The preseason prior is worth eight games

Texas Tech climbed 18 rating points across the 2025 season in our replay, which starts from 2024’s final SP plus rather than our own composite. Vanderbilt climbed 14. Neither move came from a preseason opinion. Both came from the layer that rewrites a team’s number after every Saturday. Ohio State finished that walk at the top of the board, with Notre Dame, Texas Tech, Indiana, and Georgia rounding out the top five. A preseason rating is an opinion formed in August. The question is how fast that opinion should give way to what a team actually does on the field, and by how much after each result. We built a number for that instead of guessing.

What the update actually does

Every team enters the season at its EL Preseason Rating composite, the seven-number formula we have written up separately. After that, each week’s finals move the number. The update is the standard running-mean shrink:

err      = capped(actual margin) - expected margin
expected = r_team - r_opp + 2.5 home field (0 at neutral sites)
r_new    = r_prior + K(g) * err        (the other side subtracts the same)
K(g)     = k_mult / (n0 + g + 1)       g = games the team has already played

A team’s expected margin comes from the gap between its rating and its opponent’s, plus 2.5 points for home field and nothing at a neutral site. The error is the gap between what actually happened and what the rating expected, capped at 28 points so a 70 point FCS blowout counts the same as a four touchdown win. The rating then moves by that error times a weight, K, that shrinks every week.

Non-FBS opponents are pinned at minus 22 and never get a rating row of their own. Only the FBS side of a mismatch updates.

How fast a preseason opinion should die

The whole design lives in K(g) and one number: n0, set to 8. With n0 at 8, the preseason prior is worth about eight games of evidence. Game one moves a team by 2/9 of its error against expectation. After twelve games the same size of surprise moves the rating by about 2/21, less than half as much. A blowout in September and an identical blowout in November are not treated the same, because by November the rating has already absorbed most of a season’s worth of information and one more data point matters less.

The other two numbers, the multiplier on K and the margin cap, came out of a grid search on 2023 alone: k_mult 2.0, n0 8, cap 28, chosen because that combination minimized margin error on 679 games from week 2 on (12.619 points of MAE, against 14.152 for a rating that never moved off its preseason number). 2024 and 2025 never touched that grid. The grid’s top five settings land within 0.03 points of each other, so the exact choice of k_mult or n0 is not doing much work. The shape of the decay is what matters, not the third decimal.

Did it beat the frozen prior

We replayed 2024 (starting from 2023’s final SP plus as the prior) and 2025 (starting from 2024’s) week by week, predicting every game blind to results that had not happened yet, and scored margin error against three things: the live updater, a frozen version of the same preseason rating that never moves, and that season’s own final SP plus applied retroactively, which no real-time system could have had in September.

Season	Games (weeks 2+)	Updater MAE	Frozen prior MAE	Retroactive ceiling
2024	703	12.960	15.253	10.854
2025	693	12.373	14.431	10.416

The updater beat the frozen prior by 2.29 points of margin error in 2024 and 2.06 in 2025, and it won or tied every week of both seasons except one, a 0.004 point edge for the frozen prior in week 2 of 2025. The gap between the frozen prior and the retroactive ceiling is about 4.4 points in 2024 and 4.0 in 2025. Our in-season updater recovers roughly half of that gap in each season, 52% in 2024 and 51% in 2025. The other half is information a results-only system cannot see from final scores alone: injuries, opponent-adjusted efficiency, the things play-level data like SP plus is built to catch. Results-only updating gets you halfway there. It does not replace a model built on the plays.

What it will not claim

A 12.4 point margin error describes a power rating, not a market edge. Closing spreads sit in that same neighborhood, and nothing here is a claim about beating a line, covering a number, or any closing-line value. This table validates that the rating tracks what actually happened on the field, and it stops exactly there.

There is also a seam in the backtest worth stating plainly. The 2024 and 2025 walks had to start from the prior season’s SP plus, because our own preseason composite’s component tables do not reach back to 2023. Production 2026 starts from the composite itself, which already beat that same SP plus baseline out of sample in its own preseason test. So if anything, this backtest is the conservative floor, not the ceiling, for how the live 2026 numbers should perform.

What it powers

This is the layer that moves team ratings on our power ratings board every week of the season, starting from the preseason composite and updating after Saturday’s finals land. The full board, updated live, is at edgelabs.bet/cfb/power-ratings, the formula it starts from is broken down at edgelabs.bet/research/the-cfb-model-in-seven-numbers, and the win total board built on top of both is at edgelabs.bet/cfb/win-total-predictions. Founding access is open at edgelabs.bet/join.