How the college football model works

The foundation is a per-play efficiency rating for every FBS team, opponent-adjusted so beating a top-10 defense counts for more than beating a bottom-10 one. College football's problem is thin data — twelve games, 136 teams, massive roster churn — so early-season ratings blend in a preseason prior built from returning production, recruiting talent, transfer portal movement, and coaching changes. The prior's weight fades as real games arrive, and the Predictium 136 board shows that weight for every team.

Game margins and totals are predicted by gradient-boosted models (GBM) over the rating features. Predictions feed a key-number-aware empirical score distribution: college football margins cluster on key numbers (3, 7, 10, 14…), and a distribution fitted to real historical scores captures that better than any smooth curve. The p10–p90 range on every game card comes straight from this distribution — college football is high-variance and we show it rather than hide it.

Once books post a line, our win probabilities are anchored to the market line rather than derived purely from our own margin estimate. This is an honesty decision: across nine walk-forward seasons the consensus closing spread predicts margins better than our model (about 0.9 points of MAE), so a probability that ignored the market would simply be worse. Cards are badged Market-anchored when anchoring is active and Model-only before a line exists.

Where our number disagrees with the market, we show both — and we track publicly whether the market ends up moving toward us (closing line value) rather than claiming edges we can't prove.

Every published claim is walk-forward: models are trained on seasons strictly before the one they predict, then scored against consensus closing lines — 7,600+ games since 2017. The backtest page shows model vs market MAE season by season, without spin: the market is ahead, which is the normal state of the world for public models. The number that matters long-term is CLV — whether the market moves toward our first read of a game — and that tracking is public too.

The games board carries a picks column split by how a pick qualifies, because the two kinds mean different things. Model leans are games where our number disagrees with the market by 3+ points — the backtest shows leans are close to coin-flips against the closing line, so they carry no EV claim; they're graded on closing-line value instead. Price-value picks are the rarer case where a book's own price beats the no-vig consensus, and those do carry real EV math, shown with the price it was found at. Every pick, either kind, settles publicly on the backtest page.

Per-game detail pages (line movement, key factors, head-to-head) are the one artifact still ahead — the games board, ratings, picks, and backtest are all live now.