Regime detection is only useful if it changes position sizing before the regime does damage. This note compares five methods — trailing realized volatility, GARCH(1,1), a two-state hidden Markov model, a moving-average volatility gate, and a jump-diffusion filter — on a single metric: how much drawdown they remove per unit of return given up.
The surprising winner
The simplest method — a moving-average gate on realized vol — is within a hair of the sophisticated models on out-of-sample risk reduction, at a fraction of the parameter risk. The HMM detects regime transitions more crisply in-sample but pays for it with lag and instability when refit.
Sizing, not forecasting
The lesson is that regime detection for risk sizing rewards robustness over precision. You don’t need to nail the transition date; you need to be roughly de-risked through the high-vol stretch. Complexity buys sharper in-sample transitions and worse live behavior.