Fantasy Football Model

How it works

Six separate models, one per position, each predicting how many points a player scores in a game that has not been played, using only what was knowable before kickoff. Trained on 1999–2023, tuned on 2024, and scored once on 2025, which it never saw during fitting.

The scoring it projects

Standard 1-point PPR. These rules are implemented explicitly rather than taken from the source data, then checked back against nflverse’s own fantasy column on all 153,026 player-weeks, and they now agree on every one.

Passing yards

1 point per 25 yards

Passing TD

4 points

Interception

-2 points

Rushing and receiving yards

1 point per 10 yards

Rushing and receiving TD

6 points

Reception

1 point (full PPR)

Fumble lost

-2 points

Kicking

1 per extra point; 3 for a field goal inside 40, 4 from 40-49, 5 from 50+

Defense

1 per sack, 2 per takeaway, 2 per safety, 6 per touchdown, plus a points-allowed tier

What goes into a projection

Four families of input, all of them lagged by a game so that nothing a row predicts can leak into what predicts it.

• Recent formRolling three- and five-game averages of scoring, yardage and volume, plus the player's per-game average across the whole of last season.
• Usage and roleTarget share, carry share, air yards share, weighted opportunity rating, and snap share: how much of the offense actually runs through him, rather than how efficient he was with it.
• MatchupPoints that defense has allowed to this position, built from the weeks already played and falling back to last season's figure in September.
• Game environmentThe betting market's spread and total, converted into how many points each team is expected to score. This is the one input from outside the box score, and for kickers and defenses it is most of the signal.

A different model for each position

A kicker and a wide receiver share almost no useful inputs, so they do not share a model. Two families are fitted every time, a ridge regression and a gradient booster, and the one that does better on the tuning season is the one that ships. Which wins is a result, not a preference.