KM-ARL and KM-ADD: New changepoint detectors from ICML 2026

Changepoint detection is critical for monitoring system failures, network intrusions, and market shifts, but two classical metrics—average run length (ARL) and average detection delay (ADD)—break down when data sequences are short or irregularly spaced. Taiki Miyagawa and Akinori Ebihara tackle this by drawing an analogy between quickest changepoint detection (QCD) and survival analysis. Their new estimators, KM-ARL and KM-ADD, use the Kaplan–Meier method to model detection probabilities under sequence truncation, making them robust to finite, non-uniform observation windows.

The authors derive bias bounds and prove these estimators are asymptotically unbiased (unless extrapolation is required). Experiments on simulated and real-world datasets (e.g., network traffic, sensor logs) show that KM-ARL and KM-ADD yield more reliable rankings of detectors and enable intuitive model selection. The accompanying Python package on GitHub provides ready-to-use implementations. This work was accepted at ICML 2026 and opens practical use of QCD metrics in domains where data is scarce or irregular.