OATS: Online Data Augmentation for Time Series Foundation Models

Junwei Deng, Chang Xu, Jiaqi W. Ma, Jiang Bian

Keywords: time series foundation model, data augmentation

Abstract: Data augmentation is a key approach for enhancing the training of Time Series Foundation Models (TSFM). However, existing augmentation methods typically rely on heuristic synthetic data generation and follow a static paradigm, where synthetic data is generated once prior to training. To overcome these limitations, we propose Online Data Augmentation for Time Series Foundation Models (OATS), an algorithm to generate high-quality synthetic time series data through principled and dynamically adaptive augmentation strategies. OATS introduces methods to identify high-quality samples, guide the generation process with these samples, and scale augmentation efficiently. Experiments on TSFM demonstrate that OATS significantly improves the generalization performance of TSFM pretraining.

Submission Number: 45