Modeling Crude Oil Price Chaotic Behavior Using Machine Learning

Abstract

The oil prices and its future are one of the most prominent topics in our world nowadays. In 2014, oil prices are down sharply, losing over 50% of their value since June peak, when West Texas Intermediate was $115 a barrel and it is now below $51 [1]. This has raised a number of questions: What are the factors and reasons that led to this change in prices? What is the effect of this decline on the countries that are highly dependent on oil based economy or importing countries? Falling oil prices have both positive and negative impacts. On one hand for many people, cheaper oil means lower fuel prices and economies of importing countries may rebound a bit if the declining prices are exploited and on the other hand oil-exporting countries are very hardly hit. Similarly rise in prices have opposite effects in both directions. Prediction of oil prices is an important task and difficult challenge at the same time under a number of complex factors that influence the determination of oil prices such as political, economic, climate factors and so on. Experts and analysts indicated the importance of expected future oil prices to support the global economy, companies and institutions to hedge against surprise changes to make sound decisions and building a healthy and successful econo-my. This volatile behavior is predicted to prompt more new and interesting research challenges. In the present research, machine learning and computational intelligence ap-proaches are used to predict crude oil prices using direct prediction and combined prediction models. In this research, before constructing a computational model, several aspects of initial preparation of data were selected, which consists of 14 input as attributes to predict the West Taxes Intermediate (WTI) as output. Normalization, feature selection and data partition are used for preparation of the inputs. Then several direct prediction models that have shown good (a priori) performance on datasets similar to the prediction of crude oil prices were examined. Radial basis function neural network outperformed other methods in obtaining the prediction error. In order to improve the accuracy of the direct models, different combined prediction models were used, which include Meta learning schemes, hybrid and ensemble models. For the generalized ensemble method, V Particle Swarm Optimization method was used to determine the optimal weights and obtained the best results. Volatility Implied Equity Index (VIX), West Texas Intermediate (WTI), New York Harbor conventional gasoline spot prices (GPNY), Exchange rate (ER), and Future contracts 1 (FC1) are the most important factors to determine the crude oil price. The generalized ensemble is a good model to explore and explain crude oil market‘s rules with 80% and 20% of the data for training and testing. Comparison with different results in the literatures presented further proved the effectiveness and superiority of the gener-alized ensemble model for the prediction of the WTI crude oil price. Modern society relies on crude oil dramatically. Crude oil price is affected by many factors and coupled by an international network of thousands of producers, refiners, marketers, traders, and consumers for buying and selling physical volumes of crude oil. So there is also a great need to understand, organize, analyze and explain the behavior of the crude oil price market and different aspects of international crude oil pricing and prediction using novel information enterprise architecture for crude oil pricing and prediction based on Zachman framework were discussed. This novel infor-mation architecture leads to a deeper understanding of the comprehensive structure of the crude oil market. Keywords: Prediction crude oil price, Machine learning, Information enterprise architecture, Zachman framework.