[EM] Sports-based models for voters and votes

[Paul Kislanko]

One problem with using wins between teams as "votes" is that in this case
the "voters" are the games played, so an A>B voter is different from a C>D
voter and no "voter" ranks more than 2 alternatives. It's as good as a way
as any to put random numbers into a pairwise matrix, but it's a lot faster
to just generate 100*R and 100*(1-R) if you've got a computer.

But here's a source for ranked ballots that could be used to model any
system. Kenneth Massey has collected up nearly 100 polls and ranking systems
for American College Football. Each of these is a "ranked ballot" that
orders (possibly a subset) of the 117 teams in division 1A.

On this page the "voters" are listed across the top, and the candidates down
the side. 

The data as it exists is not very interesting, because there's really not
that much disagreement at the top and bottom of the ballot, since even
though all of the "voters" have different weights for "issues" (margin of
victory versus opponents' winning percentage, etc) they all have the same
goal of picking the most dominant team, which is by definition likely to be
very highly rated in "issue" that is relevant.

However, if you threw out all the options that appeared on any voters top
10, I think you'd have a pretty good set of data to test against various
alternatives (and lots of cycles and sub-cycles).

His site is at http://www.mratings.com/cf/compare.htm