<p>In multivariate time series forecasting, series decomposition architectures that decouple trend and residual components have become a mainstream solution for handling data non-stationarity. However, existing methods are limited by the “component independence assumption,” modeling these components independently, thereby neglecting the intrinsic constraints macro-trends impose on micro-fluctuations. Furthermore, trend modeling faces a dilemma between efficiency and expressiveness: global attention-based methods are prone to introducing non-causal noise and disrupt trend continuity, whereas simple linear mappings lack dynamic modeling capabilities. To address these challenges, we propose the Causal Trend Evolution and Adaptive Modulation Network (CTMNet), a coupled decomposition architecture that employs a modeling strategy balancing differentiation and interaction. First, we design a Causal Trend Encoder (CTE), which utilizes patch-level causal convolution and gating mechanisms to introduce a strict temporal causal inductive bias, accurately characterizing the unidirectional evolution of trends with linear complexity. Second, for the residual component, we innovatively propose an Adaptive Trend Modulation Interaction (ATMI) mechanism. This mechanism uses the causal trend state extracted by the CTE as a contextual prior to dynamically generate affine transformation parameters, to hierarchically calibrate residual features. This design not only maintains the physical consistency of trend modeling but also restores the deep coupling between trend and residual components. Extensive experiments on 10 real-world benchmark datasets, covering both long- and short-term forecasting, demonstrate that CTMNet achieves competitive or leading performance in terms of prediction accuracy and robustness compared to 7 state-of-the-art (SOTA) models.</p>