Forecasting Natural Gas Prices in the United States Using Artificial Neural Networks

Abstract

Prediction of the natural gas price is imperative to producers, suppliers, traders, market makers, and bankers involved in the natural gas exploration, production transportation, and trading. Additionally, consumers are also highly affected by the changes in the price of oil and gas products. Several attempts have been made to model the energy commodity prices over the past few decades. Stochastic differential equation, linear and nonlinear regression, auto regression, and neural networks are the main techniques that have been implemented. In this thesis, three different categories of models are examined which are, stochastic differential equations, ARIMA, and autoregressive neural networks. The results indicate that, the NAR neural network provides a better fit to the given data as compared to the other proposed models. The three-layer NAR model with 6 hidden neurons was found to have the best performance in terms of one month ahead price prediction. The accuracy of the NARX model with 6 neurons was found to be higher than that of the other models. Although, this model provides a reasonable fit to the given data, it fails to capture the price spikes effectively. The sensitivity analysis shows that CDD/HDD temperatures, extreme minimum temperature, and WTI oil prices have an insignificant effect on the results. On the other hand, total consumption, total production, and mean temperature of weather impact the results significantly.