Gold risk premium estimation with machine learning methods

This paper assesses the accuracy of several machine learning models’ predictions of the gold risk premium when using an extensive set of 186 predictors. We perform an out-of-sample evaluation and consider both statistical and portfolio metrics. Our results show that machine learning methods and forecast combinations have a limited ability to outperform the historical mean when predicting the gold risk premium. Slightly better results are obtained when predictors are used individually. More specifically, we find that several technical indicators (moving average and momentum series) have forecasting power during periods of expansion, while several business cycle variables and geopolitical risk variables help predict the gold risk premium during recessions. An economic evaluation accounting for transaction costs shows that investors using machine learning methods to estimate expected returns on gold should anticipate limited portfolio gains.