All Purpose Analytics and Extended Stats Discussion

[BobRouse]

So the analytics community is pretty much not sold on Face offs meaning much correct?

For some reason to me it does. Especially late in games and on special teams.

Or am I way off?

 

[RandyHolt]

You're way way off Rouse get with the program, man!

Is it accurate to say that the AC is not sold on our home goal scoring prowess meaning anything significant WRT to our future playoff fortunes, and specifically the most significant part is that its unsustainable, and thus should be largely disregarded?

I understand the reasoning, but the AC is still kind of new at this, and the game is evolving. The Pens once laughable D quickly became baby penguin puck movers that quietly drove them to a cup which I think teams now try to mirror (unknown delta on stats), and more recently, we have mandated downsizing of goalie pants.

I don't know how it factors in (read: kills stats) but the league also loves parity. IIRC we didn't get a PP until late last night. They don't want every game to be a blowout and will try their darndest to give you a 4-3 tight back and forth game that goes to shootout to further showcase the stars. Final Corsi 50.1 to 49.9. They want 5 teams vying for that final playoff spot on the last night of the season and manage games to deliver the best product.

 

[BobRouse]

I'm trying RH. Us curmudgeon old schoolers need time to catch up and acclimate to the AC club.

So Faceoffs are meh. Hits are bad. Zone starts don't really effect Corsi much.

Expected Goals For is my next conquest.

 

[twabby]

So the analytics community is pretty much not sold on Face offs meaning much correct?

For some reason to me it does. Especially late in games and on special teams.

Or am I way off?

Just saw this on Twitter:

http://statsportsconsulting.com/main/wp-content/uploads/FaceoffAnalysis12-12.pdf

As your intuition indicated, it seems like special teams draws are much more important and neutral zone faceoffs are much less important.

Jay Beagle is a +83 at 5v5 (+0.817 goals), and a +28 (+0.685 goals) in special teams situations (almost exclusively shorthanded), for a total of 1.502 goals added in goal differential so far due to faceoffs.

Seeing as Beagle's one of the best in the league and he's only added approximately 1.5 goals due to faceoff prowess (pro-rate it to about 2.25 goals for an entire season), it's tough to claim that it's really that important compared to other aspects of a player's game.

 

[BobRouse]

Well in 3 on 3 OT it seems its going to be a big deal too but probably doesn't add much overall value.

I hate games where we get crushed at the faceoff circle. Its frustrating if we lose draws badly to a team that is a possession team.

However against chip and chase teams it isn't as big a deal.

I have exactly zero stats to back me up. Its just a gut feeling I get when I yell at the screen "Win a damn faceoff!!!"

 

[RandyHolt]

I don't know if its accurate to say, but I feel like faceoffs are less frequent than ever before. Play just flows, they can barely get a TV commercial in some games.

Anything to be seen in Beagle's poor relative corsi? Trotz generally rolls power on power so I am not sure its 100% tough competition.

 

[BobRouse]

I don't know if its accurate to say, but I feel like faceoffs are less frequent than ever before. Play just flows, they can barely get a TV commercial in some games.

Anything to be seen in Beagle's poor relative corsi? Trotz generally rolls power on power so I am not sure its 100% tough competition.

Didn't realize how poor of a possession season Beagle is having this year. Yikes!!

But he is getting 67% dzone starts! That's a career high and just under 10% higher than last year. I find it hard to believe that doesn't have some level of impact on him.

Beagle is hovering around 44% CF.

How are Winnik and Wilson around 48%??

Seems like Beagle goes on as an extra center for many dzone draws without them. At least intuitively that's what seems.

 

[g00n]

There's also the problem with faceoff stats, as I believe tx may have pointed out in the past, that awarding a "win" can be a bit subjective and even meaningless if the puck quickly changes hands.

 

[RandyHolt]

We just don't see as many late game critical drops as we used to. It used to be one icing after another late in games, not too many decades ago, but not anymore. I won't even touch on the work of the crooked linesmen.

