Ensemble forecasting with IBM Granite time-series models

IBM's Granite Timeseries Ensemble provides a method for forecasting by combining four complementary, pre-trained models without requiring retraining on a specific dataset. This approach merges different architectural views, including a Patched Transformer, Conformer, a state-space model, and an MLP mixer, to create a more robust prediction. The post details two ways to combine the forecasts: a default Probabilistic Ensemble and an adaptive Inter Quartile Range (IQR) Ensemble. The ensemble recipe and its member models are available on Hugging Face and can be implemented using the `granite-tsfm` Python library. This technique aims to reduce the guesswork in selecting the best single model for a given time-series task.