That is what error bars are for. If your model is useful reality will fall inside the error bars.
Tom ... this is just meaningless.
Doesn't matter. Concerning climate models, even if the theory were perfect there is no way to calculate the results from first principles (see Navier–Stokes). On top of that climate is nonlinear and chaotic. Therefore the output is highly sensitive to small changes in initial conditions. Since climate data is so sparse we can't even get the initial conditions correct. If you can't make a prediction you can't test the theory.
The financial models are very much like climate models. They are nothing more that sophisticated curve fit models. IOW, they are statistical models that have no predictive power for out of sample data. Therefore "black swan" (out of sample data) always produces unpredictable results.
It isn't the complexity, semiconductor manufacturers have very complex physical models based on first principles that work very well. They also have tons of data to validate their models.
The climate models are primarily statistical models. They are tuned (curve fit) to mimic past climate. Various tuning’s include aerosols, clouds, water vapor, and albedo. The reality is that we have little to no data to start with. For example, we have no idea the volume or distribution of aerosols from the 70’s. Clouds, even if we had detailed data, are too small to model due to model grid size.
Another issue which Dr. Pat Frank expands on is error propagation which is well worth a listen. (43 minutes)