I had some time and finally took a look at the different publicly available expected goal models to see how well they correlate with actual goals scored. As far as I can tell, and I could definitely be wrong about this, the only free and available expected goal data is on Natural Stat Trick and Hockey Reference. An aside, but if any of you cabrones know of another publicly available XPG model I’ve missed, let me know and I’ll run the numbers. I have also included Fenwick, Corsi, and a proprietary tracking system (not the Bruins), to see how they compared.
This work included data for the last 3 seasons (2018-19 through 2020-21), using all NHL position players with over 500 minutes in a season in 5 on 5 situations. This gave me a sample of just over 1500 player seasons. I ran a simple linear regression looking at the relationships between expected goals and actual goals, expected goals against and actual goals against, and expected goal percentage vs actual goals scored percentage when a player is on the ice. The results below are the coefficient of determination (r^2), which gives the proportion of the variation in the dependent variable that is predictable from the independent variable, or the proportion of actual goals/goals against/goal percentage which is predictable from the expected goal models.
Interestingly, but unsurprisingly, expected goals scored has a stronger correlation to actual goals scored in all systems than expected goals against to goals against. If you think about it, this makes sense because of the huge disparity in talent level among NHL goalies.
As
@burstnbloom suggested, there is an extremely strong relationship between the Natural Stat Trick and Hockey Reference XPG data (r^2 G/GA/G% is 78.4% /72.5% /82.1%). Both Natural Stat Trick's and Hockey Reference's expected goal models track goals better than either Fenwick or Corsi. I was pretty surprised to see Hockey Reference's expected goals have a stronger relationship to actual goals than Natural Stat Trick, but it did both for each individual season, as well as for the combined data. The proprietary model I used was much stronger than either of the publicly available systems, which you would expect. This helps show how much better the data the analytically advanced teams have is (far more actions tracked and far more sophisticated).
Below is a table with the coefficient of determination for each relationship discussed above. I need to include a disclaimer here, as hockey has the highest degree of luck of any of the major sports, which, considering it's played with an odd shaped object and sticks on ice, makes perfect sense (if you're interested in seeing the numbers on this, let me know). So my confidence in the below is medium.
Model |
XPG vs G |
XPGA vs GA |
XPG% vs G% |
CF |
26.7% |
12.5% |
21.8% |
FF |
29.6% |
17.8% |
24.1% |
NST |
33.8% |
24.0% |
28.7% |
HR |
40.1% |
26.3% |
32.4% |
XX |
49.8% |
34.5% |
44.1% |
If you have any questions, let me know.