Follow-up to Jerry Palm's online BCS discussion

[Andrew R. Smith]

:--- "Andrew R. Smith" <sm...@matrix-one.com> wrote:
:> Re: SOS
:> =======
:> The SOS views playing two 6-5 teams as better than playing an 11-1
:> and a 1-11 team.

I chose a poor example. My point is that for a top-10 team, playing
another top-10 team (10-1) and a doormat (1-10) is more likely to produce a
lose, and thereby hurt the team in the polls, computer rankings, and the
BCS in general, than playing 2 average teams (say 6-5 and 5-6).

Another way of saying it is that a top-10 teams' chance of losing does not
change linearly at one goes from 12-0 to 0-12 opponents so adding W-L
records and averaging them does not make sense.

Just ask yourself which schedule is more difficult for a top-10 team to go
undefeated against, 12 6-5 teams or 6 10-1 teams and 6 2-9 teams? To me,
the first is clearly easier than the second but the BCS SOS views them as
equal!

:[How to quantify the loss risk for a different number of games played]

One method -
For a team, take each game and calculate their expected loss probability,
then add those numbers up for each game of the season. Loss probability
could be based upon W-L records, composite computer rankings, whatever, and
historical data or a reasonable theory could help determine how to generate
the probability (e.g., how often over the past 10 years has a 10-1 team
loss to a 1-10 team).

Such a method should be especially appealing if the loss component remains
in the BCS since it mirrors it - losses vs expected losses based on a
team's schedule.

?Does anyone know of an existing computer rating which does this?

:> Re: Ratings vs Rankings
:> =======================
:> This is a problem with the poll and computer components as well as
:> the SOS. The case where the difference between the #2 and #3 team in
:> either voting poll (or any of the computer polls) is 1 vote (or 0.01)
:> is seen as no different from the case where the difference is huge.
:> The polls and computer rankings should be normalized and treated as
:> ratings, not rankings.
:
:That would be too cumbersome to implement and explain.

I understand this, but it's no more cumbersome than the computer rankings
or the SOS!

For the computer rankings, just have the people who compute them normalize
them before giving them to the BCS and SKIP the process of converting them
to rankings. People understand the idea of averaging 2.85 + 3.98 + etc.
For the polls people only have to understand dividing the points received
by the number of votes.

Andy

______________________________________________________
The Nebraska Cornhuskers
National Champions - 1970, 1971, 1994, 1995, 1997, ...