[Game Takes] Flames 4 Predators 3

[Bingo]

Flames 4 Predators 3

- Fourth line with two goals
- Flames erase three Predator leads
- 25 game streak without 4 goals snapped

[DeluxeMoustache]

This is where xGA gets silly.

Perfect pass by Stamkos and perfect tip. xGA contribution is probably about .2 but a shot placed there is pretty much going in 9 times out of 10 at least

Horrible defensive coverage leaves Novak all alone for a cross crease pass which is promptly shot in off the inside of the post. Shots from a similar situation, in that statistical bucket, include thousands of shots , many off the goalie’s pads, into his logo, etc. This one was off the inside of the far post. Shot placement isn’t considered in xGA but a shot placed there goes in closer to 9 times out of 10 than twice out of 10.

Similarly with Marchessault’s goal. Perfectly executed

I’d give those 3 shots alone about 2.5 xGA, based on execution, and shot placement, never mind the contribution from the other 17 shots


Impressive showing from Vladar, especially quite solid in the last minutes of the 3rd under pressure

Loved the resilience. Huberdeau and Kadri are looking good together

Great snipe by Pachal, a pleasant surprise

Fun game

[Jay Random]

Shot placement isn't considered in xGA because that stat is meant to measure skater vs. skater play, rather than shooter vs. goalie. Good shooters will routinely outperform their xGF, good goalies will routinely outperform their xGA.

I don't complain about my spoon because it doesn't cut beefsteak.

[DeluxeMoustache]

Originally Posted by Jay Random

Shot placement isn't considered in xGA because that stat is meant to measure skater vs. skater play, rather than shooter vs. goalie. Good shooters will routinely outperform their xGF, good goalies will routinely outperform their xGA.

I don't complain about my spoon because it doesn't cut beefsteak.

Absolutely

You should complain about your steak

(True story - I went to a restaurant in Argentina once where the waiter cut the steak with a spoon)

[edslunch]

Originally Posted by Jay Random

Shot placement isn't considered in xGA because that stat is meant to measure skater vs. skater play, rather than shooter vs. goalie. Good shooters will routinely outperform their xGF, good goalies will routinely outperform their xGA.

I don't complain about my spoon because it doesn't cut beefsteak.

So xGF and xGA shouldn't be compared to GF and GA?

[Bingo]

Originally Posted by edslunch

So xGF and xGA shouldn't be compared to GF and GA?

I think it forms the basis of what the average NHL skater would do with a shot in that vicinity arriving under a specific circumstance.

But doesn't allow for the quality of the shooter, or the position of the goaltender.

Over a big enough sample size it evens out.

Smaller sample sizes would need to be viewed through that lens.

[The Cobra]

Originally Posted by DeluxeMoustache

This is where xGA gets silly.

Perfect pass by Stamkos and perfect tip. xGA contribution is probably about .2 but a shot placed there is pretty much going in 9 times out of 10 at least

Horrible defensive coverage leaves Novak all alone for a cross crease pass which is promptly shot in off the inside of the post. Shots from a similar situation, in that statistical bucket, include thousands of shots , many off the goalie’s pads, into his logo, etc. This one was off the inside of the far post. Shot placement isn’t considered in xGA but a shot placed there goes in closer to 9 times out of 10 than twice out of 10.

Similarly with Marchessault’s goal. Perfectly executed

I’d give those 3 shots alone about 2.5 xGA, based on execution, and shot placement, never mind the contribution from the other 17 shots

Thanks for that explanation. I didn't know this, although I've always known that this was not reliable over a small sample size, and you've explained why.

But over a large sample size, I like it. Just not sure how large a sample size we really need.

[Enoch Root]

Originally Posted by Bingo

I think it forms the basis of what the average NHL skater would do with a shot in that vicinity arriving under a specific circumstance.

But doesn't allow for the quality of the shooter, or the position of the goaltender.

Over a big enough sample size it evens out.

Smaller sample sizes would need to be viewed through that lens.

Somewhat.

And maybe even mostly.

But some teams are going to allow the shooters more time, more space, and more opportunity to make plays, than other teams do. So it is never going to even out. And the problem is: we don't know the degree of that problem so we can't know how close it will, or won't, come to evening out. And that makes it very difficult to draw meaningful conclusions from the numbers.

