The cost of complexity in forecasts of population abundances is reduced but not eliminated by borrowing information across space using a hierarchical approach. Issue 2 (10th December 2019)

• Record Type: Journal Article
• Authors: Chevalier, Mathieu; Knape, Jonas
• Is Part Of: Oikos. Volume 129:Issue 2(2020)
• Page Start: 249
• Page End: 260
• Publication Date: 2019-12-10
• Subjects: baseline models -- bird abundances -- forecast performance -- hierarchical models -- population dynamics -- scoring rules; Ecology -- Periodicals; 570
• DOI: 10.1111/oik.06401
• Languages: English
• ISSNs: 0030-1299

Abstract

Abstract : Anticipating ecological changes is paramount if we are to manage biodiversity and the services they provide to humanity. When forecasting population abundances, studies have shown that simple statistical models often have better forecast performance than complex models. These studies have evaluated forecasts of models fitted separately to data from single sites (single‐site approach). Here, we aim to contrast the forecast performance and forecast horizon between a single‐site approach and a hierarchical multi‐site approach where a single model is fitted to data from multiple‐sites, and to investigate how they vary with model complexity. We used 5273 population time series on 84 species from the Swedish breeding bird survey program, and found that simple models on average had better forecast performance and forecast horizon than complex models for both the single‐ and the multi‐site approach. However, the cost of complexity was considerably reduced under the multi‐site approach, while the proportion of species for which complex models had better forecast performance than simple models was also much larger than under the single‐site approach. This suggests that the multi‐site approach is useful for inclusion of more detailed processes which may benefit forecasts for some species and which are of importance for managers. Still, our results are in line with some previous studies suggesting that it is surprisingly difficult to construct complex models that, on average, beat trivial baseline forecasts.