BigQuery's AI.FORECAST adds TimesFM zero-shot forecasting

Until now, forecasting revenue or inventory in BigQuery meant building a model. ARIMA and BigQuery ML's ARIMA_PLUS both require a learning phase, which carries its own data, time, and operating costs, and ARIMA_PLUS additionally needs missing-value imputation and seasonality settings.

AI.FORECAST takes a different route. It calls TimesFM, a time series foundation model that Google Research pre-trained on large amounts of time series data, so it works across a broad range of series without domain-specific retraining. In practice you create a remote model object pointing at the timesfm-1.0-200m endpoint, then feed AI.FORECAST a subquery with the three required columns and a STRUCT for horizon and confidence_level. Multiple series — product categories, traffic sources — can be handled in one pass by keying on time_series_id, so you don't need a separate query per series.

The trade-off is visible in the design. Tuning room is limited compared with ARIMA_PLUS, accuracy tends to drop when a series is too short, and spikes from sales starts, ad-budget increases, or seasonal events aren't learned in advance. For those, ARIMA_PLUS's holiday_region option or external regressors remain the better fit. The author's own workflow hints at where this lands: save results to a table on a schedule so dashboards stay current, then track MAPE and interval coverage, and consider widening the input window or switching to ARIMA_PLUS if accuracy slips. Whether this replaces existing forecasting setups will likely hinge on how much of a team's demand is short- and mid-term, since that is the range TimesFM is positioned for.