Loading goals…
hockeygamebot

Methodology

How the numbers work.

HockeyGameBot uses play-by-play, shift, expected goal, RAPM, and production data to estimate player value. Two products: HGB WAR (what a player produced this season) and HGB Talent (how good the player is, with ice-time allocation normalized).

Hockey Game Bot is an independent project and is not affiliated with the NHL. Every number on this site is produced automatically each night from NHL data. If one looks wrong, email support@hockeygamebot.com and tell us the page and the number. We will check it and correct it if it is ours.

The site is free. If you like it and want to chip in, you can support HockeyGameBot here.

WAR vs Talent

Current season

HGB WAR

Current-season accounting. How much value did this player produce this year versus a replacement-level player? Single-season, no priors, and it counts every minute a player skated. Comparable to baseball WAR in concept: it tells you what happened, not how good the player is.

Opportunity-normalized

HGB Talent

How good the player is, with ice-time allocation normalized. Each situation is weighted by a fixed position-level share of ice time rather than the minutes a coach handed out, so power-play usage doesn't inflate the number. Multi-season history anchors it. A player can have strong Talent but a mediocre WAR (small sample this year), or the reverse.

What Goes Into WAR

EV Offense

Isolated on-ice expected goal share contribution at 5v5, via RAPM. Measures how much a player moves the needle offensively when they're on the ice, separated from their teammates' contributions.

EV Defense

Isolated on-ice expected goals against impact at 5v5, via RAPM. How much does this player's presence suppress opponent offense, independent of who they play with?

Individual Offense

Personal goals and primary assists production bonus, z-scored versus positional peers. Rewards players who drive personal counting stats above and beyond what RAPM already captures in shot volume.

PP Offense

Power play RAPM times PP TOI. How much offensive value a player generates on the man advantage, scaled by how much they're actually used there.

PK Defense

Penalty kill RAPM times PK TOI. Defensive impact on the kill, weighted by actual deployment.

Penalties

Drawn minus taken, with a flat value per opportunity of approximately 0.20 goals. Players who consistently draw power plays are adding value that doesn't show up in shot metrics.

Opp. Quality (QoC)

Average quality of opponents faced. Context only, not a WAR component. High QoC means a player faces tougher competition; low QoC means sheltered deployment.

Context · Not scored

Mate Quality (QoT)

Average quality of linemates. QoT and QoC appear on player cards to help interpret the scored components. They explain results, they don't add or subtract from the WAR score.

Context · Not scored

Early-Season WAR

WAR needs 500 five-on-five minutes to be official. Before that, a player shows a range instead of one number once he reaches 250 five-on-five minutes (350 for players with fewer than 500 five-on-five minutes last season). It is marked Early and shows his minutes.

The range is where two in three players land when they reach 500 minutes. It is wide at 250 minutes and narrows as the minutes pile up, down to a single number at 500. Everyone past the floor, early or official, is ranked in one pool against players at the same position, and the range is built around the player's place in that pool. The range appears once at least 30 players at the position have reached the floor; until then, WAR shows a dash and the minutes.

The early number uses the same formula as the official one, so a player's value does not jump when he crosses 500 minutes. Below the floor, WAR stays blank.

Expected Goals

Expected goals (xG) is the engine under almost everything here. Every unblocked shot, on net or missed, gets a probability of becoming a goal, based on where it was taken, the angle, the shot type, whether it followed a rebound and how fast, whether it came off the rush, and what happened the moment before. Add those probabilities up and you get chance quality that doesn't care whether the puck actually went in.

HGB runs four separate xG models, not one: even strength, power play, penalty kill, and empty net. A power-play shot and a 5v5 shot are different problems, with different shot locations and different goalie behavior. Push them through a single model and it over-rates power-play chances and flattens the rest. Splitting them tightens the calibration. How close predicted goals land to actual goals is under “How we check it” below.

The models are gradient-boosted, trained on every unblocked shot since 2022, the seasons after the NHL changed its shot tracking, so the location data is consistent. This is the same shot basis the public models use: unblocked attempts, missed shots included. xG feeds the even-strength and power-play numbers below and the goalie save model. Better xG, better everything downstream.

How we check it. The model is tested out of sample. Across 2023-24 to 2025-26, the model expected 16,485 even-strength goals and 16,696 were scored, 1.3% off. Every season was scored by a version of the model that had never seen it, and no single season was off by more than 4%. See every test, explained in plain English → Expected goals are raw model totals, not adjusted for score or venue. Individual xG counts all situations; on-ice columns are 5v5. The column guide on the Data page defines every column.

How RAPM Works

RAPM (Regularized Adjusted Plus-Minus) isolates individual player impact by controlling for teammates, opponents, zone starts, score state, and PP expiry. Every shift creates a set of equations where each player on the ice gets credit or blame for what happened, and ridge regression finds the player-level estimates that best explain the full dataset, while shrinking noisy estimates toward zero.

A player with a small sample gets pulled toward league average. A player with thousands of 5v5 minutes gets a more confident estimate. The model uses a two-perspective design: each shift generates one home-offensive row and one away-offensive row, keeping offensive and defensive signals separated so that a defenseman who suppresses shots doesn't accidentally look like an offensive contributor.

