Compilation of Michael Jordan’s on-off data with the Bulls

[lessthanjake]

~Thinking Basketball's Actual Jordan Playoff Plus/Minus Data~

Over the holidays, I was able to contact Thinking Basketball, who was kindly willing to share the actual plus minus data from his 'I tracked every Michael Jordan playoff game for a decade' video. This is Jordan's Plus/Minus data for every playoffs from 1988–1998. Of course, full credit goes to Thinking Basketball and the rest of the team that tracked this data. It's due to the ambitious work of basketball historians like TB and Squared2020 that we're able to get this entirely new insight into past players. And plus, learning about the old stats is just fun! So big thanks to the team that tracked this

Here's the raw data copied directly from Ben (I've added commas between column headers for readability) . First the year by year data:

Spoiler:

PS, On Net, On Min, Off Net (at check in time), Off Min, On/48, Off/48, Net/48, Year, Opp RS Net, Adj On
1988 -33 431.96 -8 48.04 -3.7 -8.0 4.3 1988 3.0 -0.7
1989 63 726.76 -41 99.24 4.2 -19.8 24.0 1989 5.7 9.8
1990 100 678.66 -47 86.67 7.1 -26.0 33.1 1990 4.2 11.3
1991 196 694.623333 3 126.376667 13.5 1.1 12.4 1991 2.8 16.4
1992 135 917.37 15 143.63 7.1 5.0 2.1 1992 4.3 11.3
1993 120 786.29 -9 140.71 7.3 -3.1 10.4 1993 5.4 12.7
1995 -15 421.76 24 63.24 -1.7 18.2 -19.9 1995 5.8 4.1
1996 199 731.16 -9 137.84 13.1 -3.1 16.2 1996 5.0 18.1
1997 136 804 -31 108 8.1 -13.8 21.9 1997 6.6 14.7
1998 158 872 -11 146 8.7 -3.6 12.3 1998 5.1 13.8
That's by year. Let me get by series...

Now for the series by series data:

Spoiler:

SERIES, On Net, On Min, Off Net (at check in time), Off Min, On/48, Off/48, Net/48
CLE88 16 217.68 -8 22.32 3.5 -17.2 20.7
DET88 -49 214.28 0 25.72 -11.0 0.0 -11.0
CLE89 22 216.36 -18 28.64 4.9 -30.2 35.0
DET89 -19 258.61 12 29.39 -3.5 19.6 -23.1
NYK89 65 240.29 -30 38.38 13.0 -37.5 50.5
DET90 -13 297.17 -8 38.08 -2.1 -10.1 8.0
MIL90 68 169.62 -30 22.38 19.2 -64.3 83.6
PHI90 45 211.87 -9 26.21 10.2 -16.5 26.7
PHI91 29 172.58 0 35.75 8.1 0.0 8.1
NYK91 48 113.643333 12 30.356667 20.3 19.0 1.3
DET91 47 160.32 -1 31.68 14.1 -1.5 15.6
LAL91 58 224.08 -9 20.92 12.4 -20.7 33.1
MIA92 56 118.77 -2 25.23 22.6 -3.8 26.4
NYK92 39 302.06 2 33.94 6.2 2.8 3.4
CLE92 20 245.97 -9 42.03 3.9 -10.3 14.2
POR92 20 250.57 24 42.43 3.8 27.2 -23.3
ATL93 55 103.39 -6 40.61 25.5 -7.1 32.6
CLE93 31 154.27 3 37.73 9.6 3.8 5.8
NYK93 25 251.09 3 36.91 4.8 3.9 0.9
PHO93 9 277.54 -9 25.46 1.6 -17.0 18.5
CHA95 22 165.42 -8 31.58 6.4 -12.2 18.5
ORL95 -37 256.34 32 31.66 -6.9 48.5 -55.4
MIA96 72 99.45 -3 44.55 34.8 -3.2 38.0
NYK96 30 216.24 1 28.76 6.7 1.7 5.0
ORL96 53 163.52 14 28.48 15.6 23.6 -8.0
SEA96 44 251.95 -21 36.05 8.4 -28.0 36.3

And finally the 3-year AuPM approximation:

Spoiler:

We can't technically know his AuPM, but here are 3-year estimates per game:

