synthetic-pretraining

Timestral is a family of four time series foundation models (5M to 209M parameters) achieving state-of-the-art probabilistic forecasting on GIFT-Eval, TIME, and fev-bench via decoupled horizon forecasting, native future covariate support, and synthetic pretraining followed by mixed post-training.

This paper proposes using controlled synthetic pretraining with the O'PRIOR task generator to make training-data attribution experimentally verifiable for tabular foundation models, showing that top-attributed task removal measurably degrades held-out performance and revealing that provenance association and interventional faithfulness can diverge.