IBM Research has published Granite Timeseries Ensemble on Hugging Face. It is a recipe that runs four pre-trained Granite time-series models together and merges their forecasts into one. Pre-trained forecasters can predict without training on your data, but you still have to choose which one. The ensemble is meant to remove that choice.
What is in it
The default recipe combines four models with different architectures: 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 the ensemble configuration, not merged weights. It downloads the four member checkpoints at runtime, so the first run takes longer. Installation needs granite-tsfm version 0.3.11 or later.
The output is a point forecast plus quantile forecasts, by default from 0.1 to 0.9. The 0.1 and 0.9 quantiles give a central interval with nominal 80% coverage. IBM says the recipe and its default Granite members are permissively licensed.
Two ways to combine
The recommended default is the Probabilistic Ensemble. It gives each member equal influence and pools their quantile forecasts, so there is no fitting step and no weights to tune. The alternative is the IQR Ensemble, which gives more weight to members with narrower forecast spread. IBM warns that a narrower interval is not necessarily more accurate, and that this weighting can favour an overconfident member.
What the benchmark shows
IBM evaluated both recipes on 97 GIFT-Eval dataset configurations without fine-tuning, using leaderboard scores captured on 20 October 2026. The results are mixed.
- On MASE, the Probabilistic Ensemble scored 0.683 and the IQR Ensemble 0.684. The best single member, PatchTST-FM-r2, scored 0.685.
- On CRPS/MWQL, PatchTST-FM-r2 kept the lowest score (0.467, against 0.471 for both ensembles).
- On average rank, the IQR Ensemble was best of the six forecasts shown on CRPS, while the Probabilistic Ensemble led on MASE.
- Among eligible systems, the ensemble approach ranks 3rd on both MASE Rank and CRPS Rank. IBM notes that some excluded systems, such as fine-tuned or non-permissively licensed ones, score better.
IBM itself calls the margin over the best single member small. The claimed benefit is robustness: comparable or better aggregate accuracy without knowing in advance which architecture suits your data.
Limits
The current recipe forecasts multivariate channels independently and does not use cross-channel information. Loading four checkpoints increases inference time and memory compared with one model. Members differ in context limits, horizons, preprocessing and missing-value handling. Neither aggregation method guarantees calibration under distribution shift, so forecasts need monitoring after deployment.
The common Forecaster interface also accepts compatible outside models, so the member mix or aggregation method can change while the output format stays the same. The repository also has expanded recipes that add other IBM Research checkpoints, but these carry separate terms, including non-commercial and research-use restrictions. Review each model card before using them.
How IBM suggests testing it
IBM recommends backtesting on historical windows that match your horizon and series. Compare the Probabilistic Ensemble, the IQR Ensemble and a strong single member on the same windows. Check quantile loss alongside MASE, and compare the empirical coverage of the 0.1 to 0.9 interval with its nominal 80%. Record runtime and memory too.
Top comments (0)