Superpower Daily
Google’s TimesFM-3 Adds Known Future Events to Zero-Shot Foreca...Microsoft Releases GigaPath-Flash, Reporting Roughly 50x Less C...Nvidia Puts $3.5B Into MediaTek to Keep Custom AI Chips in Its...Claims Adjusters’ AI Backlash Hits 98% Negative Reviews as Insu...John Ternus Takes Apple’s Top Job as AI and China Pressures Con...Replica Cyber Ships Isolated AI Agents With Runtime Controls fo...Google’s TimesFM-3 Adds Known Future Events to Zero-Shot Foreca...Microsoft Releases GigaPath-Flash, Reporting Roughly 50x Less C...Nvidia Puts $3.5B Into MediaTek to Keep Custom AI Chips in Its...Claims Adjusters’ AI Backlash Hits 98% Negative Reviews as Insu...John Ternus Takes Apple’s Top Job as AI and China Pressures Con...Replica Cyber Ships Isolated AI Agents With Runtime Controls fo...
Modelspublished

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

The 330-million-parameter model can combine related data streams with inputs such as promotion calendars and weather forecasts. Google reports leading results on three public benchmarks; performance on an organization’s own data remains the practical open question.

By 3 min read
Google’s TimesFM-3 Adds Known Future Events to Zero-Shot Forecasting
Google’s TimesFM-3 Adds Known Future Events to Zero-Shot Forecasting
Sign in to save and react

Listen to this story

The audio brief

About 1:31
0:001:31
Read transcript
Google’s TimesFM-3 is designed to forecast several related time series at once, while also using events already known to be coming—like promotions, holidays, and weather forecasts. That is the important shift from earlier TimesFM releases, which Google says were limited to univariate forecasts based on one series’ history. TimesFM-3 has 330 million parameters and was trained on more than one trillion real and synthetic time points. It groups data into patches of 32 time steps, then produces the entire forecast horizon in one pass rather than generating predictions sequentially. For every target and time step, it returns nine quantile forecasts, covering the 10th through 90th percentiles, so users get a range of possible outcomes instead of one number. Google reports first-place results against other pretrained models on three benchmarks: Gift-Eval, FEV-Bench, and Time. The comparisons included Chronos-2, Toto 2.0, and TimesFM-2.5. But this is a company-run scorecard, not proof that the model will work best on any particular organization’s data. Google’s promotion example shows roughly a 20 percent sales lift, though it is explicitly illustrative, not a production result. TimesFM-3 is available through GitHub and Hugging Face, with BigQuery integration planned. The practical question is whether its advantage survives contact with a company’s own data, calendars, and related signals.

Story brief

3 key points

Google has released TimesFM-3, a 330-million-parameter forecasting model trained on more than 1 trillion real and synthetic time points. Its main product change is support for multivariate, zero-shot predictions that incorporate future-known signals—such as promotions, holidays, and weather forecasts—while blocking unavailable future values. Google reports top scores against other pretrained models, but the...

  1. 01

    TimesFM-3 processes 32-step patches and generates the full forecast horizon in one forward pass.

  2. 02

    It returns nine quantile forecasts, spanning the 10th through 90th percentiles, for each target and time step.

  3. 03

    Google’s comparisons included Chronos-2, Toto 2.0, and TimesFM-2.5 across three named benchmarks.

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.

Chart showing a multivariate sales forecast responding to scheduled promotion days represented by highlighted covariate blocks.
Google illustrates a sales forecast that uses a planned promotion schedule as a known-future covariate. Source: research.google.

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.

Sources

  1. research.googleTimesFM-3: A zero-shot foundation model for multivariate forecasting

Related coverage

Continue reading