TimesFM-3 enables the simultaneous forecasting of multiple time series

Google’s new model combines historical data and variables known in advance to forecast several related indicators, without the need for task-specific adjustments.

Google announced TimesFM-3 on August 31. This 330 million parameter model can predict multiple series that evolve together and incorporate information like a promotional schedule. It produces the entire forecast horizon in one pass, with estimates of uncertainty.

The weights and code are available on Hugging Face and GitHub. Google reports favorable results on three public benchmarks, under the conditions described in its publication. Integration with BigQuery is announced for the following weeks. The practical interest is to exploit the relationships between indicators rather than predicting them separately.