Long-term forecast of energy commodities price using machine learning. (15th July 2019)

• Record Type: Journal Article
• Authors: Herrera, Gabriel Paes; Constantino, Michel; Tabak, Benjamin Miranda; Pistori, Hemerson; Su, Jen-Je; Naranpanawa, Athula
• Is Part Of: Energy. Volume 179(2019)
• Page Start: 214
• Page End: 221
• Publication Date: 2019-07-15
• Subjects: ANN -- Random forests -- Natural gas -- Coal -- Oil; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2019.04.077
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: We compare the long-horizon forecast performance of traditional econometric models with machine learning methods (Neural Networks and Random Forests) for the main energy commodities in the world using monthly prices provided by the International Monetary Fund (IMF). We study the case of Oil (Brent, WTI and Dubai Fateh), Coal (AU) and Gas (US and Russia). Models accuracy are measured using RMSE and MAPE and the M-DM test is applied to evaluate whether there is a statistically significant difference between the methods. We computed thousands of tests regarding the machine learning parameters combinations as there is no method to set the optimal structure for these models. The results show that machine learning methods outperform traditional econometric methods and also that they present an additional advantage, which is the capacity to predict turning points. This study adds further evidence for the discussion on the use of machine learning algorithms for the development of more accurate forecasts to support policymakers and help the decision-making process in the international energy market. Highlights: Two machine learning techniques and traditional econometric methods are compared. Random forests outperforms neural networks and econometric methods. Great results using random forests without exogenous variables for long-term forecast.