ZAAS: Zonal Aware Anomaly Score for Time Series

Abstract

Time series anomaly detection plays a critical role across domains from industrial monitoring to cybersecurity use cases. But its evaluation remains challenging. Traditional window level F1 score overweights long anomaly intervals, while heuristic “point‑adjusted” variants introduce bias by extending single detection across entire zones. We propose a Zone Normalized F1, which treats each true and each predicted anomaly interval as a unit, macro‑averaging precision and recall over intervals rather than windows. This eliminates length bias and yields a fairer comparison of detectors. We formalize the metric, illustrate its behavior on toy and real examples, and show how it complements existing protocols.