And yet every day, we see people take these stats from individual games, and apply them like they are what's what.

[GioforPM]

Originally Posted by Enoch Root

Somewhat.

And maybe even mostly.

But some teams are going to allow the shooters more time, more space, and more opportunity to make plays, than other teams do. So it is never going to even out. And the problem is: we don't know the degree of that problem so we can't know how close it will, or won't, come to evening out. And that makes it very difficult to draw meaningful conclusions from the numbers.

And yet every day, we see people take these stats from individual games, and apply them like they are what's what.

It's even tougher for goalies and fancy stats, even with, say, a season long sample. Only one other goalie plays under roughly the same conditions as Vladar over a season, and that's Wolf. All other goalies face the same shooters more or less. BUt they don't all play behind the same team defence as the Flames' goalies.

More than any other position, I think you need to rely on eye test for goalies, but over a long period of time.

[Bingo]

Originally Posted by Enoch Root

Somewhat.

And maybe even mostly.

But some teams are going to allow the shooters more time, more space, and more opportunity to make plays, than other teams do. So it is never going to even out. And the problem is: we don't know the degree of that problem so we can't know how close it will, or won't, come to evening out. And that makes it very difficult to draw meaningful conclusions from the numbers.

And yet every day, we see people take these stats from individual games, and apply them like they are what's what.

That's where you lose me.

The stats work in a large sample size.

But they also work in a one game sample size if that's the extent of the conclusion.

You shouldn't take 11 minutes of five on five ice time, and a terrible xGF% split and say the player sucks.

But if you want to say he had a tough night that's valid. If the opposition generates more of league average bad events against you than you generate for you it's likely a tough night. There are nuances for sure, but the objective assignment of expected goals in both directions is from a model. And that model would suggest a tough night if the other team had more instances than you did by a large margin; even in just one game.

[Bingo]

Originally Posted by GioforPM

It's even tougher for goalies and fancy stats, even with, say, a season long sample. Only one other goalie plays under roughly the same conditions as Vladar over a season, and that's Wolf. All other goalies face the same shooters more or less. BUt they don't all play behind the same team defence as the Flames' goalies.

More than any other position, I think you need to rely on eye test for goalies, but over a long period of time.

Not sure I agree.

That's taken into account by expected goals against.

It could get better, but it's objective.

If Calgary's team defense (or defensemen) are worse than the rest of the league, the Calgary goaltenders would have their expected goals against rise, meaning they could give up more actual goals and still be flat in goals saved above average.

[GioforPM]

Originally Posted by Bingo

Not sure I agree.

That's taken into account by expected goals against.

It could get better, but it's objective.

If Calgary's team defense (or defensemen) are worse than the rest of the league, the Calgary goaltenders would have their expected goals against rise, meaning they could give up more actual goals and still be flat in goals saved above average.

It depends on what the particular defence allows because the stat doesn’t account for quality of shot and shooter. A good D may allow a lot of shots and shot attempts, even in the “good” locations but in reality the shots might not be as dangerous. Or they can be like Calgary two years ago and stifle shots unless they are breakaways and two on ones to the opposition’s best players.

[Bingo]

Originally Posted by GioforPM

It depends on what the particular defence allows because the stat doesn’t account for quality of shot and shooter. A good D may allow a lot of shots and shot attempts, even in the “good” locations but in reality the shots might not be as dangerous. Or they can be like Calgary two years ago and stifle shots unless they are breakaways and two on ones to the opposition’s best players.

Expected goals models account for location and circumstance.

If you give up shots from the home plate off of rebounds, deflections and passes that cross the royal road you get a high danger event and a greater % chance of a goal. A shot from the same location without the deflection, rebound or pass are just a scoring chance and are assigned a lower goal value. Outside the home plate than moves down accordingly with distance and angle.

So it definitely takes into account what a team defense gives up.

The quality of the shooter is unique and can't be accounted for, but it's not like the Flames defenseman are playing in a different league. If one division has all sputtering offences you may see a weighting issue on a season for sure, but I'm not sure that's the case.

