Contributed talk at the 12th Mining and Learning from Time Series workshop, co-located with KDD 2026 in Jeju, South Korea.
Many recent multivariate time series anomaly detection models model cross-channel dependencies, assuming the anomalies in our benchmarks actually carry cross-channel structure. The talk walks through the per-segment diagnostic framework we use to test that assumption on eight public benchmarks, the synthetic sanity check that shows the framework does catch cross-channel-only anomalies when they exist, and what follows for how we should be evaluating these models.
Slides 17 onwards are the backup slides used during questions.