A random walk through the trees: Forecasting copper prices using decision learning methods

Abstract

We investigate the accuracy of copper price forecasts produced by three decision learning methods. Prior evidence (Liu et al. Resources Policy, 2017) shows that a regression tree, a simple decision learning model, can be used to predict copper prices for both short-term and long-term horizons (several days and several years, respectively). We contribute to this literature by evaluating more sophisticated decision learning methods based on trees: random forests and gradient boosting regression trees. Our results indicate that random forests and gradient boosting regression trees significantly outperform regression trees at forecasting copper prices. Our analysis also reveals that a random walk process, recognized in the literature as one of the most useful models for forecasting copper prices, yields competitive out-of-sample forecasts as compared to these decision learning methods.