Ensemble forecasting with IBM Granite time-series models
IBM's Granite Timeseries Ensemble combines four pre-trained Granite forecasters of different architectures to output point and quantile forecasts without retraining. On GIFT-Eval, the Probabilistic Ensemble had the lowest MASE of six forecasts compared, by a small margin.
What the article says
IBM Research published Granite Timeseries Ensemble on the Hugging Face blog on October 22, 2026: a forecasting recipe that combines four pre-trained Granite time-series models (PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1 and TTM-r3). It avoids having to pick a single model for a given series and needs neither retraining nor fitted ensemble weights.
The recipe returns a median forecast and quantile forecasts, 0.1 to 0.9 by default. Two aggregation methods exist: probabilistic (recommended) and IQR, which weights members by forecast spread. On GIFT-Eval (97 configurations), both ensembles beat each of the four members on aggregate MASE. The authors note limits: multivariate channels are forecast independently, inference time and memory increase, and calibration is not guaranteed. They recommend backtesting on representative data before use.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Hugging Face. The content and rights belong to the publisher: please read the original article.