IBM Research published the Granite Timeseries Ensemble on 22 October 2026: a recipe that combines four pre-trained Granite models for time-series forecasting and returns, in a single call, a central forecast and quantile intervals — with no retraining and no fitted ensemble weights.
The practical problem is model choice. Pre-trained time-series models can forecast without training on the user's data, but users still have to pick one. A model that handles smooth industrial signals well may struggle with noisy retail demand or long seasonal horizons. The ensemble turns that choice into a sensible default, in the same spirit as weather forecasting ensembles, where several models offer different views of what comes next.
The four models in the recipe
The default recipe combines four complementary architectures, which capture trends, seasonality and local variation in different ways, so their errors need not coincide: Granite PatchTST-FM-r1 (patched Transformer), PatchTST-FM-r2 (patched Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (Tiny Time Mixers, an MLP mixer). The recipe stores only the ensemble configuration, not merged weights; the four checkpoints are downloaded on the first run. It requires granite-tsfm version 0.3.11 or later and the Hugging Face recipe ibm-granite/granite-timeseries-ensemble-r1.
Two ways to combine forecasts
The recommended default is the Probabilistic Ensemble: every member has equal influence, quantiles are pooled in probability space, and there is no fitting step or weight hyperparameter for a new series. The alternative IQR Ensemble is adaptive: it gives more influence to members with narrower intervals, with weights varying by sample, target and horizon step. IBM itself warns that a narrower interval is not necessarily more accurate and that this weighting can favour an overconfident member.
What the numbers say — and what they do not
On 97 GIFT-Eval dataset configurations, without fine-tuning, the Probabilistic Ensemble recorded an aggregate MASE of 0.683, against 0.685 for the best single member, PatchTST-FM-r2 — a small margin. IBM's honest reading is that the main benefit is robustness: comparable or better aggregate accuracy without knowing in advance which architecture will suit each series. On aggregate CRPS/MWQL, PatchTST-FM-r2 kept the best value (0.467), while the IQR Ensemble had the best average CRPS rank (47.619) among the six forecasts compared. According to IBM, in the 20 October 2026 comparison the ensemble ranked 3rd among eligible systems on MASE Rank and CRPS Rank.
The stated limits matter: multivariate channels are forecast independently, without cross-channel information; loading four models costs more memory and inference time than one; and neither aggregation method guarantees calibration or robustness under distribution shift. Expanded recipes adding other IBM Research checkpoints may carry different licences, including non-commercial or research-only restrictions, so each model card should be checked before use. IBM recommends backtesting on representative data and measuring the empirical coverage of the 0.1-to-0.9 interval, whose nominal coverage is 80%.
For demand, telemetry, equipment monitoring or financial signals, the value is not a new giant model, but an open, permissively licensed default that removes the upfront architecture choice. The real test, as IBM itself acknowledges, is backtesting on the user's own data — not the benchmark average.
Sources: IBM Research's original publication on Hugging Face, “Ensemble forecasting with IBM Granite time-series models” (https://huggingface.co/blog/ibm-research/granite-timeseries-ensemble-r1).
Image credit: GIFT-Eval ranking chart published by IBM Research on the Hugging Face blog, accompanying the Granite Timeseries Ensemble announcement.
Transparency: this article was prepared with the assistance of artificial intelligence, based on public sources on the internet, including IBM Research's original publication. Readers should consult the original sources to confirm details before any technical decision.
Muse
