Time Series Foundation Models Meet Variational Autoencoders for Anomaly Detection

Time Series Anomaly Detection (TSAD) is a well- known data mining task that is relevant to several domains, ranging from manufacturing to healthcare. Since the anomalies observed in time series from different domains exhibit peculiar characteristics developing general-purpose yet effective solutions is particularly challenging even within the univariate setting. On the one hand, recent advances in Time Series Foundation Models (TSFMs) offer new avenues for TSAD; on the other hand, classical models, such as autoencoders networks retain interest as they show higher-than-expected robustness to complex anomalous patterns. This paper presents a new TSAD approach that combines an established TSFM approach, designed for time series forecasting, with a latent variational autoencoder model. Specifically, in the proposed solution the variational autoencoder is trained to reconstruct the transformer embeddings extracted from the TSFM backbone. Experiments run on the TBS-AD benchmark reveal a clear complementarity between forecasting-based and reconstruction- based scoring. While the base TSFM variant excels at detecting pointwise anomalies, our proposed VAE approach demonstrates superior performance on sequence anomalies, achieving highly competitive results against state-of-the-art models across multi-domain datasets (e.g., VUS-PR +0.869 on SED, +0.363 on LTDB). These findings demonstrate that a pretrained probabilistic forecasting model serves as a strong TSAD backbone, provided the scoring head is matched to the expected anomaly structure.