IBM Research releases Granite Timeseries Ensemble for zero-fine-tuning probabilistic forecasting

IBM Research released Granite Timeseries Ensemble, a recipe combining four pre-trained models (Granite PatchTST-FM-r1, Granite PatchTST-FM-r2, Granite FlowState-r1.1, and Granite TTM-r3) to provide median and quantile forecasts without retraining.
On the GIFT-Eval benchmark (October 20, 2026 snapshot), the Probabilistic Ensemble achieved a MASE of 0.683 and a MASE Rank of 47.175, outperforming the best single model, Granite PatchTST-FM-r2, which recorded a MASE of 0.685.
The IQR Ensemble variant achieved the lowest average CRPS Rank of 47.619 among the six evaluated systems, utilizing adaptive weighting based on prediction spread rather than fixed uniform probabilities.
The system supports a Forecaster interface for adding external compatible models and is available via Hugging Face and the granite-tsfm GitHub repository, though it requires loading four checkpoints which increases inference time and memory usage compared to single models.

Aggregation Methods

Probabilistic Ensemble uses uniform probability pooling in the probability space, avoiding the need for fitting or hyperparameter tuning.
IQR Ensemble assigns weights based on the interquartile range of predictions, giving more influence to members with narrower prediction intervals, and uses isotonic regression to prevent quantile crossing.

The recipe forecasts multivariate channels independently and does not utilize cross-channel information.
Users are advised to backtest on representative data and compare the ensemble against strong individual members like Granite PatchTST-FM-r2 to verify calibration and runtime costs.