Google Releases TimesFM-3, a 330M-Parameter Model That Forecasts Sales Using Weather and Discount Data

Google Research has released TimesFM-3, a time series forecasting model that predicts future outcomes—like retail sales—by combining historical data with related variables such as weather forecasts and planned discounts. The model is available now on GitHub and Hugging Face, with Google planning to add it to BigQuery in the coming weeks.

What the model does

TimesFM-3 has 330 million parameters and was trained on more than one trillion real and synthetic time series data points, according to Google. Like its predecessors, it works zero-shot, requiring no additional training to handle a new forecasting task.

Google illustrates the model's approach with an ice cream sales example: a useful forecast shouldn't just extrapolate from past sales, but also account for related products (waffle cones, syrup), historical foot traffic, weather, discount campaigns, and holidays. TimesFM-3 handles three categories of supplementary data—forecasting multiple related variables simultaneously, incorporating factors known only historically, and using known future events like scheduled discounts or weather forecasts.

Rather than producing a single point estimate, the model outputs nine values per time step to represent a range of uncertainty around each prediction.

Architecture changes

TimesFM-3 is built on a Transformer architecture, consistent with earlier versions in the family. It groups 32 consecutive data points into a single patch and normalizes each series to a common scale, allowing it to compare measurements of different magnitudes directly. The model processes data along two axes: within a single time series (looking only at past values, to prevent information leakage) and across series at a given point in time, to learn relationships—such as how a discount on one product affects sales of a related item.

Google says earlier TimesFM versions predicted the future one block at a time, an approach it describes as slow, compute-heavy, and prone to compounding errors as each prediction built on the last. TimesFM-3 instead marks all future time steps as blanks and fills them in a single pass, a method Google calls one-shot forecasting.

Benchmark claims

According to Google, TimesFM-3 ranks first among pretrained forecasting models on three benchmarks—Gift-Eval, FEV-Bench, and Time—in both point accuracy and uncertainty calibration. Competing models include Amazon's Chronos-2, the Toto-2.0 family, and Google's own prior release, TimesFM-2.5. Google claims that even when restricted to single-variable forecasting, TimesFM-3 matches or beats competitors, and the performance gap widens when multivariate data is added. On FEV-Bench, Google says Chronos-2 comes close to TimesFM-3's univariate mode but trails the full multivariate version significantly.

These are Google's own reported results; independent verification of the benchmark scores has not been published.

Availability

TimesFM-3 is available now on GitHub and Hugging Face. Google says it will integrate the model into BigQuery in the coming weeks; the AI.FORECAST command there currently runs TimesFM-2.5, which handles only single-variable forecasting. TimesFM-2.5 was released in September 2025, and Google says all versions through 2.5 were limited to processing one data series at a time—making TimesFM-3's multivariate support the primary architectural shift in this release.

Google says the TimesFM family, launched in 2024, has already seen deployment in retail, finance, manufacturing, healthcare, and scientific research.

What this means

TimesFM-3's real advance is architectural, not just accuracy: moving from single-variable, step-by-step prediction to one-shot multivariate forecasting addresses a known weakness in autoregressive time series models—error compounding. Whether the benchmark leads hold up under independent testing remains to be seen, since Gift-Eval, FEV-Bench, and Time results here come solely from Google. For teams doing operational forecasting—retail demand planning, capacity planning, financial forecasting—the ability to feed in known future events like discount calendars without retraining is the practical selling point. The BigQuery integration, once live, will matter more for adoption than the open-weights release, since it puts the model directly in front of enterprise data teams already running SQL-based analytics pipelines.