1990 +7.7
1991 +8.7
1992 +7.1
1993 +6.3
1995 +5.5
1996 +5.5
1997 +5.6
1998 +6.7

Now just for readability, here's a summary that's in a slightly more legible format:
Playoffs: Net On/off per 48 minutes (On per 48 minutes, Adjusted On relative to opponent's average MoV per 48 minutes)
1988: +4.3 (-3.7, -0.7)
1989: +24 (+4.2, +9.8)
1990: +33.1 (+7.1, +11.3)
1991: +12.4 (+13.5, +16.4)
1992: +2.1 (+7.1, +11.3)
1993: +10.4 (+7.3, +12.7)
1995: -19.9 (-1.7, +4.1)
1996: +16.2 (+13.1, +18.1)
1997: +21.9 (+8.1, +14.7)
1998: +12.3 (+8.7, +13.8)

3 year AuPM (estimate)
1990: +7.7
1991: +8.7
1992: +7.1
1993: +6.3
1995: +5.5
1996: +5.5
1997: +5.6
1998: +6.7
*note: year listed is the final year, so 1990 means 1988–1990. Why are these an estimate and not the true value? AuPM takes a player's net plus minus data over a timespan (e.g. a playoff run) and estimate their APM/RAPM, without the full play-by-play data needed to actually calculate APM/RAPM. Pros: it's much stabler in small samples, like a playoff run (unlike APM/RAPM), and it does a pretty good job at approximating the long term APM/RAPM. To do this, AuPM needs to approximate the 'adjustment' part of APM/RAPM, without the play-by-play. This requires knowing the total plus minus data of a player's teammates, but since only Jordan's plus minus data was tracked in these playoffs, we don't have that information. Specifically, we apply an adjustment if a player's teammate has a higher value than them. See original article for details (https://thinkingbasketball.net/2017/09/18/augmented-plus-minus-evaluating-old-pm-data/). We can estimate this (e.g. it's super unlikely that someone would have a better on/off in the 1990 postseason... it's super likely that someone has a better on/off in the 1995 postseason) to get realistic AuPM estimates, so that would inform the estimate. Thus this is a most-likely estimate of 3 year AuPM, until we can get the plus minus data for all Jordan's teammates too.

3-year data:

Spoiler:

3-year On per 48 minutes, 3 year off per 48 minute, 3 year Net on/off per 48 minutes:
1990: +3.4, -19.7, +23.1
1991: +8.2, -13.1, +21.3
1992: +9.0, -3.9, +13
1993: +9.0, +1.0, +8.0
1995: +5.4, +4.1, +1.3
1996: +7.5, +0.8, +6.7
1997: +7.9, -2.5, +10.3
1998: +9.8, -6.2, +16.1
*note: these are calculated by me, but the calculation's fairly simple. Feel free to check if you doubt.

To close, I'll add a little personal analysis here.

Spoiler:

So it looks like the playoffs would rank:
Net/48: 90 > 89 > 97 > 96 > 91 > 98 > 93 > 92 > 88 > 95
Adj On: 96 > 91 > 97 > 98 > 93 > 92 > 90 > 89 > 95 > 88
3-year AuPM: 89–91 > 88–90 > 90–92 > 96–98 > 91–93 > 93–95 > 94–96 > 95–97

-Consistency: Impressive that only 1995 had a negative Net/48. We also have Net/48 estimates for 1985 and 1986, and those are both positive, so all that leaves is 1987 missing (note: I wrote this before Djoker added 1987... it looks like Dojoker found 1987 was positive too!). If ’87 is positive, Jordan would join Curry as the only modern top 10 candidate to have just one negative Net/48 playoff run. LeBron, Shaq (post-97), and Kobe have one negative Net/48 playoff run in their prime, and more pre/post-prime. Duncan and Garnett had multiple negative Net/48 playoff runs in their prime. Many might consider 1995 Jordan to be not fully healthy / in shape, so if you discount unhealthy years, only Jordan and Curry have positive playoff Net/48 in every prime healthy year.

-Peak: 1989–1991 stands out as his peak in AuPM, especially 1990 by Net/48 and 1991 by Adj On & BPM. Surprising dip in 1988, which looks like one of his worst playoffs — I wonder how much is this is noise or signal.

