The Sports PageMichigan Nearly Lost to Western Michigan. Eight Days Later It Beat Oklahoma.
A team that badly misses the number tells you almost nothing about what it does next. We measured it across 236 teams and two weekends: last week explains three tenths of one percent of this week.
The question: a team wins ugly when it should have won big. Does that tell you anything about next week?
Almost nothing. Across 236 teams, the correlation between one week’s miss and the next is +0.06.
Why that is surprising. Everybody treats an ugly win as evidence. They should have lost; they are frauds; the ranking is a lie. That reasoning assumes the ugliness carries forward. It does not.
What you get: a way to tell a real signal from a loud one, by Saturday night.
Michigan, in seven days
Week one, Michigan hosted Western Michigan as roughly a 27-point favorite and won 13–12. Not a scare. A near-disaster, and by any honest reading they were outplayed.
You know what was said that night, because it is said every night: they are overrated, the ranking is fiction, this team is a fraud.
Week two, Michigan hosted Oklahoma as an underdog and won 17–10. Not a survival. They took it to them, wore them down, and won going away from the fourth quarter on.
Same roster. Same coaches. Eight days apart.
The ugly win and the statement win were the same team. Only one of them was ever evidence, and it was neither.
The ProfessorSo we measured all of it
One team proves nothing, so here is every team.
For each club we took how far it finished from the betting line — the crowd’s number, not ours — in week one, and again in week two. If a bad week meant anything, the two would move together.
That cloud has no tilt. The correlation is +0.06, which squares to 0.3% — the share of this week you can explain with last week. The fitted line rises about a twentieth of a point for every point of last week’s miss, and you should read that as flat.
Watch what happens to the teams that missed by the most:
| Biggest week-one misses | Week 1 | Week 2 |
|---|---|---|
| Rutgers | −45.3 | −4.0 |
| Massachusetts | +45.3 | +3.5 |
| Mercer | +36.5 | −22.5 |
| Nevada | +35.7 | −8.2 |
| LSU | +31.0 | −3.7 |
| Appalachian State | +31.2 | +9.5 |
Rutgers missed by 45 points, then by 4. Massachusetts missed by 45 the other way, then by 3. Two of the six flipped sign entirely.
That is regression to the mean, and it is worth naming because it is the most useful idea in this newspaper. An extreme result is usually a middling team having an extreme day. The team persists; the day does not.
The crowd misses too, and by less than we do
Fair is fair. If we are going to say the crowd overreacts, we should check our own homework against theirs.
On the 49 rated games of week two, the closing line missed the final margin by 10.8 points on average. Our own ratings, frozen on the Thursday before kickoff, missed by 12.0. The crowd was closer in 26 of 49 games.
That difference is 1.2 points with a standard error of 0.6, so we cannot claim the gap is real at this sample. What we certainly cannot claim is the opposite. The market is not beating us because it is smarter; it is beating us because ten to twelve points of miss is simply what this sport does. Both numbers are noisy for the same reason the Michigan result was noisy.
What to take home: separate the day from the team
When a team wins ugly, ask one question before you file it as evidence: is there a mechanism, or just a margin?
A margin is a number that already regressed before you finished talking about it. A mechanism is different — a quarterback who left on a cart, a line missing two starters, a coach who changed. Those persist, and they are what actually moves next week.
Michigan had no mechanism. It had a bad day, and then it had a good one. If you had sold them on Saturday night you would have sold at the worst possible moment, which is what regression does to people who mistake a day for a trend.
Hold onto this the next time an announcer says a team was exposed. Ask what exactly persists. Usually the honest answer is: the schedule.
Baseball is ranked by championship leverage: we simulate the rest of the season 60,000 times, then split those seasons by who won each of today's games. College football cannot be simulated that way, so it is ranked by how close to a coin flip the ratings make it — a lopsided game teaches you nothing. Tomorrow this block reports what happened to today's pick.
Notes & sources
Games and scores from the CollegeFootballData API. Closing spreads are the mean across that API’s books for each game, taken as a historical record of what was believed before kickoff — unlike a rating, a past line cannot be quietly updated by the result.
A team’s miss is its final margin minus the margin the line implied, signed from that team’s perspective. Both clubs in a game therefore carry equal and opposite misses. Weeks 1 and 2 of 2026, every game where a line existed: 253 team-games in week 1, 240 in week 2, and 236 teams appearing in both.
Correlation between a team’s week-1 miss and its week-2 miss: r = +0.058, t = +0.89, r-squared 0.3%. Fitted slope +0.051 points per point. Mean absolute miss: 13.4 points in week 1, 11.5 in week 2.
Crowd against model, on the 49 week-2 games that had both a line and a rating for each side: mean absolute miss 10.76 for the closing line, 12.00 for SP+ frozen on 2026-09-10 by scripts/snapshot_ratings.py. Mean difference 1.24 points, standard error 0.64, t = 1.95 — not separable from zero.
Three limits. Two weekends is a small window, and early-season lines are the least settled of the year, so the true correlation is probably somewhat higher than 0.06 once the market has more information. A correlation near zero measures only the LINEAR relationship; a team that changes for a real reason mid-season would not show up here. And this measures margin against expectation, not quality — a club genuinely transformed by an injury is exactly the case the method cannot see, which is why the piece asks for a mechanism rather than a number.
