2016 RAPM/RPM/etc. *Full 2016 RS + PS RPM & RAPM Updated 6/24*

[Knosh]

Your on court/off court numbers don't show that Jordan didn't have negative impact on the Clippers defensive rebounding they show that the team got a higher percentage of defensive rebounds when he was on court than when he was off court which isn't the same thing.

I'm having a hard time reconciling this statement. From a defensive rebounding standpoint, the goal is to obtain the defensive rebound/prevent the opposing team from getting the offensive rebound. So getting a higher % of the defensive rebounds is, in essence, "doing better" on the defensive boards.
Thus, the way I'm interpreting your above statement is: Jordan has a negative impact on their defensive rebounding, but their defensive rebounding is better when he's on the court.
These assertions seem at odds with each other.


Now it could be that the rebounding rates of his four teammates on the court go down when he steps on to the court (i.e. his rebounds are coming mostly at the expense of his teammates). That is often the case of many (if not most) high-volume rebounders, I should think. Perhaps more so with Jordan than with others, idk.
But still, if he's grabbing so many boards as an individual (even if many of them are at the expense of his teammates) that the team defensive rebounding rate goes up, I don't understand how that can be labeled a net negative impact on their defensive rebounding.

It could be that if he was blocking out instead of "chasing" boards, that the TEAM defensive rebounding rate would be even better (even if his individual defensive rebounding rate was a little worse) when he's on the court; but that's not the same thing as having a negative impact.


So I'm still kinda confused by your above statement. Perhaps you could elaborate for me.

The on/off DREB% you posted from bbref is basically like raw +/-. DREB% is better with him on court, so that's like a positive unadjusted +/-.

The FFAPM rebounding numbers are the equivalent of adjusted +/-. Those factor in his teammates and the opponents impact on rebounding numbers. And as you said, Jordan's impact is worse than 58% of the league, so below average.

It's like a player having a positive unadjusted +/- but having a negative RAPM. The team is outscoring the opposition while he is on the court, but the player still has a negative impact.

[SideshowBob]

Here's the table GotBuckets.com provides for approximation of percentile values (also just saw that all the FFAPM models are 2-year).

Link

DJ's 42 DREB% score corresponds to a -0.43% shift in team DREB%.

[Texas Chuck]

I still think we have to be careful to not just go by these numbers. Tactics change base on who is on the floor. When Jordan is on the floor gobbling up lots of rebounds I promise you that the guards don't crash the defensive boards nearly as hard and guys start leaking out which hurts the team percentage because those same players all hit the boards when he is on the bench because they realize they need to help pick up the slack.

It's pretty much impossible that Jordan is personally having a negative impact on the team rebounding. This is a coaching issue/teammates issue more than a Jordan issue.

[trex_8063]

.

.

Your on court/off court numbers don't show that Jordan didn't have negative impact on the Clippers defensive rebounding they show that the team got a higher percentage of defensive rebounds when he was on court than when he was off court which isn't the same thing.

I'm having a hard time reconciling this statement. From a defensive rebounding standpoint, the goal is to obtain the defensive rebound/prevent the opposing team from getting the offensive rebound. So getting a higher % of the defensive rebounds is, in essence, "doing better" on the defensive boards.
Thus, the way I'm interpreting your above statement is: Jordan has a negative impact on their defensive rebounding, but their defensive rebounding is better when he's on the court.
These assertions seem at odds with each other.


Now it could be that the rebounding rates of his four teammates on the court go down when he steps on to the court (i.e. his rebounds are coming mostly at the expense of his teammates). That is often the case of many (if not most) high-volume rebounders, I should think. Perhaps more so with Jordan than with others, idk.
But still, if he's grabbing so many boards as an individual (even if many of them are at the expense of his teammates) that the team defensive rebounding rate goes up, I don't understand how that can be labeled a net negative impact on their defensive rebounding.

It could be that if he was blocking out instead of "chasing" boards, that the TEAM defensive rebounding rate would be even better (even if his individual defensive rebounding rate was a little worse) when he's on the court; but that's not the same thing as having a negative impact.


