Anomaly Detection Tutorials
A sequence of runnable notebooks that teaches time-series anomaly detection from simple z-scores up to modern deep learning models, with honest evaluation and real-world case studies.
Anomaly Detection Tutorials is a curriculum of notebooks progressing from methods you can compute by hand to deep models used in production: starting from z-scores, moving statistics, and STL residuals, through classic machine learning methods such as Isolation Forest, One-Class SVM, and LOF, autoencoders and their adversarially trained and probabilistic variants, the training-free Matrix Profile method, forecasting-based detection, a GAN-based detector, graph neural networks, transformers, self-supervised representations, streaming detection with the Robust Random Cut Forest, and zero-shot detection with foundation models, ending with a chapter on evaluating detectors correctly and a planned case study on detecting silent regressions in LLM telemetry.
Every notebook runs top to bottom on CPU in minutes, is seeded for reproducibility, pairs each method with its failure modes rather than only its successes, measures results against ground-truth labels where they exist and says clearly where they do not, and uses one consistent, colorblind-safe visual style. Notebooks can be run locally after installing with uv, or opened directly in Google Colab.