Google adds multivariate forecasting to BigQuery SQL

Google has introduced a preview of TimesFM 3.0 in BigQuery, adding multivariate time-series forecasting to SQL workflows.

Developed by Google Research, the model is available through BigQuery's AI.FORECAST function. It extends earlier TimesFM versions, which were limited to univariate forecasting from a single historical series.

Users can now forecast multiple targets at once and include related variables in the same query. The model supports both covariates known only from historical data and covariates known in advance across the forecast horizon.

In practice, this covers forecasting tasks where outcomes depend on several related measures and external drivers such as holidays, promotions, or weather. That mirrors many operational forecasting problems, where business variables move together rather than in isolation.

Broader scope

Previous TimesFM releases in BigQuery, first based on TimesFM 2.0 and later updated to TimesFM 2.5, let users generate forecasts directly in SQL without building or maintaining custom models. TimesFM 3.0 keeps that workflow in place while expanding the model from single-target forecasting to single or multiple targets.

BigQuery users can now specify multiple target columns in one AI.FORECAST call, along with past and future covariate columns. Historical covariates can be used when values are available only up to the forecast cut-off, while future covariates can be supplied across both historical and forward-looking rows when those values are already known.

The model also serves as a direct upgrade for existing univariate forecasting in AI.FORECAST and anomaly detection in AI.DETECT_ANOMALIES. Existing workflows can switch to the new model by specifying TimesFM 3.0 in the query.

How it works

Google illustrated the feature with a BigQuery query using the public New York Citi Bike dataset. In that example, the model forecasts daily subscriber and casual trips over seven days, while using average trip duration as a past covariate and a weekend indicator as a future covariate.

The input uses a single-table format that combines historical observations with forecast-horizon rows. Historical rows contain populated target metrics and covariates up to the cut-off date. Future rows contain known future covariates, while target columns and past-only covariates are set to NULL.

That structure allows the model to estimate several related outputs jointly. In the Citi Bike example, subscriber and casual ridership are forecast together rather than through separate model runs for each rider group.

The output returns a structured prediction for each target column, including a point forecast and lower and upper prediction interval bounds. In the sample output, BigQuery produces separate forecast values and confidence intervals for subscriber and casual trips for each day in the seven-day period.

Benchmarks

Google said TimesFM 3.0 is a 330-million-parameter foundation model and ranks first on the public time-series leaderboards GIFT-Eval, FEV-Bench, and TIME. It also said the model achieved the top zero-shot accuracy on both covariate and multivariate forecasting tasks compared with existing foundation models.

Those benchmark claims are likely to matter for BigQuery users weighing whether to rely on general-purpose foundation models rather than build custom forecasting systems. The model is part of Google's wider effort to bring more advanced analytical methods into the standard SQL environments used by data teams.

Natural language

Google is also extending the model to BigQuery Conversational Analytics. In that interface, users can ask for a multivariate forecast in natural language, and the system generates and runs the underlying AI.FORECAST query.

In the Citi Bike example, the system can produce seven-day predictions from a plain-language request, along with forecast charts, prediction intervals, and a written summary of the results. That could widen access to forecasting tools for users who do not write SQL directly but still work with BigQuery datasets.

The release adds another forecasting option to Google's analytics stack as cloud providers increasingly embed foundation models into database and warehouse products. The main shift is that multivariate forecasting now moves into the same SQL-based workflow BigQuery users already use for querying and analysis.