IBM Research ships four-model Granite forecasting ensemble

IBM Research has published Granite Timeseries Ensemble on Hugging Face, a recipe that runs four pre-trained Granite time-series models and merges their forecasts. It outputs point and quantile forecasts (0.1–0.9); on 97 GIFT-Eval configurations its MASE was 0.683–0.684 versus 0.685 for the best single member.

• Ensemble combines PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1 and TTM-r3
• MASE is 0.683–0.684 versus 0.685 for the best single model
• Requires granite-tsfm version 0.3.11 or later
• Output includes 0.1–0.9 quantiles with nominal 80% coverage