Sabres Trade Alex Nylander to Chicago for D Man Henri Jokiharju

[LGR4GM]

20 minutes ago, triumph_communes said:

https://arxiv.org/pdf/1006.4310.pdf

 

Pages 34-37

A snippet:  

i.e., when a player spends the majority of his time with another skater, the player's stats become indistinguishable from each other.

 

Casey spent all year with Okposo and in limited time with Thompson, and Thompson otherwise spent his year with Sobotka outside a few games.  As a result, their charts are going to mimic those players to a large degree, a digression explicity noted by the creator of the statistics.

 

The terms in the regression are built off of a massive dataset, so the error by smaller minutes played put into the model is relatively low.

At 5v5: 

Casey spent 326minutes (38%) with Okposo and 533minutes (62%) without Okposo. (minutes are rounded to nearest whole number) 

Kyle and Casey both see their CF% increase when they are away from eachother. 

Tage spent 354mins with Sobotka and 361 without, Tage sees a significant rise in his CF% without Sobotka and Sobotka sees a significant dip without Tage. 

Edited July 10, 2019 by LGR4GM
error

[North Buffalo]

2 minutes ago, LGR4GM said:

At 5v5: 

Casey spent 326minutes (38%) with Okposo and 533minutes (62%) without Okposo. (minutes are rounded to nearest whole number) 

Kyle and Casey both see their CF% increase when they are away from eachother. 

Tage spent 354mins with Sobotka and 361 without, Tage sees a significant rise in his CF% without Sobotka and Sobotka sees a significant dip without Tage. 

Interesting... and add another year of maturity and some better players and see what kind of changes happen.  At this point, what choice do we have but an off season hope and a prayer.

Edited July 10, 2019 by North Buffalo

[LGR4GM]

For fun, take a look at Erod with Sobotka versus without. Sobotka is an offensive black hole of sadness. Erod without Sobotka: 53.46% CF versus with 45.76%

Erod with Tage: 60% CF but only 47min TOI. 

[triumph_communes]

3 minutes ago, ... said:

Are you saying Casey spent 92% of his time with Okposo and TT 92% of his time with Sobotka?  

Since the league is littered with pairs, are you arguing the fancy stats aren't built to account for the effect one might have on the other within a normal context?

The highest errors in the model come when percentages are above 60% for players with >700 minutes.

 

 

And yes, quite literally, the author of these fancy stats states:

 

Quote

By not including interaction terms in the model, we do not account for interactions between players. Chemistry between two particular teammates, for example, is ignored in the model. The inclusion of interaction terms could reduce the errors. The disadvantages of this type of regression would be that it is much more computationally intensive, and the results would be harder to interpret.

 

[LGR4GM]

I'm sorry, I read the number incorrectly, Tage is about the same away from Sobotka. 

Tage with Sobotka: 46.17 CF%

Tage w/out Sobotka: 47.93 CF%

(CF% - Percentage of total Corsi while that combination of players is on the ice that are for the selected team. CF*100/(CF+CA))

Edited July 10, 2019 by LGR4GM

[Curt]

17 minutes ago, triumph_communes said:

Casey spent all year with Okposo

Do you have any data to back this up, because I'm pretty sure that the 2nd half of the season, Casey didn't play with Okposo at all and Okposo was with Girgs/Lars.

[darksabre]

3 minutes ago, LGR4GM said:

I'm sorry, I read the number incorrectly, Tage is about the same away from Sobotka. 

Tage with Sobotka: 46.17 CF%

Tage w/out Sobotka: 47.93 CF%

(CF% - Percentage of total Corsi while that combination of players is on the ice that are for the selected team. CF*100/(CF+CA))

Tage being about the same with Sobotka as without might honestly be a testament to Tage. If he has any talent it at least allowed him to avoid being sucked into the Sobotka black hole.

[woods-racer]

12 minutes ago, LGR4GM said:

At 5v5: 

Casey spent 326minutes (38%) with Okposo and 533minutes (62%) without Okposo. (minutes are rounded to nearest whole number) 

Kyle and Casey both see their CF% increase when they are away from eachother. 

Tage spent 354mins with Sobotka and 361 without, Tage sees a significant rise in his CF% without Sobotka and Sobotka sees a significant dip without Tage. 

