IBM releases Granite Timeseries Ensemble, combining four open forecasting models

IBM Research has released Granite Timeseries Ensemble, a permissively licensed recipe that combines four pre-trained Granite architectures—Patched Transformer, Patched Conformer, state-space, and MLP mixer models—into a unified forecasting interface. The ensemble returns median and quantile forecasts with no retraining required. Users can choose between a Probabilistic Ensemble with equal fixed weights or an adaptive IQR Ensemble that weights models by forecast spread. The recipe is available through the granite-tsfm Python package and downloads member checkpoints from Hugging Face at runtime.

How the ensemble combines forecasts

• Four architectures: PatchTST-FM-r1 (Patched Transformer), PatchTST-FM-r2 (Patched Conformer), FlowState-r1.1 (state-space model), and TTM-r3 (Tiny Time Mixers/MLP mixer)
• Probabilistic Ensemble (default): uniform probability pooling with equal fixed weights, no fitting step or hyperparameters
• IQR Ensemble: adaptive weighting based on each member's forecast spread, giving narrower intervals greater influence
• Recipe stores configuration, not merged weights: member checkpoints download at runtime from Hugging Face
• Outputs: point forecast via prediction_outputs and quantile predictions (0.1 to 0.9 by default) via quantile_outputs
• Requires: granite-tsfm>=0.3.11, context and prediction lengths are configurable parameters

Why it matters

The ensemble approach reduces the operational burden of model selection for time-series forecasting, which has historically required evaluating multiple models per dataset. By providing a permissively licensed default that works out-of-the-box, IBM lowers the barrier for practitioners who need robust forecasts without deep model-tuning expertise.