Advances on TSLies: Time Series AnomaLIES, an efficient machine learning based anomaly detection framework

Anomaly detection in time series is an important task in scientific and industrial applications, where traditional statistical methods are limited when processing high-volume, high-dimensional, and non-stationary data. This proceeding describes methodological advancements to the TSLies (Time Series anomaLIES) framework, a machine learning-based Python toolkit designed for detecting anomalous events in temporal data. Building upon our previous work applying machine learning techniques to the Fermi Gamma-ray Space Telescope Anti-Coincidence Detector (ACD) data, we address fundamental limitations of deterministic neural network approaches through the introduction of probabilistic modeling techniques. The enhanced framework now incorporates implementations of Bayesian Neural Networks (BNNs) with TensorFlow Probability for uncertainty quantification in model predictions. Furthermore, we develop spectral domain learning capabilities through custom loss functions operating on Fourier transforms of the time series, enabling an accurate modeling that is frequency-dependent. We demonstrate the framework's capabilities through applications to two distinct domains: continuation of the analysis of Fermi ACD data and application to daily insurance policy time series from Intesa Sanpaolo, demonstrating cross-domain applicability.