All four Tiljanders passed validation. ]]>

For example, one could visualize a reconstruction wiggle matching the temperatures exactly, but with amplitudes which differ substantially from the actual temperatures. This explains why other methods such as RE and CE exist.

And if the reconstruction wiggle-matches very well in the calibration period (r2=0.688), but looses this capability during verification (r2=0.00003) it is better to hide the r2🙂

]]>Although there are problems with these latter two methods in regard to choosing levels for deciding whether the “fit” is a good one, in the case of CPS, it is not to difficult to show (from “first principles”) that the RE for the calibration period satisfies the identity: RE = 2*Correlation – 1. For other types of reconstructions, I don’t think any simple relationship exists.

]]>If we have two series x and y that haven’t been calibrated against each other (i.e we don’t know a & b) then measuring correlation makes sense, since correlation is independent of a and b.

Yes, you need statistically significant slope to obtain satisfactory confidence intervals.

Once we’ve calibrated the series though, we want to know the combined effect of x, a, and b. At this point correlation is useless because it’s unaffected by the algorithm’s choice of a and b.

Why that makes correlation useless?

]]>