Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval

Fanpu Cao, Lu Dai, Jindong Han, Hui Xiong

TL;DR

The paper tackles the limitation of fixed historical windows in multivariate time series forecasting by introducing the Global Temporal Retriever (GTR), a lightweight, plug-and-play module that learns a global cycle embedding and retrieves aligned global segments to fuse with local sequences using a 2D convolution. GTR is backbone-agnostic and can be mounted on top of simple backbones like MLPs, with RevIN stabilizing non-stationary data and a complexity of $O(NT^2)$ for the retrieval component, yielding overall scalable performance. Empirical results on six real-world datasets show state-of-the-art performance for both short- and long-term forecasting, with minimal parameter and compute overhead and robust improvements across diverse domains; notable gains occur on Solar-Energy and PEMS03, and the method remains effective even with shorter look-back windows. The work also demonstrates ablations validating the 2D fusion mechanism and RevIN, and discusses extensions like the Global Token Aggregation to capture inter-channel dependencies, highlighting GTR's practical impact for efficient, global-period aware MTSF in resource-constrained settings.

Abstract

Multivariate time series forecasting (MTSF) plays a vital role in numerous real-world applications, yet existing models remain constrained by their reliance on a limited historical context. This limitation prevents them from effectively capturing global periodic patterns that often span cycles significantly longer than the input horizon - despite such patterns carrying strong predictive signals. Naive solutions, such as extending the historical window, lead to severe drawbacks, including overfitting, prohibitive computational costs, and redundant information processing. To address these challenges, we introduce the Global Temporal Retriever (GTR), a lightweight and plug-and-play module designed to extend any forecasting model's temporal awareness beyond the immediate historical context. GTR maintains an adaptive global temporal embedding of the entire cycle and dynamically retrieves and aligns relevant global segments with the input sequence. By jointly modeling local and global dependencies through a 2D convolution and residual fusion, GTR effectively bridges short-term observations with long-term periodicity without altering the host model architecture. Extensive experiments on six real-world datasets demonstrate that GTR consistently delivers state-of-the-art performance across both short-term and long-term forecasting scenarios, while incurring minimal parameter and computational overhead. These results highlight GTR as an efficient and general solution for enhancing global periodicity modeling in MTSF tasks. Code is available at this repository: https://github.com/macovaseas/GTR.

Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval

TL;DR

The paper tackles the limitation of fixed historical windows in multivariate time series forecasting by introducing the Global Temporal Retriever (GTR), a lightweight, plug-and-play module that learns a global cycle embedding and retrieves aligned global segments to fuse with local sequences using a 2D convolution. GTR is backbone-agnostic and can be mounted on top of simple backbones like MLPs, with RevIN stabilizing non-stationary data and a complexity of for the retrieval component, yielding overall scalable performance. Empirical results on six real-world datasets show state-of-the-art performance for both short- and long-term forecasting, with minimal parameter and compute overhead and robust improvements across diverse domains; notable gains occur on Solar-Energy and PEMS03, and the method remains effective even with shorter look-back windows. The work also demonstrates ablations validating the 2D fusion mechanism and RevIN, and discusses extensions like the Global Token Aggregation to capture inter-channel dependencies, highlighting GTR's practical impact for efficient, global-period aware MTSF in resource-constrained settings.

Abstract

Multivariate time series forecasting (MTSF) plays a vital role in numerous real-world applications, yet existing models remain constrained by their reliance on a limited historical context. This limitation prevents them from effectively capturing global periodic patterns that often span cycles significantly longer than the input horizon - despite such patterns carrying strong predictive signals. Naive solutions, such as extending the historical window, lead to severe drawbacks, including overfitting, prohibitive computational costs, and redundant information processing. To address these challenges, we introduce the Global Temporal Retriever (GTR), a lightweight and plug-and-play module designed to extend any forecasting model's temporal awareness beyond the immediate historical context. GTR maintains an adaptive global temporal embedding of the entire cycle and dynamically retrieves and aligns relevant global segments with the input sequence. By jointly modeling local and global dependencies through a 2D convolution and residual fusion, GTR effectively bridges short-term observations with long-term periodicity without altering the host model architecture. Extensive experiments on six real-world datasets demonstrate that GTR consistently delivers state-of-the-art performance across both short-term and long-term forecasting scenarios, while incurring minimal parameter and computational overhead. These results highlight GTR as an efficient and general solution for enhancing global periodicity modeling in MTSF tasks. Code is available at this repository: https://github.com/macovaseas/GTR.
Paper Structure (39 sections, 1 theorem, 30 equations, 7 figures, 12 tables, 1 algorithm)

This paper contains 39 sections, 1 theorem, 30 equations, 7 figures, 12 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Problem Statement and Overview
  5. Global Temporal Retriever
  6. Backbone and Projection Structure
  7. Method Details
  8. Experiment
  9. Experiment Setup
  10. Experiments Results
  11. Long-term Forecasting
  12. Short-term Forecasting
  13. Ablation Studies and Analysis
  14. Limitations and Future Work
  15. Conclusion
  16. ...and 24 more sections

Key Result

Theorem D.1

Given the above assumptions, if $\sigma_\varepsilon^2 < \sigma_\eta^2$, then: where $\mathrm{corr}(x_n, x_m)$ is the correlation of raw observations, and $\mathrm{corr}(z_n, z_m)$ is the correlation after GTR module processing.

Figures (7)

  • Figure 1: Pearson correlation matrix of time series segments from the Electricity dataset. We divided the series into several sub-series by the local-cycle length. The series demonstrates both the local-cycle pattern (e.g., $Corr(S_{12}, S_{13})=0.94,Corr(S_{12}, S_{14})=0.88$) and global-cycle pattern (e.g., $Corr(S_{12}, S_{5})=0.96$). Critically, the global-cycle pattern is stronger than local-cycle pattern (e.g., $Corr(S_{12}, S_{5}) > Corr(S_{12}, S_{13}), Corr(S_{12}, S_{14})$).
  • Figure 2: Overview of the Global Temporal Retriever (GTR): a plug-and-play module compatible with any MTSF forecaster. GTR operates in three stages: (1) retrieves corresponding segments from global temporal embedding; (2) aligns them with the input and uses 2D convolution to jointly model local and global periodicity; (3) fuses the result with the original input via residual connection.
  • Figure 3: Influence of look-back length. Forecasting horizons are fixed at 336 for Electricity and Solar-Energy, and 96 for PEMS03 and PEMS04. GTR consistently outperforms other models across varying look-back lengths, particularly at the shortest window length.
  • Figure 4: Visualization for multivariate correlation analysis on 4 datasets. The visualization is implemented based on the Pearson Correlation Coefficient. The Ground Truth denotes the correlation computed across the entire dataset. The two columns on the left denote the correlation between the variables before and after the GTR module, respectively. It shows that the model drives the learned multivariate correlations closer to the global correlation structures through the GTR module.
  • Figure 5: Performance of GTR under different random seeds on several datasets. The look-back length $T$ and forecasting horizon $S$ are fixed at 96.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Theorem D.1
  • proof