Richard Betts joined in the conversation about climate models today, making some interesting comments on validation:
As I've mentioned before, the earlier climate models used in the 1970s were used to make estimates of warming over the next 30 years which were fairly close to what happened ... BH asks for tests of the projections made 10 years ago, but the problem is that with internal variability in the system you need longer than that to test the models, unless you specifically initialise the models with the conditions of (say) 2001 using data assimilation techniques, and that kind of thing was not available then, we only started doing it 5 years ago.
So yes, out of sample testing on timescales relevant to GHG rise is an important point but by definition difficult with the latest models!
One minor point is that I had said I would have been more convinced had the story of model versus data in the last ten years been different - I agree with the 30 years figure for falsification. However, more interesting is a point made in a recent post by Nir Shaviv:
From the first IPCC report until the previous IPCC report, climate predictions for future temperature increase where based on a climate sensitivity of 1.5 to 4.5°C per CO2 doubling. This range, in fact, goes back to the 1979 Charney report published by the National Academy of Sciences. That is, after 33 years of climate research and many billions of dollars of research, the possible range of climate sensitivities is virtually the same! In the last (AR4) IPCC report the range was actually slightly narrowed down to 2 to 4.5°C increase per CO2 doubling (without any good reason if you ask me). In any case, this increase of the lower limit will only aggravate the point I make below, which is as follows.
Because the possible range of sensitivities has been virtually the same, it means that the predictions made in the first IPCC report in 1990 should still be valid. That is, according to the writers of all the IPCC reports, the temperature today should be within the range of predictions made 22 years ago. But they are not!
Go and take a look at the graph at Nir's site. This seems a reasonable point to me.