IBM Granite time-series ensemble sets a four-model default
The default recipe returns median and quantile forecasts from four Granite models, without retraining or fitted ensemble weights.
| Granite Timeseries Ensemble | |
|---|---|
| members | four open Granite models |
| context_length | 1024 |
| prediction_length | 24 |
| quantile range | 0.1 to 0.9 |
| nominal coverage | 80% |
| license | permissively licensed |
| target column | OT |
| dataset | ETT-small ETTh1.csv |
A Hugging Face page lists the default members as Granite PatchTST-FM-r1, Granite PatchTST-FM-r2, Granite FlowState-r1.1, and Granite TTM-r3.
The example installs granite-tsfm>=0.3.11, reads ETT-small ETTh1.csv, uses target column OT, context_length 1024, and prediction_length 24.
The stated limit is nominal 80% coverage for the central interval formed by the 0.1 and 0.9 quantiles.
The page does not measure actual coverage, GIFT-Eval rank, or error against reference systems.
For a buyer, the recipe narrows model selection and gives a probabilistic interface, but it does not prove that the ensemble beats a single strong baseline on a named workload.
The user must verify coverage and rank before treating the default as a performance guarantee.
The packet leaves out a specific benchmark score, training-data composition, and exact license identifier.
A later correction can add measured coverage, GIFT-Eval rank, or reference-system error without changing the recipe.
The default is a usable recipe, not a verified performance claim.
After arxiv.org. We did not report this. The pictures, if any, are theirs.
