Development of Model Building Procedures in Wavelet Neural Networks for Forecasting Non-Stationary Time Series
Suhartono Suhartono, Subanar Subanar · European journal of scientific research · 2009
The aim of this research is to study further some latest progress of nonlinear time series analysis, particularly about Wavelet Neural Networks (WNN). There are three main issues that are considered further in this research. The first is some properties of scale and wavelet coefficients from Maximal Overlap Discrete Wavelet Transform (MODWT) decomposition, particularly at non-stationary time series. The second is about development of model building procedures of WNN based on the properties of scale and wavelet coefficients. Then, the third is empirical study about the implementation of procedures that have been developed and comparison study about the forecast accuracy of WNN to other models. The results show that scale coefficients of MODWT at non-stationary has trend pattern and wavelet coefficients are stationary. The results of model building procedure development yield four procedures for non-stationary time series. In general, these procedures accommodate input lags of scale and wavelet coefficients introduced by Renaud et al. (2003) and other additional lags. The result of empirical study shows that these procedures work well for finding the best WNN model for time series forecasting. The comparison study of forecast accuracy show that the first procedure (with stepwise) of nonstationary time series yields the best forecast compared to ARIMA, MAR and WNN models by using other procedures. It’s showed by the smallest RMSE at testing data.