Regime Detection

A strategy that prints money in a trending market can hemorrhage in a choppy one. Regime detection is the attempt to know which world you’re in now. A common approach runs a Hidden Markov Model (HMM) over realized volatility and return features, classifying the market into a small number of latent states — say calm, trending, and high-stress.

The states aren’t observed directly; the model infers the most probable current regime from recent data and can flag when a transition is underway. That single classification is enormously valuable because almost every other parameter — sizing, stop width, whether to make or take — should depend on it.

For an autonomous agent, regime is the top-level switch. An agent can run different sub-strategies per regime and cut risk on a detected shift to stress, rather than discovering the change through losses. Because the classifier is quantitative, “reduce gross exposure when regime = stress” becomes an enforceable policy rather than a discretionary hunch.