Time-Series Anomaly Detection: Benchmark Flaws, Evaluation Metrics and Strong Baselines

Abstract

Dozens of deep architectures for time-series anomaly detection have been published since 2018, most of them reporting higher scores than their predecessors on the same handful of datasets. A series of critical studies now suggests that much of this progress is an artefact. The datasets have flaws that make them easy or mislabelled. The most common scoring protocol, point adjustment, inflates results so strongly that random scores can compete with trained models. And when classical or deliberately simple methods are run under the same conditions, they often match the deep ones. This review assembles that evidence. We describe the method families briefly, then examine the benchmark critiques, the evaluation metrics and their failure modes, and the large comparison studies. A small worked example, constructed by hand and labelled as such, shows the arithmetic of point adjustment. We end with a reporting checklist and with the open question of whether pretrained time-series models change the picture. Apart from that worked example, no new result is reported.