CFB Predictor — A college football model that shows its work.

What this is for

Most public predictions are a number with no history attached. This site is the opposite: every pick is posted before kickoff and graded after it, and the running record sits at the top of the picks page whether it flatters the model or not.

The goal isn't to beat the bookmakers on every game — nothing does. It's to find the handful of games each week where the model and the market genuinely disagree, and to be honest about the rest.

Anatomy of a pick

Every game on the picks page is one card. This is a real published pick — Army at Navy — with each part labelled.

Live example

Final Off by 3

Army 16

@Navy Win 17

Vegas Navy −6.5 ✓ ATS

Line moved ▼ 1.5 toward Navy

Total 50.5 ✕ O/U

{# Row-level, not page-level: with a free tier the same page carries both full and withheld cards, and the data layer has already stripped the withheld ones. Presence of the number is the honest test. #}

Model Navy −4.5 60% win prob

Enhanced Check Details →

1

Kickoff & status

Before kickoff: the local start time, or TBD when the TV window isn't set, plus an Upcoming chip. Afterwards: Final and how many points the model missed the margin by. A ◆ Toss-up chip appears when Vegas has the game inside 3.5 points — historically the model's best spot against the spread.

2

The matchup

Away team on top, home team below with an @. Before the game the side the model picks carries a gold bar and a Pick tag; once it's played, the bar and a Win tag move to whoever actually won, beside the final score.

3

Vegas — the market line

The sportsbook spread the model saw when it made the pick, favourite first: −6.5 means that team is favoured to win by 6.5. Before kickoff this row also carries Edge, the gap between the model's number and the market's. Afterwards it becomes ✓ ATS, ✕ ATS or = Push.

4

Line — where the market moved

The line when we first recorded it, then the line now, sampled every six hours. Green means the market is drifting toward the side the model backs; red means it's moving against it. The row is hidden when the line hasn't moved.

5

Total — combined points

A separate model from the one that predicts the margin. It projects the two teams' combined score, and whether that lands Over or Under the posted total. It doesn't guess at either team's individual score — nothing here predicts that.

6

Model — the pick itself

Written in the same notation as the Vegas line above, so the two compare directly: Navy −4.5 means the model has Navy winning by 4.5. Beneath it, the calibrated win probability — when the model says 70%, that side has historically won about 70% of the time.

7

Going deeper

Enhanced Check sends an AI agent to verify the pick against ESPN, DraftKings and FanDuel. Details opens the full matchup page with the team stats behind the number. Both are live on the picks page; here they're only for show.

How the model works

What it learns from

Team-strength features built from play-by-play going back a decade: opponent-adjusted EPA for passing and rushing on both sides of the ball, success rate, points per scoring opportunity, havoc rate, and an ELO rating that carries form across seasons.

How it picks

A ridge regression predicts the final margin from the two teams' adjusted profiles — a deliberately simple model, because with this much noise per game a heavier one mostly learns the noise. A second model projects the combined total, and a calibrated win probability turns the margin into odds: when it says 70%, that side has historically won about 70% of the time.

When it's at its best

Early-season picks lean on last season's numbers, regressed toward the league average, so treat September with some skepticism — by around week 5 the features are mostly this season's play, and the model is at its best from there on.

It reads lopsided games conservatively, and understates them. The efficiency features measure raw per-play performance, so a team that feasted on a soft schedule can rate ahead of a stronger one — only the ELO rating carries strength of schedule. Ridge regression then pulls extremes further toward the middle. On games the market lines at 21 points or more it has been calling about 17 where the market says 30, so read a big gap on a heavy favorite as a known blind spot rather than a lean.

The sweet spot is the close game. Across two seasons of backtesting, its picks in games the market lines inside 3.5 points covered 53% against the spread — the one spot it has beaten the books' breakeven in both seasons.

Measured against the market

Each pick is published beside the sportsbook spread and total, along with how far the line has moved since we first recorded it. Games the market calls close are where the model earns its keep — and the against-the-spread record is tracked separately, so its edge is a number rather than a claim.

Checked by an AI agent

Enhanced Check sends a single game to an AI agent that pulls current injuries, weather, and live odds from ESPN, DraftKings, and FanDuel, then says whether it agrees with the model and why. It catches the things a season-long statistical profile can't see — a quarterback ruled out on Friday night, 30 mph wind.

See this week's picks