Google Research has introduced TimesFM-3, a new AI model designed for time series forecasting. The model utilizes a Transformer architecture and features 330 million parameters. It was trained on a dataset comprising over one trillion real and synthetic data points.
The model introduces multivariate support, allowing it to process multiple related variables simultaneously, such as sales figures, weather data, and discount schedules. It replaces step-by-step forecasting with a one-shot approach, filling in all future time steps in a single pass.
TimesFM-3 outputs nine values per time step to account for prediction uncertainty. According to Google, the model outperformed competitors, including Amazon's Chronos-2 and the Toto-2.0 family, across the Gift-Eval, FEV-Bench, and Time benchmarks.
The model is currently available on GitHub and Hugging Face. Google plans to integrate TimesFM-3 into BigQuery in the coming weeks. Previous versions of the TimesFM family have been deployed in sectors including retail, finance, and healthcare since 2024.
