What the words mean

What a player was worth, in goals and then in wins.

WARWins above replacement

How many wins a player added compared with a replacement — someone freely available to any team, a minor-league call-up or a waiver claim. A WAR of 2.0 means the team won about two more games than it would have with that player replaced by the next man up. Contributions are measured in goals first, then converted at the rate the standings say a goal is worth that season.

Not a measure of talent in the abstract. A good player who misses half a season has a small WAR, because WAR counts what happened rather than what was possible.

GARGoals above replacement

The same idea one step earlier, before goals are converted into wins. Everything the model computes is built in goals; GAR is that total, and WAR is GAR divided by the goals-per-win rate for that season.

SPARStandings points above replacement

WAR expressed in standings points rather than wins, which is the currency a playoff race is actually run in. Roughly double WAR, because a win is worth two points.

Replacement levelThe baseline everything is measured against

Not an average player — a freely available one. The level is solved so that the value of every player in the league, added up, matches the wins the league actually produced above that baseline. An average NHL regular is comfortably above replacement; a thirteenth forward is close to it.

Goals per winThe conversion rate, refitted every season

Found by regressing team wins against team goal differential over a five-year window centered on each season. It runs from about 5.2 to 6.2 across the sixteen seasons here — a goal was worth less in the high-scoring years than in the dead-puck ones, and using one number for all of them would misprice an entire era.

The numbers on the player pages and the game pages.

RatingA percentile out of 100

Where a player sits among NHL players at his position. A rating of 82 means better than 82% of them. Defensemen are ranked against defensemen and forwards against forwards, so the two are never compared directly.

Not a score out of 100 in the school sense. A 50 is a genuinely average NHL player, which is a long way from a failing grade.

Game gradeA letter for one night

Every skater who dressed, graded against the other skaters on the ice that night. A C is an ordinary NHL performance. The grade weighs production, even-strength play, defense and special teams, scaled by how much a player was asked to do.

Not a measure of whether the team won. A player can grade an A in a loss, and often does.

Component ratingsEven strength, power play, shorthanded

Each rating compares a player only with others who take that role, so the four are not measured against the same field. A fourth-liner who never kills penalties has no shorthanded rating rather than a bad one.

The measures the model is built from.

Expected goalsWhat a shot was worth

Every unblocked shot is assigned a probability of becoming a goal, from where it was taken, what kind of shot it was, and what happened just before it. Add them up and you get what a player or a team should have scored given the chances they generated.

Not a prediction of the future, and not a claim that the actual goals were wrong. It is a description of chance quality.

RAPMRegularized adjusted plus-minus

A regression that separates a player from the company he keeps. It looks at every shift in the league and asks what happens when this player is on the ice, holding constant who he played with and against. This is what stops a depth forward inheriting a credit for a star linemate's season.

Not plus-minus. Ordinary plus-minus counts goals while someone was on the ice and attributes all of them to him.

FinishingGoals against chances

The gap between the goals a player actually scored and what his shots were worth. A shooter who turns 20 expected goals into 28 did something the chance model did not see. Scaled so the league nets to zero each season, so a player is compared with his own contemporaries.

GSAxGoals saved above expected

For goaltenders: the shots that went in, against what those shots were worth. The goalie equivalent of finishing, and the main thing separating a goalie who faced hard shots from one who faced many.

CF% and FF%Corsi and Fenwick share

The share of shot attempts a team took while a player was on the ice — Corsi counts blocked shots, Fenwick does not. Crude next to expected goals, but they accumulate faster and settle sooner in a season.

Zone startsWhere a player's shifts began

The share of faceoffs a player was on the ice for that happened in the offensive, defensive or neutral zone. A player started in his own end repeatedly is being asked to do something harder than the raw numbers suggest.

Not as influential as it is often assumed. Zone starts explain about 3% of the variation in how a line performs.

Quality of teammates and competitionWho else was out there

The strength of the players alongside and against someone, averaged over his ice time. The model adjusts for both rather than reporting them as excuses.

What kind of player someone is, and how combinations work.

ArchetypeWhat kind of player he is

Value says how much a player is worth; an archetype says what he does. Each player is scored on twelve attributes — shooting volume, finishing, playmaking, play-driving, defensive suppression, blocks, hits, fighting, penalties drawn, and special-teams and faceoff usage — and the label is whichever stands furthest out. Sniper, Playmaker, Power Forward, Two-Way, Volume Shooter, Energy Grinder, Enforcer for forwards; Two-Way, Offensive Creator, Shutdown, Physical, Stay-At-Home, Shooter, Enforcer for defensemen.

Not a judgement of quality. A Sniper can be elite or replaceable; the label says what he does, not how well.

Complete and BalancedWhen nothing stands out

A player whose attributes are level gets one of these rather than a specific archetype. Complete means strong at everything; Balanced means notable at nothing. The difference is level, not shape.

Line chemistryHow a combination is built

Every forward trio and defense pair since 2010-11 — about 178,000 of them — grouped by what they are made of. A combination is graded on what structurally similar units have done, not on what this particular one did.

Not a claim about whether three players like each other. It is a statement about the shape of a combination.

Structure percentileThe headline chemistry number

Where a combination ranks among all possible combinations. A structure in the 90th percentile is built the way successful lines tend to be built.

ConfidenceHow much history is behind a grade

A combination seen four times reports something close to league average, because four lines cannot say more than that. Rare structures are quieted in proportion to how rare they are rather than discarded.