TimesFM-3: A zero-shot foundation model for multivariate forecasting
Google says its new 330M-parameter TimesFM-3 can forecast multiple related time series at once, zero-shot, in a single forward pass.
The model uses target series plus historical and known-future covariates, such as promotions, holidays, or weather forecasts. Google says alternating temporal and cross-series attention lets it capture dependencies across coevolving signals without task-specific fine-tuning. In its tests on Gift-Eval, FEV-Bench, and Time, TimesFM-3 ranked first among pre-trained foundation models for both point and probabilistic forecasting. It is available on GitHub and Hugging Face, with BigQuery integration planned “in the coming weeks.” Google Research's note
The model uses target series plus historical and known-future covariates, such as promotions, holidays, or weather forecasts. Google says alternating temporal and cross-series attention lets it capture dependencies across coevolving signals without task-specific fine-tuning. In its tests on Gift-Eval, FEV-Bench, and Time, TimesFM-3 ranked first among pre-trained foundation models for both point and probabilistic forecasting. It is available on GitHub and Hugging Face, with BigQuery integration planned “in the coming weeks.” Google Research's note
score 6