Abstract: The formation mechanism of gold prices is intricate, reflecting the interaction of commodity and financial attributes over various time scales. Traditional univariate time series models struggle to capture the rich multidimensional information inherent in such complex systems. To address this limitation, this study proposes a novel forecasting framework grounded in the "decomposition–reconstruction–ensemble" paradigm, integrating a robust decomposition strategy and a memory-driven hybrid regression model for enhanced gold price prediction. First, a Stable Variational Mode Decomposition (SVMD) technique is developed to address boundary distortion issues commonly overlooked in previous decomposition studies. SVMD ensures stable component boundaries by refining the optimization process, thereby mitigating boundary effects and supporting continuous feature learning. Second, the Hurst exponent is employed as a reconstruction criterion to quantify the memory properties of decomposed components. Based on memory strength, the components are categorized into short-term, medium-term, and long-term elements. Each class is then modeled using a hybrid linear regression (HLR) approach that combines the dimensionality reduction advantages of Scaled Principal Component Analysis (SPCA) and the variable selection efficiency of Minimax Concave Penalty (MCP) regression. This design enables targeted feature extraction from financial variables at different time scales while filtering out redundant covariates. Third, a hierarchical ensemble strategy is introduced to integrate predictions from the original layer, decomposition layer, and reconstruction layer. Unlike traditional nonlinear ensemble methods prone to overfitting, this ensemble leverages in-sample structural information and does not require additional supervised learning, enhancing robustness and generalization capacity. Empirical evaluations on two international gold price datasets (London and Chicago) validate the effectiveness of the proposed model. Compared with 13 benchmark models, the proposed HLR-HE-minT model achieves the lowest Mean Absolute Error (MAE, around 11.7) and the highest directional accuracy (exceeding 60%). Results from Diebold-Mariano (DM) tests further demonstrate that the proposed model significantly outperforms both traditional univariate models and single-layer decomposition models in terms of forecasting precision. This study contributes theoretically and methodologically to the decomposition–reconstruction–ensemble forecasting paradigm. It introduces a stable decomposition method that corrects boundary distortions, a memory-guided hybrid regression strategy for model assignment, and a novel ensemble method that ensures consistency across layers.