Tex is right. In addition, a faceoff win in the D zone, often puts the puck behind the goal line... going backwards lets say, with 3 opposition that were near the dot with a planned forecheck and closing fast. It's not a given the puck is getting out any time soon.

What is the % of a DZone faceoff wins leading to possession in center?

 

[BobRouse]

There's also the problem with faceoff stats, as I believe tx may have pointed out in the past, that awarding a "win" can be a bit subjective and even meaningless if the puck quickly changes hands.

Probably can say that about "scoring chances/high danger chances" too right? And the latter is part of the equation of the xGF stat correct?

The podcast I was listening to the other day said the Caps xGF stat isn't great but that stat is strange in itself as it doesn't account for the kind of shooters a team has.

Ovechkin alone breaks that stat since his scoring range rides beyond the top of the circles.

Faceoffs are also easier to win on the PP and harder to win a man down. I doubt they track those stats right? Also Faceoffs are only about halfway on the guy who takes them.

He could win it pretty good but his winger could drop the ball and get outworked thus resulting in a "loss"

 

[artilector]

So, OPP CF%. Alzner's is 50.6, Orlov's is 49.8. On its face, that seems like a minor difference. But if you look league-wide, it turns out its really not. Amongst the 137 defensemen who've played 750+ minutes this season (so, top 4 defensemen plus the best 3rd pairings), Alzner ranks 13th in OPP CF% while Orlov ranks 112th. If you wanna use OPP GF%, Alzner ranks 8th and Orlov ranks 107th. Going by either of those metrics, you get the same conclusion as with the TOI based ones: Alzner gets some of the toughest usage in the league, while Orlov's getting some of the easiest usage of any top 4 defensemen. Basically, you cannot make the claim that the difference between Alzner and Orlov's usage is insignificant, without simultaneously making that claim that in the entire NHL there is no significant difference between the usage of shutdown and sheltered pairings, and that's a bold claim that runs counter to league-wide dogma.

You keep arguing against claims that I'm not making and going off to argue for different things -- I was specifically talking about Alzner and Niskanen in context of possession, and you want to illustrate a different point by bringing in Orlov and something called "OppCF% rank" in context of usage, which remains a vague term.

I was NOT making the claim that "the difference between Alzner and Orlov's usage is insignificant" -- I don't even know what units of measure you have in mind for general "usage".

[As an aside: you quote rank of OppCF% as indication of usage. But what quantifiable information about strength of opposition can you extract from rank? When I'm talking about possession, I'm talking in terms of Corsi. In terms of what units of measure are you talking about? If there are no units of measure, than the whole discussion about stats is utterly pointless.]

However, I would make a claim that the difference between Alzner's and Niskanen's opposition is relatively insignificant specifically *in terms of possession effects*. The difference is right there in the numbers -- if #2 faces opposition with OppCF% of 50.18, and #27 faces opposition with OppCF% of 50.57, then the difference between their opposition *in terms of possession performance" amounts to 50.57-50.18=0.39 Corsi events (per 100).

Maybe Alzner indeed faces somewhat stronger opponents than Niskanen's in terms of usage (whatever that actually means), but purely *in terms of possession*, this is a miniscule 0.39 difference per 100 corsi events/units. Why is this insignificant (again, in terms of possession)? Because -- Alzner's presence drags Niskanen's own CF% from 53.8 (overall) to 50.5 (with Alzner) -- that's a difference of 3.3 per 100 Corsi events.

So again, the "possession difference" between Alzner & Niskanen's opposition amounts to 0.39 (per 100) Corsi events;
On the other hand, the "possession difference" between Niskanen's performance overall and his performance with Alzner amounts to 3.3 (per 100) Corsi events.
These are both "apples" that can be compared. And we're seeing basically an effect that's an order of magnitude greater (3.3 vs 0.39). Alzner's opposition may be stronger on average in some overall sense (that remains to be quantified, btw), but it is only marginally stronger in terms of possession -- so the fact that (in relative terms) Niskanen's possession drops by a giant amount with Alzner can only be attributed to Alzner, IMO. And as the WOWY graphs indicate "at a glance", the same will be true pretty much for every single Cap. Alzner makes everybody's possession worse -- taking opposition's possession into account.

