This paper addresses key issues in stock price prediction, including weak trend capture, poor model generalization, and insufficient noise resistance. Here, a novel deep learning framework is proposed, combining trend capture, frequency domain feature decoupling, and adaptive feature fusion. The study introduces a dual-channel trend capture method, using the Moving Average Convergence Divergence (MACD) Indicator to build multi-dimensional time-based features. Then, differential normalization is applied to enhance sensitivity to relative trend changes. Moreover, Fourier frequency domain decomposition is used to separate long-term trends and short-term noise. Finally, a feature fusion layer is constructed based on the Kolmogorov-Arnold Network (KAN), which achieves dynamic coupling between MACD indicators and price sequences through its B-spline function mechanism. Experiments show this method reduces mean squared error (MSE) by 0.37%–19.73% across 9 stock datasets compared to SOTA time series algorithms. This study provides new insights into feature decoupling for financial time series prediction, with the developed stock price forecasting model demonstrating superior predictive accuracy that holds practical value for quantitative investment applications.