Forecasting copper price using VAR and the XGBoost model: an experiment with a relatively small dataset

Abstract: Given the importance of copper prices to investors, governments, and policymakers, this paper investigates short-term price predictability using VAR and XGBoost models. All models are trained with historical data from November 2021 to December 2022 and using MSE, RMSE and MAE for evaluating the model performance. The results show that the XGBoost model outperforms VAR models, implying that machine learning models are more robust than traditional statistical models. However, specific scenarios with a lag of 1to predict one day ahead(h=1day) show similar performance between XGBoost and VAR, indicating that traditional statistical models can still be competitive in certain situations. Therefore, it is critical not to dismiss traditional statistical models entirely, as they provide benefits in terms of interpretability and computational simplicity. Moreover, we also find that the selection of lag values for models is demonstrated to be empirical, with different lag values resulting in varying model performance. Thus, practitioners are encouraged to experiment with different lag settings in order to find the best model for their specific tasks and dataset sizes.