Google releases TimesFM 3.0 time-series model in BigQuery preview, adding multivariate forecasting

Google DeepMind

Google's TimesFM 3.0, a 330-million-parameter time-series foundation model, is in preview in BigQuery. The update adds forecasting for several linked series at once, which earlier versions could not do, and lets users feed in covariates that the model can use.

• Existing AI.FORECAST and AI.DETECT_ANOMALIES queries can switch to it by setting the model to 'TimesFM 3.0'.
• Earlier versions forecast a single series only; version 3.0 accepts single or multiple target columns.
• Past-only covariates and past-and-future covariates are supported; versions 2.0 and 2.5 supported neither.
• Google says the model ranks first across the GIFT-Eval, FEV-Bench and TIME leaderboards; the page gives no numeric scores.
• The page gives no general-availability date and no pricing, and the model may change while in preview.