So I'm still kinda confused by your above statement. Perhaps you could elaborate for me.

The on/off DREB% you posted from bbref is basically like raw +/-. DREB% is better with him on court, so that's like a positive unadjusted +/-.

The FFAPM rebounding numbers are the equivalent of adjusted +/-. Those factor in his teammates and the opponents impact on rebounding numbers. And as you said, Jordan's impact is worse than 58% of the league, so below average.

It's like a player having a positive unadjusted +/- but having a negative RAPM. The team is outscoring the opposition while he is on the court, but the player still has a negative impact.

OK, that makes sense.

Hypothetically a player could have a negative impact on his team's defensive rebounding while still being a defensive positive (even with things like forced turnovers, foul rates, and blocks being equal), if his efforts force the opposition to lower% shots and lower eFG%. But surprisingly Jordan was only the 46th percentile in eFG% last year.

It strikes me odd that Jordan still rates out as a +2.09 defensive FFAPM when he's 46th percentile in eFG%, 42nd percentile in DREB, and 9th percentile in TOV rate. I assume that's because this metric is still partially informed by boxscore metrics?

[laika]

The more I look at RPM and RAPM the more I'm convinced that those two stats are fatally flawed.
They produce too many weird values. But if you know some values are wrong how can you trust the rest of the ratings?
In theory a properly calculated APM would be the ultimate stat. But I don't think anyone has managed to overcome the statistical problems so far.

I know it's RPM instead of RAPM, but the discrepancies are just too big for poor box score production to make up the difference. Just some of the latest unlikely values-

Rank--------Unadjusted +/-
16 Crowder neg 0.1
19 Bosh neg 7.3
25 Ezeli neg 1.2
36 George Hill neg 5.0
38 Fournier neg 6.8
49 Iguodala pos 14.6
51 Bogut neg 9.7
81 David Lee neg 9.4
101 Asik neg 10.8
104 Burke neg 6.5
112 Mozgov neg 12.2
113 Deng neg 7.5
133 Chalmers pos 19.8, pos 15.1
150 Pleiss neg 38.8
167 Korver pos 14.9
173 Hollis-Jefferson pos 14.0
204 Pressey pos 14.4
208 Ish Smith pos 16.3
218 Dellavedova pos 20.8
225 Richaun Holmes pos 13.2
258 Speights neg 23.5
309 Evan Turner pos 8.3
314 Hollis Thompson pos 19.2
330 Winslow pos 16.1
358 Raymond Felton pos 7.2
365 Devin Booker pos 4.8
381 Kobe Bryant pos 5.3

[Doctor MJ]

OK, that makes sense.

Hypothetically a player could have a negative impact on his team's defensive rebounding while still being a defensive positive (even with things like forced turnovers, foul rates, and blocks being equal), if his efforts force the opposition to lower% shots and lower eFG%. But surprisingly Jordan was only the 46th percentile in eFG% last year.

It strikes me odd that Jordan still rates out as a +2.09 defensive FFAPM when he's 46th percentile in eFG%, 42nd percentile in DREB, and 9th percentile in TOV rate. I assume that's because this metric is still partially informed by boxscore metrics?

You're missing the 4th of the 4 factors: Free throws. Jordan had a percentile of 100 on that front.

[Doctor MJ]

The more I look at RPM and RAPM the more I'm convinced that those two stats are fatally flawed.
They produce too many weird values. But if you know some values are wrong how can you trust the rest of the ratings?
In theory a properly calculated APM would be the ultimate stat. But I don't think anyone has managed to overcome the statistical problems so far.

