TiRex-2: Generalizing TiRex to Multivariate Data and Streaming

2026-07-01

AI Summary

This work addresses the high computational complexity of Transformer-based models in multivariate time series streaming forecasting, their need for full-history recomputation, and the challenge of jointly leveraging future covariates while preserving strict causality. To overcome these limitations, the authors propose a recurrent foundation model based on xLSTM, featuring a memory-centric design that achieves constant per-segment inference cost and supports arbitrarily long streaming inputs. The architecture integrates a bidirectional temporal mixer and an asymmetric grouped-attention variable mixer to effectively capture cross-variable dependencies. A novel pretraining strategy based on dynamically synthesized multivariate samples is introduced, enabling zero-shot state-of-the-art performance on GIFT-Eval and fev-bench. The model activates 38.4M parameters in univariate mode and an additional 44.1M parameters in multivariate mode.

Abstract

We introduce TiRex-2, a recurrent xLSTM-based time series foundation model that generalizes the univariate TiRex to multivariate forecasting with both past and future covariates. Real-world forecasting is inherently sequential: observations arrive continuously, variables evolve jointly, and a subset of covariates is known ahead of time. Existing Transformer-based time series foundation models capture cross-variate dependencies but incur quadratic complexity in context length and require full-history recomputation as new observations arrive. TiRex-2 addresses these limitations through a memory-centric recurrent design that operates at constant per-patch cost under streaming. The model combines a bidirectional time mixer with an asymmetric grouped-attention variate mixer, enabling the integration of future-known covariates while preserving strict causality over target variables. To our knowledge, this is the first time series foundation model that achieves this combination of properties. To support scalable multivariate pretraining, we propose a synthetic coupling pipeline that composes diverse multivariate samples on the fly from large univariate corpora. Empirically, TiRex-2 achieves state-of-the-art zero-shot performance on GIFT-Eval and fev-bench, remains stable when streamed to arbitrary context lengths, and maintains constant inference cost per patch. The model uses 38.4M active parameters in univariate mode, with an additional 44.1M parameters activated for multivariate forecasting.

Problem

• multivariate time series forecasting
• streaming prediction
• future covariates
• quadratic complexity
• time series foundation model

Innovation

• xLSTM
• multivariate forecasting
• streaming time series
• future covariates
• memory-centric recurrent model