Skip to content

Explainers / Methods reportUpdated 2026-09-30

From every shot to WAR $: how our numbers are made

Our player measures start from 1.9 million recorded shots and every shift since 2007–08. Here is each step, what it assumes, how we test it and what it can't see.

The short version

Every measure on a player page comes from one chain of calculations. We rate the danger of every unblocked shot, work out how much each skater changed the danger of the chances at both ends, and add four more parts of his game. We convert all five from goals into wins and add one adjustment for the depth player who would otherwise take his minutes. The result is wins above replacement, or WAR. WAR $ puts that WAR into today's cap dollars at one fixed rate.

Here is the chain, one step at a time:

  • Expected goals (xG). Each unblocked attempt gets a probability of becoming a goal, from where and how it was taken and what happened just before. Adding those probabilities gives expected goals.
  • Five components, each against an average player. Process is a skater's effect on the balance of chances at five-on-five. Finishing is how much his own shooting added beyond the chances he had. Power play, penalty kill and penalties (drawn minus taken) complete the set. Each is measured in goals, then converted into wins, so the total is wins above average (WAA).
  • One replacement adjustment. Replacement is a lower bar than average: it's what teams get from their depth players. Adding one adjustment for the minutes a player played turns WAA into WAR.
  • Rates. Player listings show WAR per 82 covered games over the last three seasons. Goalies get WAR per 60 covered games.
  • WAR $. WAR times $8.3 million, the rate that prices a player's share of league WAR at the same share of teams' skater spending above minimum salaries.

Connor McDavid's 2025–26 season shows the whole chain. He added 1.22 wins through process, 0.07 through finishing, 0.62 on the power play, 0.16 on the penalty kill and 1.00 through penalties: 3.06 wins above average. His 1,309 five-on-five minutes earn a replacement adjustment of 0.25, for 3.31 WAR. At $8.3 million per WAR, that's $27.4 million of WAR $. Throughout this report, figures are rounded, so parts may not add exactly to their totals.

Connor McDavid's 2025–26 season, from five components to WAR and WAR $.
Connor McDavid's 2025–26 season, from five components to WAR and WAR $.

The rest of this report takes each step in turn. Every section starts with the short answer, then explains how it works, the numbers behind it and its limits. Technical notes with the exact settings sit at the end.

The data: public records of every shot and shift

Everything starts from the NHL's public game records: every shot, goal, penalty and faceoff, and a record of who was on the ice for each second of play. We rebuild those records into our own shot and shift tables, season by season, and measure players from them.

What the records contain. The play-by-play logs every event with its time, the players involved and, for shots, where on the ice it was taken and what kind of shot it was. Shift records place every skater and goalie on the ice, so we can split each game into stretches of play with the same players on the ice. We call those stretches stints. Schedules, results and player biographies supply game outcomes, positions, handedness and birth dates.

How far back it goes. Skater measures cover 19 regular seasons, 2007–08 through 2025–26. For 2007–08 to 2009–10 we use the SFU NHL database, an archive of the league's play-by-play from those years. It records who was on the ice but not where shots were taken, only how far out, so those three seasons use a simpler shot model (see expected goals). Goalie WAR starts in 2018–19 (see goalies).

What we countTotal
Regular seasons, skaters19 (2007–08 to 2025–26)
Games with shift records22,831
Skater-seasons measured17,125, covering 3,189 skaters
Unblocked attempts at a goalie1,919,364, including 124,450 goals
Five-on-five stints2.13 million, 17,668 hours of play
Five-on-four and four-on-five stints388,000, 4,150 hours of play
Goalie-seasons, 2018–19 to 2025–26803, covering 210 goalies and 586,993 shots on goal

What's left out. We measure regular-season play only. Blocked shots are recorded where they were blocked, not where they were taken, so expected goals uses goals, saved shots and misses. Shots at an empty net, penalty shots and shootouts sit outside every measure: an empty-net goal mostly follows the game situation, and penalty shots and shootouts are separate contests.

Filling gaps in the records. Our play-by-play source was missing who was on the ice for 174 games in 2019–20 and for a week of 57 games in April 2025. We rebuilt the 2019–20 games from the NHL's official shift charts and game reports: 51 are complete, 123 are partial, and together they recover 90% of those games' regulation time. The 2024–25 week came from SportsDataverse's full play-by-play file. Eleven older games, all between 2008–09 and 2013–14, still have no usable shift record.

Contracts are a separate record. Cap hits and terms come from CapWages, checked against PuckPedia, team announcements and the NHL's roster lists. A cap hit is a contract's yearly charge against the salary cap, usually its average annual value (AAV): the total value divided by the number of years. The current snapshot is September 29, 2026: 818 signed players across 32 teams. It is not a certified opening-night roster.

