Hunting the outliers: Machine learning for anomalous time series detection

The increasing availability of large-scale time series datasets from modern astronomical surveys, such as those provided by Gaia and the forthcoming Legacy Survey of Space and Time (LSST) to be conducted with the Simonyi Survey Telescope at the Vera C. Rubin Observatory, is transforming time-domain astrophysics, enabling the systematic study of variable and transient phenomena across billions of sources. However, the sheer volume and heterogeneity of these data present significant challenges for traditional analysis techniques. Feature-based representations have emerged as a powerful solution, allowing the application of machine learning methods for efficient characterization and classification of astrophysical sources, including the reliable identification of Active Galactic Nuclei (AGNs). In this work, we introduce a general-purpose methodology for anomaly detection in time series data that transfers this feature engineering framework to the financial domain, where time series display complexities, such as noise and irregular patterns, closely resembling those in astrophysics. By combining domain-informed features with unsupervised algorithms (specifically Isolation Forests and Autoencoders), our approach effectively detects anomalous time series relative to the sample studied, demonstrating strong performance and highlighting its cross-domain transferability. Moreover, we propose an extension based on adaptive temporal windows to localize anomalies at a finer temporal resolution, further enhancing detection capabilities. Finally, we discuss the potential for reapplying this adaptive strategy to astrophysical time series, aiming to improve the identification of rare or unexpected behaviors in future studies.