S2TX: Cross-Attention Multi-Scale State-Space Transformer for Time Series Forecasting

Time series forecasting has advanced with multi-scale models designed to capture both long- and short-range patterns, yet existing approaches often process variates independently and learn global and local representations in isolation. To address these limitations, we introduce S2TX (State Space Transformer with cross-attention), which integrates a Mamba model for long-range cross-variate context with a Transformer using local window attention. Through cross-attention, S2TX enables richer variate-level interactions and effective communication between global and local features. Extensive experiments on six benchmark datasets demonstrate that S2TX achieves robust, state-of-the-art performance while maintaining low memory overhead.