Three different dates. A player page combines three clocks. The model was last rebuilt and adopted on September 30, 2026. Performance runs through the 2025–26 regular season. Rosters and contracts are dated September 29. An article can report a newer trade or signing without changing any of the three.

Covered games: the games we can measure

A covered game is one in which our shift records place a player on the ice. Rates divide by covered games, and season rows count only those games.

Because nearly every game has a complete shift record, covered games usually match games played. They can differ where the record is missing or incomplete: the eleven older games above, and the partial 2019–20 reconstructions, where a player's time outside the recovered stretches isn't measured. An uncovered game isn't a game with zero contribution. We simply haven't measured it.

A goalie's covered game is one in which the shift record has him in net. His WAR per 60 divides by those games.

Expected goals: what each chance was worth

Expected goals gives every unblocked attempt a probability of becoming a goal, based only on what the record shows about the shot: where it came from, what kind of shot it was, the manpower and score, and what happened just before. It doesn't know who took the shot, who was in goal or whether it went in.

What the model sees. The shot's distance, angle and position; its type (wrist, snap, slap, backhand, tip-in and so on); the skaters on each side; the score before the shot; the period and game time; how long ago the previous event happened, how far away and what it was; whether the shot followed a quick rebound and how far the puck moved across the ice; time since the last faceoff and who won it; and the shooter's position and handedness. The score comes from our own running tally of goals, not from a field that may already include the shot being rated.

What it deliberately doesn't see. The shooter's and goalie's identities are left out, so finishing and goaltending can be measured separately against the same baseline. So is anything recorded after the shot, along with the penalty-clock and shift-length fields, which can quietly reveal that a goal was scored.

How it's built. A gradient-boosted tree model (a standard machine-learning method that combines many small decision trees) learns from 724,069 regular-season attempts at a goalie from 2018–19 to 2024–25, including 48,487 goals. Recent seasons count more, because the game and its record-keeping change: each season back counts 45% as much as the one after it.

How well it works. The 2025–26 season was left out of training, so we can compare the model's probabilities with what happened on its 111,068 attempts. (It isn't a perfectly untouched test: it was one of the seasons used to choose the recency weighting.)

  • Ranking chances. Given a random goal and a random non-goal, the model rated the goal as more dangerous 77% of the time (an AUC of 0.771).
  • Level. It expected 7,887 goals; 7,572 went in. That's 4% too many, spread across most of the range, and one reason we re-level each finished season (below).
  • Earlier seasons. Trained each time only on the seasons before, it posted AUCs of 0.750, 0.761 and 0.773 for 2022–23, 2023–24 and 2024–25, with expected goals between 0.4% and 5.3% above the actual total.
Held-out 2025–26 attempts grouped by predicted chance: in most groups the share that scored ran slightly below the prediction.
Held-out 2025–26 attempts grouped by predicted chance: in most groups the share that scored ran slightly below the prediction.

Re-levelling a finished season. When we describe a completed season, we adjust the probabilities so that, within each shot type and manpower situation, they add up to about the goals actually scored. This corrects season-to-season drift, including changes in how scorers label tip-ins, without changing much about which chances rank as dangerous. The chance values that feed process and finishing also carry a small adjustment for where on the ice the shot came from and whether the shooter was on his off wing. For 2010–11 through 2017–18 they include a small team-level adjustment for how often each team's shots went in over those seasons.

Older seasons. For 2010–11 through 2017–18 we use a version of the same model trained on all seasons from 2010–11 to 2025–26, then re-levelled season by season. For 2007–08 to 2009–10 the archive has no shot locations, so a separate model uses distance, shot type, zone, manpower, score and a rebound flag. Arenas recorded distance differently in those years, so we correct each home rink's distances against the same team's road games. On a reused test season, 2013–14, that simpler model scored an AUC of 0.750.

Limits. The record can't show screens, passes, goalie position or how clean a look the shooter had. Rebound chances are rated too low through 2022–23 and too high from 2023–24 on, even after re-levelling, which can shift finishing credit between players who get many rebounds and players who don't.

Process: who drives the chances at five-on-five

Process is our estimate of how much a skater changes the chances his team creates and allows at five-on-five, after accounting for who he plays with and against, the score, where his shifts start and home ice. It measures chance creation and chance prevention separately, then nets them.