[GioforPM]

Originally Posted by Bingo

Expected goals models account for location and circumstance.

If you give up shots from the home plate off of rebounds, deflections and passes that cross the royal road you get a high danger event and a greater % chance of a goal. A shot from the same location without the deflection, rebound or pass are just a scoring chance and are assigned a lower goal value. Outside the home plate than moves down accordingly with distance and angle.

So it definitely takes into account what a team defense gives up.

The quality of the shooter is unique and can't be accounted for, but it's not like the Flames defenseman are playing in a different league. If one division has all sputtering offences you may see a weighting issue on a season for sure, but I'm not sure that's the case.

I’ve never been convinced that circumstance has or even can be accounted for with accuracy. The point is that the Flames D doesn’t play against different players. But they may allow different things. Other teams’ D have different players with different strengths. So the most accurate comparison for Wolf is Vladar and not Skinner.

[Bingo]

Originally Posted by GioforPM

I’ve never been convinced that circumstance has or even can be accounted for with accuracy. The point is that the Flames D doesn’t play against different players. But they may allow different things. Other teams’ D have different players with different strengths. So the most accurate comparison for Wolf is Vladar and not Skinner.

Yeah I have no idea what the accuracy level is ... but I'm pretty sure the grey area can be assigned to all teams equally.

So the model can improve (and will), but over a season I see the sample size enough to judge what a team gives up as they are living in the same model equally.

I can't think of a way a defense core could play worse than what those numbers suggest consistently in order to wreck the summary.

[Enoch Root]

Originally Posted by Bingo

Not sure I agree.

That's taken into account by expected goals against.

It could get better, but it's objective.

If Calgary's team defense (or defensemen) are worse than the rest of the league, the Calgary goaltenders would have their expected goals against rise, meaning they could give up more actual goals and still be flat in goals saved above average.

This is where I think you're wrong, and is exactly what we're arguing against - the stat takes into account location and situation (to an extent), but it cannot take into account the quality of the situation. Team defense can be weak by being poor positionally (which would get captured somewhat), but they can also be weak by being slow, or simply allowing shooters too much time. Same shooter, same place, same 'circumstance', but if the shooter has more time, they are getting off a better quality shot. The stats do not capture this (and other similar things).

The fundamental question is: does this all average out, or are some teams more prone to it? And I don't think there's any question that some teams are more prone to it.

[Bingo]

Originally Posted by Enoch Root

This is where I think you're wrong, and is exactly what we're arguing against - the stat takes into account location and situation (to an extent), but it cannot take into account the quality of the situation. Team defense can be weak by being poor positionally (which would get captured somewhat), but they can also be weak by being slow, or simply allowing shooters too much time. Same shooter, same place, same 'circumstance', but if the shooter has more time, they are getting off a better quality shot. The stats do not capture this (and other similar things).

The fundamental question is: does this all average out, or are some teams more prone to it? And I don't think there's any question that some teams are more prone to it.

I'm suggesting they play the same sport in the same fashion with a different level of skill.

Not sure the model would fail to capture those affects in a counting exercise.

So I wouldn't say I'm "wrong"

I do admit that a unique situation is always different, but I just don't think one team can find a way to measure higher in a model consistently than their peers to change the outcome.

That seems like a huge stretch to me.

You'd have to be bad in way that shows counts aren't that bad for bad events, while also making the bad events even worse, and you'd have to find a way to do that consistently.

[Enoch Root]

Originally Posted by Bingo

That's where you lose me.

The stats work in a large sample size.

But they also work in a one game sample size if that's the extent of the conclusion.

You shouldn't take 11 minutes of five on five ice time, and a terrible xGF% split and say the player sucks.

But if you want to say he had a tough night that's valid. If the opposition generates more of league average bad events against you than you generate for you it's likely a tough night. There are nuances for sure, but the objective assignment of expected goals in both directions is from a model. And that model would suggest a tough night if the other team had more instances than you did by a large margin; even in just one game.

And we'll just have to agree to disagree on the small sample size issue - hockey is much more subject to random bounces than most sports - weird shaped ball, ice, skates, sticks... it all adds up to lots of bounces/luck. And that equates to there being far too much noise in small sample sizes for them to be useful.

