Method May 2026
Notes on regime detection
Separating trend from noise across shifting volatility regimes, and why most switches are seen only in hindsight.
Every model that trades a trend is implicitly betting on a regime — a stretch of market behaviour stable enough that yesterday's relationship still holds today. The hard part was never detecting the trend. It is knowing, in something close to real time, when the regime that made the trend tradeable has quietly ended.
Most regime labels are obvious in hindsight and useless in the moment. The chart makes a break or a grind look clean once you can see what came next. Live, the same data is ambiguous — and a detector tuned to be confident after the fact will whipsaw you before it.
Trend, noise, and the cost of being wrong
We treat regime detection as a cost-asymmetry problem rather than a classification one. Calling a calm market volatile is cheap — you trade smaller and miss a little. Calling a volatile market calm is expensive — you size up into exactly the conditions that punish leverage. The threshold should reflect that asymmetry, not a symmetric notion of accuracy.
A regime model that is right on average and catastrophic at the turns is not right. The turns are the only part that pays.
Signals that lead, signals that confirm
We separate detectors into two roles and refuse to mix them:
- Leading measures — dispersion, correlation breakdown, liquidity withdrawal — are noisy but early; they tell us to reduce conviction.
- Confirming measures — realized volatility, trend persistence — are reliable but late; they tell us a regime has actually changed.
Acting on leading signals and sizing on confirming ones keeps us from the two failure modes that ruin regime models: trading every false alarm, and recognizing the new world only after it has already cost us.
Method note
Detectors are validated out-of-sample across regimes the model never saw in fitting. A signal that only separates regimes in-sample is treated as decoration, not evidence.