Interpreting Elastic Net Factor Scores in Rolling Multivariate Regressions
Summary
This explanation describes a factor evaluation score derived from rolling Elastic Net regressions that include all submitted factors together. The model predicts each stock’s next-day return after cross-sectional standardization, fits on rolling windows, and scores a factor by the mean absolute value of its fitted weights relative to their variation across windows. A high score therefore reflects weights that are both substantial and stable; it is a measure of a factor’s contribution within the submitted group, not an isolated test of its standalone predictive ability.
The note explains why adding factors can lower existing scores: correlated factors can divide weight or alternate selection across windows, a stronger factor can absorb overlapping information, and L1 regularization can set redundant weights to zero. More factors can also leave too few observations for some windows, reducing the usable sample. It recommends reviewing factor correlations and selection rates, testing a factor alone when assessing standalone value, and judging the full portfolio when building a combination. The description explains the scoring procedure but provides no independent validation of predictive performance.
Key ideas
- The score comes from a joint rolling Elastic Net regression rather than separate single-factor tests.
- It rewards factor weights that are large on average and stable across windows.
- Correlated or stronger new factors can dilute an existing factor’s weight or make its selection inconsistent.
- L1 regularization and reduced usable window counts can also make scores fall or become unstable.
- Factor correlations and selection rates help diagnose overlap, while portfolio construction calls for evaluating the combined model.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.