# What actually gets a college football coach fired? — methodology

reelgorithm.py · data through week 8 of the 2026 season (cut 2026-10-25), published 2026-10-26

**NC State fired Dave Doeren on October 24, 2026, after 98 wins and 74 losses. What actually makes a college
football coach get fired?** We trained a machine-learning model on every FBS head coach since 2009 and tested it only
on seasons it had not seen.

Three files ship with this write-up:

| file | rows | what it is |
|---|---|---|
| `coach-firing-odds-2026.csv` | 138 | every FBS head coach at week 8 of 2026, with the model's chance he is fired |
| `coach-firing-history.csv` | 2,108 | every FBS coach-season at week 8, 2009–2025 (2020 left out): record, years in the job, what happened, and for 2016–2025 the chance the model gave him before that season was seen |
| `coach-firing-methodology.md` | — | this file |

---

## 1 · The question, and what it is not

One row is one FBS head coach at **week 8** of a season, the point the 2026 season had reached when we built this.
The model is asked: given only what was known that week, will this coach be **fired after week 8**, in the rest of
the season or the offseason that follows it?

It is **not** a claim about why any one athletic director decided anything. It measures which situations have
been followed by a firing, and how often.

---

## 2 · Sources

- **Game results:** ESPN's public scoreboard feed (`site.api.espn.com`, FBS, every regular-season and postseason
  week, 2000–2026), pulled 2026-10-25. Scores, dates, conferences and rankings. Results are facts; nothing of ESPN's
  beyond them is redistributed.
- **Who was fired, and when:** English Wikipedia's "NCAA Division I FBS football coaching changes" table for each
  season (2007–2026), and its "List of current NCAA Division I FBS football coaches" for each current coach's first
  season. Wikipedia text is CC BY-SA 4.0; the kit uses the facts in those tables (team, coach, date, reason).

Every number re-derives from cached copies with the scripts in section 9.

---

## 3 · The setup

- **Outcome.** *Fired* when Wikipedia's table gives the reason as fired, dated after that season's week-8 cut.
  *Resigned* is left out of the data entirely, because some resignations are firings and some are not, and we
  cannot tell which. *Left for another job* counts as not fired.
- **Who is in it.** Every FBS program's head coach at the week-8 cut. A coach already gone before the cut is not a
  row (in 2026, Mike Locksley on Oct. 21 and Doeren on Oct. 24 are scored on the week-8 situation they left; see 5).
  A preseason interim counts as that season's coach. **2020 is left out:** COVID shortened the season and schools
  held off.
- **Inputs (28), all known at week 8:**
  - **this season:** win %, conference win %, point margin per game, current losing streak, the last three games,
    below .500, projected wins;
  - **the rest of the season:** projected final win % and the chance of reaching 6 wins, both from a
    schedule-adjusted team rating and the published schedule;
  - **last season and before:** last season's win %, two seasons ago, losing seasons in a row, last season's rating;
  - **expectations:** the program's own 5-year norm, this season vs. that norm, this season vs. last season, Power
    conference, ranked in the preseason, ranked now, the rating and its change;
  - **the coach:** years in the job (capped at 6), first year, second year, his record in this job, vs. the
    program's record before he arrived, vs. the program's norm.
- **The model:** the average of a logistic regression and shallow gradient-boosted trees (depth 2). Its settings
  were fixed on 2016–2021 before the final comparison.

---

## 4 · How it was tested

**Walk-forward.** Every season from 2016 to 2025 is predicted by a model trained only on the seasons before it.
2016 is predicted from 2009–2015, and 2025 from 2009–2024. Nothing a test season contains is seen in training.
That gives **1,140 tested coach-seasons, 100 of them fired**, a base rate of 8.8%.

| measure | this model | win % only | chance |
|---|---|---|---|
| ranks the fired coach above a safe one (AUC) | **81%** | 74% | 50% |
| of each tested season's top 10, how many were fired | **30 of 90** | 14 | 7.9 |
| average precision (base 0.088) | 0.294 | — | — |
| Brier score | 0.0708 | — | — |

"Picked the fired coach over a safe one 81% of the time" is the AUC. Take one coach who was fired and one who was
not, from the tested seasons. The model gave the fired one the higher chance 81% of the time.

The 2026 numbers come from a final model trained on every labelled season, 2009–2025. It has never seen 2026.

---

## 5 · The numbers

**What it leans on.** We shuffled one group of inputs at a time inside each tested season and measured how much
ranking ability (AUC) was lost:

| group | AUC lost |
|---|---|
| **Years in the job** | **0.155** |
| Rest of the season | 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 close to a free pass.** We took the 2025 coaches, set every one of them to the same years in the job,
and averaged the model's chance:

| year in the job | 1 | 2 | 3 | 4 | 5 | 6+ |
|---|---|---|---|---|---|---|
| chance fired | **1.0%** | 4.3% | 10.5% | 11.3% | 11.7% | 11.9% |

**Doeren.** At week 8 of 2026 (3–4, his 14th season) the model gave him **20.2%**. That is **7th of the 138**
coaches it scored, against an average of 6.6%. A year earlier the walk-forward model had him at 12.8%, 20th of 129.

