Google Cloud announced the preview of TimesFM 3.0 in BigQuery, a 330-million-parameter time-series foundation model developed by Google Research. It is the first TimesFM version in BigQuery to support native multivariate forecasting, while also serving as a drop-in upgrade for existing univariate workflows via AI.FORECAST and AI.DETECT_ANOMALIES. The post states that TimesFM 3.0 ranks #1 on three public time-series leaderboards — GIFT-Eval, FEV-Bench, and TIME — for zero-shot accuracy on covariate and multivariate forecasting tasks.

Earlier TimesFM versions in BigQuery were limited to univariate forecasting, predicting future values from the history of a single series. TimesFM 3.0 addresses the more common enterprise case where multiple related targets and external drivers jointly shape outcomes. Through AI.FORECAST, it adds three multivariate capabilities: forecasting multiple correlated target columns in one call, using past-only covariates that are not known in advance, and using future covariates with known values across both historical and forecast horizons.

The announcement illustrates the workflow with the NYC Citi Bike public dataset, forecasting subscriber_trips and casual_trips over a seven-day horizon. Input is a single table combining historical observation rows and future horizon rows; target columns and past-only covariates are set to NULL in future rows. TARGET_COLS captures cross-series dynamics, PAST_COVARIATE_COLS gives historical context for shocks versus trends, and FUTURE_COVARIATE_COLS conditions predictions on known calendar events. Output includes structured predictions for each target with point forecast, prediction interval bounds, confidence level, and status.

Multivariate forecasting is also available through BigQuery Conversational Analytics by asking a natural-language question, which generates and executes the underlying AI.FORECAST query and returns charts and insights. TimesFM 3.0 is in preview; Google recommends reading the research blog post, exploring the docs, following the tutorial, and testing model => ‘TimesFM 3.0‘ with AI.FORECAST, AI.DETECT_ANOMALIES, and AI.EVALUATE.