I tried all kinds of things Z. I'll spare everyone the full list, but based on what worked in the end it's sort of a manual approximation of a neural network. I actually started with the approximation based on what was already in my head and then had a couple profs say that a NN should work. Definitely in over my head trying to explain exactly why, but there probably just wasn't enough data to use a true NN and they weren't able to find the same relationships.
The final model more than doubled the value of draft position as a predictor though -- which was pretty cool. It only uses players who came into the league before 2012 though, and you can see in the intervening years that the league as a whole is already adopting some of the things I found (the Development piece below for sure).
If you're into these things...here's the Reader's Digest version:
I broke most of the predictor variables into sub-variables based on inflection points in lowess line graphs, and broke the WRs up into subgroups based on build.
Then ran stepwise regression for each WR sub-group using the sub-variables to generate two new variables -- a development score (using draft position, NCAA volume, NCAA efficiency, age), and a physical score (using HT, BMI, Speed, Explosion, Agility). Probably forgetting something there, but close enough. I didn't know it when I was actually doing it, but using the line segments is called piecewise regression.
Then the results of the Developmental and Physical models were regressed against my measure of career performance.
What was kind of cool about it was that the model I ended up with was pretty close to what was in my head from working with this stuff informally for a long long time. The hard part was figuring out how to formally model what I thought I already knew.
That's pretty much what I came up with, too.