Note: if you look at what happens to Niskanen with Alzner compared to without (as opposed to total with all players), the results are even worse -- Niskanen's CF% drops from 55.9 to 50.5, that's a 5.4 Corsi event (per 100) drop, as opposed to the 0.39 difference in opposition possession level expressed in Corsi events per 100.

 

[Zoidberg Jesus]

You keep arguing against claims that I'm not making and going off to argue for different things -- I was specifically talking about Alzner and Niskanen in context of possession, and you want to illustrate a different point by bringing in Orlov and something called "OppCF% rank" in context of usage, which remains a vague term.

I was NOT making the claim that "the difference between Alzner and Orlov's usage is insignificant" -- I don't even know what units of measure you have in mind for general "usage".

[As an aside: you quote rank of OppCF% as indication of usage. But what quantifiable information about strength of opposition can you extract from rank? When I'm talking about possession, I'm talking in terms of Corsi. In terms of what units of measure are you talking about? If there are no units of measure, than the whole discussion about stats is utterly pointless.]

However, I would make a claim that the difference between Alzner's and Niskanen's opposition is relatively insignificant specifically *in terms of possession effects*. The difference is right there in the numbers -- if #2 faces opposition with OppCF% of 50.18, and #27 faces opposition with OppCF% of 50.57, then the difference between their opposition *in terms of possession performance" amounts to 50.57-50.18=0.39 Corsi events (per 100).

Maybe Alzner indeed faces somewhat stronger opponents than Niskanen's in terms of usage (whatever that actually means), but purely *in terms of possession*, this is a miniscule 0.39 difference per 100 corsi events/units. Why is this insignificant (again, in terms of possession)? Because -- Alzner's presence drags Niskanen's own CF% from 53.8 (overall) to 50.5 (with Alzner) -- that's a difference of 3.3 per 100 Corsi events.

So again, the "possession difference" between Alzner & Niskanen's opposition amounts to 0.39 (per 100) Corsi events;
On the other hand, the "possession difference" between Niskanen's performance overall and his performance with Alzner amounts to 3.3 (per 100) Corsi events.
These are both "apples" that can be compared. And we're seeing basically an effect that's an order of magnitude greater (3.3 vs 0.39). Alzner's opposition may be stronger on average in some overall sense (that remains to be quantified, btw), but it is only marginally stronger in terms of possession -- so the fact that (in relative terms) Niskanen's possession drops by a giant amount with Alzner can only be attributed to Alzner, IMO. And as the WOWY graphs indicate "at a glance", the same will be true pretty much for every single Cap. Alzner makes everybody's possession worse -- taking opposition's possession into account.

Note: if you look at what happens to Niskanen with Alzner compared to without (as opposed to total with all players), the results are even worse -- Niskanen's CF% drops from 55.9 to 50.5, that's a 5.4 Corsi event (per 100) drop, as opposed to the 0.39 difference in opposition possession level expressed in Corsi events per 100.

In going back and rereading our argument, I think I have misunderstood what you were arguing and gotten a bit of tunnel vision. I blame being in the middle of midterms. Sorry about that.

Anyway, I think your analysis proves that Orlov is a better possession player than Alzner and Niskanen is a better possession player than Carlson, but I don't think it proves that Alzner has dropped off and Niskanen has improved. You mentioned earlier in this debate that "last year Niskanen was 50.3 overall and 50.1 with Alzner, but this year he is 53.0 overall and 50.4 with Alzner." That's disingenuous though, because Niskanen played far fewer minutes away from Alzner last season, which skews the numbers. In fact, he had a 57.5 CF% away from Alzner last year compared to a 50.1 with him, while this year he has a 56.1 away and a 50.4, so the difference has actually shrunk. I don't think there's any proof that Alzner has gotten significantly worse from last year, and that was my point when I started this latest round of debate.