-1st 3-peat vs 2nd 3-peat: it’s slightly surprising 91–93 has a lower AuPM than 96–98. It looks like 1996 and 1997 are right near his peak level, while 1992 (and to a lesser extent 1993) are lower in Net/48 and Adj On. The stats are close enough that noise could explain the difference. But if there’s signal, I wonder whether this says something about 92/93 Jordan’s value, the Bulls’ rotations, or their dependence on Jordan’s role in either of the 3-peats.

If the AuPM estimates here are in per game units, that would give Jordan the 1st, 4th, and 9th best 3-year AuPM runs ever (Duncan’s 2nd at +8.5). Dominant stuff either way!

.

Hey Jake! Feel free to update the OP if you'd like! Using these actual numbers should be more accurate than our by-eye estimates of the plots. Let me know if you have any questions!

This is fantastic info! Thanks for getting it! I will aim to eventually fully update my OP with it. It’d take me a fair bit of time to actually fully do that, because the data is in per-48-minutes terms and I’d tried to convert that to per-100-possessions terms in the OP—which requires pulling some data and making some additional calculations (I unfortunately didn’t keep my original calculations saved in an organized fashion, so that’d make it take longer than it should). And I’m actually not sure anymore that that’s the best way of doing things (since I’ve started to become a bit skeptical of Basketball-Reference’s pace-related data). So if I update the OP with this more precise data, I’m not sure exactly what form it should take (i.e. just leave it in per-48-minute terms or try to convert it to per-100-possession terms). I’m inclined to leave it in per-48-minute terms and just report out the total playoff on-off fully on a per-48-minute basis, in order to avoid potentially serious issues regarding conversion of the data. But that could also be a bit confusing, since some of the other data I’ve cited in the OP for regular season is on a per-100-possession basis, so the post would be using two different types of measuring sticks (not to mention that people are used to per-100-possession on-off values, so it’s not immediately clear exactly how good a number is when reported on a per-48-minute basis). That could be confusing. So I have to think about how best to do it. If you’ve got any particular views on that, I’m certainly happy to hear them! In the meantime, I’ve calculated the per-48-minute data (as per the below) and added that info as a note in my OP (crediting you and Djoker with the extremely helpful info you’ve provided), along with an explanation about how I’m still deciding whether and when to try to convert that to per-100-possession terms. I’ve also added notes in a few places in the OP pointing to the data you reported, as more accurate with regards to the Thinking Basketball data, and noting that the numbers in the rest of the post haven’t actually yet been updated to account for those more accurate numbers.

For now, as alluded to above, we can actually calculate Jordan’s career playoff on-off in per-48-minute terms:

To do this, I’ve taken: (1) the “on” plus-minus, “off” plus-minus, “on” minutes, and “off” minutes you provided above from Thinking Basketball, (2) the “on” and “off” values derived from Squared’s data from the 1984-1985 playoffs (I think he’s taken that data down now—perhaps as a result of arguments on these forums—but I do have record of that, which showed +10 “on” and therefore -32 “off”), (3) the “on” and “off” values for 1986 and 1987 from Djoker’s post (with the correction that Djoker later made after your post—see https://forums.realgm.com/boards/viewtopic.php?p=110650813#p110650813); and (4) the minutes on and off for 1985, 1986, and 1987 from Basketball-Reference.

And I’ve used that data to compile a full per-48-minute on-off for Jordan in the playoffs:

Michael Jordan Career Playoff On-Off Per 48 Minutes
- On: +6.6829 (+1044 in 7498.58333 minutes)
- Off: -7.3914 (-178 in 1155.936667 minutes)
- On-Off per 48 minutes: +14.0743

Very impressive stuff! And, of course, it’s important to recognize that this is in per-48-minute terms. Since there was virtually always fewer than 100 possessions per 48 minutes on average, it is highly likely that the per-100-possession on-off would be a bit higher—without doing the full calculations, a good guess would probably be around a +15 or +16 on-off per 100 possessions (which is consistent with Djoker having calculated a +16.15 per 100 possessions before we got this more precise data).

[VanWest82]

It was both.

Not really. It was peak Jordan with Pippen and Grant playing the best they'd played to that point. I went through the 92 season and sorted average MOV by quarter. Here are the results.

