TiRex-2: Generalizing TiRex to Multivariate Data and Streaming (arxiv.org)

Researchers have announced TiRex-2, an advanced recurrent xLSTM-based model that enhances the original TiRex’s capabilities for forecasting by targeting multivariate time series data in streaming scenarios. This model addresses significant limitations of existing Transformer-based approaches, which struggle with high computational complexity and the need to recompute full histories as new data is received. TiRex-2 employs a unique memory-centric design featuring a bidirectional time mixer and an asymmetric grouped-attention variate mixer, allowing it to efficiently handle both past and future covariates while enforcing strict causality. The introduction of TiRex-2 is noteworthy for the AI/ML community as it establishes a new benchmark for real-time multivariate forecasting, demonstrating state-of-the-art zero-shot performance on evaluation benchmarks like GIFT-Eval and fev-bench. The model's architecture utilizes 38.4 million active parameters for univariate mode and offers an additional 44.1 million for multivariate forecasting, all while maintaining a constant per-patch inference cost. This positions TiRex-2 as a transformative tool for applications requiring real-time data analysis, offering researchers and practitioners a powerful solution for complex forecasting tasks.