How the rankings are made
A model sets the order. Analysts can move a player, in writing. A human signs it. The results are graded whether they flatter us or not. This page is the whole of the model; model version ur-proj-1.
1. Opportunity
Targets, carries and pass attempts per game, as a weighted average: each game this season counts 1, each game last season counts 0.25 (half that if he has changed teams). A player with less than three games' worth of history is blended toward a baseline for his depth-chart role. A team's backup quarterback is projected at almost nothing.
2. Efficiency
Catch rate, yards per target and per carry, touchdown rates, completion rate, yards per attempt and interception rate. Each is the player's own rate pulled toward the league rate at his position — harder for noisy stats. A touchdown rate needs hundreds of targets before it is mostly his own; a catch rate needs far fewer.
3. Matchup
Fantasy points the opponent allows to the position per game, relative to the league, pulled toward average and capped at ±8%. It is capped that tightly because, tested honestly, matchups predict less than most rankings pretend.
4. Game environment
How many points the team is expected to score, taken from the market's closing numbers where available, relative to 22.5. Volume moves with its square root; touchdowns move with it directly. We use it because it predicts well. We never publish odds, and nothing on this site is betting advice.
5. Availability
From the official injury report: Out 0%, Doubtful 25%, Questionable 85%, and 90% for a player who did not practise and has no game status yet. Bye weeks are not projected.
Scoring
Passing yard 0.04, passing TD 4, interception −2, rushing and receiving yard 0.1, TD 6, fumble lost −2, two-point conversion 2. A reception is worth 1 in PPR, 0.5 in half-PPR and 0 in standard. These match nflverse's published fantasy points exactly.
The analysts
Each synthetic analyst may move up to six players up to eight places, with a reason. Every number in that reason must appear in the player's evidence — his stat lines, last season's per-game numbers and his projection — or the database refuses to sign the ranking. Moves are shown next to the player, with the model's original rank.
Does it work?
Before publishing anything, the model was run over the 2025 season, weeks 2–17, using only data available before each week. Among the players a manager is choosing between, its order agreed with actual finishes with a rank correlation of 0.394, against 0.292 for simply ranking by season average.
| Position | Rank correlation |
|---|---|
| QB | 0.332 |
| RB | 0.514 |
| WR | 0.440 |
| TE | 0.291 |
Weekly fantasy scoring is noisy, and no honest model gets close to 1. What matters is beating the obvious baseline, every week, in public. The live record is here.
Grading
After the games, every ranked player's points and finish are recorded from nflverse and can never be changed. Only players whose game had not started when the ranking was signed are graded. Two numbers are published per position: the hit rate — of the players we ranked inside the start line (QB12, RB24, WR36, TE12), the share who finished there — and the rank correlation between our order and the real one.
Data and licences
Everything the model reads is free, open and re-downloadable by anyone who wants to check our work.
- Weekly player statistics — nflverse, calculated with nflfastR, licensed CC BY 4.0. Our projections, rankings, trade values and player notes are calculated from these numbers; they are ours, not nflverse's.
- Weekly rosters and injury reports — the NFL's own records, published through nflverse.
- Schedules, results and closing game totals — nfldata by Lee Sharpe. The market numbers are used only as a model input and are never published.
- Depth charts — via nflverse. Used only as a model input for players with little history, and never published.
- Waiver trends — Sleeper's public API: how many Sleeper leagues added or dropped a player in the last 24 hours, rounded. Shown beside the waiver calls and on player pages, and given to the analysts as evidence. Not used by the projection model.