Change-point detection

Change-point detection identifies the specific points in a time series where its underlying behavior, such as its mean, variance, or trend, shifts abruptly to a new regime, rather than drifting smoothly. A retailer's daily sales series might show a change point on the exact day a pricing policy changed, distinct from ordinary day-to-day noise around a stable level.

Methods range from simple approaches that scan for the split point that best separates two segments with different means, to more elaborate sequential algorithms that flag a change as soon as enough evidence accumulates, useful for near-real-time monitoring. This differs from anomaly detection, which flags individual points that deviate from an established pattern; change-point detection instead identifies where the pattern itself permanently changes, and a single anomalous spike is not, by itself, a change point.

Change-point detection matters for spotting policy changes, system failures, or shifts in customer behavior that a smoothed trend line can obscure, and it is a useful diagnostic before fitting a time series forecasting model, since a change point violates the stationarity many models assume. It is closely related to statistical process control, which monitors a process for exactly this kind of shift, and typically relies on the same autocorrelation diagnostics used elsewhere in time-series work. A common pitfall is over-sensitive detection that flags ordinary volatility as a regime change, producing frequent false alarms that erode trust in the monitoring system.