IBM Research has released a forecasting recipe that combines four open Granite time-series models behind one interface. The goal is to give developers a reliable starting point when they do not know which architecture best fits their data.
The default Probabilistic Ensemble gives each member equal influence and pools their quantile forecasts without fitting weights on a new dataset. An alternative uses forecast spread to adjust each model’s influence, although IBM warns that narrower intervals can reward an overconfident model. Both return a central prediction and probability ranges.
Across 97 GIFT-Eval dataset configurations, the default ensemble recorded a mean absolute scaled error of 0.683, narrowly improving on the best individual member’s 0.685. Both ensembles beat all four members on aggregate point error and average point-forecast rank, but one individual model retained the best aggregate probabilistic score. The benefit is consistency rather than a large accuracy leap.
The recipe and default model checkpoints are permissively licensed and available through Hugging Face. Running four models increases memory use and latency, and the current setup forecasts multivariate channels independently. IBM recommends backtesting accuracy, calibration and cost on representative data before deployment.