Forecasting of jack mackerel landings ( Trachurus murphyi) in central-southern Chile through neural networks.

In the present study, the performance of neuronal networks models in monthly landing forecasting of jack mackerel ( Trachurus murphyi) in central-southern Chile (32°S-42°S) was assessed. Thus, monthly estimations for 10 environmental variables, fishing effort (fe) and jack mackerel landings for the period 1973-2008 were used. A preliminary analysis was done in order to remove strongly correlated variables. Sea surface temperature ( SST) and fe are established as input variables, then, a non-linear cross correlation analysis was performed to estimate the lag between the input variables and jack mackerel landings. Two models were adjusted: model one includes both training and testing cases randomly selected using all data involved in the analysed period; for model 2, the data is divided into two time series: the first from 1973 to 2002 used for training and the second between 2003 and 2008 used for validation. The external validation process for model 1 showed an explained variance of 92%, with a standard forecasting error of 30%. The explained variance for model 2 was 81%, with a standard forecasting error of 38%. Finally, the sensitivity analysis for both models showed the fe as the most influential variable to jack mackerel landings, which presents functionality depending on anthropogenic effects rather than environmental conditions. [ABSTRACT FROM AUTHOR]