Why there's no ranking for QB, RB, WR and TE
Because we tested one and it lost. We built a projection model for returning players β 20 inputs: usage, expected points, touchdown luck, age, seasons played, draft capital, contracts and cap share, team changes β and graded it against the draft market with both lists on the same players. The market won at every position, .598 to .544. Finding 25 β
Then we settled it in 9,000 simulated leagues with full drafts and seasons. Drafting by the model instead of ADP roughly halved title odds (10.8% against 20.7%), and every blend of the two finished behind the plain board. Finding 32 β
We built a rookie model too, on draft capital, combine numbers and landing spot, deliberately without fantasy ADP. It lands on top of NFL draft order almost exactly. Its projections are honestly calibrated β rookies it put at 8+ points a game averaged 10.0 and were startable 86% of the time, against 1% for its bottom tier β but that tier information is just what round a player was drafted in, which is public and already priced into every draft board. Finding 31 β
So the honest advice for skill players is the boring one:
take your platform's ADP in order and don't reach. What did
survive from that work is descriptive rather than predictive β
expected points, touchdown-luck regression, and the
position-room readout in the draft tool β and it's below or
still shipping. The code and the dead ends are kept in
archive/season-projection-model/.
Expected points β the "2025 expected" column
Not a prediction β an audit of last season. For every 2025 play it asks what a typical player would have scored from the same chance β who got the ball, where on the field, down and distance β and adds it up. A player who scored far above his chances usually got there on touchdown luck, which doesn't repeat; the chances themselves do. That gap is the strongest buy/fade signal in our research at WR and TE, and the player model above already weighs it. Finding 16 β Finding 17 β
Kicker model
Predicts: kicker points, weekly and for the season. Reads: the points Vegas expects his own offense to score β that one number is nearly the whole model β plus the venue: a dome is worth about +0.7 points a game, wind costs about β0.5. A kicker's own history predicts almost nothing. Finding 28 β
Defense (D/ST) model
Predicts: defense points, weekly and for the season. Reads: the points Vegas expects the opponent to score β that does about 97% of the work; each point off the opponent's total is worth about +0.38 D/ST points. The season list sums it over the whole 2026 schedule. Finding 27 β
Weekly player model β live in September
Predicts: this week's points for QB, RB, WR and TE. Reads: recent scoring, expected points from recent usage (its strongest signal), how many points the opponent's defense has been giving up to his position, and the Vegas line for the game. It powers the in-season weekly boards once real games start.
Injury model
Predicts: availability, from the 2018β25 injury reports: the chance a player listed on Friday actually plays Sunday, what playing hurt costs him, how long each injury type keeps a player out, and the rust in his first game back. Finding 21 β
Vegas offense table
Not our model β the betting market's. Sum the lines for all 272 games on the 2026 schedule and you get expected points per game for every offense. It feeds the kicker and defense models above. It is not a tiebreak between players: everyone can read the same lines, so it's already in each player's draft price. Finding 30 β
Every model's code and data are public: github.com/Zinkelburger/Fantasy-Football-Tool