This is matching the NAO model on 3 different training intervals. The lunisolar tidal forcing is essentially the same across each. What does change slightly is the high-K wavenumber coefficient (and its amplitude scaling) that is used to sharpen the peaks. The lower K coefficients vary as well but do not contribute as much to the amplitude.

**1950-1973** CC in training 0.76

**1973-1996** CC in training 0.72

**1996-2019** CC in training 0.78

This is the value of the K wavenumber which is in arbitrary units. The NAO amplitude does seem to grow over time.

Low Mid High
k(wn) 61429 61614 62660
Amp 1.370 1.697 2.161

It's impossible to tell at the moment whether this variation is overfitting to noise or whether it is actually capturing the variation. Yet the transitions slightly outside the training intervals does suggest that the fit for the training does extend a few years outside each interval.