TimesFM

Forecasts regularly sampled time series without training a new model for each dataset.

Field
Cross-domain
Method
Foundation model
Data it takes
time-series
License
Apache-2.0

TimesFM forecasts a time series it has not been trained on. It is a decoder-only model handling context windows up to 16,000 points, and can return prediction intervals rather than a single trajectory.

No training run or per-series fitting is required. It suits streamflow and climate reanalysis series, and gap-filling in eddy covariance and soil moisture records. Where sufficient history exists to fit a dedicated model, a trained approach such as NeuralHydrology generally performs better.

Upstream project →

Catalog entry last checked 2026-08-17. All models →

Run TimesFM on your data

The lab runs this on its own compute, in a container, with the inputs and parameters recorded alongside the result. Initial scoping conversations are free.

Contact the Lab