Methodology
How the rankings work
Enough to use them well and know when to ignore them.
Reading a ranking
- Set your scoring first. Full PPR, half, or standard — the buttons at the top. A ten-catch receiver is ten points different between the ends, and the order changes to match.
- Tiers matter more than ranks. Two players in the same tier are close enough to start whichever you prefer. The gap between tiers is the real information.
- Rank 4 and rank 7 are not meaningfully different. No weekly ranking is that precise, including this one. Use it to narrow a call, not to make it.
- Anyone ruled out is already gone. The injury report is checked before every list goes up.
What goes into a projection
Three things. What he's been doing — how much work he gets and what he turns it into, with recent games counting more. The game he's walking into — the betting line, the roof, the wind forecast, home or away, and which teammates are out. And Vegas: when the sportsbooks post a line on a player, it counts for most of his projection, because a number with money behind it knows about the trade and the comeback before a stat sheet does. The books fill in over the week — touchdown odds first, yardage lines by Thursday or Friday — and a player's line only counts once his yardage is posted, so Wednesday's list leans on the market less than Sunday's does.
Quarterbacks get a second opinion from a machine-learned model trained on three seasons, averaged with the first. It earned its place by beating the hand-built version on two seasons it had never seen. Nothing else has cleared that bar, so nothing else is in.
What it doesn't know
Anything from the offseason that Vegas hasn't priced. How good a defense is in week 1, because nobody has played. And why — it has arithmetic about a player, not an opinion. If you know something it can't, you're probably right. Override it.
How good is it
Tested the only way that counts: run it against seasons that already happened, one week at a time, letting it see only the games that came before. On that test it beats every simple baseline at every position, and the scorecard holds it against the two things you could start from instead every week: the betting market's lines and the plain box score. Week one it edged both by a point or three. That's an honest place to be, and the scorecard will show whether it stays there.
That's also less impressive than it sounds. In a typical week, about half the quarterbacks it ranks top-twelve actually finish there. That is true of every weekly list anyone publishes, and it is why the tiers matter more than the ranks. The scorecard shows the real number every week, including the weeks it loses.
Rest of season, and trades
The weekly ranking answers who scores on Sunday. A trade asks who scores from now to the playoffs, so the rest-of-season projection starts from the same model with the week switched off: no opponent, no spread, no weather, no home edge. That rate is then blended with what the betting market has said about the player over his lined weeks this season, because the model prices usage and has no idea who a player is, and the market does. Each remaining week is scaled by his chance of playing it, measured from every absence by a quarterback, back, receiver or tight end from 2021 to 2025: Out and Doubtful men almost never play that week, a Questionable starter plays 70% of the time, injured reserve waits the league's four games and then returns at the rate players with that injury did, and a torn ACL is the season.
The trade analyzer turns that into a value: points above what the last starter at his position in your league would score in the same games, for your league size, one or two quarterbacks, and your scoring. Each player carries a "give or take", the model's measured miss at his position times his games, and a deal whose gap is inside that reads as even. The model alone was tested week by week over 2024 and 2025 against what players actually averaged the rest of the way, and beat a plain season average at back, receiver and tight end in both seasons. The market blend and the injury curves cannot be tested that way, because no line archive exists for past seasons, so they are graded forward: every week's values are saved and scored against the second half of this season. The numbers are in the model log.
The other tools
Usage and stats are the model's inputs and the box score, for everyone who took a snap, with expected points beside actual so luck is visible. Trends compares each player's last two games with his earlier ones, and only calls it a trend when the role grew and the points came with it; a role that grew without the points gets its own list. Compare puts two players' usage side by side, week by week. The games put every source on one card per game, with the weather shown rather than modeled: only wind moves a projection, because rain was tested over two seasons and did not reorder anyone once the betting lines were in.
Survivor
A separate model. Each team can be used once, so the question is not who wins this week but which team to spend in which week. Win chances come from the sportsbooks — averaged, with their cut taken out — for this week and next; weeks beyond that use our own team ratings until the odds appear. We then solve the whole remaining season at once: every week matched to a team so the chance of surviving to the end is as high as it can be. This week's picks are the ones that spend the least of that season for the most safety today. Backtested on 2019–2025 against "biggest favorite you have not used": about one percent more survival a week, compounding. It is graded every Tuesday like everything else.
Sources
- Play-by-play, weekly stats, schedules, injuries — nflverse, used under CC BY 4.0 and adapted: every number here is derived from it, none is reproduced as published.
- Weather — the National Weather Service, public domain.
- Rosters, depth charts and injury reports — nflverse, same licence; the depth charts are ESPN's as nflverse publishes them.
- Betting lines — via The Odds API. An input to the rankings; the survivor page shows win chances derived from them. Nothing on this site is betting advice.
Corrections
A wrong number gets fixed and the page says it was fixed. Quietly editing a bad take is the same kind of dishonesty as inventing a stat.