Ensemble forecasting with IBM Granite time-series models
Web page · https://huggingface.co/blog/ibm-research/granite-timeseries-ensemble-r1
First seen by Oriel on .
AI summary
Granite Timeseries Ensemble combines four pre-trained Granite forecasters into point and quantile forecasts, offering a no-retraining default that performs competitively across 97 GIFT-Eval configurations.
- The recommended Probabilistic Ensemble pools member forecasts with equal influence and fixed weights; the adaptive IQR Ensemble weights narrower predicted intervals more heavily without fitting to historical errors.
- On aggregate MASE, the Probabilistic Ensemble scores 0.683 versus 0.685 for the strongest individual member, PatchTST-FM-r2; PatchTST-FM-r2 retains the best aggregate CRPS/MWQL score.
- The recipe loads four checkpoints at runtime, increasing inference time and memory compared with a single model; forecasts treat multivariate channels independently and do not use cross-channel information.
- The article recommends backtesting both recipes and a strong individual model on representative data, checking quantile loss, empirical interval coverage, runtime, and memory.
#time series forecasting #ensemble learning #granite tsfm #probabilistic forecasting
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/ibm-research/granite-timeseries-ensemble-r1 | https://huggingface.co/blog/ibm-research/granite-timeseries-ensemble-r1 Read discussion |