SAFT: Learning Scale-Aware Inter-Series Correlations for Time Series Forecasting

Abstract

Multivariate time series forecasting faces the fundamental challenge of modeling complex temporal dependencies while capturing cross-variable relationships, especially in real-world applications where data exhibits intricate patterns across multiple scales. Existing models often overlook the necessity of explicitly incorporating scale awareness when modeling cross-variable correlations, leading to potential overfitting and instability. To address these issues, we propose Scale-Aware Forecasting Transformer (SAFT), which introduces a novel scale-aware multi-head attention mechanism to model cross-variable dependencies across different time scales. SAFT progressively integrates information from coarser to finer scales, enabling robust modeling of complex temporal dynamics. Extensive experiments demonstrate that SAFT achieves overall state-of-the-art performance in both long-term and short-term forecasting tasks.