Comparison of the forecasting accuracy of neural networks with other established techniques

Abstract

A comparison of the forecast accuracy of artificial neural networks is made to other more established forecasting methodologies. Eight different types of forecasts were developed on a daily basis for five months and results analyzed. The MAPE (mean absolute percent error) was computed for each model. The series being forecast was the total system load for the Puget Sound Power and Light Company. The performance of the neural nets was disappointing with all but one of the other techniques outperforming them. Although the neural nets did not do well in this competition, this may be caused by a lack of forecasting experience by the neural net developers rather than limitations in the abilities of nets themselves. Forecasts made with neural nets using the same inputs showed dramatic improvements but the performance was still not as good as the best regression forecast.<>