Why it can't just be on-ice totals. Counting the expected goals for and against while a player is on the ice gives his linemates much of the same credit, and rewards anyone who plays with stars or against weak opponents. Process separates them by comparing thousands of stints with different combinations of teammates and opponents.

How it works. For every five-on-five stint with both goalies in net, we take the expected goals each side created, as a rate per 60 minutes. One regression, fitted season by season across every stint, gives each skater two numbers: his effect on his team's chances when on the ice, and his effect on the opposition's. It accounts at the same time for which team was at home, the zone of the faceoff that started the stint (or a change on the fly) and the score: tied, up or down one, or up or down two or more. Longer stints count for more.

Pulling thin evidence toward average. A fourth-liner with 300 minutes, or two defencemen who never play apart, can't be separated as reliably as a regular who plays with everyone. The regression uses a statistical penalty (ridge regression) that pulls every player's estimate toward zero, meaning average, and pulls hardest where the evidence is thinnest. We chose its strength by testing which setting best predicted chances in games left out of the fit, using 2024–25, and use the same setting in every season. This is the method public analysts call regularized adjusted plus-minus (RAPM).

From a rate to wins. Take McDavid in 2025–26. We estimate that he added 0.44 expected goals per 60 minutes to his team's chances and 0.12 to the opposition's. The net, +0.32 per 60, over his 21.8 hours of five-on-five play comes to 6.9 goals. Divided by 5.64 goals per win (see from goals to wins), that's 1.22 wins above average. Defence can carry a player too: in 2023–24, Jaccob Slavin added 0.35 expected goals per 60 in creation and prevented 0.07, for 1.64 wins.

The numbers. In 2025–26 that meant about 124,000 five-on-five stints and 1,030 hours of play, with the league average at roughly 2.5 expected goals per team per 60 minutes. About 1% of 2025–26 attempts couldn't be matched cleanly to a stint and its players. The stints involved are left out of the fit rather than counted as if nothing happened.

Limits. Players who always share the ice stay hard to separate, and the pull toward average moves their estimates toward each other. The pull toward average understates genuinely extreme players, especially in one season. For 2010–11 through 2017–18, the team-level adjustment in the chance values means a team's conversion rate over those years has a small effect on its players' process. Process credits chances, not goals: finishing is separate.

Finishing: scoring beyond the chances

Finishing measures how many goals a skater added, or cost, by converting his own attempts better or worse than their expected-goal values, adjusted for how noisy shooting is. It isn't simply goals minus expected goals.

Why adjust. A shooting season is a small sample. One more post or save swings the raw gap, and much of any hot or cold streak doesn't repeat. So we don't credit the whole gap. We estimate the shooter's effect on his scoring odds, then pull that effect toward zero by an amount that depends on how much evidence his attempts carry and on how much real finishing talent varies across all shooters of his era. Statisticians call this empirical-Bayes shrinkage.

How it works. For each skater-season we compare his goals with the chance values of his attempts and turn the gap into an effect on the log-odds of scoring, the scale on which a probability can't go below zero or above one. The typical real spread in that effect across shooters is estimated from everyone with at least 150 attempts, separately for 2018–19 onward and for earlier seasons. Each player's effect is shrunk toward zero accordingly. We then apply the shrunk effect to every one of his attempts and add up how much it changed their scoring chances. That sum is his finishing, in goals.

An example. In 2023–24 Auston Matthews scored 67 goals on attempts worth 49.4 expected goals, not counting his two empty-net goals. The raw gap is +17.6. Our finishing credits him with +9.5 goals, or 1.68 wins. A shooter with fewer attempts, or an era with less real variation in finishing, would keep a smaller share of the same gap. The share kept isn't a probability that the skill is real.

Limits. Shrinkage can understate a genuinely exceptional finisher and can't say which goals were luck. The rebound calibration drift noted under expected goals affects players with many rebound chances. For 2007–08 to 2009–10, finishing is estimated from road games only, then scaled to the full season, because home arenas recorded shot distance so differently.

Special teams: power play and penalty kill

The power-play component is a skater's effect on his team's net chances with the man advantage; the penalty-kill component is his effect on the opposition's advantage while shorthanded. Both are measured against average, at five-on-four and four-on-five, and accumulate over the minutes he played there.

How it works. The same regression approach runs on five-on-four stints. The response is the power-play team's net chances: what it created, minus any shorthanded chances it allowed. Power-play skaters get an offensive effect and penalty killers a defensive one, with the same score, home and zone-start adjustments and its own pull toward average, again chosen by testing on held-out games. Each effect is multiplied by the player's minutes in that role. Five-on-three and other manpower states sit outside these two components.

