I think if you add a
d/(100-x) term it will improve the fit near the top. Perhaps use
d*(100-x)^e for the slow/accurate version, although you can't use simple linear algebra to find the best
e then, a genetic approach would probably be better.
I like the idea of this, just for historical reasons to have a decent mapping of APS to ELO, and would consider adding a column in the LiteRumble page for it. I would calculate this on the fly as the page loads, since for me with Python on App Engine the slow part of the code is the interpretation of the code, not the individual math operations the code calls =)Of course I'd have to test to make sure it doesn't slow it down too much, but it should be fine, the fast version if nothing else.