August 31, 2024, opening Saturday of the college football season. A power-conference team with 81% of its prior-year offensive production returning, plus the addition of a portal quarterback rated as the eighth-best in the country, walks into a neutral-site game as a three-point underdog against an opponent that lost six starters to the NFL Draft, including its starting quarterback and three All-Conference defensive players. By the fourth quarter, the underdog has rolled up 38 points on 467 yards and won by 17. The handful of analysts who had quoted SP+’s preseason projection — which had favored the underdog by four points — get a quiet moment of vindication, mostly on Twitter and mostly to small audiences. The poll voters, who had ranked the opponent six spots higher, are now hastily revising. The model had known something the human consensus did not. The thing it knew was a number called returning production, and it is, by some distance, the single most predictive variable in college football.
The relationship between returning production and next-season performance has been studied seriously in college football analytics since the early 2010s, when Bill Connelly and Brian Fremeau at Football Outsiders first began publishing the percentage of each program’s prior-season offensive and defensive production that was coming back. The intuition was straightforward enough. College football, unlike the NFL, has structural roster turnover every year — graduating seniors, NFL Draft departures, increasingly the transfer portal — and the teams that retain more of their production from year to year tend to do better than teams that have to rebuild. What the data revealed, however, was that the predictive power was much larger than most pundits assumed. Returning production correlates with year-over-year change in SP+ at roughly r = 0.55 across two decades of data — the strongest single-variable correlation any public CFB analytical model has surfaced.
I have been writing about football analytics since 2018, with a focus on college football’s particular chaos, and the variable I find myself reaching for first when projecting any team’s upcoming season is the one this article is going to unpack. Returning production — what it measures, why it matters so much in college football specifically, where it breaks, and how to use it without losing your sense of nuance, is the subject of this article.
The origin: where returning production came from
The concept of returning production has a fairly clear origin in the work of Bill Connelly at Football Outsiders in the early 2010s. Connelly, then writing the S&P+ and FEI ratings work, noticed that the year-over-year persistence of program quality varied dramatically depending on how much production a team retained. A team that returned 90% of its offensive output (measured by snap-weighted contribution to the team’s prior season) tended to start the new season at or above its prior SP+ level. A team that returned 50% tended to drop ten or more SP+ points.
Connelly’s original framing measured returning production separately for offense and defense, with each side weighted by the contribution of returning players to the prior season’s production. An offensive line that returned three of five starters but had lost the All-American left tackle would score lower than a unit that returned five role players. A receiving corps that returned three depth-tier receivers but lost its 1,200-yard primary target would not get full credit for those returnees. The math, in other words, is weighted by actual contribution rather than raw counts of returning bodies.
By the mid-2010s, the variable had become a standard input to public SP+ preseason projections. Connelly published returning-production rankings each spring, and the rankings became one of the most-cited preseason indicators in serious college football coverage. ESPN’s FPI incorporated the variable into its own projections. The relationship between returning production and next-season performance, by 2018, was one of the few stable, broadly-accepted truths in the public CFB analytics conversation.
The transfer portal era, which began in 2018 and accelerated dramatically through the NIL changes of 2021, complicated the original formulation. A team could now subtract substantial returning production via outbound portal departures while adding equally substantial production via inbound portal arrivals, all in a single offseason. The variable’s structure had to expand to account for both flows. By the 2024 cycle, returning production was effectively a net measure — production retained plus production added via portal, weighted by quality, minus production lost via NFL Draft and outbound portal.
How returning production works: in plain language
The basic calculation is structured around the prior season’s production. Imagine a 13-game college football season in which a team’s offense averaged 35 points and 450 yards per game. The model attributes those numbers, on a snap-by-snap basis, to the players on the field at the time. A wide receiver who played 800 snaps and caught 75 passes for 1,100 yards is responsible for, very roughly, an identifiable fraction of the offensive output during his on-field time.
