Impact rating

How many points per 100 possessions a player adds to his team's margin, after accounting for the nine other players on the floor. Split into offense and defense.

Impact ratings could not load right now. Try again in a minute.

How this works

The method

Every game is cut into stints: stretches where the same ten players are on the floor. For each stint we count the points each side scored per 100 possessions. One regression then asks: which players, when they are on the floor, go with more points for their team and fewer for the other team, after accounting for everyone else out there? That is the offense (O) and defense (D) number; Impact is the sum.

With about 500 players and a lot of noise, the plain answer overreacts, so every player is pulled toward average (a ridge regression). How hard to pull is picked by testing on games the model did not see: we split games into 5 groups, fit on 4 and score the fifth. Longer stints count more because they have more possessions.

The range comes from refitting the model 50 times on reshuffled sets of games. 8 in 10 of those fits land inside the bar.

What we measured

Model numbers appear after the first fit.

What it cannot see

One season is noisy. Two players a point apart are usually not different; read the ranges, not the ranks.
It uses no box score at all: no points, rebounds or steals. A player gets credit only for what the scoreboard did while he played, so bench players who share the floor with the same teammates are hard to separate.
It does not know about role, matchups, garbage time or injuries, and players with few minutes stay close to 0 by design.
Players under 250 minutes are stored but shown as "not enough minutes".