TiRex: Zero-Shot Time-Series Forecasting with xLSTM
TiRex (xLSTM-based Zero-Shot Forecasting Model) · Also known as: Time-series xLSTM Forecaster, TiRex Zero-Shot, xLSTM Time-Series Model, Zaman Serisi Sıfır-Atım Tahmincisi
TiRex is a pretrained zero-shot time-series forecasting model introduced in 2025 by the NX-AI xLSTM team (Auer et al.). Built on the Extended Long Short-Term Memory (xLSTM) architecture, TiRex is trained at scale on diverse time-series corpora and can forecast unseen datasets without any fine-tuning. Its core idea is to exploit enhanced in-context learning: the model reads the entire available history as a context and produces forecasts for both short and long horizons directly from that context.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
TiRex is well-suited when labeled training data for a specific time series is scarce, unavailable, or costly to obtain, making zero-shot generalization valuable. It handles both short and long forecast horizons in a single model. Key assumptions are that the input series are sufficiently regular for patch-based tokenization and that the pretraining distribution covers similar domains. It is less appropriate when domain-specific fine-tuning is feasible and when interpretability of individual model components is required. Alternatives include Chronos, TimesFM, or domain-specific LSTM/Transformer baselines.
Strengths & limitations
- Zero-shot forecasting eliminates the need for dataset-specific training or fine-tuning
- xLSTM backbone provides efficient sequential processing with high-capacity matrix memory cells
- Supports both short-horizon and long-horizon forecasting within a single unified model
- Probabilistic output head enables uncertainty quantification alongside point forecasts
- Pretrained on a fixed corpus; out-of-distribution series with unusual dynamics may yield degraded accuracy
- Patch-based tokenization assumes locally coherent temporal structure, which may not hold for highly irregular or sparse series
- Large pretrained model size can increase inference latency and memory requirements compared to lightweight baselines
- In-context learning capacity is bounded by the model's context window length, limiting very long historical inputs
Frequently asked
Does TiRex require any fine-tuning on my dataset before use?
No. TiRex is designed for zero-shot use: the pretrained model is applied directly to unseen series at inference time without any gradient updates. The in-context learning mechanism adapts the forecast implicitly by reading the input history, so no labeled target data is needed for deployment.
How does TiRex differ from Transformer-based forecasting models like PatchTST?
TiRex uses xLSTM recurrent blocks rather than self-attention layers. xLSTM processes the patch sequence recurrently and compresses context into a hidden state, whereas attention-based models compute pairwise interactions across all patches. This recurrent design can be more memory-efficient for long contexts and aligns with TiRex's in-context learning paradigm.
Can TiRex produce prediction intervals, or only point forecasts?
TiRex includes a probabilistic forecasting head that outputs a predictive distribution over the forecast horizon. This enables generation of quantile forecasts and prediction intervals in addition to point estimates, allowing users to communicate forecast uncertainty explicitly in downstream decision-making processes.
Sources
- Auer, A., Podest, P., Klotz, D., Böck, S., Klambauer, G., & Hochreiter, S. (2025). TiRex: Zero-shot forecasting across long and short horizons with enhanced in-context learning. arXiv preprint. link ↗
How to cite this page
ScholarGate. (2026, June 2). TiRex (xLSTM-based Zero-Shot Forecasting Model). ScholarGate. https://scholargate.app/en/deep-learning/tirex
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
Compare side by side →