Google’s TimesFM-3 Adds Known Future Events to Zero-Shot Forecasting

Google’s new TimesFM-3 promises a more useful kind of zero-shot forecast: one that can model connected series and account for events already on the calendar. The company reports first-place benchmark rankings, but those results are not yet evidence of how the model will perform on a particular business’s data.

Released August 31, TimesFM-3 is a 330-million-parameter time-series foundation model pre-trained on more than 1 trillion real-world and synthetic time points. It supports zero-shot multivariate forecasting, meaning it can predict multiple related targets without task-specific fine-tuning.

That changes the scope of the TimesFM line. Google says releases through TimesFM-2.5 were limited to univariate forecasting, which uses one series’ own history. TimesFM-3 can instead take multiple targets, past-only inputs such as foot traffic, and dynamic inputs whose future values are known, such as promotions, holidays, or weather forecasts.

A forecast can see the calendar

Google’s illustration uses ice-cream sales. The model can learn the historical relationship between promotions and sales, then use a forthcoming promotion schedule when forecasting future days. Google’s chart depicts an approximately 20% sales lift on promotion days; it is an illustrative example, not a disclosed production result.

One pass replaces a sequential loop

The model groups data into patches of 32 time steps. It alternates attention over earlier points in the same series with attention across related series at the same point in time. The first is restricted to earlier data, preventing the model from using future values it would not have when making a real forecast.

Rather than generating each future patch in sequence, TimesFM-3 uses contiguous patch masking to fill the full horizon in one forward pass. It returns nine quantile forecasts, from the 10th through 90th percentiles, for every target at every forecast step. Those ranges provide several possible outcomes instead of a single estimate.

A public release, and a company-run scorecard

Google reports that TimesFM-3 ranked first among the pre-trained foundation models it evaluated on Gift-Eval, FEV-Bench, and Time, for both point forecasts and probabilistic forecasts. Its comparisons included Chronos-2, the Toto 2.0 family, and TimesFM-2.5; Google also says multivariate mode improved results when cross-series information and covariates were available.

TimesFM-3 is available on GitHub and Hugging Face. Google says a BigQuery integration is planned for the coming weeks; until then, it directs BigQuery users with univariate tasks to TimesFM-2.5.