The Use of Generalized Additive Models for Forecasting the Abundance of Queets River Coho Salmon

We examined three types of models for preseason forecasting of the abundance of Queets River coho salmon Oncorhynchus kisutch: (1) a simple model in which estimates of smolt production are multiplied by projected marine survival rates, (2) a Ricker spawner-recruitment model, and (3) a regression model relating log-transformed adult recruitment to smolt production. Each type of model was formulated with and without environmental variables that influence production and survival. We attempted to use a nonparametric generalized additive model (GAM) to guide the selection of the environmental variables and the form of the regression model. The GAM model was derived through a stepwise selection strategy based on the Akaike information criterion. Parametric approximate models were developed for each selected GAM model, and their performance was compared with postseason estimates of abundance using three criteria: the mean absolute percentage error, the largest absolute percentage error, and the probability of being included in the 90% prediction interval. This paper shows that the GAM approach is useful in constructing forecasting models by identifying promising relationships with predictor variables and improving abundance forecasts through the incorporation of environmental variables.