Enhancing Gold Price Forecasting through Decomposition Algorithms and Deep Learning: A Comprehensive Comparative Analysis and Hybrid Model Matching Approach

Gold price forecasting has an important research value in the field of macroeconomic decision-making and financial market investment. This study compares the effects of time series decomposition methods and deep learning models in gold price prediction, constructing and evaluating 84 hybrid model combinations. The empirical analysis based on the daily closing price data of AU99.99 from the Shanghai Gold Exchange shows the forecasting performance of different technical combinations: Empirical Mode Decomposition (EMD) demonstrated the best adaptability compared to methods such as Wavelet Decomposition and Variational Mode Decomposition (VMD). The Bidirectional Gated Recurrent Unit (BiGRU) combined with Attention Mechanism (AM) performs well in time series modeling. Furthermore, a hybrid modeling framework based on frequency feature matching is proposed, where BiGRU-AM is applied to model the low-frequency components and Bidirectional Temporal Convolutional Networks combined with Attention Mechanism (BiTCN-AM) are used for the high-frequency components. Experiments demonstrate that the method effectively improves the prediction accuracy of non-stationary financial time series, providing a new idea for quantitative analysis in complex market environments.