Anomalies in Multivariate Time Series Benchmarks Are Mostly Univariate

In three linesAn arXiv study analyzes 8 benchmarks for multivariate time series anomaly detection. A diagnostic framework shows 79-100% of anomalies are univariately detectable on 6 datasets. Cross-channel models provide no measurable gain. Current benchmarks fail to validate multi-channel modeling capabilities.

Summary generated by Claude — human-verified