Ensemble forecasting with IBM Granite time-series models

Why it matters

IBM has introduced Granite Timeseries Ensemble, a permissively licensed recipe that merges four pre-trained Granite time-series models into a single forecaster, available through Hugging Face and loadable via the Granite TSFM library. The motivation is practical: zero-shot forecasting still forces users to pick one model, and no single architecture suits every kind of series. Rather than training a new model, the recipe stores a configuration and combines the members' quantile forecasts at runtime, returning both point and quantile predictions without retraining or fitted weights. Two combination methods are offered. The default Probabilistic Ensemble pools member forecasts with equal influence, drawing on forecasting research showing simple equal-weight combinations often hold up well. The adaptive IQR Ensemble instead weights members by forecast spread, though IBM cautions that narrower intervals do not guarantee better calibration. IBM evaluated both recipes across many GIFT-Eval configurations without fine-tuning, comparing against the constituent models. The design echoes ensemble practice in weather forecasting, where multiple models are merged for seasonal prediction.