**Most likely to be fired next** (still in the job at week 8 of 2026):

| | coach | school | record | chance |
|---|---|---|---|---|
| 1 | Scotty Walden | UTEP | 1–6 | 43% |
| 2 | Deion Sanders | Colorado | 2–5 | 31% |
| 3 | Bryant Vincent | Louisiana Monroe | 1–6 | 30% |
| 4 | Gerad Parker | Troy | 3–4 | 21% |
| 5 | Joe Moorhead | Akron | 1–7 | 20% |
| 6 | Fran Brown | Syracuse | 3–4 | 20% |
| 7 | Sean Lewis | San Diego State | 3–5 | 19% |
| 8 | Chris Creighton | Eastern Michigan | 4–5 | 18% |
| 9 | Tony Sanchez | New Mexico State | 3–4 | 18% |
| 10 | Jeff Lebby | Mississippi State | 4–3 | 17% |

These are chances, not predictions that anyone will be fired. A 43% chance means it usually does not happen.

---

## 6 · Data traps, stated rather than fixed quietly

- **Years in the job come from the change history.** That history starts with the 2007 table and lists only FBS
  moves. Two consequences:
  - In 2009–2011, a coach with no change since 2007 is counted as in his job since 2007, which can understate a
    long tenure. Those seasons are training only, and the model caps tenure at 6 anyway.
  - A program new to FBS reads as "no change since 2007", so its coach's tenure in its first FBS seasons can be
    overstated.
  - For 2026, every coach still in the job takes his first season from Wikipedia's current-coaches list instead.
    *Correction before publishing:* a first pass ranked Sacramento State's first-year coach as a 20-year veteran
    (41% instead of 3%) and put Doeren 8th; he is 7th.
- **Resignations are out.** Some were forced, so the model may understate the risk where pushes are dressed as
  resignations.
- **"Fired" is Wikipedia's word.** It is the reason given in its coaching-changes table, checked against the dates.
- **2026 is in progress.** The 2026 chances use games through Oct. 25. Dave Aranda (Baylor) was fired on Oct. 25,
  the day of the cut; the model had him at 11.2%, 26th.
- **Doeren's record.** His NC State game log (ESPN) is 98–74, and Wikipedia's season-by-season rows also sum to
  98–74. Wikipedia's career total line and some outlets say 98–73.
- **Association, not cause.** The tenure curve is what the model learned, averaged over real coaches. It does not
  show that a long tenure causes a firing.

---

## 7 · The charts in the video

- **The hook's ring:** one line per opponent Doeren faced at NC State (54 teams). Its width is proportional to the
  games played. Its colour is his win share against that team, on a diverging scale: green means he won most of
  them, red means he lost most, neutral means an even split.
- **"Every firing since 2009":** one logo per coach fired after week 8, stacked by season. 2009–2015 were only ever
  trained on; 2016–2025 were each tested.
- **The radar:** his percentile among the 2026 coaches on six inputs, where outward is better for him. The dashed
  shape is the median percentile of the coaches who were fired in the tested seasons.
- **"How good is it?"** and **"Year one is safe"** are sections 4 and 5. **"Most likely fired next"** is the table in
  section 5.

---

## 8 · Columns

**`coach-firing-odds-2026.csv`**:
- `rank_of_138`, `coach`, `school`, `record_at_week_8`, `years_in_job`, `power_conference`, `firing_chance_pct`;
- `status` ("in the job", or "fired <date>");
- `rank_among_still_in_job` (the video's list).

**`coach-firing-history.csv`**:
- `season`, `school`, `coach` (blank where the tables do not name him), `record_at_week_8`, `win_pct_at_week_8`,
  `years_in_job`;
- `last_season_win_pct`, `program_5yr_norm_win_pct`, `power_conference`;
- `outcome_after_week_8` (fired / stayed / left for another job);
- `used_as` (training only / test);
- `model_chance_pct`, `model_rank_in_season`, `coaches_in_season` (test seasons only).

---

## 9 · Images in the video

- **Dave Doeren, hook (rendered as ASCII):** "2025-1219 Dave Doeren.jpeg", Bobak Ha'Eri, CC BY 3.0, Wikimedia Commons.
- **Dave Doeren, features scene:** "321GreaterGoodCoach.jpg", Dallasoliver, CC BY-SA 4.0, Wikimedia Commons, cropped.
- **The pointing hand (rendered as ASCII):** "Index finger = attention.JPG", Pogrebnoj-Alexandroff, CC BY-SA 3.0,
  Wikimedia Commons. The 3D hand and ASCII rendering are adaptations, under CC BY-SA 3.0.
- **The brain (rendered as ASCII):** NIH BioArt, public domain.
- **School logos:** used only to identify each program.
- The 3D models behind the turns were made with TripoSG (VAST-AI-Research, MIT licence).

---

## 10 · Reproduce it

```
../.venv/bin/python data/coachfire/fetch.py     # ESPN scoreboards + Wikipedia tables (cached)
../.venv/bin/python data/coachfire/build.py     # games.csv, changes.csv, teamseason.csv
../.venv/bin/python data/coachfire/model.py     # walk-forward test + 2026 → model.json
../.venv/bin/python data/coachfire/kit.py       # these files; refuses if they disagree with the video
```
