Power play metrics

[Master_Of_Districts]

That said, 16% is probably the modern limit for sustainablity of on-ice PP shooting%. If a team is above that on the year like Edmonton was, there's a good chance they're going to fall.

Perhaps for individual players.

For teams, the upper limit is surely lower - the skill standard deviation is a little less than 0.5%, so a talent level of 16% would be over seven standard deviations above the mean (the talent distribution is approximately normal).

 

[overpass]

Perhaps for individual players.

For teams, the upper limit is surely lower - the skill standard deviation is a little less than 0.5%, so a talent level of 16% would be over seven standard deviations above the mean (the talent distribution is approximately normal).

What set of years do you get that skill standard deviation from?

 

[Master_Of_Districts]

What set of years do you get that skill standard deviation from?

I was just going by memory from when I looked at this type of thing in the past.

Actually checking, it looks like the skill standard deviation for the group of seasons spanning 2005-06 to 2009-10 (inclusive) was 0.0048.

Not sure why I didn't include 2003-04 and 2010-11 at the time, both of which I have data for. If I add those seasons, I get 0.0064.

I remove empty net goals, so if you've calculated something different, that might explain it.

 

[Talks to Goalposts]

Perhaps for individual players.

For teams, the upper limit is surely lower - the skill standard deviation is a little less than 0.5%, so a talent level of 16% would be over seven standard deviations above the mean (the talent distribution is approximately normal).

You're probably right, I was just positing an upper limit for true talent based on that one observation.

 

[mindmasher]

I don't think anyone would disagree that SH/60 is a better predictor than SH. But why use either when you can use goals? Or better yet, GF/GA?

I honestly don't know how many times this can be gone over. Team shot volume is generally repeatable. Team shooting percentage is generally not - it tends to go back to league average.

Goals are a result of these two factors, and by looking at the maintainable and repeatable statistic, we have a better idea of future performance without adding randomness from shot quality and shooting percentage.

In addition, there are far more shot events than goal events and because of this it is a lot more impervious to small sample variations. You could easily have 10 PP goals in a year by blind, fluke luck, but you'd be hard pressed to generate 20% of your entire shot volume by luck.

I see a lot of people constantly talking about GF and goal differential being more useful than shots. The above reasons are why people with statistical backgrounds tend to favour shooting rates.

 

[Czech Your Math]

I honestly don't know how many times this can be gone over. Team shot volume is generally repeatable. Team shooting percentage is generally not - it tends to go back to league average.

Goals are a result of these two factors, and by looking at the maintainable and repeatable statistic, we have a better idea of future performance without adding randomness from shot quality and shooting percentage.

In addition, there are far more shot events than goal events and because of this it is a lot more impervious to small sample variations. You could easily have 10 PP goals in a year by blind, fluke luck, but you'd be hard pressed to generate 20% of your entire shot volume by luck.

I see a lot of people constantly talking about GF and goal differential being more useful than shots. The above reasons are why people with statistical backgrounds tend to favour shooting rates.

Shots may be miscounted, as it appears happens in some arenas.

I would like to see the studies and data that correlate PP shots/60 to PP% (PPG/PPO) & PP Shooting%.

 

[overpass]

I honestly don't know how many times this can be gone over. Team shot volume is generally repeatable. Team shooting percentage is generally not - it tends to go back to league average.

Goals are a result of these two factors, and by looking at the maintainable and repeatable statistic, we have a better idea of future performance without adding randomness from shot quality and shooting percentage.

In addition, there are far more shot events than goal events and because of this it is a lot more impervious to small sample variations. You could easily have 10 PP goals in a year by blind, fluke luck, but you'd be hard pressed to generate 20% of your entire shot volume by luck.

I see a lot of people constantly talking about GF and goal differential being more useful than shots. The above reasons are why people with statistical backgrounds tend to favour shooting rates.

While some of us don't wish to dismiss the percentages, that doesn't mean we're statistically ignorant.

Shot volume and shooting percentage are the same in several ways. Both are a skill - different players have different abilities, and since teams are just collections of players, different teams have different abilities. Both are subject to random variation.

The difference is a difference in degree, not in kind. Shot volume stabilizes more quickly at the team or individual's "true talent" than shooting percentage.

What to do? Some people would have us evaluate players based solely on shot rates, to minimize the effects of random variation. This is one option, but not the only one. Here are some others:

If the time period is sufficiently long, we could look at goal rates instead of shot rates. What is a sufficiently long length of time? This is an empirical question that can be answered by finding the point at which the variation in shooting percentage is more signal than noise.

We could regress shooting percentage partially, depending on the length of time we are looking at. What is the appropriate amount of regression for a given time period? Again, a question for which the answer can be found in the data.

