Forecasting Iranian Crude Oil Price Using Artificial Neural Network and ARIMA Models

The aim of this study is to introduce more optimal models to forecast Iranian crude oil prices. The study uses weekly data for the period 1987-2010, to separately forecast 10, 20 and 30 percent of data variables. The study applies 4 Artificial Neural Networks and one ARIMA regression model. The selected Artificial Neural Networks are Feed-Forward Back Propagation, Cascade Back Propagation, Elman Back Propagation and Generalized Regression. The experimental functions are Levenberg-Marquardt and Quasi-Newton BFG. The findings indicate that for 10 percent of price data networks of Generalized Regression and Quasi-Newton BFG based Cascade Back Propagation networks give the best forecasts with errors of less than 1 and 2 percent respectively. To forecast 20 percent of Iranian Crude oil prices Feed-Forward Back Propagation and Elman Back Propagation networks based on the Levenberg-Marquardt experimental functions had the best performance. In the case of 30 percent of price data also Feed-Forward Back Propagation was found more optimal. The results also indicate that as the percentage of data forecast increases prediction accuracy tends to decrease and this happens most markedly when we increase the percentage of data used from 10 percent to 20 percent. The study also reveals a lower forecasting power for the ARIMA model compared to all the other models.