Correcting Long-Side Failure in High-Frequency Equity Factors
Summary
The document describes a problem in equity factor models: high-frequency factors can show strong overall information coefficients and long-short returns after being orthogonalized against common factors, yet fail to improve a portfolio when used for stock selection. It attributes this gap to different relationships between factor values and future returns on the long side versus the factor overall.
It proposes weighting factor groups differently when calculating IC, with greater emphasis on the long side, to better assess long-side usefulness. It also describes adding higher-order polynomial terms or radial basis function features to capture nonlinear exposure-return relationships; a linear model on the expanded features can approximate piecewise regression. The summary reports improved portfolio performance from polynomial terms but gives no detailed data or test design. It cautions that feature expansion can destabilize estimates, make exposure matrices rank-deficient, and increase overfitting, especially with limited monthly rebalancing history.
Key ideas
- A high overall IC or long-short return may conceal weak predictive value on the long side.
- Weighting groups differently in IC calculations can emphasize factor effectiveness among long candidates.
- Polynomial terms can represent nonlinear relationships between factor exposures and expected returns.
- Radial basis features paired with a linear model can approximate piecewise regression.
- Feature expansion can increase estimation instability, rank problems, and overfitting risk.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.