The bar on each card is the projected
score split, not a win probability. Fathom publishes no win probability, and
drawing one purely to fill a bar would put a number on the page that has never been
measured. The bar simply shows how the projected points divide between the two
teams.
No projected totals. The margin
between the two numbers is validated; their sum is not. Fathom's projected total
measures worse than simply assuming each team scores its season average, so it is
left off rather than published as though it carried the same weight.
Market is the median posted spread, shown as an implied margin so it reads in the
same direction as ours. 59 of 62 games are priced.
How accurate these forecasts have been
Walk-forward on 1,659 held-out games, 2023–2025, weeks 5+,
both teams FBS. Every game is predicted using only what was known before it.
Margin error — mean absolute error in points, lower is better
Market spread12.02
Fathom projection12.54
Season avg point differential13.88
Home team by home-field edge15.27
Every game a tie15.52
Better than the best rule available for free by
1.34 points, and within 0.52 of the closing line — a number set
by a market with real money behind it.
Picking the winner — accuracy, higher is better
Fathom projection72.0%
Season avg point differential69.3%
Better win percentage67.6%
Home team always57.2%
1,194–465. Ahead of season point differential by 2.7 points,
the better record by 4.3, and always taking the home team by 14.8 — all
statistically significant, none of them close calls.
Why the comparison is to those rules and not to a rival model
A forecast is only worth something if it beats what you could
do without it. Backing the home team costs nothing. Ranking teams by their average
point differential costs nothing. A model that cannot clear those is decoration.
Every number above is measured against them on the same held-out games, and the
comparison is reported whether it flatters the model or not — which is why
projected totals are missing from the other tab.