You take a sport like baseball... every play is isolated, and started fresh. We can measure a pitcher's velocity, spin rate, release angle, etc. And all of those stats are essentially as valid in small sample sizes as they are in large samples. Same for the hitter - how did they fare against fast balls, sliders. inside, outside? All measurable, and all essentially equally measurable from game to game.

Hockey isn't like that at all. In order to compare shot statistics, you need large samples. Any random player has a one-timer - it could be 60 MPH, or it could be 95. It all depends on how the puck laid down for his stick. Then it might hit the net, it might not, again, largely depends on the positioning of the puck on the ice at contact. So much randomness.

The other factor is the one-game samples are too small. Sometimes we see a player at 25% (or whatever) for expected goals for and against. Sounds like he got caved in. But in reality, he was probably on the ice for only a very few chances either way. And maybe one of his linemates should have shot once, or had a puck bounce over their stick. That one play would have completely changed the player's xGF/A number. One turnover by another player, and the xGF/A gets hammered. Larger sample sizes, these things get averaged out. But in a single game - especially for a low-event team, and even more so for a very low event fourth line, the total scoring chances might be 2-1 for the game. That's a 33% result, where a single bounce would have made it 50%. A single bounce each way, and it's 66%.

All this would be fine (is what it is), if everyone understood that. But they don't. Every game we see "this person had a great game, that person got caved in"

[Enoch Root]

Originally Posted by Bingo

I'm suggesting they play the same sport in the same fashion with a different level of skill.

Not sure the model would fail to capture those affects in a counting exercise.

So I wouldn't say I'm "wrong"

I do admit that a unique situation is always different, but I just don't think one team can find a way to measure higher in a model consistently than their peers to change the outcome.

That seems like a huge stretch to me.

You'd have to be bad in way that shows counts aren't that bad for bad events, while also making the bad events even worse, and you'd have to find a way to do that consistently.

Well, the standings suggest some teams are bad consistently.

When we see bad records with high goals against, but the stats say their goals against shouldn't have been that high, there are two possible reasons: one, it's just been luck so far, and it should even out, or two, the stat isn't capturing what's happening.

Both are possible, but to suggest the latter isn't likely, is where we are going to disagree.

[Bingo]

Originally Posted by Enoch Root

And we'll just have to agree to disagree on the small sample size issue - hockey is much more subject to random bounces than most sports - weird shaped ball, ice, skates, sticks... it all adds up to lots of bounces/luck. And that equates to there being far too much noise in small sample sizes for them to be useful.

You take a sport like baseball... every play is isolated, and started fresh. We can measure a pitcher's velocity, spin rate, release angle, etc. And all of those stats are essentially as valid in small sample sizes as they are in large samples. Same for the hitter - how did they fare against fast balls, sliders. inside, outside? All measurable, and all essentially equally measurable from game to game.

Hockey isn't like that at all. In order to compare shot statistics, you need large samples. Any random player has a one-timer - it could be 60 MPH, or it could be 95. It all depends on how the puck laid down for his stick. Then it might hit the net, it might not, again, largely depends on the positioning of the puck on the ice at contact. So much randomness.

The other factor is the one-game samples are too small. Sometimes we see a player at 25% (or whatever) for expected goals for and against. Sounds like he got caved in. But in reality, he was probably on the ice for only a very few chances either way. And maybe one of his linemates should have shot once, or had a puck bounce over their stick. That one play would have completely changed the player's xGF/A number. One turnover by another player, and the xGF/A gets hammered. Larger sample sizes, these things get averaged out. But in a single game - especially for a low-event team, and even more so for a very low event fourth line, the total scoring chances might be 2-1 for the game. That's a 33% result, where a single bounce would have made it 50%. A single bounce each way, and it's 66%.

All this would be fine (is what it is), if everyone understood that. But they don't. Every game we see "this person had a great game, that person got caved in"

I don't think we are too far apart on this.

A fourth liner that almost never generates any offence shouldn't be analyzed by a single game xGF%. Totally agree there.

But if a player in the top six has a high danger count of 1 for and 9 against in a hockey game I feel pretty comfortable suggesting he had a rough game.