 

[Coldplay619]

Ovi-Kuz-Oshie have played 70 minutes together this season, they scored five goals and were on the ice for
none.

In those 70 minutes they generated 58.37% of the scoring chances on ice.

Per 60 minutes they created 9.54 scoring chances for to 6.81 against.

Last season they played 178 minutes together, they scored 10 goals and were on the ice for six.

In those 178 minutes they generated 68.71% of the scoring chances on ice.

Per 60 minutes they created 11.73 scoring chances for to 5.34 against.

So in 248 minutes the last two seasons, 8-92-77 have scored 15 goals and let in six, and in total have created 45 scoring chances for to 23 against.

Now let's compare:

Ovi-Backstrom-Oshie have played 321 minutes together this season, they've scored 20 goals and were on the ice for 12.

In those 321 minutes they generated 52.19% of the scoring chances on ice.

Per 60 minutes they create 9.61 scoring chances for to 8.55 against.

Last season they played 620 minutes together, they scored 29 goals and were on the ice for 14.

In those 620 minutes they generated 53.25% of the scoring chances on ice.

Per 60 minutes they created 7.69 scoring chances for to 6.75 against.

So in 942 minutes the last two seasons, 8-19-77 have scored 49 goals and let in 26, and in total have created 130 scoring chances for to 114 against.

If we extrapolate 8-92-77's numbers over those same 942 minutes we'd have 170 scoring chances for, 87 against and 57 goals for to 22 against.

That's eight more goals for, four less against, 40 more scoring chances for and 27 less against.

 

[Midnight Judges]

Good post. Do you think it is safe to assume 8-92-77 could sustain that pace over a longer stretch? Does it reduce Trotz's ability to turtle?

What are the ramifications on the second line?

 

[twabby]

Good post. Do you think it is safe to assume 8-92-77 could sustain that pace over a longer stretch? Does it reduce Trotz's ability to turtle?

What are the ramifications on the second line?

Coldplay619 can answer the first questions since he brought it up, but I'd like to address the bolded question.

Simply flipping Backstrom and Kuznetsov would have a second line of 90-19-14. This line has been fantastic over the past two seasons. Some stats on 90-19-14 (all stats are score- and zone-adjusted except goals for):

TOI: 154.38 minutes
Corsi for %: 60.41%
Expected goals for %: 57.10%
Scoring chances for %: 57.72%
Goals for %: 69.23%
PDO: 102.48

These are fantastic numbers. Alternatively you could return to the Tre Kronor second line of 90-19-65 that was good earlier this year, but Burakovsky seemed to gel really well alongside Connolly and Eller before he was injured.

I think 90-19-14 could be an effective shutdown line as well. Match them up against first lines as much as possible and let 8-92-77 feast on easier competition.

 

[Stewie G]

From your keyboard to Trotz's ears/eyes/consciousness/whatever.

 

[twabby]

What other reason is there? Everyone on D is willing to block shots, Carlson/Orlov don't have any problems outmuscling players and Niskanen is pretty cerebral. I'd need to see something explicit to convince me they can't PK reliably. There's god knows how many instance of Orpik messing up his assignment and plowing a guy into Holtby as the puck goes in the net and no one holds it against him.