Examples. McDavid's 310 power-play minutes in 2025–26 were worth 3.5 goals, or 0.62 wins. Slavin killed penalties for 254 minutes in 2024–25, worth 2.7 goals prevented and 0.47 wins; across the last three seasons his penalty killing ranks in the top 5% of qualified defencemen who kill penalties.

Limits. This is the least clearly identified component. Power-play units are small and stable, so the same five players share almost all the credit, and a coach's system and deployment are hard to separate from the players. Totals also reflect opportunity: little power-play time isn't average power-play play. For ratings, a player needs at least 1.5 power-play minutes or 1 penalty-kill minute per game to be compared at all (see ratings).

Penalties: drawing calls and taking them

The penalties component is the power-play opportunity a player earns for his team minus the opportunity he gives the opposition, at the average worth of a power play.

How it works. Each penalty counts in two-minute units: a minor is one unit, a double minor two, a major two and a half. Penalties called at the same moment on both teams cancel each other, minute for minute, so a fight or matching minors create no advantage. Misconducts aren't counted. Each season, a unit is worth the league's net power-play goals (goals scored with the advantage, minus shorthanded goals allowed) divided by its net penalty units: 0.179 goals per unit in 2025–26, and between 0.14 and 0.18 in every season since 2010–11.

An example. McDavid drew 48 units and took 16.5 in 2025–26. The difference, 31.5 units at 0.179 goals each, is 5.6 goals, or 1.00 win. Over the last three seasons penalties supplied 2.06 of his 7.52 wins above average: drawing calls is a real part of his game, not a quirk.

Limits. This uses the average value of an opportunity, not what happened on each power play, so a player isn't credited for his teammates' power-play goals. About 7% of penalties have no recorded drawer, and those earn no one credit. We don't calculate a penalty component for 2007–08 to 2009–10, so those seasons have four components, not five.

From goals to wins

Every component is measured in goals, then divided by the number of goals it takes to add one win. In 2022–23 to 2025–26 that's 5.64 goals.

How it works. For each era, we compare every team's winning percentage with its goal differential per game, excluding empty-net goals, and find how much the winning percentage rises for each extra goal. Goals per win is the inverse of that slope. Each season converts at its own era's rate:

SeasonsTeam-seasonsGoals per win
2007–08 to 2009–10905.21
2010–11 to 2013–141205.15
2014–15 to 2017–181214.74
2018–19 to 2021–221255.21
2022–23 to 2025–261285.64

Why leave out empty-net goals. Teams score into an empty net because they're already winning, and concede because they're already losing. Counting those goals made goal differential look more spread out without adding any wins, and raised goals per win by about 9%. None of our components counts empty-net goals, so the conversion doesn't either.

Why a straight line. We tested the main alternative, a Pythagorean form that relates goals for and against to winning percentage, on seasons left out of the fit. It improved accuracy by 0.4%, far short of the 3% we required to switch.

Replacement and WAR

WAR compares a player with the depth players a team uses when its regulars aren't available. We measure that level from real depth players: those outside a team's top 13 forwards or top 7 defencemen by five-on-five ice time in a season.

Why replacement, not average. An average NHL regular is valuable: teams can't replace him from the minors for nothing. Measuring against average treats that player as worth zero. Measuring against replacement gives him credit for being a regular.

How it works. For each era and position, we pool the wins above average and the five-on-five hours of every depth player, as defined above, on his main team that season. The result is a rate per 60 minutes. In 2022–23 to 2025–26, depth forwards played 0.0113 wins per 60 minutes below average (from 800 player-seasons) and depth defencemen 0.0208 below (from 391). A player's replacement adjustment is that rate times his own five-on-five hours: what a depth player would have given up in his minutes.

Applied once. The five components stay measured against average. The replacement adjustment is added once, to the total, not to each component. McDavid's 21.8 five-on-five hours in 2025–26 earn 0.25 wins. Slavin's 22.0 hours in 2023–24 earned 0.46, because the defence rate is nearly twice the forward rate.

On team pages. A team page's Replacement adjustment adds up its current players' adjustments for 2025–26. The league average is about 4.5 wins per team.

What we tested and rejected. Players in their first NHL season with limited minutes look like an obvious stand-in for replacement. They mix fringe players with prospects on their way up: the forward group came out at average, not below it.

Limits. Replacement level is an estimate. Its 90% range in the current era runs from 0.0068 to 0.0160 wins per 60 below average for forwards and 0.0167 to 0.0248 for defencemen. It describes depth players who actually played, not every player a team could have found. The rates were last estimated in September 2026 and are held fixed between rebuilds, so a change in WAR reflects the player, not a moving baseline.

