Penalty Shots Elo Ratings

[More Hockey Stats]

I created a new metric on my website: Penalty Shots Elo Ratings. For both goalies and skaters.

This time I went a little bit further and created personal player cards - how they fared in PS so far. I also added an option to evaluate the probability of scoring between certain players and goalies.

 

http://morehockeystats.com/players/penaltyshots

 

Here's a screenshot of the skaters' table:

And here's an output of evaluation:

[Podein25]

Elo - decent band

[More Hockey Stats]

1 minute ago, Podein25 said:

Elo - decent band

 

They're ELO.

Elo is a last name.

[Podein25]

2 minutes ago, More Hockey Stats said:

 

They're ELO.

Elo is a last name.

 

I stand corrected.

[WordsOfWisdom]

1 hour ago, More Hockey Stats said:

I created a new metric on my website: Penalty Shots Elo Ratings.

 

Part of statistical analysis involves presenting meaningful and easy to understand data to users. Some complaints that I have with your "Elo" ratings are:

• What does Elo mean? If it's just a last name thing, I would strongly suggest a new name for it. Something that describes what the stat is. For example, my biggest complaint with Corsi/Fenwick is that it says nothing about puck possession in the name. The name of the stat is meaningless by itself, and I can never remember which one factors in blocked shots and which one doesn't. 
• What do the values mean? What is a score of 1975.05? How do I quantify that? The number looks like something out of a NASCAR driver points standings list.
• What is G, M, and T?

If you have a value of 100 to represent league average, then values above 100 could be better than average for example. Values below 100 would be below average. That would give you a statistic represented as an index.

 

If we're dealing with raw numbers here, then let's present the raw numbers. When it comes to shootout stats, I look for goals and attempts, and then derive a shooting percentage from that. What more do I need? For goalies, I want goals allowed and total attempts faced. I can calculate the goalie's save percentage for a shootout from that data.

 

Just some feedback / suggestions. 

 

[More Hockey Stats]

Elo is the last name of Arpad Elo, the inventor of the rating system widely used in board games.

https://en.wikipedia.org/wiki/Elo_rating_system

 

It starts from a number, usually around of 2000 and the compares an expected performance to the actual performance. If you exceed the expectation, your rating grows, but so does your next expected performance. So the high rating is harder to maintain. The expected performance is based on the difference between the performer and the opponent ratings. Elo reflects streaks, so two players who started with equal rating, but one missed five and then scored five has higher rating than the one that scored five and then missed five. More here:

https://morehockeystats.blogspot.com/2016/12/on-players-evaluation-part-ii-elo.html

 

Usually in my Elo evaluations I start with the rating of 2000, however here, since the average success level of a skater is about 33% rather than 50%, there is a built-in rating difference of 120 points. So for a goalie the initial rating is 2060, for a skater - 1940. So when a skater of rating A is attempting a PS against goalie of rating B, the chance of a goal is a function of the difference between A and B. If the ratings are equal, the chance is 50%. If the the goalie's rating is higher by about 400 points, the chance is close to zero, it's not linear.

 

Rating of 1975 says that the forward will have a 33% chance to score against a goalie with rating 2095, but about 36% against a goalie with rating 2075, and about 24% against a goalie with rating 2175.

 

There are tooltips available for the table headers, but I agree I need a better explanation there, thanks. G means goals scored, M - Misses/Saves, T - total. SW - Shootout Wins, SL - Shootout Losses.

 

Thanks for the feedback!

[WordsOfWisdom]

On 3/22/2017 at 5:09 PM, More Hockey Stats said:

It starts from a number, usually around of 2000 and the compares an expected performance to the actual performance. If you exceed the expectation, your rating grows, but so does your next expected performance. So the high rating is harder to maintain. The expected performance is based on the difference between the performer and the opponent ratings. Elo reflects streaks, so two players who started with equal rating, but one missed five and then scored five has higher rating than the one that scored five and then missed five.

