Prediction and Trading in Crude Oil Markets Using Multi-Class Classification and Multi-Objective Optimization

Shangkun Deng,Chenguang Wang,Haoran Yu,Fangjie Ma,Xiaoru Huang,Jiashuang Shen,Tianxiang Yang,Hongyu Tian

doi:10.1109/access.2019.2960379

Abstract

Crude oil price direction forecasting presents an extremely challenging task that attracts considerable attention from academic scholars, individual investors and institutional investors. In this research, we proposed an integration method by adopting the Multi-Class Support Vector Machine (MCSVM) and the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) for forecasting and trading simulation in two well-known crude oil markets. Firstly, the proposed approach applied the MCSVM to train a multi-class classification model, and it adopted the NSGA-II to optimize the threshold values of trading rules. Then, the trained MCSVM model was used to forecast the movement direction and magnitude levels. Next, the proposed method forecasted the direction of crude oil price movements one week later and executed trading simulation according to the direction and magnitude level predictions. Finally, after a testing period lasted for four years, the performances of the proposed approach were gauged in terms of direction prediction correctness and investment yields. Experimental results demonstrated that the proposed approach produced outstanding results not only on hit ratio and accumulated return but also return-risk ratio. It indicates that the proposed approach can provide beneficial suggestions for individual investors, institutional investors, as well as for government officers engaged in energy investment policies making.

Highlights

Crude oil is recognized as one of the essential energy sources, and it is one of the largest and most actively traded commodities in the world [1]
HIT RATIO RESULTS To measure the performance of crude oil price direction forecasting for the proposed method Multi-Class Support Vector Machine (MCSVM)-Non-Dominated Sorting Genetic Algorithm II (NSGA-II), the ANN, Support Vector Machine (SVM), MCSVM and MCSVM-GA based methods were chosen as the benchmarks, and their experiments were conducted on the crude oil price of Brent and West Texas Intermediate (WTI)
From the hit ratio results of the proposed approach and benchmark methods for WTI market shown in Table 6, we can observe that the average hit ratio over four testing years produced by the ANN, SVM, MCSVM and MCSVM-GA were 48.62%, 51.07%, 38.75%, and 53.67%, respectively, while our proposed method MCSVM-NSGA-II produced the highest average hit ratio of 62.16%, which was significantly better than other benchmark methods; Within the four years from 2015 to 2018, the proposed method performed best in two years (2017 and 2018)

Summary

Introduction

Crude oil is recognized as one of the essential energy sources, and it is one of the largest and most actively traded commodities in the world [1]. West Texas Intermediate (WTI) and Brent are the two most influential ones. A large group of individual and institutional investors, such as governments, energy-related companies, market investors, as well as hedge funds are all involved in these two crude. If the direction of crude oil price fluctuation can be predicted correctly to some extent, it will be of great benefits for them, for instance, it would be extremely beneficial for the government to formulate policies, helpful for energy-related enterprises to make decisions, and it could provide investment advice for individual investors and institutional investors. As a traditional time series analysis method, ARIMA (Autoregressive Integrated Moving Average) model [3] has been applied to crude oil prices prediction by a lot of researchers around the world. Moshiri and Foroutan predicted crude oil future prices using an ARIMA

Methods

Results

Conclusion