Rates, windows and workload

Player listings show WAR per 82 covered games over the last three seasons, 2023–24 to 2025–26. Goalies get WAR per 60 covered games over the same seasons. A skater needs 60 covered games in that window for a rate, and a goalie 30.

How WAR per 82 works. We add up a player's WAR and his covered games across the three seasons, then divide and multiply by 82. We don't average three season rates, which would give a 20-game season the same weight as an 82-game one. McDavid has 8.20 WAR in 225 covered games: 2.99 per 82. Matthews has 6.20 in 208: 2.44 per 82. Adam Fox is at 2.97 and Slavin at 1.92.

Goalies use 60, not 82. Goalie WAR per 60 divides WAR by covered games and multiplies by 60. The 60 is games, not minutes: roughly a starting goalie's season. Andrei Vasilevskiy has 4.38 WAR in 173 games, 1.52 per 60. Igor Shesterkin has 1.60.

Why three seasons. One season is noisy: Matthews's WAR went from 4.69 in 2023–24 to 0.51 and 1.00 in the next two. Three seasons balance what a player is doing now against a sample large enough to trust. An article may use a different window, and says so when it does.

Minimum workload. Below 60 covered games for a skater, or 30 for a goalie, across the three seasons, a player keeps his searchable page, contract and season history but gets no recent rate, rating or place in a comparison group. Among current players, 392 forwards, 211 defencemen and 64 goalies qualify; 85 forwards, 56 defencemen and 8 goalies fall short.

Who gets a page. Every signed current skater has one, as does every signed goalie with NHL games in our records. Former skaters need 500 covered games in our records, and former goalies 60 covered games from 2018–19 onward. Season tables always show totals, not rates.

WAR $: WAR in today's cap dollars

WAR $ multiplies WAR by one fixed rate: $8,274,108 per WAR, in 2026–27 cap dollars. The rate prices a player's share of league WAR at the same share of what teams can spend on skaters above minimum salaries.

How the rate is set. Start with the 2026–27 salary cap: $104 million for each of 32 teams. Reserve a $850,000 minimum salary for 20 roster spots per team. That leaves $2.784 billion of spending above the minimum across the league. Nine-tenths of it, $2.506 billion, is allocated to skaters; the rest is set aside for goalies, who get no dollar figures. Divide the skater money by the positive WAR of a fixed reference group: the 384 forwards and 192 defencemen with the most WAR in 2025–26, as many as 12 and 6 per team would give, with 302.82 WAR between them. That gives $8,274,108 per WAR.

The idea in one line. If teams spent $100 million on skaters above the minimum and those skaters produced 50 WAR, each WAR would be worth $2 million. A player who produced 5% of the WAR would be priced at 5% of the money.

How it's applied. Each season's WAR is multiplied directly by the rate. No minimum salary is added, and nothing is capped: McDavid's 2025–26 comes to $27.4 million, above the $20.8 million maximum any player can be paid. Negative WAR stays negative. Every season uses the same rate, so a 2012 season and a 2025 season are priced on the same scale, and a career total adds them up. WAR $ per 82 is WAR per 82 times the rate: $24.7 million for McDavid.

What it isn't. WAR $ prices past play. It isn't a salary, a contract estimate or a forecast. Cap hit on a player page is a separate reported fact. The rate also works in reverse: on this scale, a $10 million cap hit corresponds to about 1.2 WAR. That's arithmetic on a declared scale, not a prediction that a player will reach it.

The main choice. Giving skaters 90% of the money follows the fact that two of a team's 20 regular spots are goalies. It's a convention, not a measurement of how much goaltending is worth. At 85% the rate would be $7.8 million per WAR, and at 80%, $7.4 million. Every player's WAR $ would scale by the same proportion, so rankings wouldn't change.

Goalies: stopping more than the shots suggest

A goalie's contribution starts with goals saved above expected (GSAx): the goals he prevented compared with what the shots he faced were expected to produce, adjusted for the shooters and pulled toward average. We convert that to wins and add a replacement adjustment to get goalie WAR, from 2018–19 onward.

Which shots count. Shots on goal (saves and goals) in the regular season, with an identified goalie in net and not at an empty net. Penalty shots are excluded. From 2023–24 on, we also restore the goalie on a small number of shots whose on-ice record left him out, when the event itself names a goalie from that game's roster and nothing contradicts it.

