Abstract:

The pricing mechanism of gold assets is particularly complex， and it is difficult to fully reveal the rich multidimensional information contained in its inherent feature space through a univariate time series analysis framework. A hybrid regression model with robust decomposition and hierarchical integration strategies for gold price prediction is proposed from the perspective of decomposition-reconstruction-ensemble， effectively exploring the synergistic effects of mixed financial influencing factors at different time scales. Firstly， a stable variational mode decomposition （SVMD） technique is developed to extract the stable center frequency component boundaries of the gold price sequence for continuous feature learning. Then， using the Hurst exponent as a memory reconstruction index， the decomposition boundary is reconstructed into short-term， medium-term， and long-term scale components. Subsequently， utilizing the advantages of feature selection from scaled principal component analysis regression and minimum maximum concavity penalty regression， a hybrid linear regression （HLR） is constructed to extract important financial features for prediction at different time scales， thereby improving the overall prediction generalization ability. Finally， the hierarchical ensemble method integrated the prediction results of the original layer， reconstruction layer， and component layer to obtain a reconciliatory gold price prediction value. The effectiveness of the proposed model in the three steps of decomposition， reconstruction， and ensemble is validated on the international gold futures price dataset， and the advantages of the proposed model are compared with various prediction models and decomposition modeling strategies in existing research.

Key words: gold price prediction, hybrid regression model, decomposition and ensemble, variational mode decomposition, hierarchical ensemble