Q1: +3.1829
Q2: +1.1219
Q3: +4.0854
Q4: +1.9756

1st and 3rd quarters give us the best representation of the starters and 2nd quarter gives us the best representation of the bench. 92 Bulls had some blowouts so 4th quarter numbers might be a bit misleading. Either way, even though it's not definitive, it sure looks like starters carried that team, and bench players at best held serve. By contrast, here are the 94 team's MOV by quarter.

Q1: +0.8780
Q2: +1.4512
Q3: +0.2439
Q4: +0.6585

Here we see that Bulls bench is the one propping up the starters.

It also had a lot to do with Jordan’s replacement as starter being utterly abysmal. Do you think Kerr and Wennington were actually producing much on their own, or were they getting the benefit of the team’s best players while not being dragged down by its weakest links? Yet again, care so much about “context”, just not enough to really engage with it.

In the other Jordan plus/minus thread that got locked, I used squared's data to calculate on/offs for 91 and 93. It showed that Bulls offense fell off a cliff with Jordan on the bench (97 ORTG in 91, 100 ORTG in 93). These would've ranked dead last by a mile (91) and 2nd last (93). This was Bulls major issue leading up to Jordan's first retirement. They just couldn't muster NBA level offense when Jordan wasn't on the court, but were still ok minus Scottie and Horace (105-106 ORTG) .

You're arguing that Bulls went from not being able to sustain NBA offense to Scottie and Horace carrying bad line ups even though they hadn't been able to do it prior which we can even see in games Jordan missed from 91-93 where they go 1-5 with Scottie and Horace playing. Meanwhile, Kerr, Kukoc, and Longley are all known as offensive players. Kukoc, in particular, shows up repeatedly in impact data from that period. Myers wasn't good, but neither was Paxon in 93. Cartwright wasn't good but he wasn't good anymore in 93 either. Both those guys looked very washed and we can see just how much their on-offs fell off from the graph I posted. So No, I don't see a reasonable case for Scottie and Horace carrying bad line ups offensively; more like, the new players in 94 were quality offensive players who added significant value on that end and helped buoy Scottie and Horace who hadn't shown an ability to carry offensive line ups to that point. I do think it's fair to credit them for 94 Bulls defense, however.

Your argument for role as a way to handwave Scottie and Horace's improving on-offs in the mid 90s doesn't hold water either. If anything, the fact they matched their 91 on-offs in 94 playing a greater role likely means their impact was greater, and having the largest on-offs of their careers in 95-96 playing essentially the same roles they played in early 90s also supports that they continued to get better. Claiming they were just as good in early 90s and therefore always ready to soak up additional role and attention is a massive leap given what we saw from them in 93.

You've accused me of going off vibes, using strawmen even though it was you misrepresenting my point, not supporting my case with evidence, etc. As far as I can see, I'm the one actually providing data and supporting evidence. You're the one going off vibes and narratives, and incorrect ones at that. And between the two of us, it's pretty clear which one actually followed that team in real time and which one is attempting to interpret what happened way after the fact.

[AEnigma]

It was both.

Not really. It was peak Jordan with Pippen and Grant playing the best they'd played to that point. I went through the 92 season and sorted average MOV by quarter. Here are the results.

Q1: +3.1829
Q2: +1.1219
Q3: +4.0854
Q4: +1.9756

1st and 3rd quarters give us the best representation of the starters and 2nd quarter gives us the best representation of the bench. 92 Bulls had some blowouts so 4th quarter numbers might be a bit misleading. Either way, even though it's not definitive, it sure looks like starters carried that team, and bench players at best held serve.

… Yeah, that is a good bench performance...

By contrast, here are the 94 team's MOV by quarter.

Q1: +0.8780
Q2: +1.4512
Q3: +0.2439
Q4: +0.6585

Here we see that Bulls bench is the one propping up the starters.

This would be a lot easier if you could bother to pay attention rather than just railroad your preset narratives. Steve Kerr and Toni Kukoc are “bench players” playing as many minutes as “starter” Pete Myers, and you want to pretend there is some deep meaning to those starting lineups not being as successful as other lineups.

It also had a lot to do with Jordan’s replacement as starter being utterly abysmal. Do you think Kerr and Wennington were actually producing much on their own, or were they getting the benefit of the team’s best players while not being dragged down by its weakest links? Yet again, care so much about “context”, just not enough to really engage with it.

