Hi, Most of them in the beginning and few after an age. As per as my view & observation predictive models needs to be dynamic and should be trained,test & validate in frequent interval depending on the accuracy & senstivity of the model the age has to defined on the basis of domain & application.
Models are not developed to reproduce reality -- they are there to effectively summarize experience along some relevant dimensions. The predictions say, in essence, that experience suggests that situations within the range of the model's dimensions should produce the indicated results within an error range. further, the indicated prediction is the best available given the data.
There are better statistical descriptions of what is going on (grab your graduate statistics text), but the above explanation usually works for clients.
Ms. Clarks statement has over time been taken somewhat out of context. In the original article she states that "Model users frequently forget that all models are based on simplifying assumptions, and therefore all models are wrong. Models attempt to replicate reality, but they are not reality".
In a sense she is right in that we should never presume that we have access to all variables involved in a model.