there were quite a number of papers demonstrating the results of model predictions of land use or climate change. Nearly all of these predictions were made deterministically, as if it was possible to predict accurately the response of the hydrological system under current conditions and to estimate accurately the effects of change on the effective parameter values of the models. The models varied from lumped conceptual models, to semi-distributed models, to fully distributed models complete with links to GIS databases for the input of the required data and the (impressive) visualization of the distributed results. Almost none of the papers made any attempt at all to evaluate the reliability of the predictions.
This would appear to be a little dangerous. There will almost certainly be some cases where predictions of the hydrological impacts of some change will be able to be checked in some form of postaudit analysis. This type of audit has already been happening in groundwater modelling-and the predictions have generally been shown to be poor. Such analyses led to the paper by Konikow and Bredehoeft (1992) titled 'Groundwater models cannot be validated' and considerable consequent discussion that does not seem to have seeped over much into more generally hydrological modelling. Why is that? Is it just that the evaluation of reliability is too difficult, too costly, or too difficult to understand? Why is it that deterministic predictions of the impacts of change are still so pervasively persistent when they stand a really good chance of being wrong, regardless of model type?
We should really be evaluating model reliability or predictive uncertainty as a matter of routine now. For some models there are indeed issues of available computer power in doing so, but for many models this is no longer the case. It also certainly seems likely that increases in computer power are going to arrive much more quickly than either theoretical or measurement breakthroughs. So, looking ahead, there should be an increasing range of models for which uncertainty estimation even by crude unsophisticated Monte Carlo simulation could be realized. This is important because therein lies a defence against being wrong or, at least, of being wrong more gracefully. Evaluating the potential range of outcomes will not ensure that the actual outcome will be within that range (especially since the new boundary conditions or forcing data will generally be poorly known, as well as any changes in representing the characteristics of a catchment), but it will certainly be more likely.