Board
Every pick with a real line against it, grouped by game. A game holding both a quarterback and one of his own receivers is marked stackable: those two move together, which is the one pairing that helps a multi-pick entry.
Legs
What every leg has to clear
An entry pays only if all of it lands, so the bar is per leg. A single bet at a sportsbook needs 52.4%.
Results
Two different things get called results, and mixing them up is how a model gets oversold. Accuracy is how close the predictions land. The record is whether the side the model actually took came in. Only the second could show an edge.
The record
Every prop on the board with a real line is recorded when it is put up, and graded once the game is played. Nothing is staked and no odds are shopped, so there is no profit here to report — only whether the call landed.
Accuracy on unseen seasons
Average miss per game, on seasons the model was never fitted to. Lower is better. The bar shows how much of a flat guess's error each model removes.
| Market | Games | Model | His own average | Flat guess | Better by |
|---|
Are the stated chances honest?
Landing close is one thing; quoting an honest price is another. A model that says 70% should be right about seven times in ten. Every game in the test seasons was priced against a spread of made-up lines above and below its own prediction, which sweeps the whole range from near-certain to near-hopeless, and the claims were then sorted into bands and checked against what actually happened. On the diagonal is perfect; below it the model is overconfident.
Method
Train on the past, test on the future. Decisions are made by predicting each season from the seasons before it alone. 2024-2026 is read once, at the end, and never used to choose anything. No split is random.
Volume splits into role and pace. A receiver's share of his team's targets is stable; his team's passing volume is not. Share correlates 0.55 with his next game, against 0.25 for a quarterback's attempts and 0.72 for a back's share of carries. So share and volume are predicted apart and multiplied.
Probabilities come from real misses, not a bell curve. The model's errors lean right: a quarterback can throw twenty more than expected chasing a game, but not twenty fewer without getting hurt. A normal curve would price both tails wrong, so the actual distribution of past misses is used instead.
What it cannot see. Injuries, depth-chart changes and snap counts. In the week this page was first published it had a running back at 6.3 carries against a line of 13.5, because it had not noticed he was now the lead back. Where the model disagrees loudly with a line, the model is usually the one that is wrong.
A lean it still has. On pass attempts the model claims a shade too much on the over: through the middle of the chart above, a stated 45% happens about 43% of the time. That was measured over 26,200 priced cases and chased down. It is not the prediction, which runs only a third of an attempt high, and it cannot be fixed by nudging the prediction down — the pool of past misses re-centres by exactly as much, so the probability does not move. It comes from the middle of that pool sitting about a third of an attempt above the misses these seasons really produced. Using every season of misses in hand, rather than a pool frozen in 2023, cut the lean by more than half. The rest of it needs something no one knows before a season starts.
What it has not shown. That it beats a sportsbook. Against a bookmaker's own implied probabilities it still sits several points high on pass attempts, on two cards, which is far too few to separate the lean above from noise. On the other markets there are only a handful of priced cards so far either. Nothing here is a recommendation to bet. If you do, risk only what you can lose.