So in a nut shell. 

This past year the Sabres where a mix of players ( predominately the middle six) that had the worst possible impact on each other when on the same line.

Even scuffling them just mixed the bad mix for the same results. The bad mix being TT and Sobs. No matter where they where in the line up they drug down their line mates.

 

So the hope is by adding even just 2 players with *average* possession and CF% numbers will greatly elevate the middle six exponentially?

Is my simple assumption of these stats an over-simplification?

 

Edited July 10, 2019 by woods-racer

[ddaryl]

2 hours ago, SwampD said:

Because people are drooling over themselves about what a steal this trade is, while at the same time we’re being told that Nylander is the guy with more talent, he just isn’t motivated, all while we know (rumored to know) that he had been in a toxic environment that maybe he feels the organ-eye-zation knew about. Does anybody actual talk to these guys?! We know GMTM/DB didn’t.

I dream of a day when the Sabre hold on to talent instead of this addition by subtraction method of team building that hasn’t worked out so well for us.

If he was THIS talented/motivated he would of played more games in a Sabres uniform. Until that talent shows up on NFL ice it doesn't exist. A change of scenery was perfectly fine and the return we got made it a little sweeter for Sabres fans.

You cannot keep players in your system if they don't make the jump in a pre determined amount of years. Nylander never lit it up consistently in the AHL while others have made a bigger impact. Might as well get an asdet that's a little younger and has already shown some success in the NHL.

As for the whole toxic environment. Meh not interested in hyperbole that happens when fans need to manufacture reason for the failures. Reasons for our failures are pretty straight forward. Started with Rigas, peaked with Golisano, and then exhausted all of us with Pegula's trying to tear it down and do a full head to toe rebuild. But we are on the other side of this, and fully believe the toxic environment issues are in the past, the team just needs to come of age now and grow. 

Many of you have lost patience. Mistakes were made, lessons learned. We are moving forward IMO not backward. Just got to stay patient ?

[rakish]

Where was Nylander during the prospects camp?

[LGR4GM]

26 minutes ago, WildCard said:

 

 

Link?

https://theathletic.com/573862/2018/10/07/video-session-breaking-down-henri-jokiharjus-impressive-game-shift-by-shift/

[North Buffalo]

1 minute ago, rakish said:

Where was Nylander during the prospects camp?

Obviously at the Chicago Merc.

Edited July 10, 2019 by North Buffalo

[...]

30 minutes ago, triumph_communes said:

The highest errors in the model come when percentages are above 60% for players with >700 minutes.

And yes, quite literally, the author of these fancy stats states:

You did read this: https://arxiv.org/pdf/1201.0317.pdf Right?

And this: https://hockey-graphs.com/2019/01/14/reviving-regularized-adjusted-plus-minus-for-hockey/ too?

Because those more aptly apply to the metrics we're discussing.  You're using paper #1 to make your argument.

Quote

Additionally, we will use a technique called “regularization” in the linear regression (this is where “Regularized” in “Regularized Adjusted Plus-Minus” comes from). Regularization in a linear regression comes in two main forms – ridge regularization (also known as Tikhonov regularization or L2 regularization) and LASSO regularization (“Least Absolute Shrinkage and Selection Operator”, also known as L1 regularization). The main purpose is to address multicollinearity that is present in the data. Why do we care about multicollinearity? Well, when a pair of players play together for a significant amount of time (the classic example is Henrik and Daniel Sedin who spent over 90% of their career time on ice together) the coefficient estimates in a traditional OLS regression will be extremely unstable (and therefore unreliable). Regularization combats this by adding some amount of “bias” into the model (Gaussian “white-noise”) to decrease the variance in the coefficient estimates [more info here]. What this means, essentially, is unstable coefficient estimates are “penalized” (or “shrunk”) based around a Gaussian distribution where 0 is the mean. Ridge regularization will pull coefficients towards 0 (but never exactly 0). LASSO regularization will pull coefficients toward 0 and also “zero” some coefficients. 

 

Edited July 10, 2019 by ...
...edited this to make it less dense.

[Brawndo]

10 minutes ago, rakish said:

Where was Nylander during the prospects camp?

