Quantitative predictions are ubiquitous in ecology, yet there is limited discussion on the nature of prediction in this field. Herein I derive a general quantitative framework for analyzing and partitioning the sources of uncertainty that control predictability. The implications of this framework are assessed conceptually and linked to classic questions in ecology, such as the relative importance of endogenous (density dependent) versus exogenous factors, stability versus drift, and the spatial scaling of processes. The framework is used to make a number of novel predictions and reframe approaches to experimental design, model selection, and hypothesis testing. Next, the quantitative application of the framework to partitioning uncertainties is illustrated using a short-term forecast of net ecosystem exchange. Finally, I advocate for a new comparative approach to studying predictability across different ecological systems and processes and lay out a number of hypotheses about what limits predictability and how these limits should scale in space and time.
The first principles framework partitions forecast uncertainty into five components: endogenous, exogenous, parameter uncertainty, parameter variability, and process error. Endogenous uncertainty grows or declines exponentially according to classic stability thresholds and is expected to increase in importance at larger scales. Exogenous uncertainty is expected to increase through time but average out across space, though this is slowed by autocorrelation. Parameter uncertainty will dominate data limited problems, but declines asympotically with sampling, while parameter variability and process error do not decline with sampling but can decline with scale, again depending on autocorrelation. As illustration, a 16-day forecast of NEE was driven by parameter and process errors, with minor exogenous uncertainty and negligible endogenous uncertainty. Looking forward this framework needs to be applied to many ecological processes to identify common patterns in what drives forecast uncertainty in ecology and determine how the limits to predictability are related to temporal and spatial scaling.