Statistical validation

Abstract

Validation is a necessary step for model acceptance. No single combination of validation tests will be applicable across the diverse range of models and their uses. Choice of technique is important, as some contain problems and inconsistencies. Subjective assessment can be useful as a guide. Within visual techniques, observed vs. predicted plots are shown to have superior diagnostic capabilities compared to the more widely-used time-series plots. Mean absolute error is demonstrated as a more robust deviance measure than mean absolute percent error, and within the statistical tests a nominated sub-set of some simpler statistics should reveal most of the required information. The modelling efficiency is proposed as the best overall measure of agreement between observed and simulated values. © 1993.