Premium
Evaluation of non‐traditional modelling techniques for forecasting salmon returns
Author(s) -
McCormick J. L.,
Falcy M. R.
Publication year - 2015
Publication title -
fisheries management and ecology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.693
H-Index - 55
eISSN - 1365-2400
pISSN - 0969-997X
DOI - 10.1111/fme.12122
Subject(s) - overfitting , spurious relationship , oncorhynchus , econometrics , regression , artificial neural network , fishery , computer science , statistics , feature selection , artificial intelligence , machine learning , mathematics , fish <actinopterygii> , biology
Forecasting adult salmon abundance is problematic when the number of observations is small relative to the number of potential explanatory variables. Machine learning and other non‐traditional techniques employ algorithms designed to prevent model overfitting. Data from 18 coho salmon, O ncorhynchus kisutch (Walbaum), and seven Chinook salmon, O ncorhynchus tschawytscha (Walbaum), populations on the Oregon coast were used to evaluate the forecast performance of artificial neural networks, elastic net, least absolute shrinkage and selection operator, principal component regression ( PCR ) and ridge regression ( RR ) compared to several more traditional techniques. In general, the non‐traditional modelling techniques evaluated in this study performed similarly to the traditional techniques with the exception of sibling regression. This suggests that they have merit for improving actual predictions. Among the non‐traditional techniques, PCR resulted in the lowest prediction error for the coho salmon populations, and RR predicted Chinook salmon returns most accurately. The techniques explored are not an easy solution to a difficult problem. Spurious conclusions about the processes that generate salmon returns still may result as evidenced by the inclusion of an unrelated variable in many of the non‐traditional models.