Time Series Analysis for the COMEX Gold Spot Price Forecasting by Using NARX DE/SVR and DE/GPR Techniques

In this study, three different strategies (direct multi-step scheme, recursive multi-step scheme and direct-recursive hybrid scheme) for automatic lag selection in time series analysis are proposed. The two novel hybrid methods use here are the nonlinear autoregressive with exogenous variable support vector regression (NARX SVR) and Gaussian Process regression (GPR) in combination with Differential Evolution (DE) metaheuristic optimizer, two powerful machine learning tools that can identify nonlinear patterns effectively thanks to the introduction of a kernel function: radial basis function (RBF) kernel. This optimization technique DE involves hyper-parameter setting in the SVR and GPR training procedures, which significantly influences the regression accuracy. This article examines the forecasting performance of the DE/SVR and DE/GPR approaches (with or without NARX) using published data of gold spot prices from the New York Commodities Exchange (COMEX). The numerical results obtained have shown a better performance of the NARX DE/SVR hybrid technique than the other machine learning techniques according to RMSE statistic. The findings of this research work are in line with some previous studies, which confirmed the superiority of SVR and GPR models over other classical techniques in relative research areas.

Automated summary

In this study, three different strategies for automatic lag selection in time series analysis are proposed. The two novel hybrid methods, NARX SVR and GPR, combined with DE optimizer, show better performance in gold spot price forecasting compared to other machine learning techniques.