The Advanced Stats Thread Episode IV: A New Hope For Advanced Stats

[Crease]

'fish, what goes into xG? Is it location of the shot? Who takes it? Where other guys are when the shot takes off?

 

[silverfish]

'fish, what goes into xG? Is it location of the shot? Who takes it? Where other guys are when the shot takes off?

Shot type (Wrist shot, slap shot, deflection, etc.)
Shot distance (Adjusted4 distance from net)
Shot angle (Angle in absolute degrees from the central line normal to the goal line)
Rebounds (Boolean – Whether or not the shot was a rebound)
Rush shots (Boolean – Whether or not the shot was a rush shot)
Strength state (Boolean – Whether or not the shot was taken on the powerplay)5

http://www.corsica.hockey/blog/2016/03/03/shot-quality-and-expected-goals-part-i/

 

[Zil]

'fish, what goes into xG? Is it location of the shot? Who takes it? Where other guys are when the shot takes off?

https://hockey-graphs.com/2015/10/0...predictor-of-future-scoring-than-corsi-goals/

http://www.corsica.hockey/blog/2016/03/03/shot-quality-and-expected-goals-part-i/

 

[Ail]

Ola, that is my only issue with the "statistical analysis" trend (if you want to call it that - 90% of this board engages in no more than simple counting.) It completely ignores context or any sort of nuanced understand of the real world hockey x's and o's that drive these counts.

NYR these last four games without Girardi & Klein are generating 42% of shot attempts, 39% of scoring chances, and 40% of high danger chances. Down from 48%, 49%, and 51% respectively.

I wouldn't use that to infer that Girardi & Klein are particularly effective players. Nor would I wretch and moan all season about the need to replace G/Klein with "warm bodies off the street" as we have seen in every 4th post in every PGT. The narcissitic mentality of "how can all these old boyz not know how to count like me??" is insufferable.

Speaking of statistical analysis I'm curious to see how you arrived at your 90% figure.

 

[silverfish]

This is a good read on some of the black box stuff we can't scrape from NHL PBP data:

https://www.nhl.com/news/john-chayka-building-coyotes-by-the-numbers/c-287424100

That's what happened with Dvorak. He was showing signs of becoming an impactful player because his rates of puck touches per shift, per minute and per 60 minutes were rising. Puck touches is the key stat Chayka uses for Dvorak because the center is expected to be involved in the play all over the ice.

"As the season wore on he was becoming more and more involved," Chayka said. "As that continued to evolve his ability to create offense, create scoring chances for both himself and his teammates, getting second and third opportunities in the offensive zone, have continued to stand out. From where he was to where he is now is a significant improvement in that area."

Similarly, albeit in a negative way, the underlying numbers told Chayka enough about forward Anthony Duclair to send him to Tucson of the AHL on Jan. 19.

Duclair, who scored 20 goals in 81 games last season, has three goals in 42 games this season. The lack of production wasn't as alarming as the lack of scoring chances, the one statistical measure that fits Duclair's game best, Chayka said.

"He was extremely reliant on other players to generate, which has its benefits when you're playing with elite players, especially elite centermen, but when you're not what are you bringing to the table on a consistent basis?" Chayka said. "If he's not acquiring pucks and hanging out around the slot looking for passes, that's not ideal."

The bolded is very similar to what I've been saying about Vesey.

 

[JimmyG89]

Every. Single. Regular. Poster. In. This. Thread. Understands. That. Context. Matters. And. Doesn't. Only. Use. Stats. To. Justify. Their. Positions. But. Uses. Stats. To. Have. More. Detailed. Opinions. And. Inferences. And. Analysis.

C'mon, you should know by now that none of us watch the games, follow it on hockeystats.ca, read the charts and infer the X player sucks and Y player is amazing.

Here's the thing, I don't need a chart to tell me who is bad. I need a chart to confirm that someone is bad based on what I saw.

 

[Raspewtin]

C'mon, you should know by now that none of us watch the games, follow it on hockeystats.ca, read the charts and infer the X player sucks and Y player is amazing.

