Enhancing Gold Price Forecasting Using Machine Learning Models Optimized with Metaheuristic Algorithms

Forecasting gold prices is essential for supporting informed decision-making among investors, policymakers, and financial analysts. However, due to their non-linear and volatile behavior influenced by complex economic and geopolitical factors, predicting gold prices remains a significant challenge. This study evaluates the forecasting performance of three traditional machine learning models—Random Forest (RF), Multi-Layer Perceptron (MLP), and XGBoost—on a monthly dataset spanning from January 1991 to December 2023, using macroeconomic and commodity-related indicators obtained from IndexMundi. To enhance predictive accuracy, RF and MLP were optimized using metaheuristic algorithms including Particle Swarm Optimization (PSO), Differential Evolution (DE), Simulated Annealing (SA), and Genetic Algorithm (GA), while XGBoost was fine-tuned using Grid Search. Two ensemble strategies were developed to further improve performance: a weighted ensemble based on inverse error metrics and a boosting ensemble that sequentially combined top-performing models. The results show that combining traditional models with metaheuristic optimization significantly improves forecasting accuracy. The best performance was achieved by the boosting ensemble integrating RF-PSO and optimized XGBoost, attaining an R² of 0.9654 and a Root Mean Square Error (RMSE) of 0.0433, representing an improvement of 11.1% in RMSE over the best single optimized model. This research demonstrates that effective and scalable financial forecasting systems can be developed using established machine learning techniques, offering valuable decision-support tools in dynamic financial markets.