We could also incorporate additional information beyond the time period we are looking at. Even if we are just looking at a small number of games where shooting percentage has an unacceptably low signal-to-noise ratio, we don't have to assume all players and teams have equal shooting talent. We could use information from past seasons to estimate shooting talent, or we could make an estimate based on the shooting talent of similar players or players in similar roles. Any of these could provide a better estimate of shooting talent than assuming all players and teams are equal. We could also use these estimates instead of league averages when partially regressing the numbers.

 

[mindmasher]

Shots may be miscounted, as it appears happens in some arenas.

I would like to see the studies and data that correlate PP shots/60 to PP% (PPG/PPO) & PP Shooting%.

Yes, shots may be miscounted, but these are the kinds of stats that we can easily spot check. There simply isn't a large enough error, in my opinion, for this to really enter the debate.

As for your second sentence, I'm not sure what you would expect to see in comparing shot volume to shooting %? Low shot volume, high shooting percentage? I'd be willing to make a bet without looking at the correlation coefficient that there is no serious link. I'm curious what you expect to find there though.

As for shot volume to PP%, that, I think, WILL correlate with PP%.

 

[mindmasher]

While some of us don't wish to dismiss the percentages, that doesn't mean we're statistically ignorant.

Shot volume and shooting percentage are the same in several ways. Both are a skill - different players have different abilities, and since teams are just collections of players, different teams have different abilities. Both are subject to random variation.

The difference is a difference in degree, not in kind. Shot volume stabilizes more quickly at the team or individual's "true talent" than shooting percentage.

What to do? Some people would have us evaluate players based solely on shot rates, to minimize the effects of random variation. This is one option, but not the only one. Here are some others:

If the time period is sufficiently long, we could look at goal rates instead of shot rates. What is a sufficiently long length of time? This is an empirical question that can be answered by finding the point at which the variation in shooting percentage is more signal than noise.

We could regress shooting percentage partially, depending on the length of time we are looking at. What is the appropriate amount of regression for a given time period? Again, a question for which the answer can be found in the data.

We could also incorporate additional information beyond the time period we are looking at. Even if we are just looking at a small number of games where shooting percentage has an unacceptably low signal-to-noise ratio, we don't have to assume all players and teams have equal shooting talent. We could use information from past seasons to estimate shooting talent, or we could make an estimate based on the shooting talent of similar players or players in similar roles. Any of these could provide a better estimate of shooting talent than assuming all players and teams are equal. We could also use these estimates instead of league averages when partially regressing the numbers.

I don't disagree. It would be nice to factor in team shooting ability. The problem is to get a stable analysis of team shooting ability it may require more data points then the powerplay roster is stable for. Part of the problem in modelling any teams future results is doing is with a small enough data set that you can ignore a huge amount of team roster churn (including coaches).

Alternatively, it's possible we have a case of diminishing returns. We might be able to statistically model team shooting% variations amongst different PP's (and use it to modify 'team PP ability'), but if it nets us an additional 9% information for 300% more effort, I think we can agree that in most circumstances having 91% of the predictive power is sufficient for most discussions.

 

[Czech Your Math]

Yes, shots may be miscounted, but these are the kinds of stats that we can easily spot check. There simply isn't a large enough error, in my opinion, for this to really enter the debate.

As for your second sentence, I'm not sure what you would expect to see in comparing shot volume to shooting %? Low shot volume, high shooting percentage? I'd be willing to make a bet without looking at the correlation coefficient that there is no serious link. I'm curious what you expect to find there though.

As for shot volume to PP%, that, I think, WILL correlate with PP%.

What are you basing your opinion on that Shots/60 on the PP is a better predictor than GF/60 on the PP?

What studies and data are you using to conclude this? Where is Shots/60 data found?

 

[overpass]

I don't disagree. It would be nice to factor in team shooting ability. The problem is to get a stable analysis of team shooting ability it may require more data points then the powerplay roster is stable for. Part of the problem in modelling any teams future results is doing is with a small enough data set that you can ignore a huge amount of team roster churn (including coaches).

Right. Maybe this is where observation/watching the games can play a part, as long as it's combined with an understanding of the talent distribution for team on-ice shooting %.

Alternatively, it's possible we have a case of diminishing returns. We might be able to statistically model team shooting% variations amongst different PP's (and use it to modify 'team PP ability'), but if it nets us an additional 9% information for 300% more effort, I think we can agree that in most circumstances having 91% of the predictive power is sufficient for most discussions.

I agree. I just have the feeling that some analysts today would be downgrading 19-year old Mario Lemieux or Wayne Gretzky based on their "unsustainably high percentages", and those guys went on to create a ton of value out of those percentages. It looks like Sidney Crosby creates a lot of value from the percentages as well. Maybe Ryan Nugent-Hopkins will in future. We should be open to recognizing value no matter how it is created.