How this works
Edge is our probability minus the book's probability after the vig is stripped out,
expressed in percentage points. A +5.0 means we think the bet wins five
points more often than the fair price implies.
For yardage props we project a number, then use the historical spread of our own errors to get a probability of going over the line. On the book side, a −110/−110 pair implies 105%, not 100%, so we normalise it back to a fair price. The difference is the edge.
Touchdowns are counts, not yardage, so they use a Poisson model with red-zone and goal-line usage as the main inputs — carries inside the five predict who scores far better than season yardage does. Results are recalibrated against realised scoring rates.
Anytime TD only has a "yes" side at most books, so there is nothing to strip the vig against. Those edges keep the book's margin in them, which makes them understated rather than inflated. They are marked with an asterisk.
Rolling averages of each player's recent usage, lagged so no row ever sees its own game: targets, target share, air-yards share, aDOT, snap share, carries, red-zone and goal-line opportunity, plus Next Gen Stats separation and cushion. The only forward-looking inputs are market-derived — the implied team total and the spread.
| Injuries and inactives | Not in the model. If a WR1 is ruled out, our WR2 projection does not move. |
| Weather | Not used. |
| Depth-chart and role changes | Only visible after games are played. |
| Offseason moves | Week 1 form is carried from last season, so a player who changed teams carries his old offence's usage. |
| Slot vs outside alignment | Not available in public data. |
Defensive scheme matchups — man/zone rates, middle-of-field open/closed, box counts — were built, pooled by player archetype, and tested across five prop markets. Every version made the model worse out of sample. The effects are real (running backs gain against man coverage; tight ends lose over a yard per target) but they are already absorbed by usage and game-script features.
On the game side, the closing spread beat our margin model in every season tested, and against the spread the model hit 49.0% versus the 52.4% needed to break even.
Nothing here is validated against results. No prop line history existed before August 2026, so we cannot yet say whether these projections beat closing numbers. An edge means the model disagrees with the book — not that the model is right.
Research artifact. Not betting advice.