Opportunity Normalization

WAR measures what a player produced; Talent estimates how good he is. One major difference is deployment. WAR credits every minute a player skated, so heavy power-play time gives a player more opportunity to accumulate WAR. Talent normalizes that: each situation (even strength, power play, penalty kill) is weighted by a fixed position-level share of ice time rather than the player's own minutes. Extra power-play minutes don't mechanically increase Talent; they provide more evidence about the player's per-60 performance.

Talent leans on history. Each component blends the current season with up to three prior seasons, weighted toward the recent ones (roughly 30 / 16 / 8 across the last three years) and adjusted for age. How far the current season pulls the estimate depends on sample size. The thresholds below mark where the current season and the prior carry equal weight.

Losing a role doesn't erase ability. A player dropped from the power play keeps his established power-play prior instead of collapsing to zero; the missing minutes lower his confidence rather than his score. Players with thin history (under 1,000 even-strength minutes across the last three seasons, or a player with no usable prior history) are flagged limited sample on the card.

Where current season and prior weigh equally

Situation 50/50 point
Even strength ~300 minutes
Power play ~200 minutes
Penalty kill ~150 minutes

Below the threshold, the prior carries more weight; above it, the current season does. Power play and penalty kill reach that balance on fewer minutes.

Impact (per game)

Impact gives every skater and goalie one number for one game. For a skater it is the sum of seven areas, each measured against an average night for his position: Scoring (goals, primary assists and secondary assists), Chances (the quality of the shots he took), 5v5 (how the team's expected goals went at 5v5 while he was on the ice), Special teams (the same on the power play and penalty kill), Defense (blocked shots and takeaways), Discipline (penalties drawn against penalties taken) and Faceoffs (forwards only). Zero is an average night in every area, so the areas add up to the Impact number.

Not every point is equal. A primary assist is worth 0.65 of a goal and a secondary assist 0.30, because secondary assists are mostly noise. Shots count by expected goals (ixG), the chance each shot had of going in, not by raw shot or attempt counts. The 5v5 piece is his team's 5v5 expected-goal differential with him on the ice, compared with a bar set before puck drop for the minutes he played. The bar is three quarters league average and one quarter his team's rate from its earlier games, and it never changes after the game. These expected goals are adjusted for score and home ice, a fifth of the piece comes from the actual goals for and against while he was on the ice, and forwards and defensemen get the same weight on it.

Seven areas, fixed shares. Each area is first put on a common scale: how far his night was from an average night at his position, in units of how much that area normally varies from game to game. Each area then has a fixed share of the number. Forwards: Scoring 30%, Chances 15%, 5v5 30%, Special teams 10%, Defense 7.5%, Discipline 3.75%, Faceoffs 3.75%. Defensemen: Scoring 30%, Chances 15%, 5v5 30%, Special teams 10%, Defense 12%, Discipline 3%, no faceoffs. The averages and the typical spread come from the 2025-26 regular season and are fixed, so they are never recomputed as a season goes on and a published number never changes when later games are added. Scoring is about 39% of the variation in a game's number, so a night driven by chances, 5v5 play, special teams or defense can rank with a night driven by points.

Impact is its own scale, not goals. A goal adds about 1.5 to a forward's number (about 2.2 to a defenseman's), and the scale is set so that the typical spread of forward games, about 1.75 either side of average, is the same as before. A forward at an average night in every area reads 0. Most skaters still read below zero on a given night, because about two thirds of skater games have no point and a point is the biggest single swing.

Special teams. Power-play and penalty-kill time is its own piece of Impact. On the power play it is the expected goals the team created while he was on the ice, compared with how the team did on its power plays when he was off it, for the same amount of ice time. On the penalty kill it is the reverse: fewer expected goals against than the team allowed without him is a plus. Early in the season both comparisons lean on the league average, then shift to the team's own numbers as the minutes add up. A night with no special-teams time shows 0.

Next to every number you will see a band: top 1%, top 5% or top 10% of games at that position, above average, average, or quiet. The band is how to read the number, and the color follows it: green for the top quarter of games at the position, neutral for the middle half, red for the bottom quarter.

Goalies are scored on goals saved above expected (GSAx), on their own scale. Their bands are Stole one, Strong, Steady and Tough night, not percentiles.

Rolling Impact. Player pages and the Impact leaderboard also show Rolling Impact: the average Impact per game, not a total, over the last 10 games and over the season. It needs at least 10 games, and under 20 games it carries an early-sample chip. Goalies are scored on their own scale and are not ranked on it.

Leverage is not in the score. A goal in a 6-1 game and an overtime winner both count as one goal. How much a goal moved the game is what Swing of the Night is for.

Provisional vs official. Right after the final horn the page computes Impact from the live play-by-play and labels it Provisional. The live data has no score and home-ice adjustment and no on-ice goals, so the provisional 5v5 piece uses raw expected goals only. The overnight run replaces it with the official number, which rescores every shot with the full overnight data, so a player's number can move and ranks near the top can change.

