Ecological forecasts reveal limitations of common model selection methods: predicting changes in beaver colony...

Area of Science:

• Ecology
• Population Dynamics
• Ecological Modeling

Background:

• Ecology increasingly emphasizes prediction, driven by information criteria for model selection.
• Few studies validate information criteria's ability to identify the most predictive and parsimonious ecological models.
• Assessing predictive accuracy is vital for robust ecological forecasting.

Purpose of the Study:

• To test information criteria's accuracy in predicting beaver colony density fluctuations.
• To evaluate the influence of density dependence and independent factors on beaver population dynamics.
• To assess model performance using an ecological forecasting framework.

Main Methods:

• Modeled beaver colony densities using a discrete-time Gompertz model.
• Compared four models: density-independent (with/without covariates) and density-dependent (with/without covariates).
• Evaluated forecasting accuracy by withholding one-third of observations for hindcasting.

Main Results:

• Both information criteria and forecasting metrics supported compensatory density dependence in beaver populations.
• The most complex model (density-dependent with covariates) showed strong within-sample fit but over-fit the data.
• A simpler density-dependent model without covariates performed nearly as well in out-of-sample predictions.

Conclusions:

• Information criteria can lead to over-fitting, potentially misinterpreting ecological dynamics.
• Evaluating ecological models with long-term data and forecast performance metrics is essential.
• Comparing forecast metrics improves understanding of factors driving population dynamics.