Anomaly Detectors for Multivariate Time Series: The Proof of the Pudding is in the Eating

Anomaly detection is a popular task in time series analytics and researchers have, therefore, developed a plethora of algorithms to solve it. While most algorithms focus on univariate time series, one family of anomaly detection algorithms specializes on multivariate data. Because existing studies benchmark on non-meaningful datasets and often only within uni- or multivariate algorithm families, it is unclear whether multivariate solutions are actually superior on multivariate data.In this study, we compare univariate and multivariate approaches on common multivariate benchmark times series to demonstrate that existing benchmark datasets cannot highlight the strengths of multivariate anomaly detection algorithms. We though demonstrate such strengths with a simple, generated dataset that contains a special type of anomaly, which we call correlation anomaly. Our experimental results, therefore, call for novel types of benchmark datasets whose anomalies actually facilitate the multidimensional nature of the data.