[ICLR 2026] Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval

"The GT-R is not a supercar for a select few; it is a supercar for everyone, built to be enjoyed anywhere, anytime, by anyone." --Nissan skyline

This repository contains the official PyTorch implementation of the (GTR) — a lightweight, plug-and-play module designed to empower any multivariate time series forecasting (MTSF) model with the ability to capture global periodic patterns far beyond the fixed look-back window.

---

📣 News

• 2026/2/4 🚀 Code has be released.
• 2026/1/26 💥💥 GTR is honored to be accepted by ICLR 2026!

---

🌟 Key Innovation: The Global Temporal Retriever (GTR)

Existing MTSF models are fundamentally limited by their reliance on a fixed-length historical window, making them unable to capture crucial global periodic patterns (e.g., weekly, monthly, seasonal trends) that span cycles much longer than the input.

GTR solves this by:

1. Maintaining a Learnable Global Representation: A parameter matrix Q ∈ R^(L×N) encodes the entire global cycle pattern for all N variables.
2. Dynamic Retrieval & Alignment: For any input sequence, GTR identifies its position within the global cycle and retrieves the corresponding segment.
3. Joint Local-Global Modeling: The retrieved global segment is stacked with the local input and processed by a 2D convolution to model dependencies across both scales.
4. Seamless Integration: The enriched representation is fused back via a residual connection, making GTR compatible with any existing forecasting backbone (MLP, Transformer, Mamba, etc.) without architectural changes.

---

📈 Performance Highlights

• State-of-the-Art Results: GTR+MLP achieves SOTA performance on 6 real-world datasets for both short-term and long-term forecasting.
• Significant Gains: On the challenging Solar-Energy dataset, GTR outperforms the second-best model by 8.2% in MSE and 6.5% in MAE.
• Plug-and-Play Enhancement: GTR consistently improves diverse SOTA models (iTransformer, PatchTST, DLinear) by up to 91.9% MSE reduction (DLinear on PEMS04).
• Extreme Efficiency: The GTR module itself adds only 40.1K parameters and 4.50M MACs. The full GTR+MLP model uses just 0.98M parameters, which is only 19% of iTransformer's size.

---

🚀 Getting Started

Prerequisites

• Python 3.8+
• PyTorch 1.10+
• Other dependencies (see requirements.txt)

Installation

conda create -n GTR python=3.8
conda activate GTR
conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt

Run

You can use the following script to obtain the prediction results (Recommended). For example, to reproduce all the experiment in the paper, you can run the following script:

bash run_main.sh

To reproduce all the ablation experiment in the paper, run the following script:

bash run_ablation.sh

📜 Citation

If you find GTR useful, please consider citing our paper:

@inproceedings{
cao2026gtr,
title={Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval},
author={Fanpu Cao and Lu Dai and Jindong Han and Hui Xiong},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=QUJBPSfyui}
}