The accuracy of a simulation depends more on how well your equations and assumptions model reality, than the resolution of the simulation (computing power, as Carlton has stated). More computing power does let you have your solution in much less time however.
A model, by definition, is not the real thing. No model can be trusted unless it is tested in the real world. This is why we still have wind tunnels even though we extensively use computation fluid dynamics(CFD) for flow modeling. Its just numbers until you have validated the model. Once the model is validated you can extend it or extrapolate it to predict conditions that you can not test. For example testing a small bomb to predict the outcome of a large bomb.
In this case future prediction may very well be restricted to modeling only; leveraging past tests. Nuclear bomb testing today has a way of impeding world peace (oh, we arent building nuclear weapons, we are just testing them) ;-)
There have been lots of test in the past to validate the models they are likely using now. So for the United States at least, I wager their models are highly accurate, even including weather and GIS data.
I just skimmed the article, but they refer to the model as a "tool", not a replacement for testing. Dont let yourself reverse causation and correlation in your thinking. They dont only model with super computers because "there is no reason to test". They only model with super computers because there is no way any (sane) government will allow them to test.