How accurate are different forecast models for predicting population dynamics? That, and how predictable various animals actually are, was tested in the study “Complexity is costly: a meta-analysis of parametric and non-parametric methods for short-term population forecasting”, by Eric J. Ward and colleagues, that is now published online in Oikos. Below is the author’s summary of the study:
Forecasting ecological data presents a unique set of challenges compared to other types of time series data (stock prices, weather) – two of the most common sources of uncertainty arise from (1) scientists not measuring populations perfectly, and (2) mechanisms responsible for population fluctuations are generally complex and not measureable at a population-wide scale (e.g. density dependence). Many ecological and fisheries models are made complex in an attempt to capture biological realism. Recent work on simulated and real datasets (Perretti et al. 2013 PNAS; Sugihara et al. 2012 Science) has shown that more accurate predictions can be made from simpler non-mechanistic models. Our paper presents the results of a forecasting competition, comparing a wide range of time series models to ~ 2400 time series, representing a range of vertebrate taxa. We found that in general, the best 1-5 year forecasts originated from simple models, such as a random walk (where the predicted population size is the current population size). Taxa that have strongly cyclic population dynamics, such as sockeye salmon, are the easiest to forecast, and warrant the use of more complex types of non-mechanistic models. Across all marine fish species, we found that longer lived species, or those with larger body size are easier to predict (presumably because they have smaller recruitment variability). Similarly, for birds, we found that higher trophic levels were also correlated with better predictions.
All of the time series included in our analysis were relatively long in ecology (25 continuous data points). The failure of many of the methods we considered suggests that improvement in forecasting ability is unlikley to come from better non-mechanistic forecasting methods or more annual population data; instead we recommend that efforts be made to better understand environmental drivers, which can be included as covariates.