 

So it's trying to analyze who is currently hot or cold?

 

On 3/22/2017 at 5:09 PM, More Hockey Stats said:

Usually in my Elo evaluations I start with the rating of 2000, however here, since the average success level of a skater is about 33% rather than 50%, there is a built-in rating difference of 120 points. So for a goalie the initial rating is 2060, for a skater - 1940. So when a skater of rating A is attempting a PS against goalie of rating B, the chance of a goal is a function of the difference between A and B. If the ratings are equal, the chance is 50%.

 

Without delving too much into this (because I actually have an upgraded version of my hockey simulator in the works which is taking forever to build because of the complexity), what you are attempting there can be achieved with an index. You can determine what the overall percentage is for shootouts (shots / attempts) and use that as the league average. If two league average players meet, then the expected result is league average.

 

S = Shooter

G = Goalie

avg = League Average

 

Let's say the numbers are as follows:

Shooter = 25% goal

Goalie = 75% save

 

Savg vs Gavg = avg = 25% chance of goal

 

We could also say Savg = 1.0 and Gavg = 1.0.

 

If we have an above average shooter, say S = 50%, then...

 

Savg+25 vs Gavg = avg+25 = 50% chance of goal

 

In this case, Savg+25 = 2.0.

 

Thus, a value of 1.0 for a shooter = 25% goal chance. A value of 2.0 = 50% goal chance. A shooter with an index value of 2.0 is twice as likely to score against an average goalie. 

 

I can't write the proper mathematical notation with subscripts in the forum (it's almost impossible to quickly write mathematical notation on computers at all for that matter), but that's just a quick idea.

 

On 3/22/2017 at 5:09 PM, More Hockey Stats said:

e are tooltips available for the table headers, but I agree I need a better explanation there, thanks. G means goals scored, M - Misses/Saves, T - total. SW - Shootout Wins, SL - Shootout Losses.

 

That makes more sense.  

 

On 3/22/2017 at 5:09 PM, More Hockey Stats said:

Thanks for the feedback!

 

No problem.

 

[More Hockey Stats]

1 hour ago, WordsOfWisdom said:

 

So it's trying to analyze who is currently hot or cold?

 

 

Without delving too much into this (because I actually have an upgraded version of my hockey simulator in the works which is taking forever to build because of the complexity), what you are attempting there can be achieved with an index. You can determine what the overall percentage is for shootouts (shots / attempts) and use that as the league average. If two league average players meet, then the expected result is league average.

 

S = Shooter

G = Goalie

avg = League Average

 

Let's say the numbers are as follows:

Shooter = 25% goal

Goalie = 75% save

 

Savg vs Gavg = avg = 25% chance of goal

 

We could also say Savg = 1.0 and Gavg = 1.0.

 

If we have an above average shooter, say S = 50%, then...

 

Savg+25 vs Gavg = avg+25 = 50% chance of goal

 

In this case, Savg+25 = 2.0.

 

Thus, a value of 1.0 for a shooter = 25% goal chance. A value of 2.0 = 50% goal chance. A shooter with an index value of 2.0 is twice as likely to score against an average goalie. 

 

I can't write the proper mathematical notation with subscripts in the forum (it's almost impossible to quickly write mathematical notation on computers at all for that matter), but that's just a quick idea.

 

It's a combination of both averages and hot/cold, because hot/cold can also be an indication of an improved/decreased skill.

 

In general, when two chess players meet, you can also try to predict the outcome with indexes. But Elo does it better. So I see no reason why in the pure 1-0/0-1 event like a penalty shot, Elo should not be used. Sure, some of the constants involved may be tweaked, but the principle should remain correct.

 

These, for example, are the evaluations of the chances in the shootouts from Thursday night (we had three):

 

I'll tweet the evaluation for tonight's NYI-PIT shootout tomorrow, when the game files are processed in my database. You're welcome to follow.