“ Foundation Models for Time Series Analysis: Foundations, Methods, and Open Challenges
”
Speaker: Prof. Hamed Yazdanpanah
Abstract :
Time series data underpin decision-making across energy, finance, healthcare, retail, and traffic domains, yet the dominant modeling paradigms have evolved through three distinct stages: local statistical models, global deep learning models, and, most recently, time series foundation models (TSFMs). TSFM are models pretrained on massive, heterogeneous corpora spanning domains, frequencies, and scales, intended to learn general-purpose temporal representations that transfer to unseen data with little or no task-specific training. This talk surveys the foundations, methodology, and open challenges of TSFMs. We formalize the foundation-model objective as learning a representation mapping rather than a direct predictor, and discuss the architectural landscape, key design axes, and the self-supervised objectives that drive representation and transfer learning. We then argue that strong average predictive accuracy does not imply statistical reliability, and discuss four open problems that scaling alone does not resolve: generalization under distribution shift, calibration of uncertainty under temporal dependence, robustness to nonstationarity, and evaluation validity in the presence of benchmark contamination. We close by outlining four directions for a statistical theory of TSFMs: transferability, uncertainty, adaptation, and scaling, needed to move the field from empirical leaderboard performance toward principled, reliable deployment.