Here's the thing, I don't need a chart to tell me who is bad. I need a chart to confirm that someone is bad based on what I saw.

I actually haven't watched a game since 2012. I move pieces on a checker board that correspond with the events that take place on hockeystats.ca (not in real life) to tell me how to make decisions. It's really intricate.

 

[silverfish]

C'mon, you should know by now that none of us watch the games, follow it on hockeystats.ca, read the charts and infer the X player sucks and Y player is amazing.

Here's the thing, I don't need a chart to tell me who is bad. I need a chart to confirm that someone is bad based on what I saw.

Yes, exactly. Or if the chart "tells you" that someone is bad that you thought was good, then you can re-evaluate your position! What did I see that the chart missed? What did I miss that the chart saw?

THEN, you have this ridiculously data + eye-test driven opinion. What a concept!

 

[JimmyG89]

Yes, exactly. Or if the chart "tells you" that someone is bad that you thought was good, then you can re-evaluate your position! What did I see that the chart missed? What did I miss that the chart saw?

THEN, you have this ridiculously data + eye-test driven opinion. What a concept!

Well, my eye test isn't wrong, so I never have to change my position on a player

 

[SnowblindNYR]

So question regarding the xG numbers. Is it possible for the number to be 0? Or since literally EVERY shot has a >0% chance of going in, there can't ever be a game with an xG of 0? If that's the case aren't the xG numbers inflated since it's perfectly possible for a team to get a shutout due to good defense? Also in that case shouldn't xG be adjusted somehow?

 

[silverfish]

So question regarding the xG numbers. Is it possible for the number to be 0? Or since literally EVERY shot has a >0% chance of going in, there can't ever be a game with an xG of 0? If that's the case aren't the xG numbers inflated since it's perfectly possible for a team to get a shutout due to good defense? Also in that case shouldn't xG be adjusted somehow?

xG has nothing to do with actual goals, which is where it has its usefulness. You can fully eliminate the play of the goalie with this metric to evaluate your defense (or offense). Or, you can evaluate your goalie based off their actual goal total vs the model-built expected total against.

Side-conversation: Offense has dried up, as predicted by the "corsi nerds". No one seems to realize that this **** matters. Baffles me.

Miller has three goals since the start of Feb.
Vesey has two.
Grabner has four (what a freak)
Hayes has two
Stepan has zero

If only we could've seen this coming!

 

[SnowblindNYR]

xG has nothing to do with actual goals, which is where it has its usefulness. You can fully eliminate the play of the goalie with this metric to evaluate your defense (or offense). Or, you can evaluate your goalie based off their actual goal total vs the model-built expected total against.

Side-conversation: Offense has dried up, as predicted by the "corsi nerds". No one seems to realize that this **** matters. Baffles me.

Miller has three goals since the start of Feb.
Vesey has two.
Grabner has four (what a freak)
Hayes has two
Stepan has zero

If only we could've seen this coming!

So it's not a predictor of actual goals? Seems like it shouldn't be called expected goals then.

 

[Irishguy42]

So it's not a predictor of actual goals? Seems like it shouldn't be called expected goals then.

It has nothing to do with actual goals because it's not looking to present actual results. Nor is it tangible like Corsi and whatnot.

It's a model made to show expected goals based on real data.

Basically used for simulations and predictive modeling.

It's saying "Here's all this data, and if we do some complex math, we think Team A will score X goals and Team B will score Y"

Run that several hundred times, and see what the common results are. Over the course of the season, you get to see how your xG model matches up.

 

[SnowblindNYR]

It has nothing to do with actual goals because it's not looking to present actual results. Nor is it tangible like Corsi and whatnot.

It's a model made to show expected goals based on real data.

Basically used for simulations and predictive modeling.

It's saying "Here's all this data, and if we do some complex math, we think Team A will score X goals and Team B will score Y"

Run that several hundred times, and see what the common results are. Over the course of the season, you get to see how your xG model matches up.

Well I never thought it was anything but what you described. But is it possible to have 0 xG? I mean isn't it in part formed from the percentage of goals scored given a shot from a certain place on the ice. Since that will never be 0%, can xG be 0?

