Google Releases TimesFM-3, a Zero-Shot Multivariate Forecasting AI
TimesFM-3 extends the TimesFM line from single-series to multivariate forecasting, letting one zero-shot model weigh related signals like weather and campaign schedules against each other, while a non-commercial license keeps it research-only for now.
Reporting from 1 source: GIGAZINE.
Google Research has released TimesFM-3, a time-series forecasting model that reads multiple changing datasets, such as sales, foot traffic, and weather, together and predicts future values. Pre-trained on over one trillion data points, it works without additional training. The model adds multivariate forecasting to the TimesFM line, which previously handled single series. Google reports top average rankings in point and probabilistic prediction across three public benchmarks. The model is available on GitHub and Hugging Face under a non-commercial license.
In the published example, TimesFM-3 reads past ice cream sales, foot traffic, weather forecasts, and campaign schedules together, and it predicted sales would rise about 20 percent on future campaign days by learning the pattern of past campaigns. The model carries about 330 million parameters and was pre-trained on a mix of real-world and synthetic data exceeding one trillion points.
It also predicts the entire forecast period in a single pass rather than step by step, which Google says cuts processing time and limits the error that accumulates when predictions are repeated. Source code is Apache License 2.0, but the pre-trained weights carry a separate non-commercial license, so commercial and production use are not permitted. Google plans to integrate TimesFM-3 into BigQuery within a few weeks.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.