Robust Time Series Forecasting via Basis-Aligned Sampling in Decycled Residual Space

Longlong Xu, Zeyan Li, Xiao He, Zhaoyang Yu, Changhua Pei, Zhe Xie, Zijun Dou, Tieying Zhang, Dan Pei

Keywords: Robust Time Series Forecasting, Deep Learning

TL;DR: A robust time series forecasting method that reduces reliance on individual data points, improving resilience to point-wise perturbations while maintaining high accuracy through global context utilization and periodicity learning.

Abstract:

Time series forecasting is crucial in domains such as finance, energy, and traffic, yet real-world data are often contaminated by anomalies and noise.
In this work, we first identify a fundamental limitation of existing approaches—their excessive reliance on specific input points, particularly the most recent observation—which makes them highly susceptible to point-wise perturbations and undermines prediction reliability.
To further address this challenge, we propose RESAM, a novel approach for robust time series forecasting that effectively mitigates the impact of point-wise perturbations while maintaining high overall forecasting accuracy.
RESAM utilizes a basis-aligned randomized sampling strategy to comprehensively exploit the global context and achieve a unified representation for irregularly sampled sequences.
Moreover, RESAM employs a learnable periodicity extraction module with a two-stage training protocol to enhance the accuracy and robustness of both periodicity and residual learning.
Comprehensive evaluations on eight benchmark datasets show that RESAM achieves competitive forecasting accuracy and significantly surpasses state-of-the-art models in robustness to point-wise perturbations.

Primary Area: learning on time series and dynamical systems

Submission Number: 5512