Player impact
Goals added over a replacement player, ranked per game so the schedule does not pick the winner.
How it works
Separate the player from the company
Five-on-five offense and defense come from a regression over every shift, measured in chances (expected goals) created and allowed with him on the ice. It adjusts for teammates, opponents, score, zone starts and home ice. Ridge regression keeps estimates from swinging too freely when the evidence is thin.
Count more than points
We add power play, penalty kill, finishing, penalties drawn minus taken, faceoffs and replacement level. Replacement level makes the comparison about value above a replacement player.
Finishing gets pulled back toward average because it barely repeats. Penalty rates also get pulled toward the league rate when there are fewer minutes to judge.
Put players on the same schedule
The board ranks goals added per game. The games-played minimum is 3 GP or 40% of the maximum. Playing a longer schedule should not be what puts a player on top.
The formula
The formula
Goals added = 5v5 offense + 5v5 defense + power play + penalty kill + finishing + penalties + faceoffs + replacement level. Divide by games played for the ranking.
Finishing kept = (goals − expected goals on his shots) × xG ÷ (xG + 110), where xG is the expected goals on his own shots.
Penalties drawn minus taken are worth 0.161 goals each. Rates shrink toward the league rate with K = 500 minutes for drawn penalties and 645 for taken penalties.
Faceoff value per offensive- or defensive-zone win = 0.007 goals; neutral-zone win = 0.0034 goals.
How we tested it
| Check | Finding |
|---|---|
| Full player ratings | r = 0.649 |
| Five-on-five part alone | r = 0.645 |
| Finishing, split-half | r = 0.06 |
The team-level test supports the combined ratings. It does not establish the same predictive accuracy for every individual player. The weak finishing result is why we shrink it.
What does correlation mean?
Correlation, written as r, describes how closely two sets of results move together. A stronger positive relationship means better results in one tend to go with better results in the other. A value near zero means little relationship. It is not a percentage of predictions that were correct.
What it can't tell you
An adjusted estimate is still an estimate. The team prediction test cannot tell you that every individual ranking is settled, and a short run of goals is weak evidence of lasting finishing skill.