IBM Research released an open forecasting recipe combining four pre-trained Granite time-series models into a unified interface, Hugging Face reported on Oct. 22, 2026. The approach generates point and quantile predictions without fine-tuning weights on target datasets.
The default recipe incorporates four distinct model architectures: PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1, and Tiny Time Mixers (TTM-r3). By default, the system produces point forecasts alongside quantile outputs spanning from 0.1 to 0.9, defining a central forecast interval with 80% nominal coverage. Rather than storing merged model weights, the package saves configuration data and downloads the constituent checkpoints at runtime.
Aggregation relies on one of two methods. The default Probabilistic Ensemble pools member quantile predictions in probability space with equal weights, eliminating hyperparameter tuning. The alternative Inter Quartile Range (IQR) Ensemble applies adaptive weighting based on predicted forecast spread, allocating higher influence to narrower intervals. The IQR recipe includes a temperature setting to shift weights toward uniformity, an optional cap on individual member weights, and isotonic regression to prevent quantiles from crossing.
Benchmark evaluations across 97 GIFT-Eval dataset configurations showed that both ensemble recipes outperformed all four individual models in aggregate point-forecast error. On test results captured on Oct. 20, 2026, the Probabilistic Ensemble achieved a mean absolute scaled error (MASE) of 0.683, compared to 0.685 for PatchTST-FM-r2 and 0.684 for the IQR Ensemble. PatchTST-FM-r2 recorded the lowest continuous ranked probability score (CRPS) at 0.467, while both ensemble variations scored 0.471. The ensemble models placed third on the GIFT-Eval leaderboard among eligible reference systems for both MASE Rank and CRPS Rank.
Runtime deployment requires version 0.3.11 or higher of the granite-tsfm library. The current framework processes multivariate channels independently, omitting cross-channel relationships, and increases latency and memory usage by loading four distinct model checkpoints simultaneously.