Returning production is the percentage of that prior-season output produced by players who are still on the roster for the upcoming season. A team with 80% returning offensive production has retained players who, collectively, were responsible for 80% of the prior year’s offensive output. A team with 40% returning has lost most of its productive players to graduation, NFL Draft, or transfer.
The defensive side is calculated similarly but uses snap counts and defensive contribution (tackles, sacks, pass deflections, coverage charting) rather than raw production. The two sides are then averaged or considered separately, depending on the model.
The portal era added two new flows. Inbound portal additions bring production from another program into the team; the contribution is weighted by their prior production and the quality differential between their old and new context. Outbound portal departures are subtracted from the returning-production calculation the same way that NFL Draft departures or graduations are. The net returning-production number, in 2026, is the algebraic combination of all four flows.
The critical component: weighting by quality, not just by count
The single most important feature of well-built returning-production measures is that they weight by quality of production rather than by raw player count. A team returning five of seven defensive backs sounds positive until you note that the two it lost were the All-Conference cornerback and the team’s leading interceptor. The five returnees might collectively account for 30% of the unit’s prior-season contribution; the two losses might account for 50%.
The naive version of returning-production — counting players, not contribution — produces consistently misleading rankings. A team that retains six of its top seven offensive linemen but lost its starting quarterback and primary running back would, by player count, look “experienced.” By contribution-weighted measurement, it would be near the bottom of the league. The quality-weighted version is the one that correlates with next-season performance. The count-based version is essentially noise.
Returning production vs the alternatives: a comparison
Returning production is one of several preseason-projection inputs that public college football models use. A short comparison:
| Variable | What it measures | Predictive strength (r vs next-year SP+) | Source |
|---|---|---|---|
| Returning production | %% of prior production returning, weighted by contribution | ~0.55 (highest single variable) | SP+, FPI, public databases |
| Prior SP+ rating | Team’s per-play efficiency last year | ~0.45 | ESPN, Football Outsiders archive |
| Recruiting class composite | Avg star rating of incoming class | ~0.30 | 247Sports, Rivals, On3 |
| Five-year SP+ average | Long-run program quality | ~0.40 | SP+ archives |
| Coaching continuity | Same HC + OC + DC vs new staff | ~0.15 | Various public databases |
The variables are not independent — a team with high returning production usually had a strong prior SP+ rating and tends to have stable coaching. The combined models used by SP+ and FPI weight these inputs to maximize predictive accuracy, with returning production carrying the heaviest weight. The honest preseason CFB writer pulls three or four of these and triangulates rather than betting on one.
What the data needs: inputs
Returning-production calculations require detailed roster and snap-count data for the prior season. The minimum inputs are game-by-game roster tracking, snap counts at the position level, and departure logs for each player (graduation, NFL Draft, transfer in, transfer out).
The data is mostly assembled by hand or through scraping public school websites, conference publications, and the NCAA’s tracking systems. 247Sports publishes transfer portal databases that are widely used. Football Study Hall and Bill Connelly’s Study Hall newsletter publish weighted returning-production rankings annually. The cfbfastR R package includes some of the underlying snap-count data for those who want to compute the metric themselves.
The cleanest public-facing version of returning production for the 2025 and 2026 seasons is the one Connelly publishes each spring at his newsletter and at ESPN. The methodology accounts for portal inflows and outflows separately, weights by snap-weighted production, and produces a single percentage for offense and defense plus a combined number. That set of percentages is, in my opinion, the single most valuable preseason number a college football writer can have on their desk.
Building the analysis: a working framework
The practical workflow:
- Pull the team’s returning-production percentages for offense and defense, including portal flows.
- Compare to the team’s prior-year SP+ rating. A team with high returning production should hold close to or slightly above its prior SP+; a team with low returning production should regress meaningfully.
- Adjust for coaching changes. A new head coach, new offensive coordinator, or new defensive coordinator typically subtracts 3-5 SP+ points from the projection even with high returning production.
- Layer in the recruiting class composite. Freshman impact is unpredictable but, on average, elite recruiting classes add measurable value, especially on the defensive side.
