Monday, February 10, 2014

A CFP Standings Model for Selection Committee

A most contentious element of the 16-year BCS era was its standings. Love it or hate it, the standings had the attention of the college football world from its midseason unveiling to the end of the regular season, when the all-important final standings were published.

And it was a most useful tool. It was transparent and predictable, even if it was flawed.

The College Football Playoff, set to kick off for the 2014 season, did away with the standings—at least not in a way that may be projected. The entire decision-making apparatus of the CFP rests with its 13 committee members, who will have complete discretion in deciding not just which four teams get to make the playoff, but also the other eight teams that will play in the major bowls in the playoff rotation.

CFP executive director Bill Hancock said four standings will be released to the public starting at the midpoint of the season, before the final pairings are revealed the day after the regular season ends. Other than that, the details are incredibly vague.

From the drip-drip information the committee has given to the public during the 2013 season, we only know these factors will be strongly considered: 1) strength of schedule, 2) winning the conference championship, 3) head-to-head results, if applicable. But just how much each criteria matters, we have no clue (and probably neither does the committee at this point).

With so much uncertainty, it only makes sense to construct a model that would be useful for the committee to consider. But more importantly, it has to be transparent and predictable, so we don't end up with an outrageous surprise come the first Sunday of December.

My standings model fulfills all these requirements. They're comprised of these elements:

1) AP Poll (20%): It's the only poll that's completely transparent, with each voter's ballot available for the public to scrutinize each week. It's also the most prestigious poll that's widely used by the media.

2) Computer rankings (40%): Kenneth Massey compiles the median and mean rankings of each team from over 100 computers each week. It's less biased than the human polls and the large sample size removes undue influence by outliers.

3) Strength of schedule (30%): While there are many models to choose from, Jeff Sagarin has the most time-tested SoS formula—including results from all Division I games, FBS and FCS—that's meticulously and promptly updated each week.

4) Conference championship (10%): It matters, but only winning it matters. Teams that win their divisions but lose in the title games don't get consideration for making an appearance.

With that in mind, this is what the final standings would've looked like at the end of the 2013 regular season (see complete standings with breakdowns).

And this is what the playoff and major bowl matchups would be had the committee followed the results of the final standings:

2013 Projected CFP Bowl Matchups
Rose BowlSemifinal#2 Stanford vs. #3 Florida St.
Sugar BowlSemifinal#1 Auburn vs. #4 Alabama
C-fil-A BowlAt-large or 'Group of 5'#8 Missouri vs. #22 UCF
Cotton BowlAt-large or 'Group of 5'#5 Michigan St. vs. #12 Oklahoma St.
Fiesta BowlAt-large or 'Group of 5'#6 Baylor vs. #7 Arizona St.
Orange BowlACC vs. SEC/Big Ten/ND#9 Ohio St. vs. #16 Clemson

Do you agree or disagree with the methodology of this formula? Please submit your comments and I'll make every effort to answer your questions. Thank you.


Chris Breisch said...

Given that SOS is already included in the computers, and should be considered by the pollsters, you weight it far too heavily. Put it at 10% and Conference Championship at 30%

Logan said...

Ok State over OU? Where were the Sooners ranked?

Can you list out your standings of the entire top 25?

Who's 10 and 11?

Demosthenes said...

Undefeated Florida State ranked third, behind a one-loss Auburn team that they went on to beat in the NC game...and a TWO-LOSS Stanford team that went on to lose the Rose Bowl as well?

Two-loss Oklahoma State into the Cotton Bowl over a Sooners team that had just beaten them...and then went on to beat two-time-defending national champion Alabama in the Sugar Bowl?

UCF included at all?

Nope. No good. As maddening as it will be to deal with holistic rankings, you now have me praying the committee just releases a top 16 based on the criteria of "Here's what we think," a la the NCAA selection committee ranking and releasing their bracket.

Demosthenes said...

I'd like to elaborate on the Florida State criticism for a minute, because it's by far the weakest result of your trial run. We can debate over which of two one-loss teams is better (e.g., Florida State and Miami in 2000) Even if one team has beaten the other head-to-head, a logical case could be made that the loser of that match still has the more impressive overall performance -- so it can't just be about head-to-head results.

But when you have a lone major-conference undefeated team, like Florida State was this last year, any plausible ultimate ranking has to put that team at #1. Yes, Florida State played a weaker schedule than Auburn and Stanford. They also destroyed those teams by high margins, and sat their starters for a lot of fourth quarters. That's exactly what you would expect a great team to do. Why penalize them because their recruiting came along at a faster pace than their schedule?

An idea -- re-introduce margin of victory. I don't have the confidence to suggest a relative weighting, but a team that runs the slate by high margins against a mediocre schedule still needs to be ranked ahead of a team that lost a game or two against a tougher schedule.

The Guru said...

Thanks for your comments. The model by no means is the last word. I'm seeking ways to improve upon it before the season opener.

Ideally, last season's final four would've been FSU, Auburn, Michigan State and Stanford, in that order. It may help to boost the importance of conference championships, or give the AP poll more weight.

As for margin of victory, that's considered by most of the computer models so it's already a component.

Again, thanks for your comments, and keep 'em coming!