IBM combines four Granite time-series models in a single ensemble recipe

What is changing?

IBM Research has detailed Granite Timeseries Ensemble R1, a recipe that combines forecasts from four complementary Granite time-series models. The default configuration downloads PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1, and TTM-r3 from their separate repositories at runtime. The release therefore provides a reusable ensemble framework rather than a new set of merged weights.

The recommended default uses uniform probability pooling, giving each model’s quantile forecasts equal influence. An alternative IQR-weighted method gives more weight to members with narrower forecast intervals. IBM cautions that narrower intervals do not by themselves indicate better accuracy and says both methods should be evaluated on representative application data.

In the Granite-only evaluation, both ensembles outperformed the collection’s strongest individual member on aggregate MASE. Loading four checkpoints, however, increases latency and memory use, while neither method guarantees calibration or robustness under distribution shift.