Research map / Time series / Foundation models for time series

TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning

NeurIPS 2025 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces TiRex, a 35M-parameter xLSTM-based (sLSTM-only) pre-trained zero-shot time series forecaster with Contiguous Patch Masking and three augmentations, achieving state-of-the-art on GiftEval-ZS and Chronos-ZS.

Problem

Zero-shot time series forecasting: transformer-based pre-trained models underperform recurrent models on long-horizon forecasting, while LSTMs lack strong in-context learning and degrade in autoregressive multi-step prediction.

Approach

Use a decoder-only sLSTM (xLSTM) backbone that keeps true recurrence for state-tracking, train with Contiguous Patch Masking so multi-patch horizons are predicted by feeding missing values rather than autoregressive point estimates, and augment pre-training with amplitude modulation, censoring, and spike injection.

Key results

• On GiftEval-ZS, TiRex reaches CRPS 0.411 (std ±0.002 over 6 seeds), versus next best zero-shot models at 0.459, 0.463, and 0.481.
• On Chronos-ZS, TiRex achieves best WQL (0.592) and rank, with second-best MASE (0.776), behind TabPFN-TS.
• TiRex (35M parameters) is over 11x faster than TimesFM-2.0 and over 4x faster than Chronos-Bolt-Base in inference.

Measured effects reported in the paper

Surprising findings

• Naive multi-patch training degrades short-term performance, and standard next-token training with autoregressive inference harms long-term performance.
• Removing amplitude modulation augmentation slightly improves GiftEval-ZS overall CRPS (0.409 vs 0.411), though it hurts Chronos-ZS WQL.

Limitations

• Focuses on univariate time series; multivariate dependencies are not modeled.
• Hyperparameters were not extensively tuned due to computational constraints.
• Some benchmark comparisons involve zero-shot leakage for baselines such as Moirai and TimesFM, and the GiftEval-ZS filtering removes 16 settings.

Experimental setup

Datasets: Chronos training datasets GiftEval pre-training data KernelSynth-style synthetic GP data TSMixup augmented Chronos data

Benchmarks: GiftEval-ZS Chronos-ZS

Metrics: CRPS MASE WQL average rank geometric mean of seasonal-naive-normalized scores

Baselines: Chronos Chronos-Bolt TimesFM Moirai TabPFN-TS TTM PatchTST TFT DLinear DeepAR N-BEATS

Builds on: xLSTM Chronos KernelSynth TSMixup GiftEval

Closest papers by meaning

• TiRex-2: Generalizing TiRex to Multivariate Data and Streaming (NeurIPS 2026 · similarity 0.92)
• TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting (NeurIPS 2026 · similarity 0.88)
• Zero-shot forecasting of chaotic systems (ICLR 2025 · similarity 0.87)
• Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.87)
• A decoder-only foundation model for time-series forecasting (ICML 2024 · similarity 0.87)
• Timer-XL: Long-Context Transformers for Unified Time Series Forecasting (ICLR 2025 · similarity 0.87)