16.2. Then 15.1. Then 12.4.
Every year, top college football programs schedule a game against a team from a lower division. The smaller program gets paid to show up — these are called guarantee games, named for the check the visiting school is guaranteed to receive. They tend to be blowouts: September mornings when the eventual playoff contenders warm up against opponents several tiers below them in the sport.
The spread on those games depends, in part, on how many points the lower-division team is expected to score. For three straight seasons, that number has dropped: 16.2 points per game in 2023, 15.1 in 2024, 12.4 in 2025.
That's a 3.8-point shift in three years, and the direction isn't random. The transfer portal is concentrating talent at the top of the sport faster than any previous cycle. The best players who would have stayed in the lower division are leaving, and the scoring follows them out. Meanwhile, the top-division side of those same games has been climbing: 39.8 points per game in 2023, 41.1 in 2024, 43.9 in 2025. Both ends are moving in opposite directions, every season.
Where the gap actually matters
The lower division in college football — formally the Football Championship Subdivision, or FCS — runs its own national championship and holds around 125 programs. It's a real tier of competitive football, and those programs play mostly against each other. The crossover games against top-division opponents are the exception, not the rule.
On a typical college football Saturday, roughly 83% of the games are top-division teams playing each other. The tier gap doesn't apply there. It only matters in the guarantee matchups, where one team is from the top division and the other isn't — and those games are uncommon and front-loaded. About a third happen in the first week of the season. Nearly three-quarters are done before the fourth week.
That's a narrow slice of the schedule, and it's the only slice where the lower-division scoring level changes the spread. If that number is wrong, the spread is wrong — and nobody's looking closely, because the first weeks of September are buried under conference play previews.
What made it hard to spot
One way to check whether a projection model is working correctly is to ask whether it's consistently off in one direction. Take every game projected, group them by how large the projected margin was, and see whether the actual outcomes match. If the projections are tracking reality, they should — not on any single game, but on average across similar-looking matchups.
Our college football projections looked reasonable at the season level. That's because five of every six games on the slate are top-division matchups, and the tier gap doesn't affect them. Those games dominate the full-season picture.
The error concentrated in the crossover games, and there were few enough of them that they disappeared in the aggregate. When we isolated those matchups — 196 of them across a recent window — the lower-division teams were scoring an average of 15.3 points in our projections against an actual average of 12.4. The betting market had the implied number at 11.4. We were 34% wrong in one direction, in a group the full-season numbers couldn't see.
A check that can't find a 34% error in one type of game isn't a check on that game. It's a check on everything else.
The portal is moving both ends
Before college football had a transfer portal, the rules required a player to sit out a year before playing for a new school. That made transfers costly enough that most players stayed where they were recruited, which kept talent more distributed across the sport. The portal removed that cost. Players can transfer freely now, and the best ones tend to move upward.
Guarantee game scoring is where that movement shows up most directly. When a player who would have been a starter at a mid-level FCS program transfers to an FBS school, the FCS team doesn't just lose one player. It loses the kind of player who makes those guarantee games competitive. Season by season, the lower division is getting thinner while the top of the sport is getting deeper.
This is why a number from last year is already wrong before this season starts. Using the 2024 measured level in 2025 would have been 2.7 points off. Using the 2023 level would have been 3.8. The trend is fast enough and steady enough that any stored number starts expiring the moment it's written.
A 2.7-point error in a sport where most guarantee-game spreads land between 20 and 45 points isn't a rounding error. It's a material gap sitting in a part of the schedule almost nobody monitors.
Re-measure it every season
The fix is to measure the lower-division scoring level from actual crossover results each season, weight the most recent years more heavily, and update it every time the model projects a game.
Three things make this harder than it sounds.
First, timing. Because guarantee games are front-loaded — roughly a third in week 1, nearly three-quarters by week 3 — the model is working from a partial sample early in the year. When there haven't been enough recent crossovers to give a stable answer, it falls back to the prior estimate and flags the game accordingly. We set that minimum at 40 crossovers.
Second, the difference between how we measure teams and what ends up on the scoreboard. We rate teams on how efficiently they move the ball relative to their opponents — not directly on points. Converting that efficiency to expected points isn't a simple multiplication. A team rated 20% better doesn't score 20% more points; the relationship curves. So rather than trying to adjust the team ratings to fix the problem, we work backward from actual scored points in real guarantee games to find the right answer, then let the team ratings follow from that.
Third, how much history to use. We deliberately look at a shorter window for the tier level than we do for team strength. Team quality carries over season to season in predictable ways; this particular trend has moved fast and in one direction for three straight years. A longer memory would lag both ends while the gap kept widening. The goal is a window short enough to track the trend while stable enough to give a useful answer.
What this doesn't say
It doesn't say the model is fully calibrated on guarantee games. Even with the tier level solved correctly, we still tend to project the final margin about three points tighter than games actually play out — we consistently underestimate how lopsided these matchups get. That's a larger problem than the tier fix addresses, and it's separate. The tier solve narrows roughly a tenth of that remaining gap; the rest is still there.
It doesn't say the lower division is the only moving target. The top-division side of guarantee games has been rising too — 39.8 to 41.1 to 43.9 — which means total scoring levels are shifting from both ends at once. The gap between the tiers is widening faster than either side is moving on its own.
And it doesn't say the portal effect will continue on this slope. The rate of talent concentration at the top of the sport is unusually high right now, and there are structural reasons it might slow: conference consolidation, changing rules around player compensation, the NIL landscape still taking shape. The tier levels we measure now might look as dated in 2028 as the 2023 levels look today.
The only honest response to that is to keep measuring. The projection re-checks the lower-division scoring level against whatever the most recent crossovers actually show. It doesn't store an answer it found once and hope the sport doesn't change.
Sources
Every number here comes from the graded ledger the boards publish. Research from a statistical model, not betting advice; no outcome is guaranteed. Method: how MatchWiz works.