TimesFM-3

Updated

TimesFM-3 is a third-generation time-series foundation model developed by Google Research, released in September 2026. It is a 330-million-parameter, decoder-only transformer model designed for zero-shot forecasting.

Key Features and Improvements

  • Native Multivariate Forecasting: Unlike previous versions of TimesFM, which were limited to univariate forecasting (a single series based on its own history), TimesFM-3 is natively pre-trained for multivariate tasks. It can jointly predict multiple coevolving time series and capture cross-series dependencies.
  • Input Support: The model natively supports three types of inputs without requiring task-specific fine-tuning:

* Multiple Targets: Jointly forecasting related time series.

* Past Covariates: Features known only historically (e.g., past foot traffic).

* Past-Future (Dynamic) Covariates: Known future signals (e.g., scheduled promotions, holidays, or weather forecasts).

  • Contiguous Patch Masking: To improve efficiency and reduce latency, the model uses a "Contiguous Patch Masking" strategy to generate the entire forecasting horizon in a single forward pass, rather than generating forecasts one patch at a time.
  • Performance: It was pre-trained on more than 1 trillion time points of real-world and synthetic data. Evaluations on benchmarks like Gift-Eval, FEV-Bench, and Time show it is a top-ranked model for both point and probabilistic forecasting.

Licensing Note

While TimesFM 2.5 remains available under the Apache-2.0 license for commercial and production use, TimesFM-3 is released under a non-commercial license (TimesFM Non-Commercial License v1.0).