Regime Detection: When to Turn a Strategy Off

A strategy works for two years, then loses money for one. The losing year wasn't bad luck — the market regime changed.

Most systematic strategies are conditional, not unconditional. They have a regime where they earn their headline Sharpe and a regime where they hemorrhage. The discipline that separates institutional systematic trading from retail is knowing which regime you're in, in real time, and acting accordingly.

What "regime" actually means

A market regime is a stable statistical configuration: a set of volatilities, correlations, mean returns, and structural relationships that persist for a non-trivial period. Regime change is the transition from one configuration to another.

Examples:

Volatility regime. 2017 (low-vol) → 2018 (mid-vol with periodic spikes) → 2020 (extreme vol followed by recovery).
Correlation regime. 2007 (low equity-cross-correlation) → 2008 (high correlation, "everything goes down together").
Carry regime. 2009–2014 (steep curves, broad carry profits) → 2015 (shifting carry rankings as policy diverged).

Detection methods

Volatility breakout

The simplest and often most useful: realized volatility breaking outside its multi-year band signals a regime shift in progress.

if realized_vol_60d > percentile(realized_vol_history, 90):
    state = "high-vol regime"

A surprising amount of strategy-on/off discipline can be expressed in this single rule. Most systematic strategies that fail in vol spikes would have been turned off if a simple vol-breakout rule had been in place.

Correlation regime

A multivariate analog. Compute the rolling 60-day correlation matrix and monitor its principal eigenvalue. A spike in the largest eigenvalue indicates a "single-factor regime" where everything is moving together — historically the worst environment for diversified systematic strategies.

Markov-switching models

Statistically rigorous; operationally fragile. A two-state hidden Markov model trained on historical returns labels each period as "normal" or "crisis." Useful for regime classification in research, less useful in production because the model lags the regime change by the smoothing window.

Strategy-level drawdown signature

The most practical detection method: compare the shape of the current drawdown to historical worst-case drawdowns. A drawdown that looks like prior drawdowns suggests "same regime, bad streak." A drawdown that doesn't match any historical pattern suggests "new regime."

Acting on detection

Detection without action is theater. Three institutional postures:

Strategy-on/off switching. Detection flags trigger flat-the-position. Conservative; sacrifices recovery alpha.
Position-size adjustment. Reduce sizing by 50–75% in suspect regimes; restore on normalization. Less binary, more nuanced.
Strategy substitution. Switch from the regime-fragile primary strategy to a complementary alternative. Trend-following and mean-reversion often partner this way.

Failure modes of regime detection

Detection lag. Most regime classifiers identify the regime after it's been in place long enough to be obvious — by which time the worst of the drawdown has happened.
False positives. The detector flags regime changes that don't actually materialize. Each false positive is a turned-off strategy missing recovery rallies.
Meta-overfitting. Tuning the detector on historical data optimizes for regimes that already happened. Real regime changes are, by definition, novel.

What we do

Sentivue's standard regime overlay is a combination of volatility breakout (binary) and strategy-level drawdown shape (continuous). When both signals fire — vol regime change AND drawdown shape unmatched — the strategy is flattened. When only one fires, sizing is reduced. When neither fires, the strategy runs at full sizing.

This rule is conservative. It costs us recovery alpha some of the time. It saves us from the failure mode where a strategy continues running into a regime that breaks it.

Practical takeaways

Most strategies have a regime where they work and a regime where they don't. Know which.
Simple volatility-breakout detection captures most of the available regime signal. Don't over-engineer.
Detection without an action plan is theater. Pre-commit to the response before the regime change.