How it works. Each shot's chance of going in starts from its expected-goal value, re-levelled to the season, adjusted for the shooter's finishing and converted to a chance given that the shot reached the net. We then estimate one effect per goalie-season: how much he changed those chances. The estimate is pulled toward average by a fixed amount, set from the estimated spread of goalie talent, so a small sample can't produce an extreme rating. Applying the effect to each shot and adding up the difference gives his GSAx, which converts to wins at the season's goals-per-win rate.

Replacement for goalies. The depth level comes from goalies who ranked third or lower on their team in playing time, pooled over four reference seasons (2018–19, 2022–23, 2023–24 and 2025–26): 114 goalies facing 28,632 shots. They allowed 0.093 more goals per 100 shots than average. A goalie's replacement adjustment is that rate times the shots he faced, converted to wins.

An example. In 2025–26 Vasilevskiy faced 1,485 shots in 58 covered games and saved 11.7 goals more than expected. That's 2.07 wins above average, plus 0.24 for replacement: 2.32 WAR. Shesterkin saved 8.3 goals above expected on 1,424 shots for 1.70 WAR.

Why goalie history starts in 2018–19. We measured 2010–11 through 2017–18 too, with the same method. Before looking at those results, we set a test for whether the depth-goalie level was stable enough to pool across those seasons. It failed: in 2016–17 the depth goalies' level differed from the pooled level by 0.24 wins per 1,000 shots, and the 95% range of that difference didn't include zero. We didn't loosen the test to pass. Those seasons stay unpublished until a different method passes.

Limits. Goalie results are noisy from year to year (see how we test it). One effect per season is shared across even-strength and shorthanded play, so we don't claim separate penalty-kill goaltending. Goalies get no dollar figures, and their WAR isn't added to skater totals.

Ratings and comparison groups

A player page rates each skater against current NHL players at his position who have at least 60 covered games in the last three seasons. Articles use a different group: regular forwards or defencemen. The two can give slightly different standings.

On player pages. The comparison group is qualified current NHL skaters: 392 forwards and 211 defencemen. The overall rating is the WAR per 82 percentile: Elite from the 92nd percentile up, Strong from the 75th, Average from the 25th, Weak from the 8th and Poor below that. Each component is ranked by its three-season total. A component is called a strength at the 80th percentile or above, or a weakness below the 20th, only if it's also worth at least 0.35 wins either way. That floor stops small totals from reading as defining traits. Power play and penalty kill are compared only for players with at least 1.5 power-play or 1 penalty-kill minute per game; anyone below that is shown as rarely used, not average.

Two examples show the rules at work. Miro Heiskanen's finishing sits around the 8th percentile of defencemen but is only 0.06 wins below average, so it's labelled typical. McDavid's penalty killing isn't ranked at all: half a minute per game is too little to compare.

In articles. Writers compare a player with regular forwards or defencemen: those with at least 500 five-on-five minutes in a season, or 1,500 across three seasons. Standings are given in steps of 5%, such as top 10% or bottom 25%, because the estimates don't support finer distinctions.

Why the two can differ. The groups include different players: the article group covers everyone with regular minutes in the chosen seasons, including players no longer in the league; the page group covers current players only. The window can differ too, and the page counts ties slightly differently. A player's standing in an article and his rating on his page describe the same evidence against different benchmarks.

How we test it

We test each step against evidence it didn't learn from, check that the arithmetic adds up exactly, and publish only what passes. Several things we built didn't pass, and we say so.

Chance values. The expected-goals model was scored on a season it never saw, 2025–26, for both ranking and level (see expected goals).

Do the components repeat? A measure of skill should carry over from one season to the next more than luck does. Using every pair of consecutive seasons from 2010–11 to 2025–26, with workload minimums set before we looked:

How strongly each measure in one season predicts the next: penalties, WAR and process carry over most; finishing, special teams and goaltending least.
How strongly each measure in one season predicts the next: penalties, WAR and process carry over most; finishing, special teams and goaltending least.
Measure, one season to the nextMinimum each seasonPairsCorrelation
Penalty units drawn minus taken, per game40 covered games6,7440.60
WAR per game40 covered games6,7440.49
Process, net per 60 minutes500 five-on-five minutes6,4620.34
Finishing, per 100 attempts100 attempts5,1060.22
Penalty kill, per 60 minutes100 shorthanded minutes2,1080.21
Power play, per 60 minutes100 power-play minutes2,9490.21
Goalie GSAx, per 100 shots (from 2018–19)1,000 shots1470.17

These correlations are modest. Hockey is noisy: one season of finishing or goaltending says much less about the next than one season of chance-driving does. That's why we shrink the noisy components, use three seasons for ratings and describe players in broad bands rather than exact ranks.

