IBM releases Granite Timeseries Ensemble for forecasting without retraining
IBM Research has released Granite Timeseries Ensemble, a default recipe that combines four open Granite forecasting models and returns median and quantile forecasts with no retraining or fitted weights. The recipe's default members are permissively licensed, according to the Hugging Face post. On the GIFT-Eval benchmark, the Probabilistic Ensemble scores 0.683 on MASE, slightly better than the best single member at 0.685, while PatchTST-FM-r2 keeps the lowest CRPS/MWQL score. The recipe forecasts each multivariate channel independently and loads four checkpoints, which adds inference time and memory.