Anomalies in Multivariate Time Series Benchmarks Are Mostly Univariate

The study presents a diagnostic framework evaluating the assumption that anomalies in multivariate time series anomaly detection (MTSAD) are distributed across multiple channels. Analysis of eight public benchmarks reveals that anomalies primarily manifest as univariate deviations, with at least half of the segments showing univariate anomalies in 89% to 100% of their timesteps. This indicates that current benchmarks inadequately assess cross-channel modeling capabilities, suggesting a need for more diverse evaluation sets, which is critical for practitioners developing robust MTSAD systems.