IBM details Granite ensemble’s small forecasting gains and runtime trade-offs

In an October 22 post on Hugging Face, IBM Research describes IBM Granite Timeseries Ensemble as a four-model recipe producing median and quantile forecasts without retraining or fitted ensemble weights, potentially reducing model-selection work. Source

IBM Research reports aggregate MASE of 0.683 for the default ensemble versus 0.685 for its strongest individual member across 97 GIFT-Eval configurations—a small point-forecast improvement, while PatchTST-FM-r2 retains the best aggregate probabilistic error among the compared forecasts. Source

IBM Research cautions that four checkpoints increase inference time and memory, channels are forecast independently, and adaptive IQR weighting can favor overconfident models. Its findings leave application-specific gains, empirical interval coverage and runtime costs unresolved; neither recipe guarantees calibration or robustness under distribution shift. A key limitation: these vendor-reported results do not establish independent confirmation or universal superiority. Source