Friday Night Model
The Cooper Predictor is a statistical high school football model covering Georgia, Texas, Florida, Alabama, Tennessee, California, Louisiana and South Carolina. Each state model is built and tested separately.
In walk-forward validation the model is trained only on seasons that came before the season being tested. That keeps future results out of the inputs and gives an honest measure of how the model would have performed at the time.
Its central feature is the power rating — an estimate of how a team would perform against a reference opponent within its own state. That rating produces:
The goal is not to restate the rankings you see elsewhere. It is to produce an independent answer, publish it, and measure how well it performs.
Not a recruiting ranking. A team does not move up for having more highly rated prospects or a more recognizable name. Talent shows up in on-field performance, but recruiting rankings and roster reputation are not model inputs.
Not a media poll. There are no ballots, reputation votes or manual adjustments to force the rankings into a more familiar order. When the model disagrees with the consensus, the disagreement stands and gets tested by future results.
Not a guarantee. A team with a 70% win probability still has a meaningful chance to lose. Favorites lose and upsets happen — that's what makes the sport fun.
Not finished. New games create new evidence and ratings move as the season develops. The largest swings come in the opening weeks, when each result adds the most information.
The model evaluates what teams have done on the field, the quality of the opponents they faced, and how those performances compare across the wider state. It uses opponent-adjusted scoring, schedule strength, classification context and — once a season is underway — recent form.
That context matters because identical records tell very different stories. Going 9-1 against a demanding schedule is not equivalent to going 9-1 against a weak one. The model reads the performances behind the record rather than the win-loss column alone.
Scoring margin matters, but extreme margins are compressed rather than rewarded point for point. Beating a strong opponent says more than adding late points against an overmatched one. The difference between winning by 20 and winning by 50 is much smaller in the model's eyes than the scoreboard suggests.
Preseason ratings are built from on-field evidence available before the new season — no recruiting rankings, no subjective roster adjustments. Once games begin, current-season results blend in and gradually take on more weight.
The win probability is the most useful week-to-week output: it quantifies how confident the model is, and how big an upset would be. In a projection of Thompson 31, Central (Phenix City) 27 with a 62% win probability, the 62% is the important number, not the exact score. Thompson is the favorite; a Central win is still a plausible outcome.
Calibration is measured across groups of historical predictions, not by replaying one game a hundred times. In the latest validation, the average gap between stated probabilities and actual win rates ranged from 0.6 to 2.3 percentage points across the covered states. Teams given roughly a 70% chance have historically won at about that rate.
Where a state's postseason format can be modeled reliably, the Predictor simulates the remaining schedule and applies that state's qualification rules. Those percentages are conditional on the schedule, ratings and rules available at the time — they are not guarantees.
Formats differ significantly by state, so the projection is not identical everywhere. California, for example, currently carries league-title odds rather than statewide playoff odds.
An independent model is most valuable when it disagrees with the consensus. The Predictor uses proprietary statistics and independently trained coefficients, so no other system will reproduce the exact ratings. Where it differs sharply from other rating systems, those are the teams whose seasons will teach us the most — a gap does not prove either system right or wrong.
Some picks will be wrong. The difference offered here is transparency: every posted pick stays public, results get tracked, accuracy gets published, and the misses get discussed alongside the wins. Quietly deleting a bad prediction would make the whole exercise meaningless. The model should earn trust through transparent results, not claims or name recognition.
Great high school football is not limited to the outskirts of Dallas, Atlanta, Los Angeles and Miami. After covering the sport professionally for nearly a decade — from the foothills of the Appalachians to the Gulf Coast and the prairies of the West — we believe every game matters. It matters to the players, families, schools and communities who gather every Friday night, so it matters to us.
If you follow a smaller program outside the national conversation, this is for you too. 1A South Carolina football is as worth discussing as the nationally ranked teams. If there's a state you wish were covered, let us know — adding one and validating it properly is labor intensive, so we appreciate the patience.