I know it's RPM instead of RAPM, but the discrepancies are just too big for poor box score production to make up the difference. Just some of the latest unlikely values-

Rank--------Unadjusted +/-
16 Crowder neg 0.1
19 Bosh neg 7.3
25 Ezeli neg 1.2
36 George Hill neg 5.0
38 Fournier neg 6.8
49 Iguodala pos 14.6
51 Bogut neg 9.7
81 David Lee neg 9.4
101 Asik neg 10.8
104 Burke neg 6.5
112 Mozgov neg 12.2
113 Deng neg 7.5
133 Chalmers pos 19.8, pos 15.1
150 Pleiss neg 38.8
167 Korver pos 14.9
173 Hollis-Jefferson pos 14.0
204 Pressey pos 14.4
208 Ish Smith pos 16.3
218 Dellavedova pos 20.8
225 Richaun Holmes pos 13.2
258 Speights neg 23.5
309 Evan Turner pos 8.3
314 Hollis Thompson pos 19.2
330 Winslow pos 16.1
358 Raymond Felton pos 7.2
365 Devin Booker pos 4.8
381 Kobe Bryant pos 5.3

To me the key thing is to excise "fatally flawed" from the list of categories you put a stat under. The way to use statistics is to figure out how it can help your analysis. To do that it doesn't need to be flawless, you just need to have a decent sense of what its idiosyncrasies are.

As far as how to deal with raw +/- having a very different value from RPM, well to be clear to me the way to use this data is to look at both and everything in between to try to figure what's causing the numbers to be what they are. And when it comes to what's causing the numbers themselves, as oppose to what the numbers say about a player, you can always figure it out if you have enough information.

The most general thing is simply that if a team tends to do well when a guy is on the court but it's because he tends to be with strong teammates or weak opponents, then his raw +/- is probably going to be overrated.

To get into more specifics, I'd need a specific player to analyze - though I will say, I'm not that inclined to spend a ton of time sussing it out in November. If you wanted to ask about numbers for a player who played a lot last year, that seems more worthwhile to me.

[Knosh]

The more I look at RPM and RAPM the more I'm convinced that those two stats are fatally flawed.
They produce too many weird values. But if you know some values are wrong how can you trust the rest of the ratings?
In theory a properly calculated APM would be the ultimate stat. But I don't think anyone has managed to overcome the statistical problems so far.

I know it's RPM instead of RAPM, but the discrepancies are just too big for poor box score production to make up the difference. Just some of the latest unlikely values-

So RPM is "fatally flawed", because there are big discrepancies to unadjusted +/-? That only makes sense if unadjusted +/- would be a good stat in some way, which I don't think it is.

[laika]

The more I look at RPM and RAPM the more I'm convinced that those two stats are fatally flawed.
They produce too many weird values. But if you know some values are wrong how can you trust the rest of the ratings?
In theory a properly calculated APM would be the ultimate stat. But I don't think anyone has managed to overcome the statistical problems so far.

I know it's RPM instead of RAPM, but the discrepancies are just too big for poor box score production to make up the difference. Just some of the latest unlikely values-

Rank--------Unadjusted +/-
16 Crowder neg 0.1
19 Bosh neg 7.3
25 Ezeli neg 1.2
36 George Hill neg 5.0
38 Fournier neg 6.8
49 Iguodala pos 14.6
51 Bogut neg 9.7
81 David Lee neg 9.4
101 Asik neg 10.8
104 Burke neg 6.5
112 Mozgov neg 12.2
113 Deng neg 7.5
133 Chalmers pos 19.8, pos 15.1
150 Pleiss neg 38.8
167 Korver pos 14.9
173 Hollis-Jefferson pos 14.0
204 Pressey pos 14.4
208 Ish Smith pos 16.3
218 Dellavedova pos 20.8
225 Richaun Holmes pos 13.2
258 Speights neg 23.5
309 Evan Turner pos 8.3
314 Hollis Thompson pos 19.2
330 Winslow pos 16.1
358 Raymond Felton pos 7.2
365 Devin Booker pos 4.8
381 Kobe Bryant pos 5.3

To me the key thing is to excise "fatally flawed" from the list of categories you put a stat under. The way to use statistics is to figure out how it can help your analysis. To do that it doesn't need to be flawless, you just need to have a decent sense of what its idiosyncrasies are.

