Good, but I don't see the analysis at all.
I mean you are betting (without any actual money which is good) on the actual auction price of something in the future.
The auction price of a work of art - by definition - is not in any way like the actual market value of anything (it will be MUCH lower or MUCH higher or JUST ABOUT right) depending on a zillion factors that have nothing to do with the actual value of the piece.
And you have so few data points that a statistical approach is mostly meaningless, let alone one based (in the case of a study) based on the differences between it and the final painting.
You could use a few sessions of flippism and get a similar value.
Or count the number of unique RGB colours in the (photos of) the paintings and give each one a value between 1,217 and 23,146 US$ depending on brightness , later multiplied by the square root of the year it was painted and arrive to a same "mathematical" estimation.
It seems to me like a back-justification of something not measurable analitically.
A 5 millions US$ bracket is not in any way an approximation.
IF you had narrowed the interval (and IF the picture is at the auction sold within it), then all your data mingling exercises might have some usefulness, otherwise it is just an arbitrary estimation like the one from Christie's.
So the similarity calculated by an arbitrarily chosen software between (photos of) study and painting of improvisation #8 is 64%.
The similarity between (photos of) study and painting of improvisation #20 is a MUCH lower 57%.
It is clear how this HUGE 7% different can easily bring down the value of the estimation by more than 50% from the improvisation #8 sale price of 23 millions for a 98x70=6,860 sqcm painting, i.e. 3,350 US$/sqcm to a mere 10 millions for the improvisation #20, a 71x99=7,029 sqcm painting, i.e. 1,423 US$/sqcm, give or take a couple of millions.