Regime-Conditional Alpha Models

A momentum signal that works well in a trending market can lose money outright in a sharp reversal, and a value signal that looks defensive in a slow grind can get run over in a liquidity crunch. A model trained once, on data spanning both kinds of periods, learns some blended-average behaviour — not because that average is a good strategy, but because that's what minimising average error across the whole sample produces. A regime-conditional model tries to do better by letting the relationship between features and returns change depending on which market state the model believes it's currently in. The appeal is obvious; the difficulty is that identifying the regime and fitting a good model within it both require data, and splitting a sample by regime shrinks the data available for each piece exactly when the model needs it most.

Two ways to condition on regime

Hard switching. Classify each period into a discrete regime — say, "trending," "mean-reverting," "high volatility" — using either external labels (VIX level, realised volatility, a market-return threshold) or a fitted regime-detection model (see Regime Detection), then train a separate model for each regime and switch between them based on the current classification. This is intuitive and interpretable but has a hard failure mode: if the classifier misjudges the current regime, the model applied is simply the wrong one, with no graceful degradation.

Soft conditioning. Instead of a hard switch, include regime indicators (or continuous proxies like recent realised volatility) as features in a single model, letting it learn how feature-return relationships shift smoothly with market conditions. This avoids the discontinuity of hard switching and lets the model interpolate between regimes it has seen combinations of, but it demands that the regime information genuinely be informative as a feature — if it isn't, it just adds dimensionality without adding structure, feeding the same overfitting risk that any extra feature does (see Overfitting the Cross-Section).

A worked example

A desk labels history into two regimes using trailing 60-day realised volatility: "calm" (below the 70th percentile) and "stressed" (above it), across 10 years of data. Stressed months make up roughly 20% of the sample.

Hard-switched approach: train separate models on calm-only and stressed-only data. The calm model, trained on 8 years of relatively similar data, performs reasonably in walk-forward tests. The stressed model, trained on roughly 2 years' worth of data concentrated in a few genuine crisis episodes (2018 Q4, 2020 Q1, 2022), overfits badly — its walk-forward performance on new stress periods is close to random, because it effectively memorised three specific crises rather than learning what's common to volatile markets generally.

Soft-conditioned approach: a single model trained on the full 10 years, with trailing realised volatility included as a feature (and allowed to interact with the other signals via a tree-based model's natural interaction handling). This model can still behave differently when volatility is high, but it borrows statistical strength from the calm-period data when learning the shape of that dependence, rather than fitting the stressed behaviour from stress data alone. In walk-forward testing across held-out stress periods, it holds up better than the hard-switched stress model, precisely because it never had to learn crisis behaviour from crisis data alone.

Picture the two market states as two classes and the model's job as drawing a boundary between how it behaves in each. Too flexible a boundary, fit on too little data per class, curls around individual training points instead of the real shape of the difference — visually, exactly what a hard-switched crisis model does when trained on three crises' worth of points.

Where it breaks in practice

The regime label itself is often built using information not really available in real time — a realised-volatility regime computed with a centred window, or a change-point detector that only becomes confident about "when the regime started" some weeks after the fact. A model that trains on regime labels defined this way, then is deployed to classify the current regime live using only trailing information, faces a harder, noisier version of the classification problem than the one it was validated on.