Applying Foundational Models for Time Series Anomaly Detection
summary
Compares two concrete pipelines: TimeGPT (forecasting, uses confidence intervals from a model trained on ~100 billion points, flags observations outside the interval) versus MOMENT (reconstruction, trained on a Time Series Pile of ~13 million series ~20 GB, uses fixed 512-window embeddings and compares reconstruction error). Shows UCR dataset results: recall often decent but many false positives, ECG gives low precision, and a practical thresholding trick uses the 98th percentile ×3 for novelty detection.
AI-written summary · about
chapters (9)
- 0:00Introduction and agenda
- 1:47Fundamentals of TSAD
- 4:08Traditional modeling approaches
- 10:17Foundational model paradigm
- 14:08TimeGPT deep dive
- 19:31Moment model overview
- 25:32Measuring model performance
- 30:48Experimental case studies
- 35:37Summary and Q&A
related talks
45mAnomaly Detection in Time Series
30mRealtime Time Series Anomaly Detection in Production
39mMastering real-time anomaly detection with open source tools
43mDetecting Anomalies in Streaming Data
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
31mFoundational Time Series Models in Practice: The Future of Forecasting, or Just Hype?
26mAnomaly Detection System for the CERN Cloud
44mMethods and Tools for Time Series Data Science Problems