Ensemble forecasting with IBM Granite time-series models
huggingface ·
IBM Research has introduced Granite Timeseries Ensemble, a tool that combines four distinct pre-trained time-series models to provide robust forecasting without requiring retraining. The ensemble integrates architectures including Patched Transformers, Conformer, State-space models, and Tiny Time Mixers to capture diverse data patterns. Users can choose between two aggregation methods: the recommended Probabilistic Ensemble, which uses uniform weighting, or the adaptive IQR Ensemble, which adjusts weights based on forecast spread. Evaluation on the GIFT-Eval benchmark demonstrates that both ensemble approaches outperform individual member models in aggregate point-forecast accuracy (MASE). While individual models like PatchTST-FM-r2 may achieve lower probabilistic error scores, the ensemble offers superior consistency and reduces the need for manual model selection. The solution is designed as an extensible interface, allowing developers to swap in compatible external models while maintaining a unified output format. Although it increases inference latency due to loading multiple checkpoints, the ensemble provides a practical default for applications where choosing the optimal single architecture is difficult.
huggingface