Regime Features

Momentum works in trending markets, mean-reversion in range-bound ones. Instead of manually switching strategies, feed regime indicators as features and let the model learn when each signal works.

The Problem

Models trained on pooled data learn average relationships. If momentum has IC of +0.08 in trends and -0.04 in mean-reverting regimes, the pooled IC is near zero. Regime features let the model condition on the current environment rather than averaging across all of them.

The Pattern

WRONG

import polars as pl

# Raw VIX level - non-stationary, scale-dependent, model cannot generalize
features = df.with_columns(regime_vix=pl.col("vix"))