Hybrid wavelet-RBFNN model for monthly anchovy catches forecasting

Abstract

A hybrid method to forecast 1-month ahead monthly anchovy catches in the north area of Chile is proposed in this paper. This method combined two techniques, stationary wavelet transform (SWT) and radial basis function neural network (RBFNN). The observed monthly anchovy catches data is decomposed into two subseries using 1-level SWT and the appropriate subseries are used as inputs to the RBFNN to forecast original anchovy catches time series. The RBFNN architecture is composed of linear and nonlinear weights, which are estimates using the least square method and Levenberg-Marquardt algorithm; respectively. Wavelet-RBFNN based forecasting performance was evaluated by comparing it with classical RBFNN model. The benchmark results shown that a 99% of the explained variance was captured with a reduced parsimony and high speed convergence. © Springer-Verlag Berlin Heidelberg 2008.