After attending three prospects camps, it becomes optional. He took that option.

[WildCard]

Just now, Brawndo said:

After attending three prospects camps, it becomes optional. He took that option.

Why wasn't Mittelstadt there? Unless of course he was

[LGR4GM]

2 minutes ago, Brawndo said:

After attending three prospects camps, it becomes optional. He took that option.

I think he took a slow boat to Chicago. 

1 minute ago, WildCard said:

Why wasn't Mittelstadt there? Unless of course he was

He had played enough NHL games? Maybe? I mean Thompson was there so maybe that isn't it. 

[...]

1 hour ago, ... said:

Explain the significance of the difference between 68 minutes TOI versus 884 minutes TOI.  I'll hang up and listen.

Since this wasn't answered, yet, I will answer it.

Quote

Above, we can see that players who play less than ~100-150 EV minutes are being “regressed” to the mean (0 for each regression) – this is the regularization pulling these players towards the mean. Not only does this help to deal with multicollinearity (to an extent), it also adds a “quasi-Bayesian” aspect to the player ratings. In other words, if a player has very few shifts in the data, they are brought closer to the league mean using a Gaussian “prior” distribution. In an OLS regression, these players would have wildly inflated per 60 ratings.

So, the 68 minute RAPM is skewed, possibly making Mitts look better.

And since I'm here, let's also note this about the RAPM charts we all know and love:

Quote

So let’s summarize what the RAPM coefficients are. They are offensive and defensive ratings for each player that are isolated from the other skaters they played with, the other skaters they played against, the score state,  the effects of playing at home or on the road, the effects of playing in back-to-back games, and the effects of being on the ice for a shift that had a faceoff in the offensive or defensive zone.

 

[jsb]

34 minutes ago, rakish said:

Where was Nylander during the prospects camp?

 

22 minutes ago, Brawndo said:

After attending three prospects camps, it becomes optional. He took that option.

I can't help but think that not showing up pushed this trade ahead. He didn't think it was important enough to show and they then thought he wasn't worth the bother either.

Edited July 10, 2019 by jsb
spelling errors and adding a quote

[TrueBlueGED]

6 minutes ago, triumph_communes said:

https://arxiv.org/pdf/1006.4310.pdf

 

Pages 34-37

A snippet:  

i.e., when a player spends the majority of his time with another skater, the player's stats become indistinguishable from each other.

 

Casey spent all year with Okposo and in limited time with Thompson, and Thompson otherwise spent his year with Sobotka outside a few games.  As a result, their charts are going to mimic those players to a large degree, a digression explicity noted by the creator of the statistics.

 

The terms in the regression are built off of a massive dataset, so the error by smaller minutes played put into the model is relatively low.

If this is what you think, then you really have no clue how this model was created.  Read the links above.

 

 

We might be talking past each other. I don't think you're drawing the proper inference from what the authors are saying with respect to how it would apply to Casey's numbers and my comment about using a 68 minute subsample to evaluate him, or you're misinterpreting what I was getting at. The Sedins have a ton of error when trying to isolate because they spend 90% of their ice time together, so there isn't enough time apart to do the work with any accuracy. That all makes sense, but it's not what I was arguing about.

You presented a 68 minute subsample of Casey's career under the pretense that it was a better representation of his on-ice impact than the full sample, because the full sample includes so much time with Okposo that he's statistically indistinguishable from Okposo. I disagree with that because, as others have noted, the majority of his ice time has come with players not named Okposo. So I don't think that introduced the same collinearity problem of the Sedins. It's also worth noting that about 30% of his ice time in 2017-18 was spent with Okposo. Or, only about 7 points less than in 2018-19. If the model screwed him because of his time with Okposo, one would think it would've happened both seasons. 

Which brings me to my larger point, which I was trying to argue in the first place. Mittelstadt has 952 minutes of ice time at even strength. You're trying to draw conclusions with 7.1% of that, while chucking the rest of the data because of Okposo, when Okposo was relevant to less than 40% of it (a huge difference from the Sedins' 90%+ shared ice time). That's weak, both statistically and theoretically. There is no way you're going to statistically distinguish that small of a subsample from the rest of his ice time with any confidence that the difference in results was anything other than happenstance. 