In the other Jordan plus/minus thread that got locked, I used squared's data to calculate on/offs for 91 and 93. It showed that Bulls offense fell off a cliff with Jordan on the bench (97 ORTG in 91, 100 ORTG in 93). These would've ranked dead last by a mile (91) and 2nd last (93). This was Bulls major issue leading up to Jordan's first retirement. They just couldn't muster NBA level offense when Jordan wasn't on the court, but were still ok minus Scottie and Horace (105-106 ORTG) .

You're arguing that Bulls went from not being able to sustain NBA offense to Scottie and Horace carrying bad line ups even though they hadn't been able to do it prior which we can even see in games Jordan missed from 91-93 where they go 1-5 with Scottie and Horace playing. Meanwhile, Kerr, Kukoc, and Longley are all known as offensive players. Kukoc, in particular, shows up repeatedly in impact data from that period.

More not bothering to track the point and more grouping Pippen and Grant together. Yes, Pippen improved. Yes, they had a rotational skew toward offence among non-starters in 1994. Quantify it.

Myers wasn't good, but neither was Paxon in 93.

Paxson started 8 games and was eighth on the team in minutes played.

Cartwright wasn't good but he wasn't good anymore in 93 either. Both those guys looked very washed and we can see just how much their on-offs fell off from the graph I posted. So No, I don't see a reasonable case for Scottie and Horace carrying bad line ups offensively;

Which as we know is the only way teams are carried.

more like, the new players in 94 were quality offensive players who added significant value on that end and helped buoy Scottie and Horace who hadn't shown an ability to carry offensive line ups to that point. I do think it's fair to credit them for 94 Bulls defense, however.

The result is what matters, not the distribution. Whether Pippen and Grant could have managed an offence with worse offensive players around them is irrelevant when those positive offensive players are being covered defensively.

Your argument for role as a way to handwave Scottie and Horace's improving on-offs in the mid 90s doesn't hold water either. If anything, the fact they matched their 91 on-offs in 94 playing a greater role likely means their impact was greater, and having the largest on-offs of their careers in 95-96 playing essentially the same roles they played in early 90s also supports that they continued to get better. Claiming they were just as good in early 90s and therefore always ready to soak up additional role and attention is a massive leap given what we saw from them in 93.

Okay, then I guess your argument falls apart because 1994 Pippen is only his fourth best on/off season. Another good job.

You've accused me of going off vibes, using strawmen even though it was you misrepresenting my point, not supporting my case with evidence, etc. As far as I can see, I'm the one actually providing data and supporting evidence.

Selective evidence designed only to push what you want to argue in the moment, yes.

You're the one going off vibes and narratives, and incorrect ones at that.

No, I am just not trying to create the rosiest possible picture using assumptions I obviously do not even care to apply consistently.

And between the two of us, it's pretty clear which one actually followed that team in real time and which one is attempting to interpret what happened way after the fact.

We both followed the team, but one of us had a built-in assumption of what would happen without Jordan and has spent the past thirty years trying to find the tidiest way to explain the disconnect, and the other was willing to adapt based on what they saw.

[DraymondGold]

~Thinking Basketball's Actual Jordan Playoff Plus/Minus Data~

Over the holidays, I was able to contact Thinking Basketball, who was kindly willing to share the actual plus minus data from his 'I tracked every Michael Jordan playoff game for a decade' video. This is Jordan's Plus/Minus data for every playoffs from 1988–1998. Of course, full credit goes to Thinking Basketball and the rest of the team that tracked this data. It's due to the ambitious work of basketball historians like TB and Squared2020 that we're able to get this entirely new insight into past players. And plus, learning about the old stats is just fun! So big thanks to the team that tracked this

Here's the raw data copied directly from Ben (I've added commas between column headers for readability) . First the year by year data:

Spoiler:

PS, On Net, On Min, Off Net (at check in time), Off Min, On/48, Off/48, Net/48, Year, Opp RS Net, Adj On
1988 -33 431.96 -8 48.04 -3.7 -8.0 4.3 1988 3.0 -0.7
1989 63 726.76 -41 99.24 4.2 -19.8 24.0 1989 5.7 9.8
1990 100 678.66 -47 86.67 7.1 -26.0 33.1 1990 4.2 11.3
1991 196 694.623333 3 126.376667 13.5 1.1 12.4 1991 2.8 16.4
1992 135 917.37 15 143.63 7.1 5.0 2.1 1992 4.3 11.3
1993 120 786.29 -9 140.71 7.3 -3.1 10.4 1993 5.4 12.7
1995 -15 421.76 24 63.24 -1.7 18.2 -19.9 1995 5.8 4.1
1996 199 731.16 -9 137.84 13.1 -3.1 16.2 1996 5.0 18.1
1997 136 804 -31 108 8.1 -13.8 21.9 1997 6.6 14.7
1998 158 872 -11 146 8.7 -3.6 12.3 1998 5.1 13.8
That's by year. Let me get by series...

