Regime-Change Detection and the Problem of False Alarms

[Shadow Trader]

Detecting that a market has shifted from one behaviour to another sounds valuable, and statistical methods exist for it, including change-point tests and hidden-state models. The difficulty is that markets generate abundant apparent changes that turn out to be ordinary variation.

Every detection rule trades two errors against each other. A sensitive rule catches real shifts early but fires on noise, causing a strategy to be switched off during normal drawdowns. A cautious rule avoids false alarms but confirms the shift only after much of the damage has occurred. No setting removes both errors; it only reallocates them.

Further pitfalls compound the issue. Parameters tuned on historical data to flag past regimes look impressive in hindsight because the labels were fitted knowing the outcome. Regimes also differ in length, so a rule calibrated on month-long episodes may miss a three-day one. And checking many indicators each day makes at least one flag almost certain by chance alone.

A more modest approach sets conditions in advance: which measurable variables, such as realised range or spread, must change by what margin, for how many sessions, before size is reduced. Predefined thresholds limit the temptation to rationalise a change after a bad week.

[FXS HongKong Scalper]

The trade-off is well put. We tested a simple change-point rule on our strategy's daily returns: switch off when the rolling 30-day mean dropped more than two standard errors below the long-run mean. It fired about four times a year, and only one of those was followed by a real deterioration.

The bigger problem was the restart. After switching off, the rule needs a separate condition to switch back on, and every version we tried missed the recovery. We now use it as an alert to review the strategy, not as an automatic switch.