IBM releases open Granite Timeseries Ensemble on Hugging Face for zero-shot forecasting

IBM Research released Granite Timeseries Ensemble on Hugging Face, a zero-shot forecasting tool that combines four pre-trained Granite time-series models into a single ensemble interface. The ensemble requires no retraining on user data and avoids fitted ensemble weights, relying instead on fixed or adaptive combination rules. The default Probabilistic Ensemble uses uniform probability pooling to merge quantile forecasts from all members, while the IQR Ensemble weights members by their forecast spread, giving narrower intervals greater influence. Both recipes are permissively licensed and expose the same interface, allowing users to switch between methods by changing a subfolder in the model identifier.

The ensemble’s four member models include Granite PatchTST-FM-r1 and PatchTST-FM-r2 (patched Transformer and Conformer variants), Granite FlowState-r1.1 (a state-space model), and Granite TTM-r3 (a Tiny Time Mixers MLP mixer). These architectures capture trends, seasonality, and local variation differently, so their errors need not coincide, reducing the risk of selecting a single poorly suited model.

To use the ensemble, users install the granite-tsfm package and load the default recipe directly from Hugging Face. The first run downloads the four member checkpoints at runtime, which may require extra time. The ensemble returns a point forecast and quantile predictions (defaulting to 0.1–0.9 quantiles), enabling applications to express forecast ranges alongside a central prediction.

IBM evaluated both ensemble recipes on 97 GIFT-Eval dataset configurations without fine-tuning, using published leaderboard scores captured on 20 October 2026. On point forecasts (MASE), the Probabilistic Ensemble ranked 47.175 and the IQR Ensemble ranked 48.773, compared to the best constituent model’s rank of 51.371. For probabilistic forecasts (CRPS/MWQL), both ensemble recipes ranked similarly (50.948 and 47.619), outperforming the best constituent model’s rank of 48.918. Lower scores and ranks indicate better performance.