Does the arithmetic add up? Every published skater-season is checked: the five components sum to wins above average, one replacement adjustment turns that into WAR, and WAR times the rate equals WAR $, to within a billionth of a win and a hundredth of a cent. Goalie GSAx, WAA, replacement and WAR reconcile the same way. A missing measurement stays blank; it's never filled with zero or average.

Forecasts. We don't publish next-season projections, forecast ranges or the odds that a contract pays off. Our tests showed recent performance is a useful guide to the next season. Adding age and availability improved accuracy only slightly, by less than the minimum we'd set in advance. Ranges meant to contain 80% of outcomes contained only 57% to 60% for star players, even one season ahead, and longer ranges failed or couldn't be tested. Until a method passes, we don't show them.

No error bars, for now. An earlier version showed measurement ranges around each season. After our September 2026 review changed how contributions are counted, those ranges no longer matched the numbers they surrounded, and we haven't yet shown that new ones would be accurate. A single number doesn't mean the uncertainty is small: a gap of a tenth or two of a win between two players is rarely meaningful.

Methods we tried and dropped.

  • An earlier shot model scored higher on paper (an AUC of 0.825) because some of its inputs could reveal that a goal had been scored. We retired it in July 2026.
  • Chaining each season's process estimates to the previous season's did no better than simply pooling seasons.
  • A more elaborate statistical model of shooter and goalie talent predicted later seasons worse than our simpler shrinkage.
  • A Pythagorean goals-to-wins conversion wasn't enough better to adopt.
  • Adding features to the shot model (rush and cross-ice pass proxies, icing, off-wing position and longer event histories) didn't improve it on held-out seasons. Some of those tests used shot records we have since corrected, so we plan to repeat them: this one is a work in progress.
  • Shifting all of an arena's 2007–08 to 2009–10 shot distances by one amount distorted close-range chances; we compare each home rink with the same team's road games instead, and estimate finishing on road games only.
  • A limited-minute replacement group mixed fringe players with prospects on their way up.
  • The 2010–11 to 2017–18 goalie history failed its stability test.

What the numbers can't see

Our estimates describe recorded play, measured against stated baselines. They can't see everything a scout can, and several limits apply to every page.

  • The record's blind spots. No passes, possession, screens, gap control or goalie positioning. Playmaking shows up only through the shots that follow it, so pure passers may be underrated.
  • Shared ice. Linemates who rarely play apart are hard to separate, and the adjustment pulls their estimates together.
  • Small samples. One season of finishing, special teams or goaltending is mostly noise. Close calls, such as 2.0 against 1.9 WAR, rarely mean one player was better.
  • Special teams are the least certain component, because units and systems are hard to separate from players.
  • Some situations aren't measured. Four-on-four, three-on-three overtime and five-on-three play count only through finishing.
  • Older seasons are coarser. 2007–08 to 2009–10 have no shot locations, estimate finishing on road games only and have no penalty component. Goalie WAR before 2018–19 isn't published. Careers that started before 2007–08 are cut off at 2007–08.
  • Some games are only partly covered, mainly in 2019–20, and a small share of attempts in every season can't be matched to the shift record.
  • No published ranges. Uncertainty exists even where no range is shown.
  • Missing is not zero. A blank is an unmeasured contribution, not a bad one.

Sources and credits

Data. The NHL's public play-by-play, shift charts, game reports, schedules and roster records. SportsDataverse's fastRhockey project publishes collected copies of the NHL play-by-play and shift data that our tables are built from. The SFU NHL database, compiled by Oliver Schulte's research group at Simon Fraser University, supplies 2007–08 to 2009–10. Contracts come from CapWages, with PuckPedia and team announcements used to check them.

Methods. Expected goals, regularized adjusted plus-minus and empirical-Bayes shrinkage are established methods developed by others in hockey and statistics. We didn't invent them. Our work is the implementation, source handling, accounting and testing that connect them. The shot model uses scikit-learn's gradient-boosted trees.

Other public work. We read the published methods of other public models, including HockeyStats' WAR and expected-goals models by JFresh and Patrick Bacon (TopDownHockey) and Dom Luszczyszyn's Net Rating at The Athletic, as comparisons. None of their figures feeds ours. Where our numbers differ, the difference can reflect a different definition or modelling choice rather than an error on either side.

