Principled identification of optimal augmentation strategies for TSFM training

Determine a principled method to identify optimal synthetic data augmentation strategies for training time series foundation models, avoiding reliance on heuristics chosen prior to training and providing a systematic approach to guide augmentation design across tasks and domains.

Background

The paper surveys existing data augmentation approaches used for time series foundation models, noting they largely depend on manually designed patterns or simple transformations that are fixed before training. Such heuristics vary in effectiveness across tasks and domains and lack a unified, principled framework.

Motivated by these limitations, the paper introduces OATS, an online augmentation strategy that uses influence-based guiding signals and diffusion models to generate synthetic data during training. While OATS offers a concrete approach, the broader question of establishing principled criteria and methods for identifying optimal augmentation strategies remains explicitly stated as open in the related works.