TiRex-2: Multivariate Forecasting

TiRex-2 – A Multivariate Time Series Foundation Model

We didn’t just optimize TiRex — we added new features. In addition to well-known capabilities such as state tracking and in-context learning, TiRex-2 can handle multivariate time series by processing covariates and enables data streaming without any loss of performance.

Mulitvariate

Data Streaming

Covariates

TiRex-2 Features

The original TiRex approach could forecast a time series based on its own history. TiRex-2 extends this approach to multivariate data. Multivariate means that the model takes additional time series into account. These additional time series are called covariates. They provide context that may be relevant to the target variable. This is crucial in industry. Machines, systems, and processes rarely generate just one signal; they often measure many variables simultaneously. A forecast becomes more meaningful when the model can leverage relationships between multiple signals.

TiRex-2 can handle multivariate time series by processing covariates and enables data streaming without any loss of performance.

Like TiRex, TiRex-2 is also based on the xLSTM architecture developed by NXAI. Unlike transformer-based approaches, xLSTM processes data sequentially and continuously updates an internal system state. As a result, memory requirements and computational effort remain stable even with long time series. This is particularly relevant for industrial streaming applications. Many current foundation models are primarily optimized for traditional chat or document applications. When dealing with continuous industrial data streams, they often reach their limits in terms of efficiency and memory requirements. TiRex-2, on the other hand, processes streaming data continuously and reliably—even over long periods of time.

TiRex-2 consistently delivers lower forecast error than Chronos-2

TiRex vs. Chronos: TiRex-2 cuts forecast error by 49% using future covariates (nearly 5x better than Chronos-2's 10% gain) by correctly learning the XOR pattern between two binary flags. — **TiRex vs. Chronos:** TiRex-2 cuts forecast error by **49%** using future covariates (nearly 5x better than Chronos-2's 10% gain) by correctly learning the XOR pattern between two binary flags.

Made for real-world industrial applications

Industrial customers are already working with TiRex technology. Customers in logistics, energy, and the manufacturing industry, in particular, are increasingly demanding models that can analyze and predict complex processes in real time—and preferably on-premises. The major advantage: Customers don’t always have to train a new AI model for different machines, because TiRex-2 generalizes across different machine types or components.

Despite its expanded multivariate capabilities, TiRex-2 remains comparatively compact and efficient to deploy. In multivariate mode, the model operates with approximately 83 million parameters. For applications without multivariate requirements, TiRex-2 can be run in a smaller configuration.

With TiRex-2, we can process a continous datastream without any loss of performance. This enables complete new industrial applications with time series.