From Pre-training to Post-training: A Survey on Time Series Foundation Models

Deep learning models have achieved remarkable progress in time series analysis. However, most models require retraining when encountering new domain data, which limits their generalization and cross-domain transferability. While prior efforts have explored adapting large pre-trained language or vision models to time series tasks, modality gaps hinder their ability to capture complex temporal dynamics. Time Series Foundation Models (TSFMs) trained from scratch have been developed to overcome these limitations, enabling effective knowledge transfer through domain-specific fine-tuning while preserving intrinsic time series characteristics. To this end, this survey offers a comprehensive review of TSFMs trained from scratch, following a unique perspective from pre-training to posttraining. Specifically, we present a taxonomy of TSFMs across three dimensions: (1) datasets, discussing the construction of large-scale source domain datasets and the impact of target domains on model evaluation; (2) pre-training, covering taskagnostic and task-specific training paradigms; and (3) posttraining, encompassing optimization via supervised, collaborative, and reinforcement-based fine-tuning. Finally, we highlight potential future research directions, including sample quality evaluation, pre-training paradigm design, and agent-based posttraining, to advance the development and practical deployment of TSFMs. Our code is available at https://github.com/ZLiu21/ awesome-tsfms-from-pre-training-to-post-training.