>>10272511aha, okay, I see now. HadCRUT4, Cotwan & Way, NOAA NCEI and GISTEMP are actual data, and CMIP3 is our predictive model.
So, firstly:
of course CMIP3 agrees with the data in the hindcast, if it didn't it would have been re-tuned until it did, and that's fair.
Now, CMIP3 predicts a confidence interval into the future (i.e. after 2000). We are then only comparing 20 years of data with the output. The CMIP3 confidence interval shows the range for the model.
Let's use an example. I could predict that the next three times I flip a coin, it's going to be heads, and through some 'model' with some parameters I pull out of my arse I arrive at a 95% confidence in that. Many times, we will see three heads in a row, mostly we won't. But you see how the 95% is entirely my own construction? Nature doesn't care about my made-up confidence interval. At which point I would go back to my model and improve it (obviously I know the answer: it is a 0.5*0.5*0.5 chance of three heads in a row, but I'm using this as a toy example).
Just as 3 coin flips is not enough, I don't think 20 years is enough. Do you see now?