As far as how to deal with raw +/- having a very different value from RPM, well to be clear to me the way to use this data is to look at both and everything in between to try to figure what's causing the numbers to be what they are. And when it comes to what's causing the numbers themselves, as oppose to what the numbers say about a player, you can always figure it out if you have enough information.

The most general thing is simply that if a team tends to do well when a guy is on the court but it's because he tends to be with strong teammates or weak opponents, then his raw +/- is probably going to be overrated.

To get into more specifics, I'd need a specific player to analyze - though I will say, I'm not that inclined to spend a ton of time sussing it out in November. If you wanted to ask about numbers for a player who played a lot last year, that seems more worthwhile to me.

I tried to avoid players who had high minute teammates with better or worse respective values. I also avoided most of the low minute players. If you want to pick just one from this list, choose Winslow. There is simply no way that 80% of the league has been better than Winslow this year. Iguodala having the same rating as Bogut also seems impossible.

[Knosh]

Btw, where are you taking those numbers from laika?

[Doctor MJ]

The more I look at RPM and RAPM the more I'm convinced that those two stats are fatally flawed.
They produce too many weird values. But if you know some values are wrong how can you trust the rest of the ratings?
In theory a properly calculated APM would be the ultimate stat. But I don't think anyone has managed to overcome the statistical problems so far.

I know it's RPM instead of RAPM, but the discrepancies are just too big for poor box score production to make up the difference. Just some of the latest unlikely values-

Rank--------Unadjusted +/-
16 Crowder neg 0.1
19 Bosh neg 7.3
25 Ezeli neg 1.2
36 George Hill neg 5.0
38 Fournier neg 6.8
49 Iguodala pos 14.6
51 Bogut neg 9.7
81 David Lee neg 9.4
101 Asik neg 10.8
104 Burke neg 6.5
112 Mozgov neg 12.2
113 Deng neg 7.5
133 Chalmers pos 19.8, pos 15.1
150 Pleiss neg 38.8
167 Korver pos 14.9
173 Hollis-Jefferson pos 14.0
204 Pressey pos 14.4
208 Ish Smith pos 16.3
218 Dellavedova pos 20.8
225 Richaun Holmes pos 13.2
258 Speights neg 23.5
309 Evan Turner pos 8.3
314 Hollis Thompson pos 19.2
330 Winslow pos 16.1
358 Raymond Felton pos 7.2
365 Devin Booker pos 4.8
381 Kobe Bryant pos 5.3

To me the key thing is to excise "fatally flawed" from the list of categories you put a stat under. The way to use statistics is to figure out how it can help your analysis. To do that it doesn't need to be flawless, you just need to have a decent sense of what its idiosyncrasies are.

As far as how to deal with raw +/- having a very different value from RPM, well to be clear to me the way to use this data is to look at both and everything in between to try to figure what's causing the numbers to be what they are. And when it comes to what's causing the numbers themselves, as oppose to what the numbers say about a player, you can always figure it out if you have enough information.

The most general thing is simply that if a team tends to do well when a guy is on the court but it's because he tends to be with strong teammates or weak opponents, then his raw +/- is probably going to be overrated.

To get into more specifics, I'd need a specific player to analyze - though I will say, I'm not that inclined to spend a ton of time sussing it out in November. If you wanted to ask about numbers for a player who played a lot last year, that seems more worthwhile to me.

I tried to avoid players who had high minute teammates with better or worse respective values. I also avoided most of the low minute players. If you want to pick just one from this list, choose Winslow. There is simply no way that 80% of the league has been better than Winslow this year. Iguodala having the same rating as Bogut also seems impossible.

Winslow:

1) Early season, low sample size.
2) Bench player, likely playing against weaker competition.
3) Not box score driven, and RPM uses box score in prior. (This incidentally is something that really bugs me.)
4) Rookie? I don't believe that RPM uses a prior from previous seasons, but if it does, such metrics have issues with rookies.

