<p>As a precious metal, gold has traditionally been used as an instrument of money exchange. Despite replacing banknotes and digital currency, people are still interested in trading with gold and investing in gold as a reliable market. However, gold price fluctuations are a source of investment risk. Thus, many researchers have tried to provide a model to accurately forecast gold prices using neural networks. This method is advantageous to other methods, but it has some drawbacks. Some papers have used metaheuristic algorithms to optimize neural networks in the training process and parameter and feature selection in attempts to tackle these drawbacks. This research forecasts the closing price of global gold using a hybrid algorithm based on the marine predators algorithm (MPA), feature selection, and the cascade-forward neural network (CFNN). Technical analysis indicators are used as the input variables. In this regard, the feature selection problem is employed to reduce and optimize input variables, resulting in higher accuracy and shorter computation time. The daily values of the global gold price per ounce from 2021/9/5 to 2024/9/5 constitute the statistical population. This statistical population is based on the daily frequency dataset including the highest, lowest, opening and closing price fluctuations for each day. The accuracy of the proposed method is compared with that of the other forecasting models in terms of the forecast error and regression diagrams. The results show that the algorithm is more accurate than the CFNN (with no algorithms applied). Also, the proposed method is more accurate than the hybrid algorithms of this neural network, along with other metaheuristic algorithms (genetic and gray wolf).</p>