Team-seasons
1,231
Athletics Research · Roster Retention
Athletics Research · Ten seasons of college football
The second recruiting battle
Seasons now turn on keeping the players you already have. Across 1,231 team-seasons at 129 programs, two things predict a college football season: the season before it, and how much of the team came back. Coach pay, athletic budget and recruiting spend register nothing. What keeps a roster together is not money.
Download the report (PDF)Programs
129
Seasons
10
Coaching changes
302
What moves your win total
Bar length is how much each factor shifts a season, compared like for like. Two factors move it. Every line a school can spend does not.
Ten seasons of college football, each program measured against its own past.
Two things predict a college football season: the season before it, and how much of the team came back. Recruited talent does not survive alongside them, because a program with talent is usually one that won last year and the record already carries that. Coach pay, athletic budget and recruiting spend register nothing.
One of those two has already happened. The other is decided every year, by players who now have somewhere else to go. That is a second recruiting battle, fought against the roster a coach already has.
How the roster is measured
Three definitions
Retention
The share of last season’s production still on the roster this season.
a typical season
A point
One percentage point of that roster. Six points is six percent of last season’s production, kept or lost.
Level and change
Both are in the model, so each is read while the other is held still.
A team that goes 6-6 then 9-3 has a level of .750 and a change of +.250. A team that goes 9-3 then 9-3 has the same level and no change.
Retention counts production rather than bodies, so losing a starter costs more than losing a walk-on.
The question
What keeps players
1,231 team-seasons, 129 programs, each measured against its own past.
What predicts the share of a roster that stays · program and year fixed effects · within R² 0.30
| Factor | Effect on retention | p |
|---|---|---|
| The record improved | +0.228 | <0.001 |
| The head coach changed | −0.050 | 0.001 |
| Retention the year before | −0.242 | <0.001 |
| How good the record was, at a given change | −0.131 | 0.005 |
| Four- and five-star share of the roster | −0.264 | 0.052 |
| The money lines | ||
| Recruiting budget (doubling) | −0.026includes zero | 0.07 |
| Athletic aid (doubling) | −0.019includes zero | 0.11 |
| Head coach pay (doubling) | −0.036includes zero | 0.15 |
| Athletic budget (doubling) | +0.015includes zero | 0.75 |
| NIL and booster money raised | +0.011only in a 186-season subset | 0.04 |
Set out as raw averages rather than coefficients, three things stand out.
.585vs.523
The coach stayed, or was replaced.
A change costs about six points of the roster.
.605vs.530
The record improved, or declined.
Teams heading the right way keep more, whatever the record.
.563vs.574
Most four- and five-star players, or fewest.
The better the roster, the more of it leaves.
Every line a school can spend sits at zero. Recruiting budget, athletic budget, coach pay, athletic aid. NIL and booster money reaches significance in one specification on a seventh of the sample and nowhere else, which is not enough to call an effect.
What does hold a roster is whether the staff is the same people, and whether the program is going somewhere. Note that winning in level runs negative while the change in record runs strongly positive: a good team that has plateaued gets picked apart, while a team on the way up holds together. And the more blue-chip talent on the roster, the more of it leaves, because those are the players other programs come after.
The players hardest to keep are the ones you most wanted to sign.
What that implies
It is a relationship, not a payment
A coaching change costs six points of a roster while doubling the recruiting budget costs or buys nothing. The plainest reading is that a player’s attachment is to people rather than to a program. He committed to a position coach who sat in his living room. When that person leaves, the thing holding him leaves too.
Which makes recruiting continuous. A staff re-recruits its own roster every year, against opponents who can see exactly which of its players are worth a call, and the evidence says the tool for that is not a bigger budget.
What the filings cannot hold
Asking the players
Survey research from Tudor Collegiate Strategies with Niche, 7,129 to 31,000 respondents per wave.
School accounting records what was spent, not whether a player felt wanted or whether the message he got sounded written for him. Those can only be asked about.
63%
Say personal, relevant contact influences whether they visit or apply
16%
Say the outreach they actually received was very personalized
Six in ten students say individual contact is what moves them. Fewer than two in ten report getting any. That gap is the one measure available of the thing the filings have just ruled out being about money.
It is evidence about prospects during a college search, not athletes at a transfer window. But the mechanism it describes, being treated as a person rather than a name on a list, is the same one a coaching change destroys.
Where this goes next
The next step
The hypothesis left standing is that what holds a roster together is the relationship work the surveys measure on the way in. Nothing here establishes that. It establishes that the paid explanations fail and that staff continuity does not.
- Ask athletes who stayed and who left, in the same season. The existing survey instrument, pointed at current players at the transfer window instead of prospects during a college search.
