# Ranking résumés from scratch: what a college football ranking should actually measure

Wed, 21 Oct 2026 00:00:00 GMT

Ranking résumés from scratch: what a college football ranking should actually measure

Part 1 of 2. Part 2 covers the math, the testing, and everything I got wrong.

Kentucky is the second-best team in college football. James Madison is tenth. The AP poll has Kentucky twelfth and doesn't rank James Madison at all.

You may think I am insane for that... and you could be right, but that's what my ranking says after week 7, and by the end of this post I want you to understand exactly why, even if you still disagree with it.

Mostly, the two lists agree. Twenty-one teams appear in both top 25s, the average team in both sits about four spots apart, and the rank correlation is 0.81.

The interesting part is where they split:

Kentucky: #2 here, #12 in the AP
James Madison: #10 here, unranked in the AP
Houston: #6 here, #14 in the AP
Georgia, going the other way: #14 here, #5 in the AP

Every one of those gaps comes from one decision I made before writing a line of code: asking the question, what is a ranking truly supposed to measure?

Best vs. earned

Polls have a structural problem. Voters fill out a ballot in August, before anyone has played, and every ballot after that is an edit of the one before. A team's October rank still carries some of its preseason weight. A team nobody rated in August has to climb past everyone who was rated, one week at a time.

I grew up on the West Coast as a USC fan, following the Pac-10 and then the Pac-12, and there was always a sense that the polls leaned toward the SEC. SEC teams got the high preseason rankings, and once they had them, losing didn't seem to cost them much. If an SEC team lost to another SEC team, that was proof of how good the winner was, not how bad the loser was.

A Pac-12 team got no such cushion. Lose to a decent conference opponent and it usually sank you, because nobody gave that opponent any credit. For most of the 2000s and into the 2010s, the national view of the conference was USC, an Oregon on the rise, and maybe one or two others.

That's where the questions started. Why do we let a preseason poll shape so much of the season? Why don't rankings go back and re-grade the games already played? If you lose in week 1 to the #1 team, and that team finishes the year unranked, that loss should count for what it turned out to be: a loss to an unranked team.

When I sat down to build my own ranking, I realized "who's the best team?" and "who has earned the most?" are different questions. A poll tries to answer both at once. I wanted to answer only the second: given the games that have actually been played, whose results are the most impressive?

That led to one rule I set before building anything: results only. Every team starts at zero, and nothing from outside this season's games gets in: no polls, recruiting rankings, roster talent, returning production, last season's results, or betting lines.

That rule has a price, and I want to name it up front: with no preseason information, the ranking is nearly blind for the first few weeks. In week 1, a team that beat an FCS opponent and a team that beat a top-10 team look a lot more alike than they should. The ranking needs a month or so of games before the ratings converge, and I've accepted that trade.

Version 1, and why I threw it out

My first version scored résumés with points, and each win's value depended on where the opponent was ranked. Beat a top-10 team, get a lot of points; beat the #80 team, get a few.

It seemed reasonable. It broke in three ways.

Rank cliffs. Because a win's value was pinned to the opponent's rank, a jump in their rank re-priced your old wins overnight. One opponent moving from #21 to #2 could swing the value of a win you'd banked weeks earlier by 35 points in a single week.

A floor. The scoring bottomed out, so by week 6, 56 of 138 teams were tied at the minimum. The ranking couldn't tell nearly half the country apart.

Byes counted against you twice, through a quirk in how the scoring shrank teams with fewer games.

Two weeks in a row this season made it clear version 1 had to go.

In the October 11 update, Alabama had just lost to Georgia, and Alabama still moved up a spot, ahead of Georgia. I patched it with a heavier loss penalty, the first version of the head-to-head rule you'll see later in this post.

The next week, a 4–2 Texas reached #1. That was a rank cliff in action, and it was the final straw. I started down the road of patching version 1 again, then stopped. Two patches in two weeks meant the problem was the design, not the details, and that's when the two-layer model was born.

The lesson underneath all three: one number was doing two jobs, ranking résumés and predicting games, and it did both badly. So I split them.

Two layers

The current ranking has two separate numbers for every team.

Power answers how good is this team? It's measured in points: a team with a Power of +10 would be expected to beat an average FBS team by about 10 on a neutral field. Power is what predicts games.

WAB, or Wins Above Bubble, answers what has this team earned? It's measured in wins, and it is the ranking.

The rule that ties them together: Power only grades the opponents. A team's own Power never touches its own WAB. Power's job is to tell me how hard each game on your schedule was. What you did in those games is the résumé.

