Prediction
We Rebuilt the Model This Week. It Still Won't Give You Nebraska by Three Touchdowns.
Husker Nation, Nebraska is finally on the schedule. Before the picks, an accounting.
Two changes to how this works. From now on we only pick games with a Big Ten team in them — Week Zero was six games with no conference involvement anywhere, which is a fine test of a model and a poor use of your time. Fourteen games this week.
And we spent this week tearing the model apart.
What Week Zero actually taught us
We went 4-2. The pick we were loudest about — TCU over North Carolina, “the one game where we’re more confident than the market” — lost, because the model credited TCU nearly three points for North Carolina’s travel in a game both teams flew to Dublin for.
That was the embarrassing bug. Digging for it turned up a worse one.
Across 175 historical games where a Big Ten-class team hosted a Group of Five team in September, this model came in 5.4 points under the true margin. Every time. It wasn’t a bad read on one game. It was a blind spot in the machinery.
The six things we changed
- Conference class. The model had no idea what a power conference was. It judged rosters on ratings and recruiting and never once asked what league anybody played in.
- Transfer quarterbacks. “Returning production” means returning to the same school, so Nebraska signing a three-year starter scored as an empty position.
- Coordinators. College football’s data providers publish head coaches and no coordinators at all. We built the file ourselves — 1,550 team-seasons back to 2015.
- Coaching track records. One season of a coordinator’s work predicts nothing (correlation +0.08). Two or more predicts real things (+0.22). The model now weighs a career, not a year.
- Talent-adjusted coaching. A coordinator’s rating credits his players and his scheme together. What moves with him is the scheme.
- Calibration. The big one. The model was systematically too timid when running on preseason information — predicting 100 where the truth was 113.
That September blind spot went from −5.4 points to −1.5. The betting market’s error on those same games is −1.8.
For the first time, on this specific kind of game, our numbers are better calibrated than Vegas.
Nebraska 34, Ohio 17
Saturday, 11:00 a.m., Lincoln, FS1. We have Nebraska by 17.3. The market says 23.5.
Before the rebuild we had this at 12. We were wrong, and here’s what changed.
Nebraska’s quarterback is not a question mark. Anthony Colandrea has started 31 games across Virginia and UNLV and was the Mountain West’s Offensive Player of the Year last season. By expected points added he was better last year than the man who left. Every model in the country scored him as nothing, because production only counts if you stay put. Ours doesn’t any more.
Ohio is not a cupcake. They were 19th in FBS in rushing, quarterback Parker Navarro is back, and in one game last season they threw eight passes against 57 runs. More telling: Ohio finished 126th in talent and 61st in defense. That is a very well coached football team.
Which brings us to Rob Aurich. Nebraska’s new defensive coordinator took a San Diego State defense that ranked 28th in the country on the 86th-most talented roster. Nebraska last year: 44th on the 20th-most talented roster.
Read those twice. Aurich beat his personnel by 58 places. What Nebraska did last year missed its personnel by 24.
That is the game. Not whether Nebraska has more talent — they have enormously more. Whether a coach who has been squeezing blood from Mountain West rosters can do it with Big Ten ones, immediately, against the one opponent built to test a run defense.
One caution, and it’s ours to make. Aurich’s other great work was at Idaho, in the FCS, where nobody publishes the ratings this model runs on. So his provable record is one season, and we deliberately discount a one-season record. If you think that undersells him, you may well be right. We are not going to pretend the data says so.
The rest of the board
Every line below is live from the book, pulled Sunday.
| Game | Us | Market |
|---|---|---|
| Ohio State over Ball State | 52-0 | 50.5 |
| Rutgers over Massachusetts | 45-7 | 29.5 |
| Iowa over Northern Illinois | 38-7 | 31.5 |
| Indiana over North Texas | 42-10 | 40.5 |
| Michigan over Western Michigan | 34-17 | 27.5 |
| Penn State over Marshall | 37-17 | 24.5 |
| Oregon over Boise State | 35-17 | 24.5 |
| Illinois over UAB | 38-16 | 28.5 |
| USC over Fresno State | 33-17 | 22.5 |
| Notre Dame over Wisconsin | 31-14 | 20.5 |
| Washington over Washington State | 31-17 | 23.5 |
| Toledo over Michigan State | 27-24 | MSU by 10 |
| UCLA over California | 27-24 | UCLA by 1.5 |
Toledo is the only game where we take the underdog outright, and at 55.5% a toss-up means we don’t know. The market has Michigan State by ten. That is the widest gap on our board and the one we’d bet against ourselves on.
Two where we’re more bullish than the book: Rutgers, by nine and a half more than the line, and Ohio State, where the model rounds to 100% — it is telling you it has never seen a gap this size, not that it knows the future.
And notice UCLA at California. Before the rebuild we had Cal winning. Now we have UCLA, agreeing with the market. A model that moves a point and a half flips games decided by a field goal, which is most of them.
Across all fourteen we still come in under the market ten times. That used to be our headline. After this week’s work we know better: it is the model’s oldest habit, we have cut it roughly in half, and what’s left is disagreement rather than a defect.
What we’re still not claiming
The model does not beat the spread and we are not telling you to bet.
Week One is its weakest week of the year, because nobody has played and every number is built on last season. And on this game specifically, the market says Ohio scores 12 while we say 16. We have tested that disagreement every way we know and it is genuine, not a bug. One of us is wrong and Saturday settles it.
Husker Nation, eleven o’clock. First real football since New Year’s Eve.