Commodity Futures Pricing Via Machine Learning: An Empirical Approach

Sammanfattning: The goal of this thesis is to use established methodologies in the field of machine learning in finance to extend the list of current applications to commodity futures, reviewing and refining the established empirical approaches to return forecasting and hyperparameter optimization. We thus investigate the out of sample predictive accuracy of tree-based machine learning (ML) techniques and neural networks applied to monthly commodity futures returns, relying on conventional regression and classification accuracy metrics. We find that a large selection of machine learning techniques cannot consistently outperform the benchmark AR(1) model when applied to monthly data, and that there is no specific ML method suitable to all the analyzed commodity series at once. While we also find potential portfolio-level investor gains from using ML techniques, the robustness of these gains is questionable. Finally, we suggest an updated approach to hyperparameter optimization and find that different commodity series have to be modelled separately, which is suggested by large differences in the optimal architectures estimated by our Grid Search and Bayesian Optimization tuner algorithms.