Google releases TimesFM-3, natively supporting zero-shot multivariate time series forecasting.
Google Research released TimesFM-3, a dataset with 330 million parameters, pre-trained on a corpus of over one trillion time points. It can jointly predict multiple time series with zero-shot capability and utilize historical and future covariates.
Google Research released TimesFM-3, the first native time series model in the TimesFM series to be pre-trained for multivariate prediction.The model has 330 million parameters, and the training corpus includes real and synthetic data, covering more than one trillion time points.
TimesFM-3 can jointly predict multiple time series with common changes without requiring task-specific fine-tuning, and supports historical covariates and known future covariates. Google claims it achieves leading average rankings on three public benchmarks: GIFT-Eval, FEV-Bench, and TIME. The model is now available on GitHub and Hugging Face, with BigQuery integration to follow.