IBM's Granite Ensemble Picks the Best Forecast So You Don't Have To
Four open-weight time-series models, one interface, no retraining — and benchmark results that hold up across 97 dataset configurations.
By The ai-compute Desk
Choosing the right forecasting model is its own forecasting problem. A model that handles smooth industrial signals may collapse on noisy retail demand, and the best pick changes with the series and the horizon. IBM Research's Granite Timeseries Ensemble sidesteps that by combining four open Granite architectures into a single call that returns median and quantile forecasts — no retraining, no fitted weights. [1]
The logic is borrowed from meteorology. [2] Time-series forecasting benefits in a similar way: four Granite architectures — PatchTST-FM-r1, PatchTST-FM-r2, FlowState, and Tiny Time Mixers — capture trends, seasonality, and local variation in different ways, so their errors need not coincide. [1]
Two recipes are available. The Probabilistic Ensemble gives each member equal influence, pools their quantile forecasts in probability space, and is the recommended default — no hyperparameters to tune for a new series. The IQR Ensemble weights members by forecast spread, giving narrower intervals greater influence, though IBM cautions that a narrower interval does not necessarily mean a more accurate or better-calibrated forecast. [1]
Evaluated across 97 GIFT-Eval dataset configurations without fine-tuning, both ensembles outperform all four individual members on aggregate MASE. The Probabilistic Ensemble achieves a MASE of 0.683 against 0.685 for the best single member, PatchTST-FM-r2. The margin is thin; IBM frames the main benefit as robustness — comparable or better accuracy without knowing in advance which architecture suits your data. [1]
The appeal is architectural, not merely numerical. A single interface that degrades gracefully across unfamiliar series is more useful in production than a top-ranked model that requires careful selection each time the data changes.
The recipe and its default Granite members are permissively licensed and loadable directly from Hugging Face. Backtest on representative data before deploying. [1]
The main benefit is robustness: you get comparable or better aggregate accuracy without knowing in advance which architecture will suit your data.IBM Research