Deep belief network for gold price forecasting

Abstract

Fluctuations in gold price have historically attracted the attention of governments, institutions and individuals alike. Accurate forecasting of gold price can effectively capture the price change trend and reduce the impact of gold market fluctuations. However, this is challenging work due to the multi-factors and nonlinear feature of the gold market. In this paper, a deep belief network (DBN) model, composed of restricted Boltzmann machines (RBM) for pre-training and a layer of supervised back-propagation (BP) for fine-tuning, is proposed for gold price forecasting. Several highly related variables of gold price were used as inputs to build this DBN model for forecasting over the years 1984–2019. The obtained results were compared with those of the traditional BP neural network model, Genetic algorithm optimization for BP neural network model (GA-BP) and the mathematical linear model of Autoregressive Integrated Moving Average (ARIMA). The empirical results show that the proposed DBN model has an outstanding performance in prediction and direction, with the lowest Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE) and Mean Absolute Error (MAE), and the highest directional statistic (Dstat). Overall, the novel DBN has improved accuracy compared to the traditional neural network and linear models, distinguishing itself as a very promising methodology for gold price forecasting.