IBM Granite Models Enable Time-Series Ensemble Forecasting

IBM has introduced Granite Timeseries Ensemble, a forecasting recipe that combines four permissively licensed pre-trained Granite time-series models into one extensible interface, reducing the need for developers to select a single architecture for each dataset. The recipe integrates PatchTST-FM-r1, a patched Transformer; PatchTST-FM-r2, a patched Conformer; FlowState-r1.1, a state-space model; and TTM-r3, a Tiny Time Mixers model. The configurations are packaged as an ensemble recipe rather than merged weights, with member checkpoints downloaded at runtime. The forecasters return point predictions and quantile outputs, enabling median forecasts and probabilistic ranges such as an 80 percent central interval. The default combination is a Probabilistic Ensemble that gives each member equal influence and pools their quantile forecasts without fitted weights. IBM also offers an adaptive Inter Quartile Range Ensemble that assigns greater influence to members with narrower predicted intervals, while using isotonic regression to preserve quantile order. The common interface can accommodate compatible external forecasters, allowing organizations to adapt the member mix to domain needs without retraining individual models. In GIFT-Eval evaluations across 97 dataset configurations, using results captured on October 20, 2026, both ensemble recipes outperformed the four constituent Granite models on aggregate mean absolute scaled error and average MASE rank. The Probabilistic Ensemble achieved the lowest MASE among the compared forecasts, while the IQR Ensemble achieved the lowest average probabilistic rank. These outcomes suggest the ensemble can provide robust starting performance when the best single time-series architecture is not known in advance, although it does not rank first across all reference systems. IBM notes that aggregate benchmark improvements do not guarantee benefits for every series. The current recipe forecasts multivariate channels independently, increases inference latency and memory compared with a single model, and does not ensure calibration under distribution shift. The company recommends backtesting both aggregation methods on representative data, monitoring deployed forecasts, and reviewing licensing terms for expanded recipes that include additional IBM Research checkpoints.