IBM Launches Granite Timeseries Ensemble: 4 Open Forecasters at MASE 0.683

Executive Overview

IBM Research has announced the release of Granite Timeseries Ensemble via Hugging Face, a time-series forecasting package that combines four open Granite models into a single scalable interface. The probabilistic ensemble achieved the best MASE score of 0.683 across 97 configurations of the GIFT-Eval benchmark, outperforming the best single member, PatchTST-FM-r2, which scored 0.685. The system requires no retraining or fitted ensemble weights and delivers quantile forecasts from 0.1 to 0.9 with a nominal 80% confidence interval. This represents a practical default for developers who need stable accuracy without prior architectural selection.

Deep-Dive Features & Architecture

Granite Timeseries Ensemble borrows a concept from weather forecasting: different models capture complementary views of what will happen, then merge them into a single forecast. The system includes four complementary architectures: Granite PatchTST-FM-r1 (patched transformer), Granite PatchTST-FM-r2 (compatible patched transformer), Granite FlowState-r1.1 (state-space model), and Granite TTM-r3 (Tiny Time Mixers). These architectures capture trends, seasonality, and local variance in different ways, so their errors do not necessarily align.

The system offers two methods for combining forecasts: Probabilistic Ensemble (unified probabilistic aggregation), the recommended default, which gives each member equal influence without a fitting step or weight tuning; and IQR Ensemble (adaptive aggregation), which weights members according to the spread of their forecasts, giving narrower intervals greater influence, with adjustable temperature and an optional cap on any member's weight. The system returns a point forecast in prediction_outputs and quantile forecasts in quantile_outputs ranging by default from 0.1 to 0.9, where the 0.1 and 0.9 quantiles represent a central prediction interval with nominal 80% coverage.

The recipe stores ensemble settings rather than combined model weights and loads the four checkpoints at runtime, requiring additional time on first run. The interface can be extended to include compatible external models without merging or retraining, easing integration into existing forecasting architectures.

Benchmark & Competitive Performance

Across 97 GIFT-Eval configurations (without fine-tuning and without using test targets to select or fit the ensemble method), both ensembles outperformed the four individual members in mean MASE and mean MASE rank. Probabilistic Ensemble recorded the best MASE at 0.683 versus 0.685 for PatchTST-FM-r2, 0.701 for FlowState-r1.1, 0.717 for PatchTST-FM-r1, and 0.724 for TTM-r3. On the probabilistic CRPS/MWQL metric, PatchTST-FM-r2 retained the best overall score at 0.467, while IQR Ensemble achieved the best mean CRPS rank at 47.619. In the October 20, 2026 comparison, the approach ranked third among eligible systems.

Industry Impact & Enterprise Adoption

The release lowers the barrier to deploying high-accuracy time-series forecasting in production. Because the ensemble requires no retraining or weight fitting, enterprises can integrate it into existing pipelines with minimal engineering overhead. The permissive license and Hugging Face availability make it attractive for rapid prototyping and research, while the shared Forecaster interface allows teams to extend the ensemble with proprietary or external models. This positions Granite Timeseries Ensemble as a practical default for demand forecasting, anomaly detection, and financial modeling where stable accuracy and quantile uncertainty estimates are critical.

Conclusion

IBM Research's Granite Timeseries Ensemble delivers a robust, open, and easy-to-deploy forecasting solution that edges out individual models on the GIFT-Eval benchmark. With its no-retraining design, quantile outputs, and flexible aggregation methods, it offers developers a reliable starting point for time-series tasks without the need for architectural guesswork. The release underscores IBM's commitment to open-source AI and practical enterprise tooling.

Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis