UniTS

UniTS (Gao et al., NeurIPS 2024) tackles the fragmentation of time-series modeling head on: instead of a different architecture per task, it trains one unified transformer with shared weights and no task-specific modules, using a prompting mechanism to specify the task. Forecasting, classification, imputation, and anomaly detection all run through the same network, and prompt tuning adapts it to new domains with few examples. Across 38 multi-domain datasets, the authors report it beating task-specialized models and repurposed language models.

The repo mims-harvard/UniTS (648 stars, MIT) ships pretraining, supervised, few-shot, and zero-shot scripts plus released checkpoints. It is a research artifact more than a product: only six commits, a heavyweight training setup, and no pip package or hosted option, so expect to work from the scripts. Still, it is one of the cleanest demonstrations that a single time-series model can genuinely cover multiple tasks, and the prompting idea has influenced later multi-task designs.

Author-reported (Gao et al., NeurIPS 2024, arXiv 2403.00131): superior performance to 12 forecasting, 20 classification, 18 anomaly detection, and 16 imputation models across 38 datasets; strong zero-shot and few-shot transfer. Independent replication is limited.