- Read the schedule. A team with low returning production faces a less forgiving early schedule (no SP+ buffer) than one with high returning production. The projection has to interact with the games being played.
Where this gets weird: common mistakes
The pitfalls in returning-production writing.
Counting players, not contribution. The most common amateur version of the analysis is to report “this team returns 15 starters” without weighting for the production those starters represented. The publicly-published “returning starters” metric is much less predictive than the weighted version.
Ignoring quarterback continuity. A team that returns 85% of its non-QB production but is breaking in a new starter at quarterback is in a different situation than the headline number suggests. Quarterback is a position with such outsized impact that the model’s general returning-production framework can understate the disruption of a QB change. Separate QB-specific returning rates are sometimes published; they should be.
Treating transfer-in additions equally to natural returners. A player coming from a Group of Five program to a Power Four team brings production that does not translate one-for-one. The cleaner models discount inbound transfer production by a quality factor that reflects the league differential. The simpler models do not.
Ignoring injury history. Returning production assumes that returnees will, in fact, play. A team that returns its starting quarterback who is coming off ACL surgery in March is not the same as a team that returns a fully healthy starter. The variable does not account for medical context; the careful writer does.
Over-relying on the single variable. Returning production is the strongest single predictor, but a 0.55 correlation still leaves substantial variance unexplained. Teams with elite returning production sometimes underperform (coaching changes, injury, off-field issues); teams with poor returning production sometimes overperform (breakthrough freshman classes, scheme shifts). The variable is a strong starting point, not a verdict.
When returning production shines: use cases
The strongest applications:
Preseason team rankings. The cleanest defense against poll-voter inertia and recency bias in preseason rankings is to anchor projections to returning production. A team coming off an 11-win season but returning 35% of its production is structurally a worse bet than a team coming off a 7-win season returning 80%. The polls rarely reflect this; the data consistently does.
Conference championship odds. When two teams in a conference have similar prior-year SP+ ratings but very different returning production, the model favors the one with more returning production in the new season. Over the long run, that signal beats the prior-year rating differential.
Transfer-portal evaluation. A team that brought in three top-25-rated portal players from peer programs has added meaningful returning-production-equivalent. Comparing the gross additions to the gross departures produces a net figure that signals direction more reliably than counts of bodies.
Early-season betting markets. Vegas opening lines for early-season college football games are, in my experience, less efficient than the SP+ projections built on returning production. The gap is meaningful enough that returning-production-based projections, applied carefully, have been a sustainable edge for the public-facing analytics community in the first three or four weeks of each season.
A working example: the 2023 LSU rebuild
LSU’s 2023 season is a clean returning-production case study. The Tigers entered 2023 with returning offensive production around 70% (Jayden Daniels back, strong receiver corps intact, two of five offensive linemen returning with starting experience), defensive returning around 45% (significant draft and portal losses on the defensive side), and the addition of meaningful inbound portal production at multiple positions. SP+ projected them as a top-15 team, slightly below their 10-win finish from 2022 but expected to remain competitive.
The season unfolded almost exactly as the model predicted. Daniels won the Heisman Trophy with one of the most efficient quarterback seasons in modern college football. The offense ranked in the top three nationally by EPA per play. The defense ranked outside the top 50, exactly the kind of unit returning-production projections had flagged as a problem. The team finished 10-3, won a bowl game, and ended ranked roughly where SP+ had pegged them in August. The returning-production framework had captured the structural reality of the roster more accurately than the polls did, the betting markets did at the opening line, or most preseason media projections did.
That is, in miniature, what the variable does well at scale. Year over year, across the FBS, the teams that returning-production rankings favor tend to perform better than the teams the polls favor. The discipline of starting with returning production and only adjusting from there is one of the cleanest editorial habits in public CFB writing.
The limits: what returning production cannot tell you
The honest version of this writing names the limits.
Returning production cannot predict freshman impact. The metric, by construction, only counts players whose prior-year production it can measure. A team that signs five-star freshmen who become immediate impact players adds production the model didn’t see coming. Returning-production projections systematically underestimate teams with elite incoming classes; SP+ tries to correct for this via the recruiting composite input, but the correction is partial.
