Multivariate time series forecasting is of critical importance across various practical domains. However, effectively capturing both temporal dependencies and inter-variable correlations remains a significant challenge. Existing methods often emphasize seasonal components while overlooking the variable relationships within trend components, thereby limiting their capacity to model correlations. To address this issue, we propose a novel MLP-based forecasting model, ETMixer. Specifically, we first construct a multiscale patch structure to enhance the local details and feature representation of the series from a multiscale perspective. In addition, we incorporate an exponential moving average decomposition module to enrich the representation of the trend component. During forecasting, the seasonal component is leveraged to model temporal dependencies, while the trend component is further utilized to extract inter-variable correlation patterns. Extensive experiments conducted on real-world datasets show that ETMixer consistently surpasses existing approaches, achieving state-of-the-art forecasting performance.