Google TimesFM-3 Adds Zero-Shot Multivariate Forecasts
The 330-million-parameter model forecasts related series and external variables without task-specific fine-tuning.
Forecasting connected signals together
Google Research has released TimesFM-3, a 330-million-parameter foundation model designed to forecast multiple related time series and incorporate external variables without task-specific fine-tuning. Model artifacts are available through GitHub and Hugging Face, making the release usable outside Google’s own services.
Earlier TimesFM versions treated each series independently. TimesFM-3 is instead pretrained for multivariate forecasting: it can predict several targets together, use covariates available only in historical data, and account for variables whose future values are already known, such as planned promotions, holidays or weather forecasts. Google trained it on a mixture of real and synthetic data containing more than one trillion time points.
The distinction matters in operational settings. Retail sales depend on related products, promotions and foot traffic; energy demand moves with weather; industrial equipment signals interact rather than failing in isolation. Google’s example shows the model using a future promotion schedule to anticipate an approximately 20% sales increase on promotion days, something an otherwise comparable univariate forecast cannot infer from sales history alone.
Google reports that TimesFM-3 achieved the best average rank among evaluated pretrained models on point and probabilistic forecasting across GIFT-Eval, FEV-Bench and the TIME benchmark. The comparison included Chronos-2, Toto 2.0 and TimesFM-2.5. These are research results reported by the model’s developer, and performance on company-specific data may differ.
Why it matters
Time-series forecasting is less visible than chatbots but directly affects inventory, staffing, energy procurement, capacity planning and anomaly detection. Supporting multiple targets and known future variables in one forward pass removes an important limitation of general-purpose forecasting models and may reduce the amount of specialist modeling needed for each deployment.
The larger opportunity is operational rather than conversational: a compact, reusable model could make sophisticated forecasting accessible to organizations that cannot maintain separate pipelines for every dataset. Adoption will depend on how reliably its zero-shot advantages survive messy, shifting real-world data.