1.5 to 1.9 points
Our college football projections ran pro-home by 1.5 to 1.9 points through the second half of the season. That's not noise. Six season-year combinations, three consecutive years, all running in the same direction.
The cause: we applied a single home-field advantage constant to every week of the season. Same value in September as in November. The home team is at home in week one, and they're still at home in week fourteen, so the assumption seemed reasonable.
It wasn't. Home-field advantage in college football fades as the schedule goes on. Teams adjust to road environments. The novelty of a hostile crowd wears off when you've played through several of them. By October, the edge a home team holds over a road opponent is real but noticeably smaller than it was in September — and our projections kept using September's number all the way through December.
That's why the home team was covering our projected spread only 46.8 percent of the time. A calibrated projection should sit close to 50/50. We were more than 3 points off that target, in one consistent direction.
How the decay shows up
Here's what actually showed up in the data.
We looked at 1,663 games between top-division programs over three seasons — measuring the gap between our projected margin and the actual final margin each time. Pool all of those games together across the whole season and the signal disappears. Break them into early, middle, and late weeks, and it becomes legible.
A positive gap means we credited the home team with more of a win than they got. A negative one means we undersold them.
Weeks one through four: noisy and inconsistent. Some years we ran high on home teams, some years low. No stable direction when grouped together.
Starting around week five, the gap locked in. Pro-home in every season-year combination from that point on. Around 1.5 points on average through midseason. Close to 1.9 points from week ten onward.
The home-field value we were building into projections was worth around 5 expected points across the full slate. The late-season games say the real late-season edge is closer to 3.2.
A season-long average of a quantity that changes through the season isn't right for the early weeks or the late ones.
The dependency
The fix seems obvious: use a smaller home-field value in October than in September. It is obvious, with one complication worth knowing.
The model converts team ratings into projected scoring differences through an efficiency scale — a multiplier that controls how hard the model reacts to any given difference between two teams. Home-field advantage goes through that same scale. The two are linked: change the efficiency scale for any reason, and you change how much home-field advantage ends up in the projected margin, even if you never touched the home-field input.
We've seen that play out directly. When a separate calibration change raised the efficiency scale, it cut the home-field value in expected points from 5.1 down to 2.6. The home-field input hadn't moved. The scale it feeds through had moved, and the effect in the projection was cut nearly in half.
Fixing home field means checking the efficiency scale at the same time. The two have to land in the right relationship to each other, not just be set independently. A correction to home field that ignores the scale can be half undone the next time anyone adjusts how team ratings translate into points.
Why the first four weeks stayed the same
The early-season data doesn't give a stable answer on home-field advantage. Trying to solve an adjustment from it would just chase noise.
When we split weeks one through four by which team we're picking — home or away — the gaps ran in opposite directions. When the home team was our call, we ran about 2.5 points high. When the away team was our call, we ran about 3.5 points low. Opposite signs, roughly similar size.
That pattern has nothing to do with home field. It's about how the model starts the season: working off last year's results, decayed across the offseason. The market prices current rosters — it knows which quarterbacks transferred, which coordinators changed, which recruiting classes overperformed. We catch up once we have this year's games. But early on, when two teams with a real talent gap meet, we're compressing it. The error runs toward whichever team we underrated, regardless of who's home.
A home-field error always runs in the same direction, no matter who the favorite is. The error we're describing flips with which team is the stronger one. The early-season data matches the second pattern, not the first.
We left the first four weeks on the existing constant. Applying a smaller home-field number there would make the aggregate look cleaner while leaving the actual problem untouched.
What the adjustment changed
We re-fit the home-field value against what actually happened on the field. The goal was straightforward: remove the late-season pro-home lean without creating a new one in the other direction. We ran a range of values and tracked where the bias crossed zero.
The result was a constant roughly 36 percent smaller than what had been in use season-wide. At that setting, the late-season bias went from 1.84 points on average to effectively zero.
The home team now covers our projected spread 50.1 percent of the time, up from 46.8. That's the most direct test of whether a projected spread is calibrated — it should sit close to 50/50 — and we moved it back into that range.
The moneyline improved too. The moderate-confidence calls — games where we projected one side winning with 70 to 80 percent probability — were coming in 3.3 percentage points short before the fix. That gap is now under half a point.
What this doesn't say
It doesn't say the NFL has been fixed the same way. The same structural question applies — does home-field advantage stay constant across the season? — and that measurement hasn't been run for the NFL. College football has more games and more variation across the schedule, which makes a pattern like this easier to find. The NFL version is worth measuring; it hasn't been yet.
It doesn't say the week-one spread gap is closed. The early-season bias is a real, separate problem. Our model starts the year working from decayed last-season results while the market prices current rosters. No adjustment to home-field advantage touches that.
And it doesn't say the team-total projections are fully clean. The home team's projected score still scatters a bit more widely than the away team's after the fix, and some of that is real — home teams do score with more variance. What's left is a measurement, not an assumption.
What found this was measuring by week rather than pooling across the whole season. When you pool, the early noise covers the late signal. Split it out, and the direction becomes clear. That's the approach now — not one answer for the full season, but one that accounts for where in the schedule each game falls.
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.