Prediction of time series using wavelet transform and neural network

Abstract

In this thesis, we propose a combination of the wavelet transform and the neural network for time series prediction. Three main computational steps were applied: the wavelet decomposition, prediction by neural networks, the wavelet reconstruction. The method was applied to three sets of data: the Crude Oil Prices, the Sunspot time series, and the Mackey-Glass time series, and the normalized mean square errors of the prediction are 0.08936, 0.10152, and 0.00045, respectively. For a comparison purpose, we predicted the same data sets using only a neural network, and the normalized mean square error were 0.10192, 0.10828, and 0.00092, respectively. In addition, we found that the correlation coefficients of the scatter plot of the prediction from our method range from 0.90 to 0.99.