Research map / Time series / Foundation models for time series

OATS: Online Data Augmentation for Time Series Foundation Models

NeurIPS 2026 · Time series › Foundation models for time series · approach: Time series foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes OATS, an online data augmentation method that uses data-attribution influence scores to pick guiding samples and a diffusion model to generate synthetic time series during TSFM pretraining, with an explore-exploit scheme to cut cost.

Problem

Static, heuristic synthetic data augmentation for time series foundation models ignores that the value of training samples changes across training stages.

Approach

Use first-order Taylor influence scores on a small reference set to identify high-value training samples at each step, condition a diffusion generator (with prototype and subset-class conditioning) on them to produce synthetic data online, and reuse cached per-subset scores via epsilon-greedy explore-exploit.

Key results

• On encoder-only TSFM at prediction length 192, OATS reaches NLL 1.627 on ETTm1 vs 1.870 for regular training and 1.725 for TSMixup.
• On decoder-only TSFM, OATS achieves best NLL on ETTm1 (1.654 vs 1.737 regular) and Electricity 192 (6.041 vs 6.049 regular), though some augmentation baselines are worse than regular on several datasets.
• Per-step overhead scales with epsilon: 0.95 s at epsilon=0.3 vs 0.74 s regular and 1.32 s at epsilon=1.0.

Surprising findings

• TSMixup and Jitter give mixed results versus regular training, sometimes hurting performance on ETTh2, Weather and Electricity.
• The best explore-exploit setting is often not epsilon=1 (full influence computation), with cached scores sometimes performing better.

Limitations

• Requires training a diffusion generator and computing influence scores, adding overhead.
• Gains over baselines are sometimes within error bars and not consistent across all datasets and decoder-only results.
• Evaluation limited to small (10M) TSFM models and six LSF datasets.

Experimental setup

Datasets: LOTSA-100M ETTm1 ETTm2 ETTh1 ETTh2 Weather Electricity

Benchmarks: ETTm1 ETTm2 ETTh1 ETTh2 Weather Electricity

Metrics: NLL MAPE

Baselines: Regular training TSMixup Jitter DD (offline diffusion augmentation)

Builds on: Moirai Chronos DDIM TSMixup ghost inner product

Closest papers by meaning

• MOMENT: A Family of Open Time-series Foundation Models (ICML 2024 · similarity 0.87)
• Towards Neural Scaling Laws for Time Series Foundation Models (ICLR 2025 · similarity 0.87)
• Learning the Context of Errors: Black-Box Online Adaptation of Time Series Foundation Models (NeurIPS 2026 · similarity 0.87)
• Empowering Time Series Analysis with Large-Scale Multimodal Pretraining (NeurIPS 2026 · similarity 0.86)
• Lightweight Online Adaption for Time Series Foundation Model Forecasts (ICML 2025 · similarity 0.86)
• Synthetic Series-Symbol Data Generation for Time Series Foundation Models (NeurIPS 2025 · similarity 0.86)