Alternative Labels and Ensemble Learning for Stock Selection Models

Summary

This study compares conventional return labels with alternative labels based on information ratio and Calmar ratio for AI stock selection models. It argues that labels incorporating returns, volatility, and drawdowns across a period may convey more information than endpoint returns alone and may produce different factor weights. The tests span 67 training-window lengths, from 6 to 72 months, and use Chinese A-share universes, including index constituents.

Alternative labels perform more favorably across the broad A-share universe, with reported gains in excess return, information ratio, and Calmar ratio; results are less consistent within the CSI 500 and CSI 800 universes. Maximum drawdown of excess returns is worse in the alternative-label models across the tests. The study also combines model predictions using equal weights, historical information coefficients, or historical factor returns, reporting steadier excess-return and information-ratio improvements. These are backtest findings, and the document summary does not establish out-of-sample robustness or live-trading performance.

Key ideas

• Alternative labels can encode path information such as volatility and drawdown, beyond period-end returns.
• The study evaluates models across many training-window lengths rather than relying on one comparison.
• Alternative labels show stronger reported results across the broad A-share universe than within index constituent pools.
• Alternative-label models also show worse excess-return maximum drawdown in the reported tests.
• Ensembling predictions with several weighting methods improves average reported 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.