Introducing TimesFM 3.0 with multivariate time-series forecasting in BigQuery
Google introduces TimesFM 3.0 in BigQuery, enabling native multivariate time-series forecasting directly in SQL without custom model training or infrastructure setup.
TimesFM 3.0, a 330-million parameter foundation model developed by Google Research, expands BigQuery’s AI.FORECAST capabilities to support multivariate forecasting, addressing real-world scenarios where multiple related time series and external factors influence outcomes. Previously limited to univariate forecasting, the model now processes correlated targets such as commuter and leisure bike rides jointly, capturing shared dynamics like weather sensitivity or weekend ridership patterns. The update maintains backward compatibility, allowing existing univariate workflows to upgrade by specifying the new model version without additional setup.
The model ranks first on public leaderboards including GIFT-Eval, FEV-Bench, and TIME, demonstrating top zero-shot accuracy for both covariate and multivariate forecasting tasks compared to existing foundation models. Users can incorporate historical context through past covariates like average ride duration, which are only available in historical data, while future covariates such as weekends or holidays are applied to refine seven-day predictions. This dual approach enables the model to distinguish temporary shocks from underlying trends without requiring future values for all features.
Forecasting is executed directly in BigQuery SQL using AI.FORECAST with the TimesFM 3.0 model, returning structured predictions for each target column including point forecasts and prediction intervals. The function supports multivariate targets via the TARGET_COLS parameter, past covariates via PAST_COVARIATE_COLS, and future covariates via FUTURE_COVARIATE_COLS, enabling comprehensive modeling of interconnected datasets without data movement or custom pipelines.
BigQuery Conversational Analytics further simplifies multivariate forecasting by allowing natural-language prompts to generate and execute AI.FORECAST queries, returning interactive forecast charts, prediction intervals, and key insights. The feature eliminates the need for manual SQL composition, providing a conversational interface to forecast multiple correlated time series while incorporating historical and future covariates directly from the NYC Citi Bike public dataset.