How the rankings are built

Rex's Rankings are built from three inputs: observed player usage weighted toward recent games, the betting market's implied point total for each offense, and game script from the spread. Snap share, target share, air yards share and backfield share drive the usage component. Players who missed games are scored on the work they actually did rather than penalised with zeroes; thin samples surface as low confidence instead.

The three scoring formats are genuinely different boards. In PPR, receiving work carries far more weight; in standard, carries and downfield yardage matter more. Quarterback rankings are identical across formats, because reception scoring does not affect them.

Tiers are detected from gaps in the underlying scores rather than assigned in fixed blocks, so a tier break represents real separation.

Confidence reflects how stable a player's role has been. Low confidence means the model has an opinion but the usage underneath has moved, or the player is in a new situation. Those are the players worth a second look before you set your lineup.

FLEX rankings cannot be produced by interleaving the position boards. Each board is scored relative to its own position, so a receiver a standard deviation above other receivers is not worth the same as a running back a standard deviation above other backs. Instead, the model converts opportunity into projected fantasy points using coefficients fitted on three seasons of results, and ranks the FLEX board on that shared scale. The projection shown on each row is that number.

The conversion differs by scoring format. In standard scoring the same opportunity is worth noticeably more at running back than at receiver, which is why the flex board tilts toward backs in standard and evens out in PPR.

Defense rankings work differently from the other positions. Defensive scoring is simple enough to project directly rather than through a fitted conversion: one point per sack, two per turnover, six per defensive touchdown, plus the points-allowed tier. That last piece is projected from the opponent's implied team total, which is why the matchup dominates. The difference between facing an offense implied for 17 points and one implied for 26 is worth more than most of the gap in defensive quality.

The three scoring formats produce identical defense rankings. Reception scoring does not affect defenses, so this is intentional rather than an oversight.

Confidence reflects how steady a defense's sack and turnover production has been. Low confidence means the projection is leaning on a recent hot streak rather than a repeatable rate.

Kicker rankings combine two things: how many field goal attempts a kicker has been getting, and how much his offense is expected to score this week. Attempts and accuracy come from the previous season; the scoring environment comes from the team's implied total in the betting line.

The relationship between offensive quality and kicker value is not straightforward. More scoring means more extra points, but the very best offenses trade field goals for touchdowns, which swaps three points for one. The model weights extra points more heavily to the implied total than it does field goal attempts, for that reason.

Like defense rankings, kicker projections are computed directly from the scoring rules rather than estimated from a fitted model, and they are identical across all three scoring formats.

Where a kicker has little NFL history — a rookie or a new signing — he is ranked on league-average volume and flagged low confidence, with the reason saying so. Leaving a team off the board entirely would be less useful than an honest estimate, since every team kicks every week.

FFPC is a genuinely different format rather than a variation on PPR, and it changes the rankings in four places.

Tight ends receive 1.5 points per reception instead of 1.0. That single rule does more than move tight ends up a few slots: when the model is refitted against FFPC results, the tight end coefficient exceeds the wide receiver coefficient, inverting the relationship that holds in every other format. Tight end is also the most predictable position in FFPC, because premium scoring makes target volume even more decisive than it already is.

The flex board reflects this. In PPR, one of the top thirty flex options is a tight end. In FFPC, five are, and the first appears around flex 16 rather than flex 28. With two flex spots in an FFPC lineup, that is a meaningful strategic difference.

Kickers score by distance — three points for a field goal inside thirty yards plus a tenth of a point for each yard beyond. Across the last three seasons that averages 4.01 points per made field goal, against 3.64 under standard distance-bonus scoring, so kickers project higher throughout.

Defenses are scored more generously. Safeties pay five points rather than two, and the points-allowed tiers are higher at every level with a shallower floor, so a defense in a difficult matchup is punished less than it would be elsewhere.

Quarterbacks earn 0.05 points per passing yard, which is slightly above the common standard, and the model's quarterback coefficient is fitted accordingly.

On quarterbacks, running backs, receivers and tight ends, the projection is a modelled estimate rather than a direct calculation. Opportunity is converted to expected points using coefficients fitted against three seasons of actual results. How well those projections hold up on seasons the model was not built on is measured separately — see the accuracy page. Defense projections work differently: defensive scoring is simple enough to compute outright, so that number is a direct calculation.

Projections change between scoring formats at every position except defense, where reception rules do not apply.

Every board can be downloaded as CSV, TSV or JSON, either one position at a time or all positions for a week at once. The files carry the model version and the timestamp they were generated, so a saved export can always be traced back to the board it came from.

The exports include the nflverse attribution required by the licence on the underlying data. If you republish or build on these files, that attribution needs to travel with them.