Rutgers Was Favoured by Twenty-Nine. It Lost by Sixteen.
Across all 47 rated games, week one missed its preseason number by 14.3 points a game. That is 28% worse than the last two seasons. The openers were not just loud — they were measurably the hardest to predict in three years.
The question: was week one genuinely strange, or does every opening weekend feel that way?
It was genuinely strange, and here is the number. Ratings missed by 14.3 points a game against 11.2 across the previous two openers.
Why that is surprising. Openers are supposed to be the easy week; big programmes play small ones and win by forty. Instead the mismatches were the games that broke.
What you get: a way to tell a loud week from an unusual one.
Start with the one nobody saw
Massachusetts went to Rutgers and won 37–21.
Our preseason model made Rutgers a 38-point favourite. Three betting markets made them a 29-point favourite. Both were wrong by more than two touchdowns in the same direction, and the club they agreed on lost outright.
Put a probability on it and it comes out near 1%. We do not get to wave that away as a bad model, because the money said the same thing.
The mismatches broke in both directions
Look at the shape of that chart. The surprises are not upsets so much as blowouts nobody ordered.
Pittsburgh beat Miami of Ohio by 45 when 13 was the number. LSU beat Clemson by 41 when 10 was the number. Texas beat Texas State by 52. Those are favourites winning — by two or three times what anyone expected.
And then Michigan, a 25-point favourite, survived Western Michigan 13–12.
So the week was not simply chaotic in the underdogs’ favour. It was chaotic in both directions at once, which is a different thing and a worse one for anybody holding a forecast.
Eight upsets is a loud week. Missing by 14 points a game is an unusual one. They are not the same measurement.
The ProfessorLoud is not the same as unusual
Here is the distinction worth taking away, and it is the reason we measure the week rather than the weekend’s best game.
Eight upsets in 47 games is 17%. That is higher than last year’s 12% and lower than 2024’s 21%. On upsets alone, this week was unremarkable.
The week only looks strange when you stop counting who won and start measuring by how much. That is where 14.3 comes from, and it is 28% above the recent norm.
Those two numbers disagree because they answer different questions. Upsets count the times the favourite lost; the average miss counts how wrong the number was every time, including the games the favourite won.
What to take home: count the misses, not the shocks
When somebody tells you a week, a market or a season was unpredictable, ask which of those two things they measured.
Shocks are memorable and rare; misses are boring and constant, and the second is the honest measure of whether anybody knew what they were talking about.
A forecaster who is never surprised but is quietly wrong by two touchdowns every week is worse than one who takes an occasional beating and is otherwise close. The first will feel more reliable. Count the misses and you will see which is which.
We will run this measurement after every week of the season, and it will be interesting when it stops being high.
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
Results and preseason SP+ ratings from the CollegeFootballData API, pulled the morning of 6 September 2026. Week one played 412 games; 47 had both clubs carrying an SP+ rating, and only those are counted, because a game against an unrated opponent has no expectation to beat.
Expected margin is the home club’s SP+ rating minus the away club’s, plus 2.5 points for home field, with the adjustment dropped at neutral sites. Surprise is the actual margin minus that. Win probability converts the expected margin through a normal distribution with a 16.5-point spread, the same setting the odds board uses so the two never disagree.
Mean absolute miss: 14.3 points in 2026, against 12.4 in 2024 and 10.0 in 2025 — a prior-year mean of 11.2, so this week ran 28% above it. Medians move the same way, 11.8 against 9.9 and 8.9, so a single blowout is not carrying the result. Outright upsets: 8 of 47 (17%), against 8 of 39 (21%) in 2024 and 6 of 48 (12.5%) in 2025.
Massachusetts 37, Rutgers 21, at SHI Stadium on 3 September, verified against the game feed and against three published spreads — Bovada −29, DraftKings −29.5 twice, mean 29.3. SP+ made it 38.1. The implied chance of the result was about 1%.
Two limits. Preseason ratings are at their least informative in week one by construction — they have no current-season evidence in them — so some elevation is expected every year; the comparison against the same week in prior seasons is what controls for that. And 47 games is a small sample for a mean, so treat 28% as a strong hint rather than a settled fact until a few more weeks agree with it.
Reproduce it: python3 scripts/cfb_week_review.py.
