Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AverageTime: Enhance Long-Term Time Series Forecasting with Simple Averaging

Gaoxiang Zhao, Li Zhou, Xiaoqiang Wang

TL;DR

This work tackles long-term multivariate time series forecasting by focusing on dependencies within sequences and across channels. It introduces AverageTime, a simple yet powerful approach that uses mixed-channel embeddings and a prediction-averaging strategy, augmented by a clustering mechanism to accelerate training. A lighter variant, LightAverageTime, groups correlated channels to share predictions and further boosts efficiency. Across eight real-world datasets, AverageTime often surpasses state-of-the-art Transformer models while maintaining efficiency on par with lightweight linear methods, demonstrating strong practical potential for scalable long-horizon forecasting.

Abstract

Long-term time series forecasting focuses on leveraging historical data to predict future trends. The core challenge lies in effectively modeling dependencies both within sequences and channels. Convolutional Neural Networks and Linear models often excel in sequence modeling but frequently fall short in capturing complex channel dependencies. In contrast, Transformer-based models, with their attention mechanisms applied to both sequences and channels, have demonstrated strong predictive performance. Our research proposes a new approach for capturing sequence and channel dependencies: AverageTime, an exceptionally simple yet effective structure. By employing mixed channel embedding and averaging operations, AverageTime separately captures correlations for sequences and channels through channel mapping and result averaging. In addition, we integrate clustering methods to further accelerate the model's training process. Experiments on real-world datasets demonstrate that AverageTime surpasses state-of-the-art models in predictive performance while maintaining efficiency comparable to lightweight linear models. This provides a new and effective framework for modeling long time series.

AverageTime: Enhance Long-Term Time Series Forecasting with Simple Averaging

TL;DR

This work tackles long-term multivariate time series forecasting by focusing on dependencies within sequences and across channels. It introduces AverageTime, a simple yet powerful approach that uses mixed-channel embeddings and a prediction-averaging strategy, augmented by a clustering mechanism to accelerate training. A lighter variant, LightAverageTime, groups correlated channels to share predictions and further boosts efficiency. Across eight real-world datasets, AverageTime often surpasses state-of-the-art Transformer models while maintaining efficiency on par with lightweight linear methods, demonstrating strong practical potential for scalable long-horizon forecasting.

Abstract

Long-term time series forecasting focuses on leveraging historical data to predict future trends. The core challenge lies in effectively modeling dependencies both within sequences and channels. Convolutional Neural Networks and Linear models often excel in sequence modeling but frequently fall short in capturing complex channel dependencies. In contrast, Transformer-based models, with their attention mechanisms applied to both sequences and channels, have demonstrated strong predictive performance. Our research proposes a new approach for capturing sequence and channel dependencies: AverageTime, an exceptionally simple yet effective structure. By employing mixed channel embedding and averaging operations, AverageTime separately captures correlations for sequences and channels through channel mapping and result averaging. In addition, we integrate clustering methods to further accelerate the model's training process. Experiments on real-world datasets demonstrate that AverageTime surpasses state-of-the-art models in predictive performance while maintaining efficiency comparable to lightweight linear models. This provides a new and effective framework for modeling long time series.
Paper Structure (20 sections, 7 equations, 5 figures, 5 tables)

This paper contains 20 sections, 7 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Problem Statement
  3. Transformer Models
  4. MLP Models
  5. CNN Models
  6. Method
  7. Overview
  8. LightAverageTime
  9. Experiments
  10. Datasets
  11. Baselines
  12. Experiments Setup
  13. Main results
  14. Detail for LightAverageTime
  15. Complexity Analysis
  16. ...and 5 more sections

Figures (5)

  • Figure 1: The overall architecture of the AverageTime model, which employs MLPs for predictions and incorporates mixed-channel embedding layer with shared parameters to capture inter-channel correlations.
  • Figure 2: The model initially computes the spearman correlation coefficients between time series. Subsequently, it employs a community detection algorithm to cluster the channels, thereby obtaining time series that are enriched with grouping information.
  • Figure 3: Different threshold on model performance and efficiency on different datasets.
  • Figure 4: Comparison of Model Complexities: We compare the complexities of different models. AverageTime and LightAverageTime demonstrate efficiency comparable to that of linear models, significantly outperforming Transformer-based models with approximately half the memory usage and 70% of the training time compared to the second-best model, TimeXer.
  • Figure 5: The model's performance varies with changes in the lookback window.