Google Releases TimesFM-3 to Predict Complex Sales Trends

Google Research has introduced TimesFM-3, a new zero-shot forecasting model designed to predict future trends by analyzing multiple related variables at once. Unlike its predecessor, TimesFM-2.5, which was released in September 2025 and could only handle single-variable forecasting, the new 330-million-parameter model supports multivariate data. Built on a Transformer architecture, TimesFM-3 groups 32 consecutive data points into a single patch and normalizes each series to a common scale. It processes data in two directions: along the time axis to find patterns within a single series, and across different series to learn how variables interact at any given moment.

The model was trained on more than one trillion real and synthetic data points. To avoid the compounding errors and high computational costs of older step-by-step forecasting methods, TimesFM-3 uses a one-shot approach that marks future time steps as blanks and fills them in a single pass. It can process three types of supplementary data, including multiple related target variables, historical-only factors like past foot traffic, and known future events like weather forecasts or discount schedules. For example, when given a promotion schedule, the model can anticipate a 20 percent boost in sales on discount days. To capture uncertainty, it outputs nine distinct values for each time step.

In evaluation, TimesFM-3 ranked first among pretrained forecasting models on the Gift-Eval, FEV-Bench, and Time benchmarks, leading in both point accuracy and uncertainty calibration. It outperformed rivals such as Amazon's Chronos-2, the Toto-2.0 family, and Google's own TimesFM-2.5. For practitioners, this means a significant reduction in the custom training typically required for complex business forecasting. The model is currently available on GitHub and Hugging Face, and Google plans to integrate it into BigQuery in the coming weeks, where TimesFM-2.5 is currently used via the AI.FORECAST command.