What happened
- On October 22, 2026, IBM Research published Granite Timeseries Ensemble on the Hugging Face blog, a recipe that combines four Granite time-series models with different architectures.
- It returns a point forecast and quantiles, by default from 0.1 to 0.9, with no retraining and no fitted weights. The 0.1 and 0.9 quantiles give a central interval with nominal 80% coverage.
- The recommended option, Probabilistic Ensemble, gives each model equal weight and pools their quantiles in probability space. The alternative, IQR Ensemble, gives more weight to models with narrower spread.
- The recipe installs with granite-tsfm 0.3.11 or later and downloads the four checkpoints at run time.
Why it matters
- Choosing a pretrained forecasting model is hard: one that does well on smooth industrial signals can fail on noisy demand. The recipe replaces that choice with a default.
- Fixed weights avoid a fitting step and its hyperparameters, which helps anyone with many series and little time.
- IBM invites users to measure with their own data whether the ensemble improves on its members before adopting it.
The number
80%. Nominal coverage of the central interval between the 0.1 and 0.9 forecast quantiles.
Context
IBM compares the idea with meteorology: ensemble systems explore a range of outcomes and combine different models, as in North American seasonal prediction. It also cites a 50-year review of forecast combination and the so-called forecast combination puzzle, under which simple averages often hold up against methods with estimated weights. The members include, according to the text’s references, models from the PatchTST-FM, FlowState and Tiny Time Mixers families. The recipe and its default members carry a permissive license.
What’s next
- The text proposes extending the same interface to other compatible forecasters.
- It recommends evaluating the recipe on each application’s data, with a dedicated section.
Bottom line
There is related work in weather forecasting: a machine learning method doubles the skill of ECMWF’s AI at weeks-ahead range. Running open models depends on chips and libraries: AWS launched TorchNeuron for Trainium and Google opened TorchTPU. And for a small team, Recursive published Flint, a coding model that runs on a notebook.
Sources
- Hugging Face blog — Ensemble forecasting with IBM Granite time-series models (October 22, 2026; IBM Research article)
- Cover image: Chris Hunkeler from Carlsbad, California, USA, CC BY-SA 2.0, via Wikimedia Commons
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
.jpg){kind=link}