What changed on November 4, 2026. Impact changed on November 4, 2026; earlier posts and cards used the previous model. Every season was recomputed, so older games show new numbers.

  • Seven areas with fixed shares. Scoring used to be about three quarters of the variation in a game's number, so the number was close to a box score with extras. Each area now has a fixed share (see above) and scoring is about 39%. Example: in Buffalo's 7-6 win over Anaheim, a defenseman with two assists and no goals now sits at about 90% of the three-goal scorer's number; before it was about 40%.
  • Defense is blocked shots and takeaways. Giveaways are not counted. They mostly measure how much a player has the puck, so counting them penalized the players who carry it.
  • Faceoffs, penalties and special teams are scored on the same scale as everything else. Before, these pieces were small next to goals and rarely moved a number. Now a strong night in one of them shows up on the card.
  • Zero means an average night in every area. The seven areas add up to the number. Defensemen's numbers spread out more than before (a typical game is now about 1.6 either side of average, up from 1.3).
  • Earlier posts and cards. Posts and images published before the change keep the numbers they showed on the day. They are not edited.

What changed in October 2026. On October 28 and 29, 2026 we rebuilt the 5v5 piece of Impact and recomputed every season back to 2010-11, so older games show new numbers.

  • The bar is set before the game. A player used to be compared with his own team's rate in the same game while he was off the ice, so his number moved with how his teammates played that night. Now the bar is fixed before puck drop, three quarters league average and one quarter the team's earlier games, and new games never change old numbers.
  • Score and home ice. The expected goals in the 5v5 piece are adjusted for the score and for home ice, the way shot totals swing when a team is chasing a game. Home teams no longer get a built-in boost. Seasons before 2025-26 have no adjustment and use raw expected goals.
  • Real goals count a little. A fifth of the 5v5 piece now comes from the actual goals for and against while he was on the ice. The team 5v5 piece now tracks real 5v5 goal differential far more closely (correlation 0.30 before, 0.72 now).
  • Forwards and defensemen weigh the same. A defenseman's 5v5 piece used to count five times a forward's, so the nightly worst-5v5 lists were all defensemen. Both now use the same weight. Defensemen still play more 5v5 minutes, so their 5v5 piece runs slightly larger.

We also tested two other changes and kept both as they were. Lowering the weight of goals and assists made every number smaller but barely changed who finished on top. A baseline that grows with ice time barely helped short-minute forwards and made the number less repeatable from game to game. Most skaters read below zero on a given night because about two thirds of skater games have no point.

Stress

Stress is how much the win-chance line moved in a game. It adds up every change in the home team's win chance, play by play, ignoring the direction. A line that goes from 50% to 70% and back to 50% moved 40 swing points. One swing point is the win chance moving one percentage point, 50% to 51%. A typical game this season moves about 143. A game that is never in doubt has a low number. A game that goes back and forth has a high one.

A game. The game page shows Stress in whole swing points, next to the league's average game this season. Both teams in a game share the same number, because it is the movement of one line. The page also shows minutes in doubt: how many minutes the home team's win chance sat between 25% and 75%, out of the minutes played.

A team. A team's Stress on /stats is its average per game: the sum of its games' Stress divided by games played, shown in whole swing points. Min in doubt is the same average, in minutes. The team's game log lists every game with both numbers, and sorting by Stress puts its most stressful games first. Both columns start in 2026-27.

What is counted. Win chance comes from the HGB model, which also drives the win-probability chart. A game's number leaves out the shootout, because those plays do not change the game, and it flattens a one-play spike that is not a goal, which is usually a data glitch. Overtime counts.

What changed on November 11, 2026. Stress replaces the Seismograph magnitude, which was the same count plus half a point of win chance for shootouts, on a log scale. The game page no longer shows a magnitude or the four bands (Asleep, Felt it, Shaking, Cardiac); it shows the number and how it compares with a typical game.

Read the article: which teams play the most stressful games →

Card Guide

HGB Talent %
Hero Tile
Opportunity-normalized ability estimate. Ranks the player against same-position peers (forwards and defensemen separately, 20-plus GP; goalies excluded) with ice-time allocation normalized and his recent multi-season track record folded in.
HGB WAR %
Hero Tile
Current-season value. How much above replacement did this player produce this year, relative to positional peers?
HGB Impact %
Hero Tile
Per-game value score: goals, assists, individual xG, blocks, faceoffs, penalties, and his team's 5v5 expected-goal differential with him on the ice, compared with a bar set before puck drop for the minutes he played (a game's score never changes later). Empty-net goals don't count. Under 20 GP it carries an LTD SAMPLE chip.
Profile Bars
3-year weighted percentile profile. Each bar shows where the player ranks among positional peers.

Above average performance
Near average performance
Below average performance
Context metrics (QoC, QoT): not good or bad, just deployment context
PP / PK Bars
Percentile among players with at least 50 minutes in that situation. Small-sample players are excluded entirely.
Hatched Bar (dash)
Below the TOI threshold for that situation. No estimate is available. This is not the same as a zero value.