Improving Gold Price Prediction Accuracy Through Variational Mode Decomposition and Bayesian-Tuned Hybrid BiLSTM-BiGRU Model

Abstract

Due to the highly dynamic and unpredictable nature of the global economy, it is very difficult to make successful predictions of gold prices. Traditional approaches for forecasting price movements will frequently fail due to the complexity of the relationships among the variables driving the gold price, such as economic indicators, geopolitical events, and market sentiment. Therefore, there is a need for modern, data-driven techniques for developing accurate models that predict gold prices. In this study, we present a new hybrid model and methodology that includes multiple levels of processing to overcome many of these obstacles. The selection of input features is guided by a Granger causality analysis used as a preliminary statistical screening step to identify variables with potential predictive relationships with gold prices. Next, we utilize Variational Mode Decomposition (VMD) to denoise the temporal data. The final phase of the methodology, the hybrid Bidirectional Gated Recurrent Unit (BiGRU) and Bidirectional Long Short Term Memory (BiLSTM), was developed using a Bayesian Optimization approach to develop a model that utilizes both BiLSTMs and BiGRUs, allowing for the combined advantages of LSTMs and GRUs. Based on previously collected and recorded prices for gold, this new predictive model provides a statistically superior forecast compared to both the traditional industry benchmarks and several current internet websites that provide gold price predictions. The results of 5-fold cross-validation and an ablation study confirmed the robustness of this model. Further statistical analysis using analysis of variance (ANOVA) and the Holm-Bonferroni post-hoc procedure provides evidence that the results of this model are statistically significantly better (p < 0.05) than those of any of the benchmark models used. Additionally, the use of an Integrated Deep Learning Architecture (IDLA), advanced signal processing methodologies, and rigorous feature validation techniques has resulted in a more accurate and reliable performance in the prediction of gold price movements.