Google Cloud previewed TimesFM 3.0, a 330-million-parameter model supporting native multivariate time-series forecasting directly within BigQuery SQL.

Google Cloud announced the preview of TimesFM 3.0, a 330-million-parameter foundation model developed by Google Research. It adds native multivariate forecasting to BigQuery’s Structured Query Language (SQL) workflows, extending capabilities previously available through TimesFM 2.0 and 2.5.

Earlier versions supported univariate forecasting, using the history of a single time series. TimesFM 3.0 supports single or multiple target columns, past-only covariates, and covariates with both historical and known future values, such as the external drivers discussed in the announcement.

Google says TimesFM 3.0 ranks first on GIFT-Eval, FEV-Bench, and TIME, covering zero-shot covariate and multivariate forecasting tasks. These are vendor-reported leaderboard results; the model also serves as a drop-in upgrade for univariate forecasting through `AI.FORECAST` and anomaly detection through `AI.DETECT_ANOMALIES` by specifying `model => 'TimesFM 3.0'`.

The article demonstrates forecasting daily commuter and leisure trips from the NYC Citi Bike public dataset over a seven-day horizon. Its single-table format combines historical observations with future rows containing known future covariates, while future target values and past-only covariates are set to `NULL`.