Properties of the gold price : an investigation using fractional Brownian motion and supervised machine learning techniqes

Abstract

This thesis investigates properties of the gold price. Two different aspects have been analyzed using two different methods. The concept of fractional Brownian motion has been utilized in the search for evidence of long memory in the gold price. Further, the machine learning techniques Gradient Boosting Machine and XGBoost are used in the investigation of the relative importance of financial and economic variables in a prediction of the gold price. Both aspects are studied across multiple time periods in order to examine the potential change. We find evidence of long memory in the gold price, however not in all examined time periods from 1979 until now. The first ten years from 1979, as well as the ten years from the start of the 2008 financial crisis to today, appear to have the property of persistence, while the years in between display no evidence of long memory. The second part of the analysis, the relative importance of variables in a prediction, is more focused on the years before and after the 2008 financial crisis. We find that variables such as crude oil and durable goods orders are relatively important before the crisis, while crude oil and US 10-year treasury bonds are relatively important after. Both parts of the analysis indicate that the properties of the gold price change over time.