AI BriefingIBM Unveils Granite Timeseries Ensemble R1, Combining Four Time Series Models
·2026.10.28 18:30
Key point
IBM has unveiled Granite Timeseries Ensemble R1, which combines four time series models.
Details
- IBM Research has unveiled Granite Timeseries Ensemble R1, which combines the quantile predictions of four pre-trained time series foundation models into a single distribution.
- On the GIFT-Eval benchmark (97 settings, Zero-shot), it achieved an average MASE rank of 47.175, outperforming all four constituent models.
- It can be used after installation via
pip install "granite-tsfm>=0.3.11"without additional training; the Hugging Face repository contains only the recipe. - The probabilistic ensemble (default) pools 9 quantiles with equal weights, while the IQR ensemble assigns weights to models with narrower prediction intervals.
- The IQR ensemble is sensitive to data units, so verifying normalization is essential; the default 4-model recipe is recommended for commercial services.
- It runs in environments with Python 3.11 or higher and PyTorch 2.10 or higher, with varying context lengths and preprocessing methods per model.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.