A market regime is a period or modelled state in which market behaviour has a recognisable set of statistical characteristics, such as return variability, trend, correlation or liquidity. For a quant, the practical question is whether a risk estimate or trading rule fitted under one set of conditions remains useful under another.
“Calm”, “stressed”, “trending” and “risk-on” need definitions. They are labels chosen by an analyst or model, rather than a universal market classification. A volatility regime need not coincide with a bull or bear market: large positive and negative moves can produce high volatility without a positive average return. Ang and Timmermann's review describes regime models with differences in means, variances and dependence, including both recurring states and lasting structural changes.
Same mean, different scale
Consider two fictional, nonoverlapping windows, each containing four simple daily returns. Use a centred descriptive standard deviation with divisor , treating each window as the complete set being summarised. This is neither the sample estimator with divisor nor an annualised forecast.
| Window | Daily returns, in observation order | Arithmetic mean daily return | Descriptive daily standard deviation |
|---|---|---|---|
| A: days 1–4 | −1%, +1%, −1%, +1% | 0% | 1 percentage point |
| B: days 5–8 | −3%, +3%, −3%, +3% | 0% | 3 percentage points |
Using returns expressed in percentage points, the means are zero and the calculations are
Window B has three times the observed dispersion of A. Calling these “lower-volatility” and “higher-volatility” windows describes this particular comparison. Four returns do not establish stable underlying states or a reliable trading rule. The arithmetic mean is not a compounded holding-period return.
Original fictional daily returns. Centred standard deviations use divisor 4 and are not annualised. Completed-window summaries describe history; neither panel predicts the next return.
A label, a rule or an unobserved state?
A descriptive label summarises a completed interval, as above. A threshold rule assigns a state from a specified indicator, lookback and cutoff. Its boundary is a design choice, possibly estimated from training data; there is no universal volatility level that separates calm from stress.
A regime-switching model instead gives different parameters to different states. In a hidden-state model, the state is latent: the analyst observes returns and infers probabilities of state membership. A Markov chain can describe the transitions. The number of states, distributional assumptions and transition structure are model choices, and estimated probabilities inherit those choices. State numbers themselves have arbitrary labels; compare fitted characteristics before treating “state 1” from different fits as the same condition.
This distinction matters when evaluating a strategy. Partitioning history into attractive-looking episodes can explain past performance without producing a decision that was available at the time.
What was knowable then?
The summary for A requires all returns through day 4; B requires all returns through day 8. Assigning B's final label to a trade made before its window closed introduces look-ahead bias.
For latent-state models, filtered probabilities use observations through the current time; smoothed probabilities also use later observations in the fitted sample. The statsmodels switching-model example makes this distinction explicit. Even filtering can leak future information if parameters or thresholds were estimated on later data. A historical trading test needs the data availability, fit and decision timing that the proposed live process could actually use.
Detection also takes time. Shorter windows react faster but are noisier; longer windows can keep old observations after conditions change. Estimated states may shift when the sample, inputs or model change. Check stability and performance on later, unseen periods, including trading costs and the consequences of switching allocations. A useful risk warning need not prove that a permanent new regime has begun.
The owned lesson Drift, Decay and Retirement develops this monitoring problem: distinguish data defects, changing conditions and lost economic value before deciding how to respond. The public preview is open; the full lesson requires membership.
Sources
- Andrew Ang and Allan Timmermann, Regime Changes and Financial Markets, 22 June 2011 working paper, §§1, 2.1–2.2: recurring versus changing regimes, latent states, transition models and estimation choices.
- statsmodels, Markov switching autoregression models, Hamilton switching-model discussion: filtered versus smoothed state probabilities and switching-variance examples.