Plus/Minus the Stat that tastes like Pie

[inkman]

Light is given by the sun. The sun exists in space. So why is space dark? 

http://youtu.be/JT0zjorR68A

[SwampD]

"they" as in "advanced" versus "non-advanced" stats? I mean yeah they both have successes. I would rather have plus-minus than no numbers at all, if I wanted numbers to look at. 

 

The rest of that, I don't necessarily feel I 

 

(a.) represent the mythical entity that is the stats community

(b.) know enough

 

to address, which is why True needs to get his butt in here.

 

Though I'm cautious of the way you phrase "we often hear about how we need large sample sizes". 

 

The fact is, noise dampens in bigger sample sizes which is why they tried starting to track things that happen 30 times per game instead of 2. It just becomes a stronger data set no matter what you're measuring. I don't think there's anything more to it than that. 

 

 

That is obviously true. It just bothers me when the noise is dismissed completely or even worse, looked down upon as uninformed.

 

Big picture numbers are great for telling me, well, the big picture. But the big picture is made up of slumps and hot streaks. If a player is in a current hot streak, that overall number (whatever it may be) is less relevent than one with a smaller sample size. It might sound like the gamblers fallacy, but we're still dealing with human beings and those pesky brains that get in the way of performance.

[North Buffalo]

That is obviously true. It just bothers me when the noise is dismissed completely or even worse, looked down upon as uninformed.

 

Big picture numbers are great for telling me, well, the big picture. But the big picture is made up of slumps and hot streaks. If a player is in a current hot streak, that overall number (whatever it may be) is less relevent than one with a smaller sample size. It might sound like the gamblers fallacy, but we're still dealing with human beings and those pesky brains that get in the way of performance.

Exactly what I was trying to get at. Hot streaks and slumps can overexagerrate a players value both from a positive or negative way. Data smooths out the noise. Or it can account for variability and historical swings in a players play. It still wont account for recovery from one time injuries short term. TRpm is a good start but flawed in evaluating a players performance and value and doesnt account swings. Sabres got in trouble with Stafford because of this... that being said a player on a hot streak is valuable when on a hot streak or not. Itd be nice to see if there was a predictor or likelihood of such play and coaches and GMs could factor in that expectation and adjust accordingly. Sam’s current hot streak.. though is it sustainable because of player growth... contract year or just a hot streak?

[calti]

Light is given by the sun. The sun exists in space. So why is space dark? 

i dunno...global warming?

[Randall Flagg]

Light is given by the sun. The sun exists in space. So why is space dark? 

Brightness falls off by d^-2

 

And there's a lot of empty d in space 

Edited March 29, 2018 by Randall Flagg

[BagBoy]

https://forums.sabrespace.com/topic/24217-sabres-prospects-2016-17/?p=963912

 

I'm definitely not a fancy stats guy, but I really like a lot of the rationale that you've come up with here.  It makes sense for me for the vast majority of players on a given team, but I'm a cynic, so I gravitate to the outliers.  Chris Drury in 2005-06 was -11, and +1 2006-07.  Those were both very good Sabres teams, and his relative TRpm against his teammates would have been poor, but that guy was the most important skater on that team.  He always matched up against the opponents' #1 center and/or line.  He took tons of D-zone draws.

 

I’ve always liked +/- as a stat vs. teammates, keeping in mind that players with tougher responsibilities/matchups would have “artificially” lower totals, and I was always fine with that.  I think TRpm is clearly superior to +/-, but like +/-, it seems like it might have a blind spot for shutdown guys, especially forwards on good teams.

[bunomatic]

What kind of pie ?

[TheAud]

Eichel's 5 assist, minus 1 night clearly shows the uselessness of assists as a statistic. 

[pi2000]

Eichel's 5 assist, minus 1 night clearly shows the uselessness of assists as a statistic.

4 PP assists.

 

I've said this before, as a GM you need to decide if a negative even strength player is worth his contributions on the PP.

[Randall Flagg]

4 PP assists.

 

I've said this before, as a GM you need to decide if a negative even strength player is worth his contributions on the PP.

Woah, I agree with pi in a general sense about something. 

 

Sorta like the decision we faced with Moulson last year, innit?

 

---------------------------------------------------------------------------------------------------

 

https://mapleleafsnation.com/2018/03/12/are-scoring-chances-better-than-corsi/

 

This is an interesting article. It concisely stated at the beginning the conclusion I was trying to get to - that team-wise, a goal differential will allow you to predict current standings quite well, but it won't allow you to predict the standings 3 months from now as well as shot-based metrics do. They clean out the noise provided by randomly including shorthanded attempts, and goaltender save percentages (which can vary wildly between two similarly used teammates, just by chance).

[SwampD]

Woah, I agree with pi in a general sense about something.

 

Sorta like the decision we faced with Moulson last year, innit?

 

---------------------------------------------------------------------------------------------------

 

https://mapleleafsnation.com/2018/03/12/are-scoring-chances-better-than-corsi/

 

This is an interesting article. It concisely stated at the beginning the conclusion I was trying to get to - that team-wise, a goal differential will allow you to predict current standings quite well, but it won't allow you to predict the standings 3 months from now as well as shot-based metrics do. They clean out the noise provided by randomly including shorthanded attempts, and goaltender save percentages (which can vary wildly between two similarly used teammates, just by chance).

So, shot metrics are a better predictor in November, but after an entire season, aren't they probably pretty close?

 

We should try it. I'll do a goal diff. bracket and you do a shot bracket in the playoff challenge thingy. Could be fun.

[Randall Flagg]

So, shot metrics are a better predictor in November, but after an entire season, aren't they probably pretty close?

 

We should try it. I'll do a goal diff. bracket and you do a shot bracket in the playoff challenge thingy. Could be fun.

For basically any meaningful sample size, so sure, in late November (though I'd prefer 41 games maybe), if you were to predict the end of season standings using shot based metrics, in general analyses that have spanned every team in many seasons of hockey (just that chance versus corsi one managed all 30 teams and 4 seasons worth of data) there is a statistically significant difference in success rates of end of season standings predicted by team goal differential versus the shot based metrics. 

 

Of course, this is stats. So that doesn't imply that they will always win for every possible subset size and time length that you can pull out. Sometimes counting the number of hangnails on every player on every team and using them to predict the standings will do better than everything else. But in general, the shot metrics do better. They correlate more strongly to standings results and wins in the future statistically, and correlation between two sets of data is absolutely something you can unambiguously calculate. This does allow for 2011 standings to better reflect goal differential, maybe. But it doesn't change the correlation difference, which will tell you that if you wait long enough, that success will be passed by the other metric. So one year goes that way, but then three would go the other, or something. 

 

Now, goal differential is better than random guessing with no information available, just pulling teams out of a hat (though I'm pretty sure not by as much as you'd expect, nature is powerful! :P ). 

 

And I'd be happy to do that on the caveat that the playoffs by nature are tiny sample sizes, prone to the fact that "anything can happen", which is why they're so fun, so I want to make it clear that it would be purely for fun, and that since sample size issues are well understood in any field that uses statistics, a loss would not be some dreaded outcome that tarnishes the reliability of a more advanced stat versus a less advanced one, a conclusion I have seen get drawn here many times, as if to use advanced stats because of better performance implies that a single failure in a small set of data will ruin everything that person stands and lives for. 

Although I'm not doing that with my official Sabrespace bracket thru NHL.com, I want that one to be my own picks, for which I will not consult a single advanced stat, I just want to look at rosters. I can make one with shot metrics and put it here to compare instead.

Edited April 7, 2018 by Randall Flagg