TS Foundation Models (huggingface.co)

The latest release from the AI community, TiRex-2, introduces a pretrained foundation model designed for zero-shot multivariate forecasting. This innovative model allows for the forecasting of multiple target variates using historical data and integrating both past and future-known covariates, such as holiday and promotion schedules. Notably, TiRex-2 operates without any task-specific training or fine-tuning, providing a seamless solution for users seeking to implement multivariate forecasting models. Significant for the AI/ML community, TiRex-2's zero-shot capability simplifies the deployment process, allowing practitioners to forecast effectively with just a single model checkpoint. This is achieved with a relatively small active footprint of 38.4 million parameters for univariate tasks, and an additional 44.1 million parameters to handle multivariate scenarios. The model's ability to adapt in real-time to new observations makes it particularly valuable for dynamic applications. Comprehensive resources, including installation guides and interactive demos, are available on the project's GitHub repository, inviting researchers to explore and integrate TiRex-2 into their forecasting workflows.