TimeRM: Multi-Expert Residual Modeling for Long-Term Time Series Forecasting

Hao Liu, Chun Yang, Yuchuan Mao, Xiaobin Zhu

Keywords: Time Series Forecasting; Non-Stationary Reconstruct; Random Fluctuation; Noise

Abstract: In time series forecasting, residuals capture unmodeled dynamics that are entangled with various temporal patterns, resulting in complex structures that challenge the model’s predictive accuracy. To address this, we propose TimeRM, a simple yet effective residual modeling framework. We decompose the time series into periodic and residual components. The Multi-Level Decoupling (MLD) module processes the residuals, extracting sub-residuals that capture hierarchical temporal patterns and supply the main model with previously neglected dynamics. We further introduce the Multi-level Sub-residual Compensation (MSRC) module, based on the Mixture-of-Experts (MoE) architecture, which captures entangled temporal patterns, quantifies their contribution to the sequence, compensates the primary residual, and thereby reduces the impact of unmodeled dynamics. Additionally, to resolve the issue of insufficient representation of expert models when handling data with complex channel characteristics, we propose the Multi-Expert Coupling (MEC) module to predict sub-residuals, enhancing the prediction of neglected channel features through a multi-level coupling approach to improve model performance. Extensive experiments demonstrate that our TimeRM framework outperforms current state-of-the-art methods.

Primary Area: learning on time series and dynamical systems

Submission Number: 5425