Loading goals…
hockeygamebot

Methodology · Expected goals

How we check our xG.

Expected goals (xG) sits under most of the numbers on this site, so it has to be right. We put it through three tests, each one scored by a version of the model that had never seen the season it was graded on.

What xG means

Every unblocked shot gets a number between 0 and 1: the chance that a shot like it goes in. A long wrister from the point is worth very little. A rebound from the edge of the crease is worth much more. Add up a player's or a team's shots and you get their expected goals, the goals the chances they created were worth, whether or not the puck went in.

Goals are what happened. xG is what usually happens from those same chances. Over a few games the two can be far apart. Over a season they should be close, and that is what the tests below check.

A test it hasn't seen

A model graded on games it learned from will look better than it is, like a student grading their own homework. So every season below was scored by a copy of the model that was retrained without that season. The model never saw the answers it was graded on. This is called testing out of sample.

Does it add up?

If xG is honest, the goals it expects should add up to about the goals actually scored. Here are all even-strength shots over three regular seasons (power play, penalty kill and empty nets are left out).

SeasonGoals scoredxG expectedOff by
2023-245,5965,376−3.9%
2024-255,4785,379−1.8%
2025-265,6225,730+1.9%
All three16,69616,485−1.3%

In plain English: over three seasons and about 270,000 shots, xG lands within 1.3% of the real goal total, and no single season is off by more than 4%. The three-season figure benefits from one season over and two under canceling out, so the per-season numbers are the stricter test.

The yearly misses follow the league, not the model. Scoring per shot moved from season to season, and the version of the model that trained on those seasons misses in the same direction by about the same amount.

Does it match who scores?

Adding up is not enough. xG also has to put the goals on the right players. For every skater-season with at least 500 minutes at 5-on-5 (1,781 of them), we compared what xG expected with what happened.

The numbers are correlations. 1.0 means xG and goals line up perfectly, player for player. 0 means they have nothing to do with each other.

What we comparedAll skatersForwardsDefense
A player's goals vs his xG0.820.610.60
Team goals while he's on the ice vs on-ice xG0.710.740.64
Goals against while he's on the ice vs on-ice xG against0.560.550.57

In plain English: the players xG says should score are, by and large, the players who do. The all-skater number is the highest because it includes the easy part: forwards score more than defensemen, and xG knows that. Within forwards alone, xG still explains much of who scores. The rest is finishing skill and luck, which no shot model can see.

Does it predict next season?

This is the test that matters most. Matching goals in the same season is useful, but the point of xG is to tell you what comes next. For 992 players with 500+ minutes at 5-on-5 in back-to-back seasons, we used one season's numbers to predict the next season's goals, using models trained only on earlier seasons.

Predicting next season's…From this season's xGFrom this season's goals
Goals a player scores0.760.69
Team goals with him on the ice0.470.44
Goals against with him on the ice0.240.18

In plain English: if you want to know how many goals a player will score next year, his expected goals this year are a better guide than his actual goals this year. Goals bounce around with hot streaks and hot goalies. Chances are steadier, so they carry forward better. That is why the site leans on xG.

Goals against

xG is weakest on goals against, in both the same-season and next-season tests. That's expected: xG measures the chances a player's team allowed, and whether those chances go in depends heavily on the goalie, who xG doesn't model. Treat a skater's defensive xG as "chances allowed", not "goals allowed".

Score and venue adjusted

Teams that trail shoot more, and home teams get slightly friendlier counting. The Raw / Score & venue adjusted toggle on the Skaters and Teams tables (5v5) reweights every shot to take that out. Each shot is weighted by the shooter's score difference (capped at ±3) and whether the shooter was home or away. The weights were fit once on 5v5 shots from 2010-11 through 2025-26 and are scaled so league totals are preserved: the adjusted numbers redistribute xG, they do not add or remove any. Goalies stay raw, because goals saved above expected compares each shot's result with its own expectation. Adjusted skater values start in 2025-26.

How the tests were run

Regular season only, 2023-24 through 2025-26.

Test 1 counts every even-strength unblocked shot (on goal or missed, short misses excluded) with empty-net shots removed: 270,664 shots and 16,696 goals. At 5-on-5 only, the three-season total is off by 1.2%.

Tests 1 and 2 retrain the even-strength model three times, each time leaving out the season being scored. Test 3 scores each season with a model trained only on the seasons before it.

Tests 2 and 3 are per player-season at 5-on-5, with rates per 60 minutes and Pearson correlation. Each player-season is its own row.

The model the site runs today scores slightly higher on every test, because it has seen some of these seasons. The numbers above are the out-of-sample ones.