SignalTrace

Regime Classification

John Montgomery6 min read

Regime Classification

Regime classification is the task of assigning the market’s current, unobservable state to one of a small set of regimes — typically bull/bear, low-volatility/high-volatility, or calm/crisis. Regimes are not directly observed; they are inferred. Two families of inference dominate. Econometric regime-switching models (Markov Regime-Switching Model, Hidden Markov Model Regime Detection) treat the regime as a latent first-order Markov state and estimate it by likelihood methods, so a regime is defined by a distinct mean/volatility/correlation block of asset returns. Machine-learning approaches instead cluster the data: Oliveira et al. (2025) classify monthly regimes by a modified k-means over 100-plus FRED-MD macroeconomic series, and the Statistical Jump Model fits a discrete state sequence with a switching penalty. Ang & Timmermann (2011, published 2012) is the canonical survey: regimes “often correspond to different periods in regulation, policy, and other secular changes” and parsimoniously reproduce fat tails, heteroskedasticity, skewness and time-varying correlations. This concept is the use mode in which Markov-family models have their strongest evidence — but “strongest evidence” and “tradeable alpha” are not the same thing, and separating them is the pivotal distinction of this vault.

The clearest, best-evidenced payoff of regime classification is risk control, not return generation. Ang & Timmermann frame the asset-allocation case explicitly as risk management: a regime-aware investor holds the high-Sharpe portfolio when the low-volatility regime prevails and shifts toward the risk-free asset when the bad regime is detected; the much-cited “cost of ignoring regimes” (Ang & Bekaert 2004; Tu 2010, roughly 2% per year) is a utility loss from mis-sized risk exposure, not a documented excess return that beats a benchmark after costs. Empirical out-of-sample tests confirm this asymmetry. Bulla et al. 2010 find a Markov-switching stocks-or-cash timing strategy stays profitable after 10 bp costs across five equity indices, but the out-of-sample excess return over buy-and-hold is modest (18.5-201.6 bp/yr) — the principal benefit is a ~41% cut in volatility. Shu Yu and Mulvey 2024 reach the same verdict comparing HMM- and jump-model-guided “0/1” strategies on US, German and Japanese indices 1990-2023: regime signals “consistently outperform in reducing risk metrics such as volatility and maximum drawdown” while raising the Sharpe ratio primarily by reducing volatility at the expense of return. During COVID-19 the online signal avoided a ~20% drawdown but missed the rebound. This is the recurring pattern in the Regime-Based Asset Allocation literature — a defensive overlay, not an alpha engine.

The case for regime classification as a source of directional alpha that beats a benchmark after costs is materially weaker, and weakens further the more honestly a study is conducted. The decisive culprit is the gap between ex-post and real-time regime labels. A model’s smoothed inference conditions on the entire observation sequence, including future data, and identifies turning points sharply with hindsight; a trader only ever has filtered (online) inference, which Ang & Timmermann describe as tracking the true regime “quite accurately, but at times miss[ing] an important regime change … and at other times issu[ing] false alarms”. Backtesting on smoothed labels is therefore Lookahead Bias from Smoothed Regime Estimates, and it manufactures apparent prescience. Zakamulin (2016) is the cleanest demonstration: he re-ran a widely cited moving-average market-timing study and showed its “too good to be true” performance was entirely an artefact of look-ahead bias in the simulation — after correcting it, the timing strategy was at best marginally better than buy-and-hold on Sharpe and worse on alpha, statistically indistinguishable from passive holding. His companion out-of-sample work (Zakamulin 2015) finds no statistically significant timing outperformance in the later half of the sample. Even with honest filtered inference, Real-Time Regime Identification Lag (a median ~25-day detection delay) and Regime Misclassification erode the edge: Dacco & Satchell (1999) prove that a small real-time misclassification rate is enough to make even the true regime-switching model forecast worse than a random walk.

Studies that do report regime-conditioned alpha exist, but they should be graded against costs and benchmarks before being read as proof. Oliveira et al. (2025) report that macro-regime-conditioned ETF portfolios beat random-regime controls and the SPY/equal-weight benchmark, with the best long-only configuration reaching a Sharpe of ~1.5 versus ~0.82 for SPY and statistically significant gains (p<0.01) — and they explicitly note the improvement comes from return generation, with drawdown changes statistically insignificant. But that backtest includes no transaction costs and no slippage, and long-short variants frequently produced negative Sharpe; the apparent alpha is a gross, cost-free result, not a substantiated tradeable edge. Mulliner, Harvey, Xia, Fang & Van Hemert (2025) likewise report “significant outperformance” from a similarity-based regime/anti-regime classification used for factor timing — promising practitioner research, but timing factor premia is itself what Asness has called the “siren song”, and the result awaits independent replication after realistic costs. The honest synthesis: regime classification is a confirmed technique for de-risking — cutting volatility and drawdown — and an inconclusive-to-weak basis for standalone directional alpha once look-ahead bias is removed and costs, turnover and benchmarks are applied. Whether any regime-conditioned strategy delivers robust, replicated, after-cost alpha remains an open research question for this vault.

Regime Classification [part-of] Hidden Markov Model Regime Detection Regime Classification [part-of] Markov Regime-Switching Model Regime Classification [supports] Regime-Based Asset Allocation Regime Classification [relates] Tactical Asset Allocation Lookahead Bias from Smoothed Regime Estimates [contradicts] Regime Classification Real-Time Regime Identification Lag [contradicts] Regime Classification Zakamulin 2016 [contradicts] Regime Classification Bulla et al. 2010 [supports] Regime Classification

Connections

Sources

Graph View

Backlinks