Iguodala vs Bogut:

Aside from all the sample size concerns, note that their offense & defense aren't the same at all. Iggy has a considerable edge on offense right now, Bogut on defense. I think the more informative analysis will be focus on offense specifically or defense specifically. Regardless these are both veterans that we already have a sense of how they "look" in +/- so to be honest I'm not really looking at them very closely with this stat right now.

[laika]

Btw, where are you taking those numbers from laika?

http://stats.nba.com/team/#!/1610612748/onoffcourt/?sort=PLUS_MINUS&dir=1

You have to go to every team and then add up the Oncourt + Offcourt for each player.

[trex_8063]

OK, that makes sense.

Hypothetically a player could have a negative impact on his team's defensive rebounding while still being a defensive positive (even with things like forced turnovers, foul rates, and blocks being equal), if his efforts force the opposition to lower% shots and lower eFG%. But surprisingly Jordan was only the 46th percentile in eFG% last year.

It strikes me odd that Jordan still rates out as a +2.09 defensive FFAPM when he's 46th percentile in eFG%, 42nd percentile in DREB, and 9th percentile in TOV rate. I assume that's because this metric is still partially informed by boxscore metrics?

You're missing the 4th of the 4 factors: Free throws. Jordan had a percentile of 100 on that front.

I wouldn't think it sufficient to pull him up to being a significant net positive when he's below average in the other three (of four) factors, and significantly so in the turnover category (he's nearly equally poor in that one as he is elite in the FT factor). Further, FT's is supposed to be the LEAST weighted in importance of the three:

http://www.basketball-reference.com/about/factors.html


In summation, his rank in this fourth factor (which I was aware of prior to my last post) doesn't seem an adequate explanation to me. It's not like he's barely above neutral, either; +2.1 is significant.

EDIT: For instance, if we weight things according to the approximations stated in the link I provided above to just sort of average out his percentile rank OVERALL (not exactly how the formula works, I know), Jordan comes out approximately in the 44th percentile. That is: below average.

Yet his defensive SSAPM is +2.09.
Obv I can investigate for myself (and will do so if necessary) the reason why. Was hoping one of you had a quick explanation, though.

[laika]

Winslow:

1) Early season, low sample size.
2) Bench player, likely playing against weaker competition.
3) Not box score driven, and RPM uses box score in prior. (This incidentally is something that really bugs me.)
4) Rookie? I don't believe that RPM uses a prior from previous seasons, but if it does, such metrics have issues with rookies.

Iguodala vs Bogut:

Aside from all the sample size concerns, note that their offense & defense aren't the same at all. Iggy has a considerable edge on offense right now, Bogut on defense. I think the more informative analysis will be focus on offense specifically or defense specifically. Regardless these are both veterans that we already have a sense of how they "look" in +/- so to be honest I'm not really looking at them very closely with this stat right now.

1.I don't think sample size explains it. A low sample size would result in more extreme values. You would not necessarily expect to see values completely reversed from what is expected.
This brings up another reason to object to RPM. The values are suspiciously constrained. This year after 17 games the high is 8.85 and the low is 6.0. Last year the high was 9.34 and the low neg 7.5. So no players are having a more extreme season after 17 games than players last year did after an entire season? That is really unlikely.
Also, I forgot to add Curry to my list. Curry is having one of the best box score seasons ever with a better +/- on the best team ever. Yet miraculously RPM thinks that Curry is worse this year than last year.

2.Winslow is not a bench player. He has the 4th most minutes on the team and plays a lot with the starters.

3.Winslow's unadjusted numbers are top 5%. Winslow's box score stats aren't great, but there is no conceivable way that a box score factor could drag him all the way down to the bottom 20%.

4.Maybe if the prior states-"all rookies are assumed to be the worst player in the league"

If you look closer at Winslow's numbers it looks even worse. Among the top 9 players by minutes on the team no one has anywhere close to the +/- number Winslow has. The next best rating is Dragic at pos 3.2. Winslow's RPM rating might make some sense if Deng was the worst player in the history of the universe since that is who Winslow replaces the most. But amazingly, RPM ranks Deng at 113-far better than Winslow.

