Skip to content
Neural Daily

ModelsNews

IBM Granite Timeseries Ensemble: four models, one forecast

IBM Research released an open recipe that combines four Granite time-series models. You no longer have to guess which architecture suits your data: the ensemble returns median and quantile forecasts with no retraining.

By: Mikola Máté Patrik · · 2 min read

Four colored forecast lines merging into one forecast with an uncertainty band, under the headline "Forecasting together"
Image: IBM Research · Press kit · Official blog post social image (og:image) · Source

What happened

IBM Research released Granite Timeseries Ensemble R1 (Hugging Face blog). It is not a new set of weights but a recipe: it merges the forecasts of four pre-trained Granite time-series models into a single median and quantile forecast. The recipe and its default members are Apache 2.0 licensed.

What it does

  • Four members, four architectures: PatchTST-FM-r1 (Transformer), PatchTST-FM-r2 (Conformer), FlowState-r1.1 (state-space) and TTM-r3 (Tiny Time Mixers).
  • Probabilistic Ensemble: every member has equal influence, with no fitted weights and no hyperparameters to tune. This is the default.
  • IQR Ensemble: members with narrower forecast intervals get more weight, varying per sample and horizon step. A temperature setting pushes weights toward uniform.
  • Output: a point forecast plus quantiles from 0.1 to 0.9 by default. The 0.1 to 0.9 interval has nominal 80% coverage.
  • Extensible: the shared Forecaster interface accepts compatible external models while keeping the same output format.

How much it helps

On 97 GIFT-Eval configurations, without fine-tuning, both ensembles beat every individual member on MASE. The Probabilistic Ensemble scores 0.683 MASE against 0.685 for the best single member, PatchTST-FM-r2. On the probabilistic CRPS metric, PatchTST-FM-r2 stays best at 0.467 versus 0.471 for the ensemble. IBM says that as of October 20, 2026 the approach ranks 3rd among eligible systems on both MASE and CRPS rank.

GIFT-Eval MASE and CRPS rank chart showing both Granite Ensemble R1 recipes among the top eligible reference models
IBM Research · Press kit · Benchmark figure from the official blog post · Source

Source: IBM Research, Hugging Face blog (GIFT-Eval, as of October 20, 2026)

Why it matters

The margin is small. The real gain is robustness: you don't have to pick the winning model for every new series in advance. The cost is running four models instead of one, so measure latency and memory. IBM itself recommends backtesting both recipes against a strong single member on representative data.

Try it

Install with pip install "granite-tsfm>=0.3.11" and load the recipe with QuantileEnsembleForecaster.from_pretrained("ibm-granite/granite-timeseries-ensemble-r1"), as described on the model card. The first run is slower because it downloads the four member checkpoints. Some members of the expanded recipes in the repo carry non-commercial licenses, so check their model cards.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles