LaMarcus > Bosh

[bobf]

Where the hell is Brandon Roy on that list?

64th....

Overall +2.3
Offsense +4.3
Defense +2.0

That positive defense means BAD defense.

But keep in mind each rating statisically has a 10% chance of being off by 2 standard errors (3.6) and higher chance of being off by one standard error (1.8).

Then again, mabye Roy's defense is actually not good?

 

[Minstrel]

64th....

Overall +2.3
Offsense +4.3
Defense +2.0

That positive defense means BAD defense.

But keep in mind each rating statisically has a 10% chance of being off by 2 standard errors (3.6) and higher chance of being off by one standard error (1.8).

Then again, mabye Roy's defense is actually not good?

I'm curious...do you ever do a sanity test, like add up all the +/- of the players on a team and see if it's a good predictor of the team's point differential? Or, not even predictive, but comes close to the point differential the team already put up in the most recent season?

Do all the Blazers players' +/- from your system sum to the Blazers' point differential this year?

 

[bobf]

I'm curious...do you ever do a sanity test, like add up all the +/- of the players on a team and see if it's a good predictor of the team's point differential? Or, not even predictive, but comes close to the point differential the team already put up in the most recent season?

Do all the Blazers players' +/- from your system sum to the Blazers' point differential this year?

Good idea, I'll take a look at that. Actually, I can't just add them up, I need to weight them by minutes played or possessions.

Because this is essentially a 4-year rating it does not measure the current year, it meansure all those years together. So it does not measure player improvements. So I'd expect it to underrate the Blazers in 2008-2009 since their players are improving.

 

[Minstrel]

Because this is essentially a 4-year rating it does not measure the current year, it meansure all those years together. So it does not measure player improvements. So I'd expect it to underrate the Blazers in 2008-2009 since their players are improving.

Right, I know it's a multi-year sample. But, from my understanding, it's a multi-year sample because that's the level of data you need for something close to "real information" (as opposed to sample size artifacts). So, even though the data is not just from this year, it should be your system's best guess as to the players' current "true talent level," right? That's why I recommended checking it against this year's team point differential.

 

[bobf]

Right, I know it's a multi-year sample. But, from my understanding, it's a multi-year sample because that's the level of data you need for something close to "real information" (as opposed to sample size artifacts). So, even though the data is not just from this year, it should be your system's best guess as to the players' current "true talent level," right? That's why I recommended checking it against this year's team point differential.

No... it is purely due to sample size. If you look at my first table, the standard errors are still kind of high even using four years of data. If I used one year of data they would be too high. Even as it is, every 10th player rating is expected to be off by 3 or more.

So I have to use 4 years of data and I weight those years equally so it is not a current year rating. EDIT: In the sense that a vastly improved player will be rated the same as a vastly declined player if the overall performace during those four years were equal.

82games claims you can weight current year very heavily against prior years in order to get a current year rating while at the same time obtaining low standard errors . I don't believe them. It's like polling 10,000 people but weighting that last 1000 people much more than the other 9,000 people combined and thinking you're going to get a MOE similar to polling 10,000 people. Then they wonder why Chris Paul's defensive rating improves by something like 6 points from last year to this year.

So unfortunately I think that is a limitation of +/-. It takes years of data to reduce noise to reasonable levels, but then you are going to miss player improvement because you can't separate one year from the other. And one player's improvement missed is going to skew other player ratings somewhat. For example, Brandon Roy's improvement this year will tend to show up partly in Rudy's rating.

What I do like about this system though, is that it is the one rating system that can sense just about all of the intangibles except for interdependancies - stuff where the whole is greater than the sum of the parts.

Anyhoo... here is what I get for team +/- weighted average ratings. It's really not a predictor though since it cheats by using this year's data.

Team	Rating	Actual
BOS	9.3	8.1
CLE	8.0	8.9
LAL	6.8	8.0
SAS	5.9	3.8
HOU	5.6	3.8
ORL	4.6	7.3
POR	3.4	3.9
DAL	3.3	1.6
PHX	2.9	2.1
UTA	2.8	3.6
PHI	2.2	0.7
DEN	2.1	3.1
NOH	2.0	2.5
DET	1.6	-0.7
TOR	-0.3	-3.3
MIA	-0.4	-0.2
ATL	-0.6	2.0
CHA	-0.9	-1.1
MIL	-1.1	-1.2
NYK	-1.3	-2.6
CHI	-1.7	-0.7
NJN	-2.1	-2.3
GSW	-3.3	-3.7
IND	-3.5	-1.9
OKC	-6.0	-5.5
MIN	-7.0	-4.8
LAC	-7.7	-8.4
SAC	-7.9	-8.4
MEM	-8.1	-6.6
WAS	-8.5	-7.8

 

[Minstrel]

No... it is purely due to sample size. If you look at my first table, the standard errors are still kind of high even using four years of data. If I used one year of data they would be too high. Even as it is, every 10th player rating is expected to be off by 3 or more.

So I have to use 4 years of data and I weight those years equally so it is not a current year rating.

I think you misunderstood me. "It's purely due to sample size" is what I was saying. The point is, you build up that sample size to get actual information. Without a large enough sample size, you have information that is too unreliable to be considered useful.

I realize that you weight all the years equally, but the result should be useful in evaluating the players today. If it isn't, to be frank, what's the use?

I think it can be useful as a "today" measure, despite using multi-year data, because even though the players are improving, a larger sample size is still the best way to measure current ability. But we'd like to test that, which is what led to my original question. The fact that it doesn't track very well with actual point differentials suggests that it may not be the best evaluation tool.

 

[bobf]

I think you misunderstood me. "It's purely due to sample size" is what I was saying. The point is, you build up that sample size to get actual information. Without a large enough sample size, you have information that is too unreliable to be considered useful.

I realize that you weight all the years equally, but the result should be useful in evaluating the players today. If it isn't, to be frank, what's the use?

I think it is useful so long as you understand the limitations and the caveats. You have to factor in that we know that player's like Brandon Roy are improving while this system is measuring his overrall past performance, not last year's performance.

The potential value of +/- over other rating systems is it's ability to see things that other sytems like PER can not see well (intangibles, defense, player A making player B look good).

For example, this system "claims" that Amare's past performance is not nearly as valuable as PER suggests. He appears to be benefiting from Steve Nash who is rated very highly by this system in spite of his defense. And maybe things like Amare's defense are worse than we imagined. So even though that is only a past assessment averaged over many years, if true that would be very valuable thing to know. I was hoping he'd get traded so we could get a better picture of that.

The fact that it doesn't track very well with actual point differentials suggests that it may not be the best evaluation tool.

I think I'd need something to compare against in order to say it doesn't track well vs other methods.

 

[Minstrel]

The potential value of +/- over other rating systems is it's ability to see things that other sytems like PER can not see well (intangibles, defense, player A making player B look good).

Yes, I know that. But if it is always a lagging indicator, then it may be useful for questions like, "Who WAS better, Ewing or Mourning?" but not, "Who is most valuable today?" or "Who would help the team most next season?"

Most +/- systems try to present an indicator of current value. They have to build it using multiple seasons of data, but the goal is measuring current value. That's why I was surprised when you disavowed your system of being about current value.

I think I'd need something to compare against in order to say it doesn't track well vs other methods.

I don't know how it tracks with other +/- methods, I just meant that it doesn't seem to track closely enough to team point differential to seem better than just the normal route of using standard individual measures along with observation.

 

[STOMP]

LMA will be a max or close to max contract guy as soon as the team extends him this summer, so that's not really all that relevant.

any extension signed this summer won't kick in until 2010-11 season

STOMP

 

[bobf]

Yes, I know that. But if it is always a lagging indicator, then it may be useful for questions like, "Who WAS better, Ewing or Mourning?" but not, "Who is most valuable today?" or "Who would help the team most next season?"

They all lag somewhat. "Who is valuable today" is still a lagging indicator on "who will be valuable next year".

So it's more like this:
PER can measure how valuable a player was for the past 30 games.
+/- can measure how valuable a player was for the past 3 years.

But the hope is that +/- does not have built in bias due to
a) inability to see defense & intangilbes
b) player A stats inflated by player B

Most +/- systems try to present an indicator of current value. They have to build it using multiple seasons of data, but the goal is measuring current value. That's why I was surprised when you disavowed your system of being about current value.

Well past value is generally a decent measure of current value except when a player is rapidly improving or declining. I was just trying to make it clear that this system is not weighted towards recent performance.

What this system can do (at least in theory) that PER can not is separate Nash from Amare over the past three years and measure Amares intangibles & defense. This system knows that LMA >> Zach Randolph but PER does not.

So if Portland were contemplating trading Amare for LMA it would be extremely valuable to know their true contributions over that past three years, and then factor on top of that our observations that LMA has been improving somewhat and is more likely to improve in the future.

I'd much rather be working from an accurate lagging indicator than an inaccurate up-to-date one.