Comparison of Some Nonparametric Methods Using Simulations

Abstract

 There was a need to compare the nonparametric methods used in predicting time series to find the most efficient method for forecasting, and this was in fact the main objective of this study. The research problem was summed up that parametric statistical methods need conditions and criteria that may be difficult to meet, so it was necessary to search for new methods of prediction other than the traditional methods, which are non-parametric methods. The comparison was through the application of the simulation method, where the data was generated by the restricted generation equation based on the original data, with the lowest value and the highest value for each type of data. The focus of the study is a comparison between the two methods of neural networks (the back propagation error network) and the kernel method (Nadaraya-Watson estimator). As for the sample sizes, the sizes (15, 30, 60 and 100) were selected. As for the criteria that were used for comparison, they were (MSE, RMSE, MAE, MAPE).The results showed, through simulation, that the neural networks give better and more efficient results than the Kernel method in most of the scales and for all sample sizes.