I'd agree with this. I'd also add that penalty killing isn't just about blocking shots and winning battles, it's about denying opportunities to begin with and part of that includes skating ability and puck skills. This isn't to say that Alzner and Orpik aren't good at penalty killing because I'd guess they are good, but how much better are they really than Orlov and Schmidt? It's tough to say and there are also a few models that seem to suggest that individual skaters simply don't have a very large impact on penalty killing:

https://hockey-graphs.com/2016/10/27/extras-blending-seasonal/

After the readjustments have been made and testing conducted, it was decided that Short Handed Defense (4) will be dropped from the WAR formula. The decision was based on its extremely minor impact and essentially zero correlation from one season to the next. This topic is worth further exploration, but it fits with previous research noted here. Short Handed Defense in the model will now be completely accounted for in a player’s ability to not take penalties, the best way to help their team’s penalty kill will be to not put them in that situation in the first place.

http://www.wingingitinmotown.com/20...-of-deployment-on-individual-shorthanded-shot

1. No single person (outside of the goalie, perhaps) is responsible for the success or failure of the penalty kill. At least, not to the extent that people will have you believe.
2. It's difficult to pin down exactly what makes a good penalty killer because of hockey's inherent randomness and complexity.

I'd guess that penalty killing is much more a product of systems and quality goaltending than anything else. As a simple example look at the Capitals PP effectiveness against western conference opponents compared to against Metro teams who likely scout them more thoroughly. As another example, compare the Flyers' PK in games 1-3 to games 4-6 when they very clearly changed their tactics to be more aggressive on Backstrom. The personnel in both cases aren't suddenly much better, but the tactics are certainly much different and yield much different results.

I'd much rather figure out even-strength and PP roles first and worry about the PK later, in terms of personnel.

 

[g00n]

the best way to help their team’s penalty kill will be to not put them in that situation in the first place.

This is kind of ridiculous but it fits the overall pattern in the community of shelving inconvenient elements of analysis when it doesn't fit some model. How many times have we seen players unfairly penalized?

This is just one of those things that doesn't fit a stat model so it gets discarded, but if you sat and watched a game with someone who knew what to look for he/she could point things out as they happen. It just makes absolutely NO sense whatsoever to conveniently assume that system and goaltending makes Pkers interchangeable plugs. Like any other skill in the world, some are better at it than others.

 

[twabby]

This is kind of ridiculous but it fits the overall pattern in the community of shelving inconvenient elements of analysis when it doesn't fit some model. How many times have we seen players unfairly penalized?

This is just one of those things that doesn't fit a stat model so it gets discarded, but if you sat and watched a game with someone who knew what to look for he/she could point things out as they happen. It just makes absolutely NO sense whatsoever to conveniently assume that system and goaltending makes Pkers interchangeable plugs. Like any other skill in the world, some are better at it than others.

The author built a model and saw that statistically speaking individuals don't impact penalty killing meaningfully and repeatably compared to other aspects. You can pick apart the model that produces the output if you don't agree with the output, but to say PKing was discarded because it didn't fit the model isn't accurate. It was discarded because the output generated from the model suggested little meaningful individual ability in PKing.

And I don't think it really matters if a player is unfairly penalized or not. Tom Wilson has a reputation and it hurts the team that he has that reputation. It is unfair, but it's something that needs to be accounted for in an evaluation of him and others.

 

[Zoidberg Jesus]

I think trying to assess PKing statistically is going to run into huge problems with differences in competition and starting situations. The guys the coach considers the best PKers are going to start every PK in their own zone, against the opponent's fresh first unit. Compare that to the second unit - they're either changing on the fly after the first unit cleared it and thus getting to contest the entry, or they're facing off in their own zone against a tired 1st unit or a fresh, but weaker second unit. Whatever the situation, it's a noticeably easier assignment, and that's going to confound the statistics. If you want to claim that PKing outcomes aren't predictive of future PKing outcomes ergo PKing skill is roughly even throughout the league, you need to work out the influence of those confounding factors first. The problem is, I don't know how you do that. In order to know how much usage affects outcomes, you need to know how much the player's skill at PKing affects outcomes in order the factor that out. To know how much skill at PKing affects outcomes, you need to know how much usage affects outcomes in order to factor that out. It's a catch-22.

