Similarity-Based Regime Detection

Similarity-based regime detection is a non-parametric approach that classifies the current period by measuring how similar it is to every historical period across a set of economic state variables — typically via a Euclidean distance on z-scored series. The most similar past periods are the current “regime.” It deliberately avoids the two core assumptions of the Markov Regime-Switching Model and Hidden Markov Model Regime Detection: it presupposes neither a fixed number of named regimes nor a transition probability matrix, and conclusions are “derived from the data itself.”

It appears in this vault as the method behind Mulliner et al. 2025 (“Regimes”), the Man Group / Campbell Harvey paper that uses it to time six long-short equity factors. The transparency of the method is a genuine strength — it is just z-scores and distances — but it carries its own hazards: the choice of state variables “induces look-ahead bias” (Man Group’s own admission), the similarity tolerance and post-regime observation horizon are free parameters, and like all regime detectors it provides Regime Classification usefulness that must be graded separately from any after-cost profitability claim.

Connections

• Mulliner et al. 2025 — proposes_model, source: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5164863
• Regime Classification — detects_regime, source: vault synthesis
• Markov Regime-Switching Model — relates, source: vault synthesis
• Lookahead Bias from Smoothed Regime Estimates — suffers_overfitting_risk, source: https://www.man.com/insights/regimes-systematic-models-power-of-prediction
• Man Group — relates, source: https://www.man.com/insights/regimes-systematic-models-power-of-prediction

Ontology

Similarity-Based Regime Detection supports Regime Classification Similarity-Based Regime Detection contradicts Markov Regime-Switching Model Similarity-Based Regime Detection relates Hidden Markov Model Regime Detection