2017-18 stats and underlying metrics thread [Mod: updated season]

[Mortimer Snerd]

Everyone is subjectively choosing their own definition for SC.
HS and Corsica are using their own definitions, just like coaches, TSN, etc. are all subjectively defining their own scoring chance.

Its funny. Its hard to define and yet we all know it when we see it.

The thing is that it is a result of several variables coming together. Shot location is only one. Some of the others are much harder to track. The path the puck takes to get to the shooter affects the positions of the other players. Where is the goalie? Is there a screen? Etc.

Maybe 'Scoring Chance' is a poor choice of words. Every shot attempt is a scoring chance. You can define more 'Dangerous Shots' with some success.

 

[Joe Hallenback]

NHL tracks shot location for their GC apps. You can scrape for that.

Issue is that looking only at certain shots ends up being worse than including all shots. Adjusting is a far more effective tool, which we use with expected goals.

In terms of predicting future performance at either player or team level, Corsi which views all shots as equal still outperforms any scoring chances or high-danger scoring chances model thus far created.

This is a common phenomenon in stats when binning a continuous variable (like a shot being relatively closer than another is relatively better) into a false categorical variable (shots in home plate vs shots outside).

Some more information on that can be found here:
http://biostat.mc.vanderbilt.edu/wiki/Main/CatContinuous


Aside:
I should point out that adjusting for just location of shots, which Corsica's xGoal model does, isn't sufficient enough to outperform Corsi.
DTM's xGoal model does outperform Corsica's xGoal model AND Corsi. It doesn't just use shot location but other variables as well (such as regressed shooter history).
I also want to point out that Brian MacDonald's xGoal model that was from way back in 2008 IIRC did also outperform Corsi, mostly because it is Corsi but weighting goals more heavily than other shots, and then introduces penalty, hits, and faceoff differentials.

Are there models that show were most goals come from? Location based. Would you know if teams are using models like that and showing players these models?

 

[Aavco Cup]

VHAC 2017.....is that something worth attending as a fan of hockey and not a serious stats guy?

 

[Whileee]

Everyone is subjectively choosing their own definition for SC.
HS and Corsica are using their own definitions, just like coaches, TSN, etc. are all subjectively defining their own scoring chance.

Good point. Probably doesn't mean much with marginal differences, but when a team has 65% or more scoring chances, as the Jets did against the Hawks, it's probably an indication that they had more dangerous shots than the opposition. Just watching that game I thought the Jets had more dangerous shots than the Hawks, even though they were close in Corsi and shots on goal.

 

[garret9]

Its funny. Its hard to define and yet we all know it when we see it.

The thing is that it is a result of several variables coming together. Shot location is only one. Some of the others are much harder to track. The path the puck takes to get to the shooter affects the positions of the other players. Where is the goalie? Is there a screen? Etc.

Maybe 'Scoring Chance' is a poor choice of words. Every shot attempt is a scoring chance. You can define more 'Dangerous Shots' with some success.

I disagree. We do not all know when we see it. There are many shots out there some would conclude is a scoring chance that others would not. The definition of a scoring chance is so ambiguous because of this.

Personally, my opinion is this:
All shots are scoring chances, but some are more chancier than others.

As to shot location being only one factor, we must also realize that we are collecting some of the impacts of other factors with the ones we trace. The rush shot having a higher percentage than one from sustained dzone pressure or rebounds having higher shooting percentage is in part due to those other reasons.

A big one is that in hockey most players that are better than average players at A+C+D+E+F+G+H..etc are also usually going to be better than average at Y+Z. Not always true, but there is something to being a better overall player.

Also, accounting for shooter history will also grab a bunch of this as well. Players who consistently finish above their shot location and the other variables do so for those reasons. Stamkos has a finishing factor of about 1.5... after accounting for all the variables we can, he consistently finishes about 1.5x the goals than we expect (the highest over a full sample we wouldn't regress). Those are because of his hardness and accuracy of shot, but also because of things he can make players and goalies do.

One thing we do realize is that with each factor you can additionally account for, every next discovery will adjust xGoal model by a lesser degree on average.

Moving from goal to shot metrics was a huge jump. Moving from Corsi to DTM's xGoals was a decent but far, far smaller jump. The next jump will likely be less as well.