 

[Zoidberg Jesus]

The author built a model and saw that statistically speaking individuals don't impact penalty killing meaningfully and repeatably compared to other aspects. You can pick apart the model that produces the output if you don't agree with the output, but to say PKing was discarded because it didn't fit the model isn't accurate. It was discarded because the output generated from the model suggested little meaningful individual ability in PKing.

And I don't think it really matters if a player is unfairly penalized or not. Tom Wilson has a reputation and it hurts the team that he has that reputation. It is unfair, but it's something that needs to be accounted for in an evaluation of him and others.

Nothing in either of the sources you linked seems to indicate that they made any effort to factor in usage before coming to the conclusion that individuals don't affect outcomes (although, as I mentioned before, I'm not sure how you could do this). It's one thing to say that individual outcomes, on their own, aren't predictive of future outcomes. I can accept that. But it's another thing entirely to use that conclusion to say that there's little individual difference in PKing skill.

 

[twabby]

I think trying to assess PKing statistically is going to run into huge problems with differences in competition and starting situations. The guys the coach considers the best PKers are going to start every PK in their own zone, against the opponent's fresh first unit. Compare that to the second unit - they're either changing on the fly after the first unit cleared it and thus getting to contest the entry, or they're facing off in their own zone against a tired 1st unit or a fresh, but weaker second unit. Whatever the situation, it's a noticeably easier assignment, and that's going to confound the statistics. If you want to claim that PKing outcomes aren't predictive of future PKing outcomes ergo PKing skill is roughly even throughout the league, you need to work out the influence of those confounding factors first. The problem is, I don't know how you do that. In order to know how much usage affects outcomes, you need to know how much the player's skill at PKing affects outcomes in order the factor that out. To know how much skill at PKing affects outcomes, you need to know how much usage affects outcomes in order to factor that out. It's a catch-22.

It's not really a catch-22 because individual players get different deployments throughout the course of games and their careers. In the first half of the season Niskanen was on PK1, now he's on PK2. Alzner wasn't always on PK1, especially in 2014-15. All 4 PKing defensemen have had plenty of time against opposing teams' first units.

You can compare these performances by different players in similar deployments and get reasonable results I would imagine. I wouldn't write off statistical modeling just because the problem seems difficult to solve.

 

[twabby]

Nothing in either of the sources you linked seems to indicate that they made any effort to factor in usage before coming to the conclusion that individuals don't affect outcomes (although, as I mentioned before, I'm not sure how you could do this). It's one thing to say that individual outcomes, on their own, aren't predictive. I can accept that. But it's another thing entirely to use that conclusion to say that there's little individual difference in PKing skill.

The entire XPM model in the first source factors in usage and deployment. It's a 5 part series of posts and is explained in part 2:

https://hockey-graphs.com/2016/10/25/expected-plus-minus/

Response Variable
-Rate at which an Expected Goals event took place
Explanatory Variables
-All players on the ice
-The coaches of both teams
-Zone the shift started in
-A score adjustment
-If one of the teams is playing a back-to-back
-A home ice advantage adjustment

 

[Ridley Simon]

PK Abilty

The entire XPM model in the first source factors in usage and deployment. It's a 5 part series of posts and is explained in part 2:

https://hockey-graphs.com/2016/10/25/expected-plus-minus/

So, if it's really a function of the goalie, and nothing else (summary), then each teams save percentage rankings on the PK (and goals against) should mirror those ES. Right?

In theory, if specific "PK Abilty" doesn't exist, then there shouldn't be a team that is abjectedly better at PK percentages and/or GAA rankings than they are at ES percentages/GAA.

Not sure if that's borne out by the numbers or not.

 

[txpd]

I am not sure how you could conclude that could replace Doughty with Chorney and there would be no fall off in quality. That you could replace Beagle with Vrana and there would be no change. That you could replace Beagle Winnik Carlson Orpik with Kuzy Burakovsky Schmidt and Chorney and see no appreciable decline in effectiveness.

I am not buying that.