THE METHOD
WHAT IT DOES
Every Thursday, the model ingests the week that just happened, retrains, and projects the coming slate. Two engines — a Bayesian model and a gradient-boosted quantile model — are blended by how well each has been calibrated recently. The output is a projection and, on three of the five props, an honest range.
WHAT DRIVES IT
Recent usage and opportunity. Snap share, target share, role, and the situation in front of the player.
WHY THE RANGES ARE HONEST
We measured the ceiling. A predictor that knows each player's true season average in advance still misses receiving yards by a wide margin every week. The gap between that predictor and ours is all the room any model has. The rest is football.
Average miss on receiving yards, in yards per game. When our 80% range on receiving yards spans sixty yards, that is not sloppiness — that is what a calibrated range looks like in a sport this volatile. Anyone showing you a tight yardage range isn't calibrated; they just haven't checked.
WHY YOU CAN TRUST THE SCORECARD
The 2025 season was held out of fitting and tuning. The model never saw it while it was being built, so when we walked it forward through the season one week at a time — projecting each week from what was knowable that Thursday — the results were a test, not a description.
And every bar was committed to git before the number that would be measured against it. You can check the order in the repo: the standard is timestamped earlier than the result, so it could not be moved to fit what came back.
The grading harness runs automatically and its writes are append-only. There is no manual override path. Bad weeks get published the same as good ones.