Comparing the Forecasting Performance of Futures Oil Prices with genetically Evolved Neura

The increase in oil price volatility over the past two decades has served as an impetus in the ongoing effort to improve the quality of forecasts.  Pronounced swings in oil prices have come to be more of the norm rather than a rare occurrence and have been attributed to fluctuations on both the supply side and to shifts in market powers on the demand side.  This in turn has rendered the task of predicting oil prices to become more challenging and prone to larger forecast errors.  Because oil price forecasts are vital to many sectors in the economy, researchers and practitioners alike are working hard at identifying models that can improve the accuracy of their predictive ability. This paper focuses on designing a hybrid model that combines genetic training algorithms (GAs) with artificial neural networks (ANNs) architecture to predict oil prices and compares it to that of oil futures forecasts for different time horizons.   The approach taken in this research relies on processing information more efficiently so as to improve oil price forecasts.  The model’s forecasting performance is evaluated and compared to that of the futures contract prices by conducting two tests: accuracy and correctness.  While accuracy measures the degree of precision of the forecast by computing mean absolute forecast errors (MAFE), correctness tests the model’s ability to predict the direction of the movement.  Correctness calculates the percentage of times the forecast is on the ‘correct side’ of the realized future price.  The significance of the second test relates to rewards which market participants can realize from hedging or speculation. The basic network architecture design used in this paper is that of the multi-layer perceptron.  Using oil prices, the network through a process of supervised learning undergoes training until a desired pre-specified accuracy level is achieved.  The learning process relies on the iterative procedure of processing inputs through the network, measuring the errors, and adjusting the weights accordingly and is called “back propagation”.   Back propagation, a variation of the Delta Rule, is the most popular technique in forecasting financial data.   Because the results obtained from the ANN model are derived from data using 20-20 hindsight, it is expected that their forecasts will be more accurate and correct compared to those of future oil prices.  Moreover, the model’s ability to process information more efficiently and to identify patterns that may be ill defined provides further support for this work.