It's not a mathematical proof, but the Winslow rating is so crazy that it's reasonable to assert that Winslow is proof that RPM is producing false values.

[Swagalicious]

It's not a mathematical proof, but the Winslow rating is so crazy that it's reasonable to assert that Winslow is proof that RPM is producing false values.

There have been a plethora of examples just like that in the past. I remember when Andray Blatche was amongst the best players in the league in early 2013 lol, etc etc. In the end, the pro-RPM crowd will dismiss inconsistencies as irrelevant. Instead of trying to figure out how to identify and adjust for the flaws that produce such skewed results, we'll get the usual "it's an outlier" type of explanation. It'd take something like LeBron right now posting negative RPM to change their minds.

[Ryoga Hibiki]

I feel it's missing, with this kind of stats, something measuring the variance, also to clarify what small sample size means.

[Doctor MJ]

OK, that makes sense.

Hypothetically a player could have a negative impact on his team's defensive rebounding while still being a defensive positive (even with things like forced turnovers, foul rates, and blocks being equal), if his efforts force the opposition to lower% shots and lower eFG%. But surprisingly Jordan was only the 46th percentile in eFG% last year.

It strikes me odd that Jordan still rates out as a +2.09 defensive FFAPM when he's 46th percentile in eFG%, 42nd percentile in DREB, and 9th percentile in TOV rate. I assume that's because this metric is still partially informed by boxscore metrics?

You're missing the 4th of the 4 factors: Free throws. Jordan had a percentile of 100 on that front.

I wouldn't think it sufficient to pull him up to being a significant net positive when he's below average in the other three (of four) factors, and significantly so in the turnover category (he's nearly equally poor in that one as he is elite in the FT factor). Further, FT's is supposed to be the LEAST weighted in importance of the three:

http://www.basketball-reference.com/about/factors.html


In summation, his rank in this fourth factor (which I was aware of prior to my last post) doesn't seem an adequate explanation to me. It's not like he's barely above neutral, either; +2.1 is significant.

EDIT: For instance, if we weight things according to the approximations stated in the link I provided above to just sort of average out his percentile rank OVERALL (not exactly how the formula works, I know), Jordan comes out approximately in the 44th percentile. That is: below average.

Yet his defensive SSAPM is +2.09.
Obv I can investigate for myself (and will do so if necessary) the reason why. Was hoping one of you had a quick explanation, though.

Well to be honest it's not something I've thought too much about. I largely use the site for their 4 factors breakdown. Here's what I would say though to your particular concern:

When the Four Factor say that something is weighted least, that doesn't necessarily mean it's going to be the smallest impactor of the total. The really all depends on the variance involved with each of the different factors. As such, it's certainly possible to have the overall FF edge simply based on being better in only 1 of the 4 factors if your edge is big enough.

And here's where we run into issues reverse engineering the Got Buckets FF work: They don't give us actual numbers, they translate it into percentiles. In doing so at the very least we lose some sense of how important each of the factors truly is. So I don't look at these numbers I don't think "that's impossible, they must have made a mistake", I just think "we can't see every thing from this view of the data".

[Doctor MJ]

Winslow:

1) Early season, low sample size.
2) Bench player, likely playing against weaker competition.
3) Not box score driven, and RPM uses box score in prior. (This incidentally is something that really bugs me.)
4) Rookie? I don't believe that RPM uses a prior from previous seasons, but if it does, such metrics have issues with rookies.

Iguodala vs Bogut:

Aside from all the sample size concerns, note that their offense & defense aren't the same at all. Iggy has a considerable edge on offense right now, Bogut on defense. I think the more informative analysis will be focus on offense specifically or defense specifically. Regardless these are both veterans that we already have a sense of how they "look" in +/- so to be honest I'm not really looking at them very closely with this stat right now.

