Regime filtering without overfitting: a practitioner's checklist

Regime models are appealing because markets visibly behave differently in different periods. They are dangerous for the same reason: there are many plausible ways to partition history, and most of them improve in-sample fit without improving out-of-sample behaviour.

The honest question

A regime model is useful when, conditional on the observed regime, the distribution of forward returns has a meaningfully lower variance than the unconditional distribution. If that reduction is small, the regime is not adding information; it is adding parameters.

Checklist

• Define the regime in terms of variables observable in real time, not revised or look-ahead series.
• Test the conditional variance reduction on rolling out-of-sample windows, not on the full sample.
• Penalise regime counts: a two-state model that works is worth more than a five-state model that works marginally better in-sample.
• Examine transition behaviour. Regimes that flicker on a daily basis are usually noise; regimes that persist for weeks are at least candidates for economic interpretation.
• Report what happens during the worst three regime mis-classifications in the sample. If the answer is 'we have not looked', the model is not ready.

Why we still use them

When the discipline above is enforced, a small number of regime variables survive and contribute. They earn their place by making downstream signals more selective — they reduce trading in conditions where the underlying edge has historically been weakest. The portfolio-level benefit is rarely a higher Sharpe in good periods; it is a smaller drawdown in bad ones.