 

[Irishguy42]

Well I never thought it was anything but what you described. But is it possible to have 0 xG? I mean isn't it in part formed from the percentage of goals scored given a shot from a certain place on the ice. Since that will never be 0%, can xG be 0?

Interesting question. Not something I ever considered, tbh.

Wonder if DTMAboutHeart/Manny/someone else would be better to ask that.

SF, what do you think?

My very very very rudimentary knowledge of advanced stats would say that it's possible in theory/computing, but it's very unlikely in practice. But then again...

 

[SnowblindNYR]

Interesting question. Not something I ever considered, tbh.

Wonder if DTMAboutHeart/Manny/someone else would be better to ask that.

SF, what do you think?

My very very very rudimentary knowledge of advanced stats would say that it's possible in theory/computing, but it's very unlikely in practice. But then again...

Yeah, so the question is if no shot has 0% chance of going in and if my understanding is right that xG is calculated in part by percentage of goals allowed from a shot, then can xG ever be 0 if there are more than 0 shots by a team. If xG can't be 0, shouldn't the numbers be adjusted? If say 0.25 is the lowest that a team can get in xG (just an examples the number isn't important), isn't then every number likely inflated? Say your xG is 1, shouldn't you do something like 1/1.25?

 

[Machinehead]

Ok but Weber still sucks right?

 

[Ola]

Watching Florida they definitely seem like a team that plays a style that is a bit too focused on possession over quality of possession. Possession and momentum is very important. Nobody can survive those prolonged shifts in your own end. But its very hard to come back from being behind in this league, goals still matters a lot.

If you look at where goals are coming from in this league, it would be very obvious that just circling around in the attacking zone will not result in too many goals. So many goals comes after tempo changes, the attacking team coming in on a defense trying to organize itself and then overloading with a D joining the play. If you want max possession, take your time through the neutral zone. But it will also mean that you don't face a D scrambling to get organized very often.

Florida is an example of this. I mentioned the other day that LAK is another. Boston under Julien. Anaheim. FLA is kinda built on false premises, but the others (LAK, BOS, ANA) is kind of stuck in what worked yesterday. Its in this area the biggest adjustment have been made in the NHL the last 12-24 month. Quenville commented on it specifically. I posted the quote somewhere a few weeks ago. Don't remember exactly what he said, but it supported this. A good coach can adjust. Q have adjusted. Sutter haven't. You score goals by going in with speed against a D, challenge it, which will get them disorganized, then just move the puck to a supporting forward after the D have engaged in the 1 on 1 situation, have a D join the play. You have 4 players attacking, the D is disorganized. This is what we and Chicago among a few others really are focusing on, creating these situations. It will cost some possession but it will also create more goals. Go to nhl.com and check like our latest 50 goals. How many of them comes from plays like this? Its a significant portion.

Is goals scored of shot-attempts per TEAM available anywhere? Who are scoring goals and who are getting stuck going up against a set defense getting shots blocked or saved? Think that would be an interesting indicator to look at right now. Is % of shots blocked available anywhere?

Ok but Weber still sucks right?

He is flawed more than overrated. Would probably struggle quite much on a lesser team. Just like Seabrook. Good on a good team? Think that is fair to say. But just like Phaneuf on the Leafs, I think like Weber or Seabrook on Colorado or a team like that would be pretty messy.

 

[silverfish]

Well I never thought it was anything but what you described. But is it possible to have 0 xG? I mean isn't it in part formed from the percentage of goals scored given a shot from a certain place on the ice. Since that will never be 0%, can xG be 0?

Interesting question. Not something I ever considered, tbh.

Wonder if DTMAboutHeart/Manny/someone else would be better to ask that.

SF, what do you think?

My very very very rudimentary knowledge of advanced stats would say that it's possible in theory/computing, but it's very unlikely in practice. But then again...

xG is built off a regression. I don't believe there is any instance where a shot attempt's inputs would release an output of 0 or negative, but I can't say for a fact, because the coefficients for these models are not public.

