Google has introduced TimesFM 3.0 in preview in BigQuery. The model lets users forecast multiple related data series in one SQL query and account for factors that may affect the results, such as weather or holidays. Earlier TimesFM versions in BigQuery forecast one series at a time.
The Google Research model has 330 million parameters. In AI.FORECAST, users can specify multiple target columns so the model predicts them together. They can also supply past covariates, whose future values are unknown, and future covariates, whose values are known across the forecast period.
Google illustrates the feature with a seven-day forecast of subscriber and casual rides in the public New York City Citi Bike dataset. The example uses past average ride duration and a weekend indicator for the forecast dates. AI.FORECAST returns a predicted value and lower and upper prediction interval bounds for each target.
The model supports natural language queries
TimesFM 3.0 also works in BigQuery Conversational Analytics. Google says users can request a multivariate forecast in natural language, and the tool generates and runs the AI.FORECAST query. It then returns a summary, interactive charts with prediction intervals, and forecast insights.
Existing BigQuery workflows can select TimesFM 3.0 for single-series forecasting through AI.FORECAST and anomaly detection through AI.DETECT_ANOMALIES. Google also lists it as an option for single-series evaluation with AI.EVALUATE. The company says the model ranks first on the GIFT-Eval, FEV-Bench and TIME public time-series leaderboards.
The preview is available in BigQuery Studio
TimesFM 3.0 is available in preview in BigQuery. Google says users can try it in the BigQuery Studio SQL workspace by setting the model to TimesFM 3.0 in AI.FORECAST.
