IBM releases an open forecasting recipe that pools four Granite time-series models

IBM Research released Granite Timeseries Ensemble, which combines four of its pre-trained time-series models into one forecast without any training on your data. You give it past values, such as daily sales or energy use, and it returns a central forecast plus a range, by default from the 10th to the 90th percentile. On the GIFT-Eval benchmark of 97 dataset setups, the default ensemble was slightly more accurate than its best single member (0.683 vs 0.685 on MASE, lower is better), and IBM says the main gain is reliability across many kinds of data rather than topping the leaderboard. The recipe and its four default models are permissively licensed and install with pip install granite-tsfm; running four models is slower and uses more memory than one, and each series is forecast on its own.