Time Series Forecasting of Gold Prices with the Help of Its Decomposition and Multivariate Adaptive Regression Splines

This research presents amethodology for the forecasting of gold prices using as input information the values of this metal in the previous months and the values of others like potash, copper, lead, tin, nickel, aluminum, iron ore, zinc, platinum and silver. The proposed methodology is based on the decomposition of each of the time series in their trend, seasonal and random components and the use of the trend information as independent variables in a multivariate adaptive regression splines model. The performance of the method was tested with the help of a database of the monthly prices of the aforementioned raw materials. The information available starts in January 1960 and goes up to September 2020. The prediction of gold prices from October 2019 to September 2020 showed in the month-by-month prediction model a mean absolute deviation (MAD) of 67.6022, mean square error (MSE) of 9403.1882, root mean square error (RMSE) of 96.9700 and mean absolute percentage error (MAPE) of 3.8803%. In the case of forecasts up to 12 months ahead, the results were a MAD of 293.4832, MSE of 284499.4718, root mean square error of 533.3849 and MAPE of 15.7366%. The results obtained were compared with those given by a multivariate adaptive regression model that made use of the original time series as input data.

Automated summary

This research presents a methodology for forecasting gold prices using historical values of metals as input information. The proposed method decomposes the time series into trend, seasonal, and random components, and uses trend information as independent variables in a regression model for predicting gold prices. The empirical results indicate that the method performs well in both short-term and medium-term forecasts.