1.I don't think sample size explains it. A low sample size would result in more extreme values. You would not necessarily expect to see values completely reversed from what is expected.
This brings up another reason to object to RPM. The values are suspiciously constrained. This year after 17 games the high is 8.85 and the low is 6.0. Last year the high was 9.34 and the low neg 7.5. So no players are having a more extreme season after 17 games than players last year did after an entire season? That is really unlikely.
Also, I forgot to add Curry to my list. Curry is having one of the best box score seasons ever with a better +/- on the best team ever. Yet miraculously RPM thinks that Curry is worse this year than last year.

2.Winslow is not a bench player. He has the 4th most minutes on the team and plays a lot with the starters.

3.Winslow's unadjusted numbers are top 5%. Winslow's box score stats aren't great, but there is no conceivable way that a box score factor could drag him all the way down to the bottom 20%.

4.Maybe if the prior states-"all rookies are assumed to be the worst player in the league"

If you look closer at Winslow's numbers it looks even worse. Among the top 9 players by minutes on the team no one has anywhere close to the +/- number Winslow has. The next best rating is Dragic at pos 3.2. Winslow's RPM rating might make some sense if Deng was the worst player in the history of the universe since that is who Winslow replaces the most. But amazingly, RPM ranks Deng at 113-far better than Winslow.

It's not a mathematical proof, but the Winslow rating is so crazy that it's reasonable to assert that Winslow is proof that RPM is producing false values.

1. Well I'm right there with you in frustration with some of the black box nature of RPM. It's been implemented and adopted because it produces more reliable results than APM/RAPM, but I prefer to understand more clearly just what's going on. However since it's the metric that we now have most access to, I certainly look at it for a starting point.

Re: don't think sample size explains it. Well obviously sample size wasn't the only thing I listed.

2. Winslow plays a lot with starters. Well he spends most of the 1st quarter on the bench for example. I'm not sure where I can find the data to do the full on the analysis I'd like, but if you're a player who comes in at a time where the other team tends to be making substitutions, even if your team isn't, then you're playing against weaker competition.

I'm not saying that RPM is right here, but so long as a player isn't starting and isn't even playing 30+ MPG as a young player without injury/stamina concerns, I tend to be pretty cautious in my +/- evaluations.

3. no conceivable way box score. I'd sure hope not, but I will say that if you looked at Russell Westbrook last year in the RAPM studies I saw, he looked nowhere near as good as he did in RPM. That had to be related to his "every play I make is for a stat" game which he had going last year, and is the case in point to why I wish we had more regular access to non-box score infected regression numbers.

4. You had me basically with you until you said "false values". What does that mean exactly? Basically everyone would be in agreement that it makes no sense to use Winslow's RPM against him right now, but you seem to be suggesting something a lot stronger.

[laika]

1. Well I'm right there with you in frustration with some of the black box nature of RPM. It's been implemented and adopted because it produces more reliable results than APM/RAPM, but I prefer to understand more clearly just what's going on. However since it's the metric that we now have most access to, I certainly look at it for a starting point.

Re: don't think sample size explains it. Well obviously sample size wasn't the only thing I listed.

2. Winslow plays a lot with starters. Well he spends most of the 1st quarter on the bench for example. I'm not sure where I can find the data to do the full on the analysis I'd like, but if you're a player who comes in at a time where the other team tends to be making substitutions, even if your team isn't, then you're playing against weaker competition.

I'm not saying that RPM is right here, but so long as a player isn't starting and isn't even playing 30+ MPG as a young player without injury/stamina concerns, I tend to be pretty cautious in my +/- evaluations.

3. no conceivable way box score. I'd sure hope not, but I will say that if you looked at Russell Westbrook last year in the RAPM studies I saw, he looked nowhere near as good as he did in RPM. That had to be related to his "every play I make is for a stat" game which he had going last year, and is the case in point to why I wish we had more regular access to non-box score infected regression numbers.

4. You had me basically with you until you said "false values". What does that mean exactly? Basically everyone would be in agreement that it makes no sense to use Winslow's RPM against him right now, but you seem to be suggesting something a lot stronger.

