Study of Feature Selection for Gold Prices Forecasting Using Machine Learning Approach

Abstract

Gold has been used as a hedge in times of extreme economic uncertainty and financial instability. Due to the multi-factors that influence the gold market, it’s hard to definitively determine which factors should be used to forecast the price of gold. Therefore, this study aims to investigate the feature selection methods for determining the influencing factors for the gold price forecasting model based on the machine learning approach. The factors related to gold prices including macroeconomic indicators, market data, and investor fear indicators were investigated through six different features selection methods: correlation coefficient (R), recursive feature elimination (RFE), stepwise feature elimination (SFE), sequential feature selection (SFS), decision tree (DT) and LASSO. To forecast gold price based on these factors, we have applied four different methods: ordinary least squares regression (OLS), Prophet, long short-term memory networks (LSTM), and multi-layer perceptron (MLP) on the validation dataset for daily periods. The experimental results indicated that the SFS method provides the optimal predictor variables. Furthermore, the forecasting models using the OLS method are suitable and outperform the others method. According to this study, the feature selection methods based on the machine learning approach can identify the influencing factors similar to the econometric methodology from various research. Consequently, applying the optimal predictors for constructing gold price forecasting models may contribute to assessing the financial risk which is beneficial for investors to make decisions and manage their portfolios.