Power, without the math

Power comes from every game's final margin, with three adjustments.

Margins are capped at 28. Winning by 56 counts the same as winning by 28, so running up the score doesn't pay. Overtime games count as a tie on margin, since they were tied after regulation.
Every team gets the same pull toward average. Until its results say otherwise, every team is treated as an average FBS team. That's the "no preseason bias" promise in practice: not "no assumption," but the same assumption for everyone.
Home field is worth a fixed 2.16 points. That's the average home edge in conference games from 2014 to 2025. (Part 2 covers why letting it float week to week didn't work.)

FCS opponents are lumped together as one team, rated −22.1 this season. There's too little data to rate them individually, and beating one tells you almost nothing anyway.

Wins Above Bubble

Here's the core idea. Take the team sitting at #12 by Power. In a 12-team playoff, that's roughly the last team in. Call it the bubble team. This week it's Ole Miss, at +11.4.

For every game a team played, ask one question: what are the chances the bubble team would have won this same game, against this same opponent, at this same site? Call that chance p.

A win earns 1 − p.
A loss costs p.

That's it. Add it up over the season and you get wins above bubble.

The four corners make the logic obvious:

Beat a strong team that the bubble team would usually lose to, and you earn nearly a full win.
Beat a weak team the bubble team almost always beats, and you earn almost nothing.
Lose to a strong team, and it barely costs you, because the bubble team usually loses that game too.
Lose to a weak team, and it costs nearly a full win, because a bubble team doesn't lose that game.

And notice what's missing: the margin. A 3-point win and a 40-point win over the same team are worth exactly the same. You won the game; that's the résumé.

Honestly, this was the hardest part for me to accept. I always felt there was value in how badly you beat a team, or how badly you lost to one. But as I've gotten older (and maybe wiser), I've come to think my dad was right all along. He has always said, "A win is a win." Leaving the margin out of the résumé honors exactly that: it doesn't matter how you do it. A win is a win.

One season, worked out: BYU

BYU is #1 in WAB this week, and its season is the cleanest way to see how the pieces fit.

Take the biggest game, week 7 against Notre Dame. Put Ole Miss in BYU's place: at home, against a Notre Dame team rated +18.35. Ole Miss (+11.36) plus 2.16 for home field still comes out about 4.8 points worse than Notre Dame. Once you account for how much a single game varies, that's about a 34% chance the bubble team wins. BYU won, so it earns 1 − 0.34 = +0.66.

Now the opposite end: week 1, 63–7 over FCS Utah Tech. The bubble team wins that game 99.9% of the time. BYU's 56-point win earns +0.00.

Run that for every game and BYU is 6–0 against a schedule where a bubble team would have expected 4.31 wins. Six minus 4.31 is +1.69 wins above bubble, the best résumé in the country.

Here's the part I like most: BYU is only 15th in Power. The ranking doesn't think BYU is the best team in the country. It thinks BYU has done the most with its schedule. That's the whole two-layer idea.

This is also how the two-layer model fixes version 1's rank cliffs: every past game is re-scored each week as opponents' Power changes. If Notre Dame keeps winning, BYU's win over them slowly gets more valuable. If they collapse, it slowly gets less. Either way it moves smoothly, not in the overnight jumps that broke version 1.

"But A beat B"

Every ranking gets the same complaint: "How is A ahead of B? B beat A!"

So there's one adjustment on top of WAB: if a team beat another team and trails it by 0.25 WAB or less, it moves just ahead of it. A small window like that settles the close calls without letting one game override a whole season. A few guardrails keep it from creating loops (A over B over C over A) or chains where one result drags a team far down the list.

I didn't pick 0.25 by gut feel. I tested window sizes against 12 seasons of games. The 0.25 window removes about a third of the top-25 "but they beat them" inversions (from 23.8% to 15.3%) and costs about a tenth of a percentage point of pick accuracy. Wider windows cost more. It's a values trade-off, but a measured one.

Does it actually work?

I tested it the hard way. I re-ran every FBS-vs-FBS game from weeks 2–13 of 2014–2025 (skipping 2020). That's 7,198 games, and each week was predicted using only the games before it. The model's few settings were fitted on the other seasons, so no season grades itself.

It beats a standard Elo rating and my old model at picking winners (70.6% vs. 69.4% and 65.6%) and on log-loss, which punishes being confidently wrong.

