Zero-Shot FMs Fail Multivariate Time Series Anomaly Detection

Multivariate Time Series Anomaly Detection (MTSAD) is crucial for industrial monitoring and financial risk management, but current methods are often application-specific and costly. The emergence of zero-shot Foundation Models (FMs) for univariate time series forecasting raised questions about their potential for MTSAD without specific training.Researchers investigated the zero-shot application of TimesFM, a univariate forecasting FM, to industrial MTSAD using the SWaT benchmark. They explored two strategies: treating TimesFM as a per-feature forecaster with thresholded prediction errors, and using its intermediate representations with standard outlier detectors. Neither approach proved competitive with established MTSAD baselines.The core issue identified was that TimesFM is too effective at capturing temporal dynamics, resulting in low prediction errors even within fully anomalous windows. This makes persistent anomalies indistinguishable from normal behavior. However, the study noted that error peaks reliably occurred at anomaly boundaries, suggesting FMs are promising for change-point detection rather than direct anomaly detection. The conclusion is that naive zero-shot FMs are currently unsuitable for MTSAD.