Forecasting crude oil market crashes using machine learning technologies

Abstract

To the best of our knowledge, this study provides new insight into the forecasting of crude oil futures price crashes in America, employing a moving window. One is the fixed-length window and the other is the expanding-length window, which has never been reported in the past. We aimed to investigate if there is any difference when historical data are discarded. As the explanatory variables, we adapted 13 variables to obtain two datasets, 16 explanatory variables for Dataset1 and 121 explanatory variables for Dataset2. We try to observe results from the different-sized sets of explanatory variables. Specifically, we leverage the merits of a series of machine learning techniques, which include random forests, logistic regression, support vector machines, and extreme gradient boosting (XGBoost). Finally, we employ the evaluation metrics that are broadly used to assess the discriminatory power of imbalanced datasets. Our results indicate that we should occasionally discard distant historical data, and that XGBoost outperforms the other employed approaches, achieving a detection rate as high as 86% using the fixed-length moving window for Dataset2.