Vegas is better. But then again, when does Vegas ever lose? Where betting lines exist, the favorite wins 74.3% of the time to Power's 70.7%. The market misses final margins by 12.3 points on average, and I miss by 13.8. Most of that gap is the first month. In weeks 2–6 the market leads 75.9% to 69.3%. From week 7 it's 73.3% to 71.7%, and the margin gap shrinks from 2.8 points to 0.7. That's the price of starting every team at zero: the model spends September learning what Vegas assumes on day one.

The AP poll is a tougher comparison than it looks. In games involving ranked teams, the AP's higher-ranked team won 76.8% of the time, and my ranking's won 73.7%. But my ranking isn't built to predict; it rewards résumés. My Power rating, which is built to predict, gets 76.4% on the same games. From week 7 on, it beats the AP outright: 77.5% to 75.8%.

So here's the honest scorecard. As a predictor, Power holds its own against the AP once the season settles, and trails Vegas mostly in September. The ranking itself isn't trying to win that contest. It answers a different question, who has earned it, and it does that with one hand tied behind its back on purpose.

Part 2 goes through how the testing works, the bugs it caught, and what I still don't trust.

Back to Kentucky

So how is Kentucky #2?

Kentucky (6–1, +1.33 WAB) doesn't have a great win. Its best is #30 LSU, worth +0.60. What it has is five wins a bubble team would lose a real share of the time: LSU, Texas A&M, South Carolina, Oklahoma, and South Alabama. The loss to #15 Alabama costs just −0.46, because a bubble team loses that game about half the time too. Kentucky's Power is only #27. It is not one of the best teams in the country. It has earned the second-best résumé.

James Madison (7–0, +0.84 WAB) got there the opposite way. No win is worth more than +0.23. Liberty is the best of them, and Marshall adds just +0.02. But a bubble team playing that schedule drops a game more often than you'd think, and James Madison hasn't dropped one. Seven small wins and no losses add up.

And here's the twist: James Madison's Power is #16, better than Kentucky's #27. By my own numbers, JMU might be the better team, and Kentucky has still earned more. Earned versus best, in two teams.

The AP has Kentucky twelfth and James Madison unranked. It's ranking who looks best, and a Sun Belt team needs to look a lot better than an SEC team to get noticed.

Georgia (6–1, +0.53 WAB) is the split going the other way: #5 in the AP, #14 here. By Power, Georgia is the sixth-best team in the country, and it has the best single win of the three: 20–17 at #15 Alabama, a game the bubble team wins only 20% of the time, worth +0.80. Then it lost to #28 Auburn, a game the bubble team wins 81% of the time, and that cost −0.81. One game erased the other. The rest of the schedule is Western Kentucky, Vanderbilt, Arkansas and an FCS team, and the 70–20 and 63–3 blowouts in there earn almost nothing. A win is a win. Georgia went from #2 to #14 in a single week.

Kentucky ten spots higher than the AP has it, Georgia nine spots lower: two SEC teams moving in opposite directions. The ranking doesn't care about the conference, only the résumé.

And Alabama, at #15? Its WAB (+0.62) is actually a hair higher than Georgia's, but Georgia beat them, so the head-to-head rule puts Georgia one spot ahead. That's the October 11 problem from version 1, solved.

One last twist: James Madison's playoff odds are 76%, and Kentucky's are just 41%, despite JMU's lower rank. The playoff guarantees spots to the five highest-ranked conference champions, and JMU has a 95% chance of winning the Sun Belt. Kentucky has a 14% shot at the SEC title, so it mostly has to earn an at-large bid. Rank and playoff odds answer different questions too.

What building it changed

One of my favorite games last season came in week 8. The AP poll had Miami at #2 and didn't rank Louisville at all. My ranking had Miami fourth and Louisville fifth. Louisville went into Miami as a 10.5-point underdog and won.

I find myself hunting for those games now. Part of it is the simple satisfaction of saying "I was better than the AP poll" (or, on a really good week, "I was better than Vegas"). But mostly it's given me a new lens to watch games through.

I also listen differently. When the expert panels on ESPN or Fox break down a team (what it does well, what it doesn't, why it's good or why it's bad), I find myself comparing their reasoning to my model's. What are they prioritizing that the model ignores? Where do they line up, and where are they miles apart? When a panel says a team "lost to a good SEC team," I can now look up exactly how good that loss was. It's a whole new way to take in the coverage, and I really enjoy it.

At the start of this post I said you might think I'm insane for these rankings. That's fine. I hope you at least understand how I got there. Every team's résumé, game by game, plus a scenario builder for "what if" games, is at stats.oasis-enterprise.com.

Part 2 is coming. It covers how I tested the model as I built it, the bugs that testing caught, and what I still don't know about it.

Maxwell Monson — Sports stats