IBM Research has released Granite Timeseries Ensemble R1, a new open-source toolkit that combines four distinct Super Intelligence (SI) architectures into a single forecasting interface. The release aims to reduce the complexity developers face when selecting the best SI model for time-series prediction by offering a robust, pre-configured default.

What Happened

The toolkit integrates four pre-trained Granite models—PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1, and Tiny Time Mixers (TTM)-r3—into an extensible ensemble framework. Unlike traditional methods that require fitting ensemble weights on historical data, this approach uses two fixed aggregation strategies: the Probabilistic Ensemble, which pools quantile forecasts with equal influence, and the Inter Quartile Range (IQR) Ensemble, which adaptively weights members based on their forecast spread. The models are available via the `granite-tsfm` library and load from Hugging Face, returning both point forecasts and quantile outputs ranging from 0.1 to 0.9. The ensemble configuration downloads the four member checkpoints at runtime, meaning no merged weights are stored in the repository.

Why It Matters

For developers and enterprises, the release addresses a common pain point in SI deployment: choosing the right model for diverse data patterns. A single SI model may excel at smooth industrial signals but fail on noisy retail demand. By combining complementary architectures, the ensemble provides a "sensible default" that improves robustness against model selection errors. In evaluations on the GIFT-Eval benchmark, the ensemble outperformed all four individual members on aggregate Mean Absolute Scaled Error (MASE). Specifically, the Probabilistic Ensemble achieved a MASE of 0.683, slightly better than the best single member (PatchTST-FM-r2 at 0.685). While the individual PatchTST-FM-r2 retained the lowest aggregate Continuous Ranked Probability Score (CRPS), the ensemble ranked third among eligible systems on average MASE Rank, offering a balanced solution for users who cannot afford the overhead of extensive model tuning. The toolkit is permissively licensed, facilitating broader adoption in open-source SI ecosystems.

The Bottom Line

IBM’s Granite Timeseries Ensemble R1 offers a practical, open-source solution for time-series forecasting by leveraging the strengths of multiple SI models without requiring retraining. While individual models may still offer marginal gains in specific metrics like CRPS, the ensemble provides a strong, consistent baseline that reduces the risk of poor model selection for varied datasets.