Reproduce “OLinear: A Linear Model for Time Series Forecasting in Orthogonally Transformed Domain” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wenzhen Yue, Yong Liu, Haoxuan Li, Hao Wang, Xianghua Ying, Ruohao Guo, Bowei Xing, Ji Shi

OLinear introduces a linear model for multivariate time series forecasting, employing a data-adaptive orthogonal transformation (OrthoTrans) and a novel linear dependency learner (NormLin). This approach achieves state-of-the-art accuracy across 24 datasets and 140 tasks while significantly improving computational efficiency compared to complex Transformer-based models.

Key insights

  • Data-adaptive orthogonal transformations can provide more effective decorrelation and disentangled representations for time series compared to generic, fixed-basis methods.
  • A simple, constrained linear layer (NormLin) can outperform complex multi-head self-attention in time series forecasting, offering superior efficiency and improved gradient flow.
  • Achieving state-of-the-art performance in multivariate time series forecasting is possible with streamlined, linear-based architectures that prioritize efficiency and data-adaptive components over increasing model complexity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.