Large time series models are deep learning foundation models pre-trained on vast and diverse collections of sequential temporal data to perform a wide range of time series analysis tasks. Analogous to large language models, these architectures, typically built upon scalable transformer frameworks, learn generalized representations of temporal dynamics and dependencies across varied domains rather than being engineered from scratch for a single dataset. By unifying distinct analytical tasks such as forecasting, imputation, classification, and anomaly detection into generalized pre-training and inference paradigms, large time series models provide high scalability, cross-domain transferability, and strong zero-shot or few-shot generalization capabilities across unseen environments.
