NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 27·6 min read
TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models
TL;DR
Time series foundation models (TSFMs)—large neural networks pretrained on massive datasets—risk evaluating on contaminated data, where test sets overlap with pretraining corpora, inflating performance metrics. TSFMAudit is the first framework to detect such contamination in time series by handling continuous, heterogeneous signals lacking corpus documentation.
✦ Why It Matters
Engineers can now validate whether time series model improvements are real or caused by contaminated evaluation data.
Key Takeaways
- Audit pretraining datasets against evaluation sets to detect contamination before trusting benchmark results.
- Account for continuous signal nature and domain heterogeneity when designing contamination detection for time series.
- Document training corpora metadata to enable future contamination auditing and reproducibility.
- Validate that performance gains on time series benchmarks reflect genuine model capability, not data leakage.
Full Summary
Time series foundation models are large pretrained neural networks increasingly used for forecasting across domains. A critical concern is data contamination: evaluation datasets may overlap with pretraining corpora, causing inflated performance estimates that don't reflect real-world capability.
Unlike text or images, time series data is continuous and heterogeneous (varying formats, scales, domains), making contamination detection harder without documented training corpora. TSFMAudit addresses this gap by formalizing pretraining contamination auditing as a distinct problem for TSFMs and proposing detection methodology.
The framework handles the unique challenges of time series: continuous signals, domain heterogeneity, and missing corpus metadata. This work establishes the first systematic approach to auditing contamination in time series foundation models, enabling researchers to validate whether benchmark improvements reflect genuine capability gains or dataset leakage.