Multiresolution radial basis model for nonlinear time series prediction
Rukun Santoso, Subanar Subanar, Dedi Rosadi, Suhartono Suhartono · AIP conference proceedings · 2016
The multiresolution radial basis autoregressive model for nonlinear time series prediction is proposed in this paper. This is a development form of the multiresolution autoregressive model. This constitutes an alternative method for nonlinear time series prediction, especially for threshold model. In the beginning, time series is decomposed by wavelet transform to produce smooth part and detail part. The both are representing the main pattern and noise pattern of time series, respectively. Some radial basis functions are performed to refine the decomposition result into homogenous clusters. This refining result becomes an input series for a neural network structure. The model is applied to make one step prediction ahead of a simulation data. The normality test for residual is accepted, and the statistic test shows that the model coefficients are significant.