Are there models that show were most goals come from? Location based. Would you know if teams are using models like that and showing players these models?

Goals come from all over. Only time I looked at where goals come from, it is approximately 50% low slot, 25% rest of "home plate area", and 25% rest of rink.

Ignoring 1/2 of the data with HD SC, or 1/4 with SC, is in part why SC models ultimately fail relative to

As an aside, here is how Corsica separates the three shots:

 

[garret9]

VHAC 2017.....is that something worth attending as a fan of hockey and not a serious stats guy?

Do you like the posts typically seen on Hockey Graphs?

It will essentially be something like that, except instead a slide presentation with talking.

The tickets will likely be limited, by the way. After reserved tickets for presenters, media, and members of organizations (scouts, managers, agents, etc.) we might have something like 70-85.

 

[garret9]

Good point. Probably doesn't mean much with marginal differences, but when a team has 65% or more scoring chances, as the Jets did against the Hawks, it's probably an indication that they had more dangerous shots than the opposition. Just watching that game I thought the Jets had more dangerous shots than the Hawks, even though they were close in Corsi and shots on goal.

Personally, I would prefer to never look at scoring chances.

Between Corsi and xGoals you get a much better picture.

 

[pucka lucka]

A weird thing about conventional hockey statistics is the poisoning of the value by arbitrarily excluding valuable information . Take individual shooting %. We should be including all shot attempts to properly measure someone actual showing prowess. Only including those shots where someone actually hits the goal doesn't tell you the whole story.

DTMs xG probably takes this all into account I believe...

 

[garret9]

A weird thing about conventional hockey statistics is the poisoning of the value by arbitrarily excluding valuable information . Take individual shooting %. We should be including all shot attempts to properly measure someone actual showing prowess. Only including those shots where someone actually hits the goal doesn't tell you the whole story.

DTMs xG probably takes this all into account I believe...

DTM's xG includes missed shots, but blocked shots aren't given much strength.
The reasoning why is we cannot adjust blocked shots for shot location, as the NHL tracks where the shot was blocked and not where the shooter was.

 

[Whileee]

Personally, I would prefer to never look at scoring chances.

Between Corsi and xGoals you get a much better picture.

Maybe xGoals. I've seen a few games this year where I thought the Corsi stats skewed the actual quality of play. Jets have faced a few teams that just indiscriminately put pucks toward the net, not even looking for better shot positioning. Good teams tend to have the talent to do a bit of both (Corsi and good shot locations / situatiins).

 

[garret9]

Maybe xGoals. I've seen a few games this year where I thought the Corsi stats skewed the actual quality of play. Jets have faced a few teams that just indiscriminately put pucks toward the net, not even looking for better shot positioning. Good teams tend to have the talent to do a bit of both (Corsi and good shot locations / situatiins).

Reminder:
Corsi has less to do with in-game sample and more about future.

In a game, HDSC or SC will correlate to goals within that game far better. Corsi will correlate to remainder of season goals far better.

Corsica's xGoals will beat HDSC or SC in correlate to goals within that game, but Corsi still better with remainder of season.
DTM's xGoals trumps all of them in all though... well except DTM's XPM is more predictive than xGoals, as XPM is xGoals adjusted for QoT, QoC, coaches, schedule, and then given past season performance as a prior.

 

[Whileee]

Reminder:
Corsi has less to do with in-game sample and more about future.

In a game, HDSC or SC will correlate to goals within that game far better. Corsi will correlate to remainder of season goals far better.

Good reminder.

 

[Gil Fisher]

Garrett, how much access do you have to DTM's xG data? Are you able to present some kind of summary of the Jets' season (trend) by xGF% - DTM vs. Corsica?

 

[Mortimer Snerd]

I disagree. We do not all know when we see it. There are many shots out there some would conclude is a scoring chance that others would not. The definition of a scoring chance is so ambiguous because of this.

Personally, my opinion is this:
All shots are scoring chances, but some are more chancier than others.

As to shot location being only one factor, we must also realize that we are collecting some of the impacts of other factors with the ones we trace. The rush shot having a higher percentage than one from sustained dzone pressure or rebounds having higher shooting percentage is in part due to those other reasons.

A big one is that in hockey most players that are better than average players at A+C+D+E+F+G+H..etc are also usually going to be better than average at Y+Z. Not always true, but there is something to being a better overall player.

