Multiscale Legendre Neural Network for Monthly Anchovy Catches Forecasting

In this paper, a Legendre neural network (LNN) combined with multi-scale stationary wavelet decomposition is used to improve the prediction accuracy and parsimony of monthly anchovy catches forecasting in area north of Chile. The general idea behind this approach is to decompose the observed anchovy catches data into low frequency (LF) component and high frequency (HF) component using the multi-scale stationary wavelet transform to separately forecast each frequency component. In wavelet domain, the LF component and HF component are predicted with a linear autoregressive (AR) model and a LNN model; respectively. Hence, the proposed forecast is the co-addition of two predicted components. We find that the proposed forecasting method achieves 99% of the explained variance with reduced parsimony and high accuracy.