Model forecasts are verified against model control simulations (
perfect model experiments), thus overcoming to some extent issues of uncertainties in the observations and / or model parameterizations.
This can involve «
perfect model»
experiments (where you test to see whether you can predict the evolution of a
model simulation given only what we know about the real world), or hindcasts (as used by K08), and only where there is demonstrated skill is there any point in making a prediction for the real world.