Yes, good point, check also http://www.climateaudit.org/?p=886, #19, #26. The latter is also a good warning for those who use calibration residuals for uncertainty range estimation.

To me it seems that inverse regression (global temp) tends to overfit, and CVM cannot handle sampling problems. Local calibration works well, and it gives directly estimate of proxy noise.

]]>yc=y(1:NCal)-mean(y(1:NCal)); % calibration data, global temperature (?)

Xc=X(1:NCal,:); % X is a matrix of standardized proxy records, calibration period 1:NCal

Bh=(inv(yc’*yc)*yc’*Xc)'; % estimate of beta

yh=inv(Bh’*Bh)*Bh’*X'; % reconstruction

I think it is not too hard to write m-file that simulates different scenariors.. Let’s see..

]]>did I understand CVM correctly:

1) average the standardized proxies

2) standardize the average, and scale so that *calibration data mean and variance* and *reconstruction data mean and variance* (within calibration interval) match?

(These RCS calculations are mathematically trivial, but the ringers make them out to be some kind of magic.)

Which tells you something about their preferred audience.

]]>On RCS. Some people have to resort to making things sound like magic in order to get them on the shelf of the marketplace of ideas. The ones who just quietly go about their science don’t get the big grants.

]]>