Jim, I am using the Mathematica wavelet scalogram in #33. Yours looks different partly because you have the period increasing on the vertical axis while mine is decreasing upward.

---

Here is an amazing machine learning run using Excel that I mentioned in my previous comment :

![machineLearn](http://imageshack.com/a/img909/2582/YJfdLh.gif)

The solver only operated on post 1980 data using the fixed periods of 2.33 years (QBO), 6.4 years (CW), and *actual TSI data* and deduced the LHS and RHS of the model DiffEq by minimizing the difference and maximizing the correlation between LHS and RHS along *only the training interval* from 1980 onwards. All the dozen parameters are found by the Excel solver using the Evolutionary selector.

Although the fit is not the greatest on the training interval, it projects backwards remarkably well over the validation interval extending back to 1880 ! Each of the peaks and valleys is identified save for a few between 1895 and 1900 and one right after 1921 -- the fact that there are a few is actually good because this eliminates an artifactual basis for the fit.

This is what I might consider excellent statistical validation for the model.

BTW, I do listen to all suggestions so that John's advice to doing more statistical eval and Nad's suggestion to take the actual TSI data was taken to heart. Thanks.

---

Here is an amazing machine learning run using Excel that I mentioned in my previous comment :

![machineLearn](http://imageshack.com/a/img909/2582/YJfdLh.gif)

The solver only operated on post 1980 data using the fixed periods of 2.33 years (QBO), 6.4 years (CW), and *actual TSI data* and deduced the LHS and RHS of the model DiffEq by minimizing the difference and maximizing the correlation between LHS and RHS along *only the training interval* from 1980 onwards. All the dozen parameters are found by the Excel solver using the Evolutionary selector.

Although the fit is not the greatest on the training interval, it projects backwards remarkably well over the validation interval extending back to 1880 ! Each of the peaks and valleys is identified save for a few between 1895 and 1900 and one right after 1921 -- the fact that there are a few is actually good because this eliminates an artifactual basis for the fit.

This is what I might consider excellent statistical validation for the model.

BTW, I do listen to all suggestions so that John's advice to doing more statistical eval and Nad's suggestion to take the actual TSI data was taken to heart. Thanks.