What is the difference between predictive error and predictive uncertainty?
“Error” is the more appropriate concept to employ when we take a single model output and act on the basis of that output. This is what happens when we calibrate a model and then use this calibrated model to make predictions of future environmental behaviour. This prediction is likely to be wrong, this “potential wrongness” arising from the fact that a model, and the parameters that it uses, are simplifications of reality (even if the calibrated model fits the historical observation dataset well), and from the fact that the historical dataset is contaminated by measurement noise. Tools available through the PEST suite allow quantification of the potential for error in predictions made by a calibrated model. “Uncertainty” is slightly different. It is an intrinsic property of the data on which basis a model is parameterized and calibrated. Conceptually, it can quantified using a Bayesian approach. However it is often easier to use error as a substitute for uncertainty, especially when wor
