Authors: Yinan Hu, Theodore B. Trafalis
Addresses: Petroleum Engineering, Mewbourne School of Petroleum and Geological Engineering, University of Oklahoma, Sarkeys Energy Center, 100 East Boyd Street, Ste T-301, Norman, OK 73019-1003, USA. ' School of Industrial Engineering, University of Oklahoma, 202 West Boyd, Rm. 124, Norman, OK 73019, USA
Abstract: Natural gas prices show a non-linear, non-stationary, and time variant behaviour. In this study, we build a regression function for daily natural gas prices using ε-SVR and v-SVR and experiment with different kernels. We compare the proposed methods with artificial neural networks, RBF networks and asymmetric GARCH models. The comparison results demonstrate that the v-SVR with sigmoid kernel is the best of the compared techniques with respect to the mean square error and squared correlation coefficient criteria. The paper is also extended to discuss the price tendency prediction and the performance of SVR without data imputation. The purpose of the paper is to provide an effective way to predict the short term gas price, which can be used as a tool to reduce uncertainty and financial risk in the energy market.
Keywords: support vector machines; SVM; v-SVR; ε-SVR; sigmoid kernel; artificial neural networks; ANNs; radial basis function networks; RBF networks; GARCH; data imputation; asset pricing; natural gas prices; price prediction; short term prices; uncertainty; financial risk; energy markets.
International Journal of Financial Markets and Derivatives, 2011 Vol.2 No.1/2, pp.106 - 120
Published online: 28 Feb 2015 *Full-text access for editors Access for subscribers Purchase this article Comment on this article