[...]

3 minutes ago, TrueBlueGED said:

Which brings me to my larger point, which I was trying to argue in the first place. Mittelstadt has 952 minutes of ice time at even strength. You're trying to draw conclusions with 7.1% of that, while chucking the rest of the data because of Okposo, when Okposo was relevant to less than 40% of it (a huge difference from the Sedins' 90%+ shared ice time). That's weak, both statistically and theoretically. There is no way you're going to statistically distinguish that small of a subsample from the rest of his ice time with any confidence that the difference in results was anything other than happenstance. 

You forgot to mention the Gaussian “white-noise”.  I'm disappointed.

[North Buffalo]

Just now, ... said:

You forgot to mention the Gaussian “white-noise”.  I'm disappointed.

Thought that was only in X-Ray and photography... ah the image is slowly coming into focus.. 

[triumph_communes]

17 minutes ago, TrueBlueGED said:

We might be talking past each other. I don't think you're drawing the proper inference from what the authors are saying with respect to how it would apply to Casey's numbers and my comment about using a 68 minute subsample to evaluate him, or you're misinterpreting what I was getting at. The Sedins have a ton of error when trying to isolate because they spend 90% of their ice time together, so there isn't enough time apart to do the work with any accuracy. That all makes sense, but it's not what I was arguing about.

You presented a 68 minute subsample of Casey's career under the pretense that it was a better representation of his on-ice impact than the full sample, because the full sample includes so much time with Okposo that he's statistically indistinguishable from Okposo. I disagree with that because, as others have noted, the majority of his ice time has come with players not named Okposo. So I don't think that introduced the same collinearity problem of the Sedins. It's also worth noting that about 30% of his ice time in 2017-18 was spent with Okposo. Or, only about 7 points less than in 2018-19. If the model screwed him because of his time with Okposo, one would think it would've happened both seasons. 

 Which brings me to my larger point, which I was trying to argue in the first place. Mittelstadt has 952 minutes of ice time at even strength. You're trying to draw conclusions with 7.1% of that, while chucking the rest of the data because of Okposo, when Okposo was relevant to less than 40% of it (a huge difference from the Sedins' 90%+ shared ice time). That's weak, both statistically and theoretically. There is no way you're going to statistically distinguish that small of a subsample from the rest of his ice time with any confidence that the difference in results was anything other than happenstance. 

Okposo+Thompson, similar drags

 

The collinearity tries to correct for individual players, but it obviously can't without data outside of it.  The lowest errors for all these models happens for players who played for multiple teams.  When looking at a teenage prospect who has had limited opportunities, you have to take it with a huge grain of salt.  When you have Mittelstadt with less than 200 minutes with a replacement level player, it doesn't matter if there's 900+ minutes of data on it, he's never had a chance to do anything but be dragged.

[TrueBlueGED]

5 minutes ago, triumph_communes said:

Okposo+Thompson, similar drags

 

The collinearity tries to correct for individual players, but it obviously can't without data outside of it.  The lowest errors for all these models happens for players who played for multiple teams.  When looking at a teenage prospect who has had limited opportunities, you have to take it with a huge grain of salt.  When you have Mittelstadt with less than 200 minutes with a replacement level player, it doesn't matter if there's 900+ minutes of data on it, he's never had a chance to do anything but be dragged.

But he has, and he's been a drag on those players. He had 426 minutes with Sheary, 162 with Rodrigues, 152 with Reinhart, and 119 with Skinner. All of them had worse corsi, goals, and xG while skating with Mittelstadt. Obviously some of that is Skinner and Reinhart got to skate with Jack, but neither Rodrigues nor Sheary spent so much time with Jack as to pump their numbers much. 

Again, like with Thompson and Sobotka, these things are not mutually exclusive. Okposo can be a drag and Casey can be bad in his own right. After all, a team doesn't finish 5th from the basement if they only have 1 or 2 problem players. 

[That Aud Smell]

[TrueBlueGED]

9 minutes ago, That Aud Smell said:

Short version, in English: Triumph blames Mittelstadt's performance on everyone but Mittelstadt. I think he stunk last year in his own right.