Research Brief

TiRex-2 introduces an efficient, recurrent xLSTM-based time series foundation model for streaming multivariate forecasting, capable of integrating past and future covariates while maintaining strict causality and constant inference cost.

This paper presents TiRex-2, a new generation of time series foundation model designed to overcome limitations of current Transformer-based approaches, particularly in handling continuous, multivariate, and streaming data. Unlike existing models that struggle with long context lengths and require reprocessing entire histories, TiRex-2 employs a memory-centric, recurrent architecture that processes data patch-by-patch with constant computational cost. It combines a unique bidirectional time mixer with an asymmetric grouped-attention variate mixer, allowing it to incorporate future-known information (covariates) without compromising the causal integrity of its predictions on target variables. To facilitate its training on diverse multivariate data, the authors developed a novel synthetic pipeline that generates complex multivariate samples from simpler, large univariate datasets. Empirically, TiRex-2 demonstrates leading zero-shot performance on established benchmarks, proving stable under extended streaming conditions, and maintaining its promised constant inference efficiency.

• Real-time financial market prediction (stock prices, forex) where multiple indicators evolve jointly and some future economic announcements are known.
• Predictive maintenance for industrial machinery, integrating sensor data streams from various components along with scheduled maintenance events.
• Energy grid load forecasting, considering historical consumption, real-time generation, and future weather forecasts or scheduled events.
• Supply chain optimization, forecasting demand and logistics with multivariate data streams and known future marketing campaigns or holidays.