My understanding is, as long as there is an unblocked shot attempt, xG will be > 0.

Yeah, so the question is if no shot has 0% chance of going in and if my understanding is right that xG is calculated in part by percentage of goals allowed from a shot, then can xG ever be 0 if there are more than 0 shots by a team. If xG can't be 0, shouldn't the numbers be adjusted? If say 0.25 is the lowest that a team can get in xG (just an examples the number isn't important), isn't then every number likely inflated? Say your xG is 1, shouldn't you do something like 1/1.25?

I believe what you're referring to here is a "ratio" metric of sorts, which Corsica has. expected fenwick shooting percentage is xGF/FF and expected fenwick save percentage is 1-(xGA/FA).

 

[Machinehead]

He is flawed more than overrated. Would probably struggle quite much on a lesser team. Just like Seabrook. Good on a good team? Think that is fair to say. But just like Phaneuf on the Leafs, I think like Weber or Seabrook on Colorado or a team like that would be pretty messy.

So....he sucks

 

[SnowblindNYR]

xG is built off a regression. I don't believe there is any instance where a shot attempt's inputs would release an output of 0 or negative, but I can't say for a fact, because the coefficients for these models are not public.

My understanding is, as long as there is an unblocked shot attempt, xG will be > 0.



I believe what you're referring to here is a "ratio" metric of sorts, which Corsica has. expected fenwick shooting percentage is xGF/FF and expected fenwick save percentage is 1-(xGA/FA).

I don't think that's what I meant. I meant that if xG at even 1 shot from the lowest percentage area is >0 then doesn't it stand to reason that all of the numbers are bumped up? In theory xG should at least have some chance of being equal to goals I would think. If the lowest it could be is more than 0, doesn't it bump up all of the numbers since actual goals start from 0?

 

[silverfish]

I don't think that's what I meant. I meant that if xG at even 1 shot from the lowest percentage area is >0 then doesn't it stand to reason that all of the numbers are bumped up? In theory xG should at least have some chance of being equal to goals I would think. If the lowest it could be is more than 0, doesn't it bump up all of the numbers since actual goals start from 0?

EDIT - When you start editing a post and then you get dragged onto a call at work that lasts an hour and then Manny just comes in and settles everything.

 

[SnowblindNYR]

I don't follow

editing....

Let's say this is the regression:

Dependent variable: xG

Independent variable: Shots quantity from a particular part of the ice

Coefficient: 0.2
Intercept: 0.05

At 0 shots, the xG would 0.05. So, I was wrong, unless the intercept is zero, xG wouldn't even be 0 with 0 shots. So if the lowest estimate you could have is 0.05. But the lowest actual is 0. Aren't the estimates inflated since 0 shots would result in 0 goals in actuality? So shouldn't you adjust the xG because the estimate will always be off and inflated?

 

[eperry]

I don't think that's what I meant. I meant that if xG at even 1 shot from the lowest percentage area is >0 then doesn't it stand to reason that all of the numbers are bumped up? In theory xG should at least have some chance of being equal to goals I would think. If the lowest it could be is more than 0, doesn't it bump up all of the numbers since actual goals start from 0?

The only way to achieve exactly zero expected goals is to take zero shots. Otherwise, you'd expect at least a fraction of a goal, however infinitesimally small, based on the minute likelihood of a goal from any area on the ice.

You can roll a die without landing on 3. But, if there was such a thing as "expected threes" you would assign every roll a value of ~0.167 based on the probability. Roll four times and you might see a drastic difference in actual threes and expected. Roll a few thousand times and it normalizes.

It comes down to whether you believe shots can have a goal probability of zero. Fundamentally, I don't think they can. In practice, the average xG of neutral zone shots (0.00648) is so small that it's unreasonable to expect any skew.

 

[silverfish]

Blerg.

Getting there...

Things I still need to solve:

1. If no player is selected, show all players
2. Option to show All teams at the same time but still have the filter option
3. Option to show all positions at the same time but still have the filter option

I have an 'All' filter for strength and score state, but those are built into the data...

Close.