Forecasting crude oil price using neural networks

Abstract

This research constructed the Artificial Neural Networks (Multilayer Feed Forward) to forecast the crude oil price (Brent). The input information was the daily price range between December 27, 2002 to March 18, 2005. Total number of inputs were 561 days. Arranging the input information into groups with 10 consecutive informations in each group, 551 groups were prepared. The model consisted of 10 neurons in the input layer and 1 neuron in the output layer. Quadratic interpolation was utilized to search for the most appropriate number of neurons in the hidden layer. The research question was how many neurons in the hidden layer that would yield the most-accurate forecasting result. The comparisons of models were justified by the 1 day ex ante forecasting results. The Mean Absolute Percentage Error (MAPE) was a measurement of the accuracy. Thirty-four rounds of the forecasting contest were conducted. The least MAPE derived from the best model was 1.98 percent with 200 neurons in the hidden layer.