Assessing the predictability of time-series of population abundance

Biodiversity is declining fast so we need robust tools to predict how biodiversity will respond to changes in land-use and climate. Available global biodiversity indicators, such as the Living Planet Index, help us prioritise conservation resources and evaluate the effectiveness of conservation and policy interventions and show progress towards environmental targets. But is it possible to accurately extrapolate these indicators spatially and/or project them into the future? And if so, how far can we reliably project them and when do predictions become too inaccurate to be useful? To assess the predictability of biodiversity trends, we apply a set of models to predict inter-annual change within time-series of vertebrate population abundance based on historical land use and climate data and assess their performance against withheld data. For this, we used a hindcasting validation approach. For a time series of length t we stepwise removed x years (ranging from 1-10) a) at the tail end of the time series, b) at the beginning of the time series and c) randomly in the middle of the time series creating gaps of maximum length x. The removed data points in each time-series represented our test set and the remaining data points the training set. We then applied random forest and linear mixed effects models to the training data, with relative population change and Relative Percent Difference (RPD) between years as response variables. As this analysis is in progress, we present a sample of results for random forest models, and discuss how we plan to progress the work in order to provide a complete assessment of predictability of vertebrate population trends. Being able to accurately predict population trends is important as population declines can be a prelude to extinction and – if we get them right - predicted trends could be used to determine a species’ extinction risk via IUCN criterion A