Do climate change models make accurate backward-looking "predictions"?

[DavidD]

I've read and heard lots about the very complex models used to predict temperature change over the next century or so. Obviously, we won't know how accurate their predictions are until we arrive at the future they seek to predict. One provisional test of their accuracy, though, would be to see how accurately they "post-dict" temperatures over the past 100 years. Has that been done, or do sufficient data not exist to do that? If the former, what have the results been?

[Dawei]

As pegminer says, this is a common practice. Scientists tweak the parameters of the models and run them over past years, then compare the results with real world observations. By doing several thousand of these tweak-and-check runs, scientists can determine which specific parameters best describe how the world works.

I've run several dozen climate models on my computer from climateprediction.net. About 95% of the models I've run are from past years. The three that I'm running now are currently at 1981, 1988, and 1971.

[keithp]

Of course. That's generally the first test they run when validating the model.

[pegminer]

For the most part, they do. That's known as "hindcasting" and it's done with virtually all forward looking models. However, it's not completely a fair test because if a model is demonstrably wrong at hindcasting it will be corrected (one way or another) to give accurate results.

[bubba]

Yes. That is how the models are calibrated - predict past climate and compare to what scientist think it actually was. Modelers can then test the sensitivity to some of the parameters to determine how sensitive predictions may be to errors in these parameter values. This is standard procedure.

Probably the best reference to see what has been done is the IPCC Working Group 1 report (and associated technical papers). How the model was calibrated should be in it somewhere along with a sensitivity analysis,