SCI: A Spectral Correlated-Variate Interaction Framework for Non-Stationary Time Series Forecasting

Multivariate time series forecasting is pivotal in real-world applications, ranging from energy consumption management and financial market trend analysis to traffic flow prediction and climate monitoring. Yet, it is critically challenged by inherent non-stationarity. The continuous temporal evolution of statistical properties drives asynchronous distributional shifts and dynamic inter-variate dependencies. Such discrepancies destabilize cross-variate correlation modeling, impede robust information propagation, and frequently induce negative transfer. To address this, we propose a novel framework, Spectral Correlated-variate Interaction (SCI), designed to robustly leverage auxiliary variates for enhanced forecasting under non-stationary interference. Specifically, we introduce the Spectral Auxiliary Bridge (SAB) to map the series into the frequency domain. Through a robust interaction mechanism, SAB explicitly transfers key evolutionary patterns from auxiliary to target variates while effectively filtering out asynchronous non-stationary noise. Concurrently, to recover vital non-stationary information typically attenuated by standard normalization, we design the Spectral Stationarity Rectifier (SSR). Unlike rigid statistical denormalization, SSR explicitly models and restores the evolving trajectories of non-stationary statistics directly within the latent space. This synergy ensures that the final predictions retain accurate trend information while integrating purified details from correlated variates. Extensive experiments on multiple benchmark datasets demonstrate that SCI effectively resolves the variable interaction dilemma under non-stationary conditions, consistently achieving state-of-the-art (SOTA) forecasting performance.