Year one is a free pass
NC State fired Dave Doeren on October 24, 2026, after 98 wins and 74 losses. To find out what actually gets a college football coach fired, we trained a machine-learning model on every FBS head coach since 2009 and graded it only on seasons it had never seen. Its biggest tell was not the record. It was how long the coach had had the job.
What the model looks at
Every row is one head coach at week 8 of a season, the point the 2026 season has just reached. The model gets only what was known that week and has to say whether he will be fired after it, in the rest of the season or the offseason that follows.
Its 28 inputs, called features, cover this season’s record and point margin, a projection of the rest of the season from the published schedule, last season and the one before, the program’s own five-year norm, preseason expectations, and the coach himself: his years in the job and his record in it. Who was fired, and when, comes from Wikipedia’s coaching-changes table for each season. Resignations are left out, because some are firings and some are not. 2020, the COVID season, is left out too.
Graded only on seasons it had never seen
A model can look brilliant on the data it learned from. So every season from 2016 to 2025 was predicted by a model trained only on the seasons before it: 2016 from 2009–2015, and 2025 from 2009–2024. That is 1,140 coach-seasons it saw cold, 100 of which ended in a firing.
| on the seasons it never saw | this model | win % only | chance |
|---|---|---|---|
| ranks the fired coach above a safe one | 81% | 74% | 50% |
| of each season’s top 10, fired | 30 of 90 | 14 | 7.9 |
Take one coach who got fired and one who didn’t. The model ranked the fired one higher 81% of the time.
Knowing only each coach’s win percentage gets you to 74%. The other inputs are worth the gap: of its ten highest-risk coaches in each of the nine seasons, 90 picks in all, 30 were fired, against 14 for the record alone and about 8 by chance.
What it leans on
To see which inputs carry the model, we scrambled one group at a time inside each tested season and measured how much of its ranking ability it lost. Years in the job came out on top by a wide margin, followed by the rest of the season. This season’s record came third.
| group of inputs | ranking ability lost (AUC) |
|---|---|
| Years in the job | 0.155 |
| Rest of the season (projection) | 0.071 |
| This season’s record | 0.027 |
| Last season | 0.023 |
| vs. the program’s norm | 0.014 |
| Schedule-adjusted strength | 0.009 |
| Expectations | 0.009 |
Year one is a free pass
Hold everything else fixed, give every 2025 coach the same number of years in the job, and average what the model says:
| year in the job | 1 | 2 | 3 | 4 | 5 | 6+ |
|---|---|---|---|---|---|---|
| chance of being fired | 1.0% | 4.3% | 10.5% | 11.3% | 11.7% | 11.9% |
A first-year coach almost never goes, whatever the record. By year three the risk has climbed about tenfold, and it keeps creeping up after that. Patience runs out on a schedule, not just on results.
Where Doeren stood
At week 8 of 2026, in his 14th season and at 3–4, the model gave Doeren a 20.2% chance of being fired. That was 7th of the 138 coaches it scored, against an average of 6.6%. A year earlier the version trained through 2024 had him at 12.8%, 20th of 129.
Who it says is next
| still in the job, week 8 of 2026 | school | record | chance |
|---|---|---|---|
| Scotty Walden | UTEP | 1–6 | 43% |
| Deion Sanders | Colorado | 2–5 | 31% |
| Bryant Vincent | Louisiana Monroe | 1–6 | 30% |
| Gerad Parker | Troy | 3–4 | 21% |
| Joe Moorhead | Akron | 1–7 | 20% |
| Fran Brown | Syracuse | 3–4 | 20% |
| Sean Lewis | San Diego State | 3–5 | 19% |
| Chris Creighton | Eastern Michigan | 4–5 | 18% |
| Tony Sanchez | New Mexico State | 3–4 | 18% |
| Jeff Lebby | Mississippi State | 4–3 | 17% |
These are chances, not verdicts. Even at the top, a 43% chance means it usually doesn’t happen. All 138 coaches are in the download.
The part that argues against us
Association, not cause. The tenure curve is what the model learned from real coaches. It does not show that staying longer causes a firing.
Resignations are out. Some were forced, so where a push is dressed as a resignation the model may understate the risk.
Years in the job come from the coaching-changes history, which starts in 2007 and lists only FBS moves. For 2026, every coach still in the job takes his first season from Wikipedia’s current-coaches list instead. A first pass missed that for schools new to FBS: it counted Sacramento State’s first-year coach as a 20-year veteran and put Doeren 8th. He is 7th.
2026 is still being played. The 2026 chances use games through October 25. Baylor fired Dave Aranda that same day; the model had him at 11.2%, 26th.
The record. Doeren’s NC State game log is 98–74, and so do Wikipedia’s season rows; some outlets give 98–73.
Get the data
Free: the model’s chance for every FBS head coach at week 8 of 2026, and every coach-season since 2009 with what happened and what the model said before it happened.
138 coaches in 2026 · 2,108 coach-seasons · the methodology write-up
Sources: game results from ESPN’s public scoreboard feed (FBS, 2000–2026, pulled 2026-10-25); coaching changes and current coaches from English Wikipedia’s “NCAA Division I FBS football coaching changes” tables (2007–2026) and “List of current NCAA Division I FBS football coaches” (CC BY-SA 4.0). Model: the average of a logistic regression and depth-2 gradient-boosted trees, tested walk-forward on 2016–2025 (2020 out). Images in the video: Dave Doeren, “2025-1219 Dave Doeren.jpeg” by Bobak Ha’Eri (CC BY 3.0) and “321GreaterGoodCoach.jpg” by Dallasoliver (CC BY-SA 4.0, cropped), via Wikimedia Commons; the pointing hand from “Index finger = attention.JPG” by Pogrebnoj-Alexandroff (CC BY-SA 3.0, Wikimedia Commons), rendered in 3D and ASCII and shared under the same licence; the brain from NIH BioArt (public domain); school logos used only to identify each program. 3D models made with TripoSG (MIT licence). Analysis and figures: reelgorithm.py.