- Player-level movement, not team totals. Who left, from which position group, after how much playing time, and to where, which separates a walk-on’s departure from a starter’s.
- What an athlete was actually paid. The accounts used here are school-level and cover facilities and scholarships too. Without per-athlete figures the money explanation is ruled out only in aggregate.
Limitations
What this study cannot tell you
- Correlation, not cause
- Coaching changes are not random. They follow bad seasons, and a program losing players may be why a coach was replaced rather than the result of it. The model holds the change in record constant, which helps, but it is not an experiment.
- Partly by construction
- Retention the year before runs negative partly because of how rosters work. Keep everyone one season and there are more seniors to lose the next.
- Most of it is unexplained
- The model explains 30% of retention, and most of that is not the measured factors. Coaching change, momentum, talent and the money lines account for about 11 points of it. The rest comes from the year (the portal era hit every program at once) and from which program it is. Roughly 70% is unexplained either way.
- The NIL proxy is rough
- Those accounts cover a school’s whole athletics nonprofit, not payments to players, and per-athlete figures are not public. Our earlier work on who funds college athletics sets out what those filings can and cannot show.
- Football, top two divisions
- Mostly the top two divisions, which is where retention and talent data exist.
- The survey is about prospects
- The survey figures are Tudor Collegiate Strategies’ published results with Niche, quoted as published, and concern prospects rather than current athletes. The newest wave located was November 2022.
- No revenue-sharing year
- Direct payments to athletes began in 2025-26. That survey publishes around spring 2027 and may change this picture.
Appendix
Data and definitions
Sources
| Source | What it provides | Coverage | N |
|---|---|---|---|
| Equity in Athletics Disclosure Act surveyUS Department of Education | Recruiting spend, athletic aid, coaching counts and salaries, participation, enrollment, revenue and expense | 2015-16 to 2024-25, 37 sports, all divisions | 20,353 school-years 2,037 schools/yr |
| 247Sports Composite ratingsvia sportsdataverse | Per-recruit star rating and composite grade; roster talent | Signing classes 2015-2026, football | 48,362 recruits |
| Published game schedulesvia sportsdataverse | Win-loss records, folded up from completed games | Seasons 2015-2026, football | 5,360 team-seasons |
| Niche / Tudor Collegiate Strategies surveysincl. the 7,129-junior wave | Student college-search behavior and how outreach was received | Classes of 2021-2023 | 31,000+; 20,045; ~12,000; 9,461; 7,129 |
Variables
| Variable | Definition | Role |
|---|---|---|
| Recruiting budget | Reported recruiting expense: travel, lodging and meals for prospects and staff on visits, phone, postage and related costs. Men’s figure in the football model, total elsewhere. | Predictor, logged |
| Class rating | Mean composite grade of a program’s signees in one class. Restricted to classes with 15 or more rated recruits, since an average over a handful is unstable. | Outcome, football |
| Athletes | Participants across all sports, men and women, as reported. A count of people, not a measure of quality. | Outcome, small colleges, logged |
| Win percentage | Wins plus half of ties, over completed games. | Predictor and outcome |
| Roster talent | 247Sports team talent composite: the summed recruiting grades of players on the roster, a four-year rolling measure rather than one class. | Predictor |
| Returning production | Share of the previous season’s production back on the roster. | Control |
| Rest of athletic budget | Total athletic expense minus the recruiting line, so the control does not contain the predictor. | Control, logged |
| Head / assistant coach pay | Average annual institutional salary per head or assistant coach, as the survey asks it. Not a payroll. | Control, logged |
| Athletic aid | Athletically related student aid awarded. | Control, logged |
| Program and year fixed effects | Absorbs what is constant about a school and what is common to a year, so estimates come from a school changing against its own past. | Design |
| Within R² | Share of the remaining variation the model explains once those fixed effects have removed the differences between schools and between years. Not comparable to an ordinary R², which is much higher here because knowing which school a row belongs to predicts most of its recruiting. | Fit |
The roster, watched
Know what is being said about the players you are trying to keep
Civly Social Listening & Clips tracks your athletes, coaches and program across 14 platforms and the press, with alerts when the conversation changes.
Civly Athletics Research · report dated August 30, 2026, published September 13, 2026. Download the report as a PDF.
Spending, participation, aid and staffing from the US Department of Education Equity in Athletics Disclosure Act survey. Football ratings from the 247Sports Composite; win-loss records folded up from published schedules. Student survey figures as published by Tudor Collegiate Strategies with Niche.
Findings describe association in observed data and are not investment, contractual or compliance advice.