OATS: Online Data Augmentation for Time Series Foundation Models
Source
- Raw Markdown: paper_oats-2026.md
- PDF: paper_oats-2026.pdf
- Preprint: arXiv 2601.19040
- Official code: microsoft/TimeCraft/OATS
Status And Credibility
OATS was posted to arXiv on 2026-01-26 and is referenced by the official Microsoft TimeCraft repository as a 2026 extension. The arXiv metadata snapshot does not list a peer-reviewed venue.
Core Claim
OATS makes synthetic data generation online and training-stage-aware for time-series foundation models. It computes time-series influence scores, conditions a diffusion generator on valuable samples, and uses an explore-exploit mechanism to refresh augmentation during TSFM pretraining.
Key Contributions
- Defines online data augmentation for TSFM training rather than static pre-generated augmentation.
- Uses Time-Series Influence Scores to select valuable guiding samples.
- Trains a guided diffusion generator on a small subset of the TSFM training data.
- Evaluates encoder-only and decoder-only TSFM architectures on six LSF datasets.
Evidence And Results
The paper evaluates ETTm1, ETTm2, ETTh1, ETTh2, Weather, and Electricity, reporting normalized MAPE and NLL. It reports that OATS improves over regular training, TSMixup, and jitter baselines for both encoder-only and decoder-only TSFMs.
Limitations
- The augmentation loop uses evaluation-like reference samples in experiments; strict separation is essential before treating the method as benchmark-safe.
- Influence-score computation and diffusion generation add training complexity.
- OATS improves passive forecasting pretraining; it does not add action, intervention, or counterfactual semantics.
- Results are tied to six standard forecasting datasets and two TSFM architecture templates.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Data diversity and synthetic pretraining | partially closes | Generates training-stage-specific synthetic samples for TSFM pretraining. | Needs leakage-safe protocols and broader domain tests. |
| Dynamic compute/data allocation | adjacent | Explore-exploit score refresh allocates augmentation effort to valuable samples. | Not a model-inference dynamic-compute method. |
| Control and counterfactuals | insufficient evidence | Forecasting augmentation does not expose controllable actions. | Needs action-conditioned data contracts. |