IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license

IBM has unveiled its latest advance in time series forecasting with the release of the Granite Time Series PatchTST-FM-r2 model. This new version builds on the highly successful PatchTST framework, offering significant improvements in performance and usability. The model boasts a modest parameter count of 385 million, yet delivers top-tier results on the GIFT-Eval benchmark, ranking second for both the Continuous Ranked Probability Score (CRPS) and the Mean Absolute Scaled Error (MASE) among replicate, zero-shot models.

This places it firmly in the top tier of forecasting tools available today. Importantly, IBM has made the model freely available under a permissive, commercial-friendly license (Apache 2.0 and OpenMDW 1.0), enabling businesses to leverage its capabilities without costly licensing restrictions. The model's architecture has been refined to efficiently capture both long- and short-term dependencies, using a combination of conformer layers and convolutional components.

This hybrid approach allows the model to effectively handle complex temporal patterns while maintaining computational efficiency. PatchTST-FM-r2 supports forecasting over extensive time horizons, accommodating up to 8,192 time steps, and can generate predictions across 99 quantiles. These enhancements make the model exceptionally versatile for a wide range of forecasting applications.

IBM has also provided comprehensive documentation, training data, and code examples to facilitate easy integration of the model into production systems. Furthermore, the release underscores IBM's commitment to advancing time series analytics through open innovation, empowering organizations to harness cutting-edge forecasting technology without barriers.