Returning production cannot capture scheme changes. A new offensive coordinator who installs a fundamentally different scheme can render the returning players’ prior-year production partially irrelevant. The same wide receiver who caught 75 passes in a quick-game offense may have a very different role in a vertical-passing scheme. The variable assumes continuity that, after a coaching change, may not exist.
Returning production cannot model injury volatility. A team that enters the season with 80% returning production but loses its quarterback to injury in week three is, by mid-October, no longer a high-returning-production team. The variable is a snapshot taken in the summer; the season is a moving target.
Returning production cannot, finally, account for the genuine chaos of college football. Conference realignment, NIL deals, transfer portal volatility, and the increasingly mercenary nature of the sport mean that the year-over-year continuity that returning production measures is itself a fading concept. The variable’s predictive power has been remarkably stable over two decades, but the 2025 and 2026 seasons are testing the boundaries of how much production retention even means in a sport where rosters can be rebuilt in March.
One additional limit: the public version of returning production data has its own measurement variance. SP+’s number, FPI’s number, and other publisher’s numbers for the same team can differ by five percentage points or more depending on weighting methodology and treatment of edge cases (walk-ons, partial-season starters, position changes). The careful writer cites the source and avoids treating any single calculation as definitive.
Frequently asked questions
What is a “good” returning production percentage?
League average for FBS teams runs around 55-60% combined returning production. An elite returner is in the 75-85% range and can expect to maintain or improve on their prior-year SP+ rating. A team returning below 40% combined is structurally rebuilding and should be projected to regress 8-15 SP+ points. The very top of the league (90%+ returning) is rare and almost always projects forward strongly.
How has the transfer portal changed the metric?
Significantly. Before 2018, returning production was almost purely a function of NFL Draft departures and graduation. Now it’s a net flow that includes inbound portal additions (which can be very large for top programs) and outbound portal departures. The net figure tells you what production a team has on its 2026 roster, but the volatility of the underlying flows is much higher than it was in the pre-portal era. Year-to-year prediction is still possible; the confidence intervals have widened.
Does returning production work for early-season betting?
In my experience, yes. Vegas’s opening lines in college football tend to be less efficient than SP+’s projections in the first three or four weeks of the season, partly because the betting market is reacting to last year’s results and partly because public bettors are anchored to poll rankings that lag the model. Returning-production-based projections have been a measurable edge in early-season point-spread markets for the public-facing analytics community. Whether that edge persists as the sportsbooks integrate more advanced models is an open question.
Where can I see returning production numbers for the upcoming season?
Bill Connelly publishes annual returning-production rankings at his ESPN columns and his Study Hall newsletter on Substack, typically in April or May after the spring portal window closes. Football Study Hall and Football Outsiders’ archive maintain historical versions. The cfbfastR R package provides the underlying snap-count data for those who want to compute the metric themselves.
Sources and further reading
- Bill Connelly’s Study Hall newsletter — the canonical source for annual returning-production rankings and the writing that contextualizes them.
- ESPN’s SP+ landing page — current season SP+ ratings that incorporate returning production as a primary input.
- Football Study Hall — the spiritual successor to the Football Outsiders college football work, with returning-production analysis.
- cfbfastR R package — the open-source backbone for college football play-by-play and snap-count data.
- 247Sports transfer portal database — the most-cited public tracker for inbound and outbound portal moves.
The neutral-site upset that opened this article — heavy returning production team, low returning production opponent, a model that knew something the polls did not — is exactly the kind of game returning-production-based projections handle well at scale. Over a season, the writers who anchor their preseason takes to the variable are right more often than the writers who anchor to last year’s rankings, by a meaningful margin. The work is unglamorous. The predictions are sometimes boring. The track record is, by every measurement I have run, the best public-facing predictive signal in college football. For the broader frame on how the SP+ model fits together, our guide to SP+ is the natural companion piece.