1.RPM probably is more reliable than RAPM. But it seems that the advanced stats community has a ways to go before producing a truly reliable all-in-one stat. I don't know enough about statistics to say whether this hasn't happened because it is difficult, or that it is nearly impossible due to statistical and sample limitations.

2.Don't know where to find opposing lineups. Can find most teammate lineups at nba.com. It seems that opposing lineups would average out to a greater degree than own team lineups.

3.If it's a black box we won't know for sure. But it would take some seriously extreme box score weighting to push Winslow down that far. It's not just that RPM is weighting players lower than I would like. It seems to be internally inconsistent. Exum got a good RAPM last year with horrible box score numbers. Yet his RPM was much better than Winslow's this year. Both of them were good defensive rookies with poor box score numbers that played extensively with the starters and bench. It seems that Winslow has played much better than Exum though. I don't see how it's possible that Winslow is rated much lower than such a similar(but significantly worse) player.

4.You can replace "false" with "unreliable" if you prefer. Whatever the semantics, it seems that RPM is not that close to an ideal stat. I'm moving to the position that there is no such thing as a truly reliable all in one stat.

[Doctor MJ]

1. Well I'm right there with you in frustration with some of the black box nature of RPM. It's been implemented and adopted because it produces more reliable results than APM/RAPM, but I prefer to understand more clearly just what's going on. However since it's the metric that we now have most access to, I certainly look at it for a starting point.

Re: don't think sample size explains it. Well obviously sample size wasn't the only thing I listed.

2. Winslow plays a lot with starters. Well he spends most of the 1st quarter on the bench for example. I'm not sure where I can find the data to do the full on the analysis I'd like, but if you're a player who comes in at a time where the other team tends to be making substitutions, even if your team isn't, then you're playing against weaker competition.

I'm not saying that RPM is right here, but so long as a player isn't starting and isn't even playing 30+ MPG as a young player without injury/stamina concerns, I tend to be pretty cautious in my +/- evaluations.

3. no conceivable way box score. I'd sure hope not, but I will say that if you looked at Russell Westbrook last year in the RAPM studies I saw, he looked nowhere near as good as he did in RPM. That had to be related to his "every play I make is for a stat" game which he had going last year, and is the case in point to why I wish we had more regular access to non-box score infected regression numbers.

4. You had me basically with you until you said "false values". What does that mean exactly? Basically everyone would be in agreement that it makes no sense to use Winslow's RPM against him right now, but you seem to be suggesting something a lot stronger.

1.RPM probably is more reliable than RAPM. But it seems that the advanced stats community has a ways to go before producing a truly reliable all-in-one stat. I don't know enough about statistics to say whether this hasn't happened because it is difficult, or that it is nearly impossible due to statistical and sample limitations.

2.Don't know where to find opposing lineups. Can find most teammate lineups at nba.com. It seems that opposing lineups would average out to a greater degree than own team lineups.

3.If it's a black box we won't know for sure. But it would take some seriously extreme box score weighting to push Winslow down that far. It's not just that RPM is weighting players lower than I would like. It seems to be internally inconsistent. Exum got a good RAPM last year with horrible box score numbers. Yet his RPM was much better than Winslow's this year. Both of them were good defensive rookies with poor box score numbers that played extensively with the starters and bench. It seems that Winslow has played much better than Exum though. I don't see how it's possible that Winslow is rated much lower than such a similar(but significantly worse) player.

4.You can replace "false" with "unreliable" if you prefer. Whatever the semantics, it seems that RPM is not that close to an ideal stat. I'm moving to the position that there is no such thing as a truly reliable all in one stat.

Just focusing on reliability:

There is reliability and there is validity. The first is about precision or consistency, the second is about true accuracy.

The reason why people gets so frustrated with +/- stats compared to the box score is that with the box score it's actually not that hard to have reliable stats, and +/- is considerably weaker on that front. So yeah, it's not an issue where the community is failing at reliability, it's reliability in and of itself is meaningless. LeBron was the same jersey number every game, but that says nothing about his impact.

In the chase for validity, +/- stats provide a window that simply can't be ignored, and so we use it, but use it carefully.