Also, accounting for shooter history will also grab a bunch of this as well. Players who consistently finish above their shot location and the other variables do so for those reasons. Stamkos has a finishing factor of about 1.5... after accounting for all the variables we can, he consistently finishes about 1.5x the goals than we expect (the highest over a full sample we wouldn't regress). Those are because of his hardness and accuracy of shot, but also because of things he can make players and goalies do.

One thing we do realize is that with each factor you can additionally account for, every next discovery will adjust xGoal model by a lesser degree on average.

Moving from goal to shot metrics was a huge jump. Moving from Corsi to DTM's xGoals was a decent but far, far smaller jump. The next jump will likely be less as well.



Goals come from all over. Only time I looked at where goals come from, it is approximately 50% low slot, 25% rest of "home plate area", and 25% rest of rink.

Ignoring 1/2 of the data with HD SC, or 1/4 with SC, is in part why SC models ultimately fail relative to

As an aside, here is how Corsica separates the three shots:

Point taken. There is a substantial bin that we would largely agree on if we are all sitting together watching the game in your living room (drinking your beer ) but outside that bin the agreement might not be as clear.

I see what I consider great SC's that sometimes don't even get a shot attempt so maybe we disagree about them too.

Edit: Is that chart limited to SOG or does it use SA?

It is no surprise that the area right in front of the net is a HD area but there is often a scrum there and the offensive player is prevented from getting a shot away. I don't know how many shot attempts would be blocked but a high % of unblocked attempts should score from there. But how many attempts are prevented from ever happening and how does that affect the appraisal of danger?

 

[Mud Turtle]

There's been a lot of discussion around Dustin Byfuglien's defensive game this year. Does anyone have stats available that show how many goals he's been on the ice for and also his average minutes per game. That should give us an idea if he's on the ice for more goals against that the other D men.

 

[garret9]

Point taken. There is a substantial bin that we would largely agree on if we are all sitting together watching the game in your living room (drinking your beer ) but outside that bin the agreement might not be as clear.

I see what I consider great SC's that sometimes don't even get a shot attempt so maybe we disagree about them too.

Edit: Is that chart limited to SOG or does it use SA?

It is no surprise that the area right in front of the net is a HD area but there is often a scrum there and the offensive player is prevented from getting a shot away. I don't know how many shot attempts would be blocked but a high % of unblocked attempts should score from there. But how many attempts are prevented from ever happening and how does that affect the appraisal of danger?

Unblocked shots. Can't do location for blocked shots:

DTM's xG includes missed shots, but blocked shots aren't given much strength.
The reasoning why is we cannot adjust blocked shots for shot location, as the NHL tracks where the shot was blocked and not where the shooter was.

 

[garret9]

There's been a lot of discussion around Dustin Byfuglien's defensive game this year. Does anyone have stats available that show how many goals he's been on the ice for and also his average minutes per game. That should give us an idea if he's on the ice for more goals against that the other D men.

I'd prefer this to goals against:
http://www.winnipegsun.com/2016/12/21/positively-and-negatively-jets-byfuglien-always-has-impact

As to his effectiveness relative to icetime:

 

[Gil Fisher]

 

[Gil Fisher]

xGF% (Corsica) update:

Little 62.7%
Wheeler 60.0%
Petan 58.6%
Lowry 54.4%
Dano 54.0%
Trouba 53.5%
Copp 53.5%
Enstrom 52.7%
Byfuglien 52.0%
Ehlers 50.7%
Morrissey 50.4%
Scheifele 49.8%
Armia 49.5%
Stafford 48.4%
Laine 48.5%

 

[garret9]

DTM's model of xGoals, so player regressed finishing talent is used.
Method in regression is found here:
https://hockey-graphs.com/2015/10/0...predictor-of-future-scoring-than-corsi-goals/

 

[garret9]

 

[Aavco Cup]

Well the Leafs certainly use Corsi....what exactly is a heavy shift? Sounds "fancy"

 

[garret9]

Love the fact that they capitalize it to CORSI to make fun of Simmons.

 

[Aavco Cup]

 

[Whileee]

Well the Leafs certainly use Corsi....what exactly is a heavy shift? Sounds "fancy"

Looks like they use shot blocks, too.