Methodology

How the numbers Keel Stats makes itself are computed, in plain words. Every one comes from game data (box scores and play by play) by documented methods; past views read only the games before their day, so nothing looks ahead.

Keel Impact: Our all-in-one NBA player rating: the points per 100 possessions a player adds over an average player, split into offense and defense. It is RAPM shrunk toward the box prior rather than toward zero, with the strength of the shrinking chosen by cross-validation on held-out games. A season's Impact is shown once a player has 250 minutes; percentiles are among players with 1,000 minutes or more. Each season's ratings are graded on the next season's game margins, where Keel Impact, PER and the box prior lead within noise. Player ratings
RAPM: Regularized adjusted plus-minus. Every stretch with the same ten players on the floor, from play by play, gives each side's points per 100 possessions; a ridge regression explains them by the five on offense, the five on defense and home court. Regular season only, garbage time left out. The pooled version adds the season before at half weight. Player ratings
Box prior: What a player's box-score rates per 36 minutes and his minutes per game predict his RAPM to be, fitted on every season we store, offense and defense apart. It is the starting point Keel Impact moves away from as the lineup data comes in. Player ratings
xFG%: Expected field-goal percentage: for every NBA shot, the chance an average shooter makes it, from what is known before it goes up: distance from the rim (threes apart, the corner three separate), the kind of shot from the play-by-play words (dunk, layup, floater, step-back…) and its context (a putback, transition, the shot clock, the end of a period). Never who shot it. Shot quality is the expected effective field-goal percentage of the shots a player takes; shot making is his points over expected per 100 shots. We have no tracking data, so open and contested shots look alike. Shot quality and shot making
EPA: Expected points added, in the NFL. Expected points before a snap are what the next score in the same half tends to be worth to the team with the ball (plus 7 for a touchdown with its try, plus 3 for a field goal, negative when the other side scores next), from down, distance, field position and the clock, by our own model fitted on our play by play. A play's EPA is the expected points after it minus those before it; a score counts as the points it changed. NFL efficiency
Success rate: The share of plays with an EPA above zero. Team offense and defense are shown for passes (dropbacks) and runs, regular season, plays from scrimmage only. NFL efficiency
Prop model: For a player and a line: his games, the newest weighted most, fitted with a negative binomial distribution, and the chance of going over, shrunk toward 50%. For the NBA it also uses his recent minutes and what the opponent allows to his position. It is shown beside the market's chance with the bookmaker's margin removed, and graded on past seasons on the accuracy page. The NFL model does not beat the market yet, so it is not shown. Trends from the data, never picks. Prop model accuracy
Game projections: Before each NBA game, the home side's expected margin and win probability, from the strength of the players available (last season's Keel Impact times each player's expected minutes), availability against the team's usual roster, form (net rating so far, shrunk early in the season), rest and home court. Projections freeze at tip-off. Graded out of sample on a whole season, they picked about 70% of winners. Projections
Matchup pages: One page for each pair of teams, such as Kings vs Lakers: every meeting we have, with the featured game's box score. NBA pages count regular-season, play-in, playoff and NBA Cup final games, never preseason or All-Star games.
Starts: Shown beside games played, which count any box-score line. NBA starts come from the box score's starter flag; seasons whose box scores never say who started (before 2025-26) show starts as unknown (–), never 0. NFL quarterback starts: the feed has no starter flag, so the quarterback who threw, or was sacked on, his team's first pass of a game started it.
Workload games: NFL games in which a player's line clears his position's bar, shown beside games played, never instead: a quarterback who started or had 10 or more dropbacks (or half his team's); a running back with 8 or more carries and targets (or 15% of his team's); a receiver or tight end with 3 or more targets (or 10% of his team's); a kicker with any field goal or extra point. There are no snap counts in our data, so there are no snap shares.