League
The Model
How the matchup predictions work, and why the numbers are not pulled out of thin air
Forecast every starter
Every starter on both rosters gets a forecast built from three ingredients. In week one it leans almost entirely on draft pedigree. By midseason, the man's actual play has taken over.
Draft expectations
what players taken at his draft slot have scored, going back nearly a decade
His games this season
his typical week, not his best one. A single monster game buys no extra credit
His workload
targets and carries. Volume is loyal. Touchdowns are not.
Build both teams' totals
The forecasts are summed into each side's best lineup, minus the injured and the bye-week absentees, adjusted for the defenses standing in the way and for each manager's lineup habits.
Replay the matchup 10,000 times
Real football is random, so the model refuses to pick a winner just once. It plays the week ten thousand times, dealing every player his share of good days and disasters. Win 6,800 of the 10,000 replays and your win probability is 68%.
The ingredients
What it weighs
Every prediction is assembled from the same list of inputs, every week, for every team. No hunches, no favorites.
The receipts
Tested, not guessed
Every accuracy number comes from replaying past seasons week by week using only information available before kickoff. The model never peeks at a result it is about to predict.
Flipping a coin
50%
The model, nine NFL seasons replayed (2016 to 2024)15,118 matchups
62%
The model, our own 2025 season50 of 77 games
65%
The dashed line at ~70% is the mathematical ceiling for any predictor. Beating it would require knowing which weeks the touchdowns land before they happen. Las Vegas, with billions on the line, picks real NFL winners about 66 to 67 percent of the time. If anyone claims 80, they are guessing or cheating.
And it grades itself every week
Each week's predictions are locked in before kickoff. When the week ends, the model compares them to what actually happened and records the verdict. The accuracy above is a running, audited track record, not a one-time boast.