Research map / Time series / Anomaly detection
When Foundation Models are One-Liners: Limitations and Future Directions for Time Series Anomaly Detection
Xiaokun Zhu, Louis Carpentier, Mathias Verbeke
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MOMENT-large instead of MOMENT-base | MOMENT reconstruction TSAD · TSB-AD-U eval | mean VUS-PR: 0.311 -> 0.313 | No clear effect |
| scale model size or context window length | MOMENT, Chronos, TimesFM, Time-MoE · TSB-AD-U eval | VUS-PR | No clear effect |
| replace per-window z-normalization with Min-Max or global z-normalization (ablation) | MOMENT · TSB-AD-U eval | correlation with variance baseline | Mixed |
| replace normalization with Min-Max, global z-norm or mean-scaling (ablation) | forecasting TSFMs · TSB-AD-U eval | correlation with baseline | No clear effect |
| use Chronos-Bolt, Time-MoE or TimesFM instead of Last-3 baseline | forecasting TSFMs · TSB-AD-U eval | mean VUS-PR: 0.284 -> 0.303 | No clear effect |
| add TSPulse reconstruction head instead of Var-96 | TSPulse · TSB-AD-U eval | mean VUS-PR: 0.42 -> 0.42 | No clear effect |
| use triangulation of TSPulse heads instead of baseline triangulation | TSPulse · TSB-AD-U eval | mean VUS-PR: 0.46 -> 0.48 | No clear effect |
| increase forecast horizon from 1 to 128 for sequence anomaly | TimesFM-1.0-512 · TSB-AD-U 371 | VUS-PR: 0.04 -> 0.68 | Improved |
| increase forecast horizon from 1 to 128 for point anomaly | TimesFM-1.0-512 · TSB-AD-U 603 | VUS-PR: 1.0 -> 0.02 | Hurt |
| ensemble anomaly scores across horizons with MAX operation | TimesFM-1.0-512 · TSB-AD-U 558 | detection of both point and sequence anomalies | Improved |
| apply kNN to hidden representations instead of reconstruction error | MOMENT-base-64 · TSB-AD-U 474 | VUS-PR: 0.003 -> 0.250 | Improved |
Closest papers by meaning
- Can LLMs Understand Time Series Anomalies? (ICLR 2025 · similarity 0.91)
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.90)
- MOMENT: A Family of Open Time-series Foundation Models (ICML 2024 · similarity 0.89)
- Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.89)
- Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.89)
- When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection (AAAI 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
