timesfm is a pretrained decoder-only foundation model for time-series forecasting. It accepts time-series data and produces time-series forecasts.
Project overview
As a decoder-only foundation model targeting time-series forecasting, timesfm applies a pretrained architecture to a problem domain where such foundational approaches remain relatively uncommon.
Deployment
Refer to project documentation
License
Apache-2.0
Best for
Data teams and AI engineers who need a pretrained foundation model for time-series forecasting.
Key capabilities
Provides a decoder-only foundation model for time-series forecasting.
Includes covariate support through XReg.
Offers a fine-tuning example using HuggingFace Transformers with PEFT (LoRA).
Provides a Flax version of the model for faster inference.
Limitations and risks
This open-source version of timesfm is not an officially supported Google product.
Getting started
timesfm is available via pip install; the first success path is running pip install timesfm.