Introducing TAM: A PyTorch-Based Additive Forecasting Framework and its Application to Explainable Anomaly Detection

This is the official slide deck for the presentation "Introducing TAM: A PyTorch-Based Additive Forecasting Framework and its Application to Explainable Anomaly Detection," delivered at the DASS Workshop (Anomaly detection for modern data) on September 8th, 2026. Overview: The presentation introduces the Time series Additive Model (TAM) v1.3.0, a PyTorch-based framework that combines the transparency of standard statistical models with exact, GPU-scaled primal resolution. It demonstrates how an explainable additive model can not merely flag an anomaly, but adapt to it through both time and space. The framework's capabilities are illustrated through two distinct case studies: Power Grids (Time): Forecasting French national electricity load through severe structural regime shifts, specifically the COVID-19 lockdowns and the 2022 energy crisis. It highlights the use of sliding-window correctors (AdaptiveTAM) and dynamic expert aggregation (OperaTAM) to track shifting mechanisms in real-time. Real Estate (Space): Detecting and mathematically deconstructing spatial price anomalies in the French housing market. It showcases TAM's new statistical layer, utilizing non-Gaussian distributional location-scale solves combined with Conformal Quantile Regression (CQR) to generate exact, explainable price intervals. Related Resources: Workshop Code & Reproducibility: https://github.com/yallioux/2026_dass_tam_for_anomalies TAM Software Framework: https://doi.org/10.5281/zenodo.22544995 FORCE Dataset: https://doi.org/10.5281/zenodo.21109134