Haar Wavelet Neural Network for Multi-step-ahead Anchovy Catches Forecasting

This paper proposes a hybrid multi-step-ahead forecasting model based on two stages to improve pelagic fish-catch time-series modeling. In the first stage, the Fourier power spectrum is used to analyze variations within a time series at multiple periodicities, while the stationary wavelet transform is used to extract a high frequency (HF) component of annual periodicity and a low frequency (LF) component of inter-annual periodicity. In the second stage, both the HF and LF components are the inputs into a single-hidden neural network model to predict the original non-stationary time series. We demonstrate the utility of the proposed forecasting model on monthly anchovy catches time-series of the coastal zone of northern Chile (18S-24S) for periods from January 1963 to December 2008. Empirical results obtained for 7-month ahead forecasting showed the effectiveness of the proposed hybrid forecasting strategy.