Version history

  • June 2026. Goals per win moved to non-empty-net goal differential, which raised every WAR by about 9% without changing any ranking.
  • July 2026. Our current shot model, which uses nothing recorded after the shot, replaced an earlier version that had used fields recorded after goals.
  • September 5, 2026. The 2007–08 to 2009–10 shot model was corrected to use the score before each shot rather than after it.
  • September 6, 2026. A full review rebuilt the accounting: one record per event, corrected player and goalie identities, a single definition of WAR $ with signed values and no cap, and the removal of measurement ranges and contract probabilities that no longer matched the numbers.
  • September 25, 2026. Missing on-ice records were restored for 174 games in 2019–20 and 57 in 2024–25; goalie identities were restored for some 2023–24 to 2025–26 shots; goalie WAR was added from 2018–19; the dollar rate became $8.3 million per WAR.
  • September 29, 2026. Rosters and contracts were refreshed to 818 signed players.
  • September 30, 2026. Our shot records were rebuilt from the play-by-play: from 2019–20 on, each shot's position relative to the event before it was corrected, as were 2019–20 skater counts. The few attempts with no recorded shot type were given the most common type, and penalty shots were removed from every measure. The shot model was refit with unchanged settings, goals per win was recalculated, and the dollar rate became $8,274,108 per WAR.

Technical notes

These are the exact settings behind the sections above, for readers who want to reproduce or compare them.

StepSetting
Shot modelscikit-learn HistGradientBoostingClassifier: 130 iterations, learning rate 0.06, 31 leaves, L2 penalty 0.05
Shot model inputs18 numeric and 6 categorical pre-shot features; no identities, post-shot, penalty-clock or shift-length fields
TrainingRegular-season, non-empty-net unblocked attempts excluding penalty shots, 2018–19 to 2024–25; season weight 0.45 raised to the number of seasons before 2024–25
Re-levellingPer season: logistic regression on the model's log-odds, with shot-type and manpower (even, power play, shorthanded) terms, fitted to that season's goals
ProcessWeighted ridge regression per season on five-on-five stints of 4 seconds or more, response = expected goals per hour for the attacking side; offence and defence terms per skater; home, zone-start and score-state controls (five score buckets); weight = stint seconds; ridge penalty (alpha) 31,623, chosen by five-fold cross-validation grouped by game on 2024–25
Special teamsSame design on five-on-four stints, response = power-play team's net expected goals per hour; ridge penalty 10,000
FinishingRaw log-odds effect = (goals − Σp) ÷ Σp(1−p); spread τ² by method of moments among shooters with 150+ attempts (0.0245 from 2018–19, 0.0235 before); shrinkage = τ² ÷ (τ² + 1/Σp(1−p)); finishing goals = Σ[logistic(logit p + shrunk effect) − p]
PenaltiesUnits = assessed minutes ÷ 2, net of simultaneous opposing minutes; value per unit = season net power-play goals ÷ season net units
Goals per winInverse of the games-weighted slope of winning percentage on non-empty-net goal differential per game, by era
ReplacementRate = Σ WAA ÷ Σ five-on-five hours for players ranked below 13th (forwards) or 7th (defencemen) in five-on-five time on their main team, by era and position; adjustment = −rate × player's five-on-five hours
WAR $ rate0.9 × 32 × ($104M − 20 × $850,000) ÷ positive WAR of 2025–26's top 384 forwards and 192 defencemen = $8,274,107.77
GoaliesEffect θ = Σ(goal − p) ÷ (Σp(1−p) + 1/τ²), τ² = 0.0077; GSAx = Σ[p − logistic(logit p + θ)]; probabilities cross-fitted in five game groups; replacement = 0.093 GSAx per 100 shots
RatesWAR per 82 = Σ WAR ÷ Σ covered games × 82; goalie WAR per 60 = Σ WAR ÷ Σ covered games × 60; window 2023–24 to 2025–26
PercentilesPlayer pages: share of qualified peers at or below the player. Articles: share below plus half of ties, out of the group size minus one

Glossary guide

Each term in our glossary is explained in full in this section of the report.

Glossary termSection
Wins above replacement (WAR)Replacement and WAR
Wins above average (WAA)The short version, From goals to wins
Replacement levelReplacement and WAR
WAR per 82 gamesRates, windows and workload
Covered gamesCovered games
Regular forwards and defencemenRatings and comparison groups
Component valueThe short version
WAR $WAR $
Cap hit and AAVThe data
Price-implied WARWAR $
Uncapped WAR $WAR $
Current cap dollarsWAR $
Expected goalsExpected goals
ProcessProcess
FinishingFinishing
Goals above expectedFinishing
Goals saved above expectedGoalies
Special-teams valueSpecial teams
Penalty differentialPenalties
ShrinkageFinishing
Regression toward averageFinishing
Central estimateHow we test it
Measurement rangeHow we test it

Glossary · quick definitions · Browse explainers