Forecasting the Price of Aluminium Using Machine Learning : Empirical comparison of machine learning and statistical methods.

Abstract

This thesis challenges statistical methods for forecasting aluminium's 3-month futures contract price with a machine learning technique. The goal is to find the superior model with a one-month horizon. The first model we apply is the traditional statistical method, random walk, as a benchmark. After that, we examine the relationship between the futures and spot price by conducting the Johansen cointegration test. This test concludes that the two prices are cointegrated. Exploiting this relationship, we conduct a forecast using the more advanced statistical method, the vector error correction model (VECM). Lastly, we compare the results to the popular machine learning technique XGBoost. To validate the models' predictions, we use the time series cross-validation technique called rolling cross-origin. After that, we evaluate the performance of the three models by comparing the root mean square error (RMSE) and mean absolute error (MAE). Our research finds that the XGBoost method stays relatively stable through multiple robustness tests in contrast to the VECM, varying significantly for the majority of the tests. In conclusion, the most reliable and accurate model to use when forecasting the 3-month futures contract price is XGBoost.