Neural Networks for Generating and Combining Weak Equity Alpha Factors

Summary

This research note describes a daily equity alpha approach that uses supervised neural networks to generate many candidate factors from designed factor units. It then transforms the original inputs into weak, mutually orthogonal factors and weights them according to the performance of their long portfolios. The motivation is that established factors can become crowded, while factor information coefficients may not reliably track portfolio returns; separating components may help retain useful signals within a crowded composite factor.

The reported experiment produced 300 weak factors with pairwise correlations mostly below 20%. Individual factors had modest RankIC, while their combined Z-score reportedly achieved higher next-day and five-day RankIC. A backtest of a CSI 500 enhancement model suggests that lower turnover held up better after 2017 under assumed next-day VWAP execution and transaction costs; the note also reports that the advantage of earlier execution narrowed in recent years. These are historical results from the described setup, not evidence of future performance. The source offers limited methodological detail here, and factor crowding, execution assumptions, and market regime changes constrain generalization.

Key ideas

• The method uses supervised neural networks to produce many candidate alpha factors from designed factor units.
• An orthogonalization step separates composite signals into weak factors that can be evaluated individually.
• The reported combined signal had stronger RankIC than the average individual weak factor.
• The historical CSI 500 enhancement results indicate that lower turnover was more resilient after 2017 under the stated execution and cost assumptions.
• The note cautions indirectly against relying on rapid execution, as its historical advantage narrowed in recent years.

Tags

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