns in the reported tests.
• Combining predictions with equal, historical IC, or historical factor-return weights improves average ensemble performance in the report.

Tags

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