Dynamic ICIR Weighting for Chinese Equity Factor Portfolios
Summary
This account develops an equity-selection process that compares equal-weighted factor rankings with weights that adapt to each factor’s recent Rank IC information ratio. It describes screening a large factor pool for missing data, outliers, redundancy, predictive behavior, and economic rationale. The portfolio selects top-ranked stocks, holds them equally, and rebalances every 22 trading days. To limit look-ahead bias, it aligns past factor values with subsequent returns; absolute ICIR sets weight size, while its sign determines score direction. A threshold can freeze weights when signals are weak, and a cap limits any factor’s share.
The reported backtest compares four variants against the CSI 300 over 2019 to 2026. Dynamic weighting had the strongest risk-adjusted result among them for the top-20 portfolio, while its top-10 version did not beat the equal-weight version in annualized return. The author also reports factor leadership changing over time and describes a sign-handling error that made one portfolio resemble the benchmark. These are single-period backtest findings; the account notes sensitivity to rapid style changes and leaves adaptive windows and interactions unresolved.
Key ideas
- Rank ICIR is used to adapt factor weights based on recent predictive strength and stability.
- Factor signs must be handled separately from weight magnitudes so negatively predictive factors can contribute in the correct direction.
- The process filters factors for missingness, extreme values, correlation, predictive behavior, and economic plausibility.
- A signal threshold can preserve prior weights when ICIR estimates are weak, while a cap limits concentration.
- Reported results favor dynamic weighting for the top-20 portfolio on a risk-adjusted basis, but the evidence comes from one historical backtest.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.