Now for the series by series data:

Spoiler:

SERIES, On Net, On Min, Off Net (at check in time), Off Min, On/48, Off/48, Net/48
CLE88 16 217.68 -8 22.32 3.5 -17.2 20.7
DET88 -49 214.28 0 25.72 -11.0 0.0 -11.0
CLE89 22 216.36 -18 28.64 4.9 -30.2 35.0
DET89 -19 258.61 12 29.39 -3.5 19.6 -23.1
NYK89 65 240.29 -30 38.38 13.0 -37.5 50.5
DET90 -13 297.17 -8 38.08 -2.1 -10.1 8.0
MIL90 68 169.62 -30 22.38 19.2 -64.3 83.6
PHI90 45 211.87 -9 26.21 10.2 -16.5 26.7
PHI91 29 172.58 0 35.75 8.1 0.0 8.1
NYK91 48 113.643333 12 30.356667 20.3 19.0 1.3
DET91 47 160.32 -1 31.68 14.1 -1.5 15.6
LAL91 58 224.08 -9 20.92 12.4 -20.7 33.1
MIA92 56 118.77 -2 25.23 22.6 -3.8 26.4
NYK92 39 302.06 2 33.94 6.2 2.8 3.4
CLE92 20 245.97 -9 42.03 3.9 -10.3 14.2
POR92 20 250.57 24 42.43 3.8 27.2 -23.3
ATL93 55 103.39 -6 40.61 25.5 -7.1 32.6
CLE93 31 154.27 3 37.73 9.6 3.8 5.8
NYK93 25 251.09 3 36.91 4.8 3.9 0.9
PHO93 9 277.54 -9 25.46 1.6 -17.0 18.5
CHA95 22 165.42 -8 31.58 6.4 -12.2 18.5
ORL95 -37 256.34 32 31.66 -6.9 48.5 -55.4
MIA96 72 99.45 -3 44.55 34.8 -3.2 38.0
NYK96 30 216.24 1 28.76 6.7 1.7 5.0
ORL96 53 163.52 14 28.48 15.6 23.6 -8.0
SEA96 44 251.95 -21 36.05 8.4 -28.0 36.3

And finally the 3-year AuPM approximation:

Spoiler:

We can't technically know his AuPM, but here are 3-year estimates per game:

1990 +7.7
1991 +8.7
1992 +7.1
1993 +6.3
1995 +5.5
1996 +5.5
1997 +5.6
1998 +6.7

Now just for readability, here's a summary that's in a slightly more legible format:
Playoffs: Net On/off per 48 minutes (On per 48 minutes, Adjusted On relative to opponent's average MoV per 48 minutes)
1988: +4.3 (-3.7, -0.7)
1989: +24 (+4.2, +9.8)
1990: +33.1 (+7.1, +11.3)
1991: +12.4 (+13.5, +16.4)
1992: +2.1 (+7.1, +11.3)
1993: +10.4 (+7.3, +12.7)
1995: -19.9 (-1.7, +4.1)
1996: +16.2 (+13.1, +18.1)
1997: +21.9 (+8.1, +14.7)
1998: +12.3 (+8.7, +13.8)

3 year AuPM (estimate)
1990: +7.7
1991: +8.7
1992: +7.1
1993: +6.3
1995: +5.5
1996: +5.5
1997: +5.6
1998: +6.7
*note: year listed is the final year, so 1990 means 1988–1990. Why are these an estimate and not the true value? AuPM takes a player's net plus minus data over a timespan (e.g. a playoff run) and estimate their APM/RAPM, without the full play-by-play data needed to actually calculate APM/RAPM. Pros: it's much stabler in small samples, like a playoff run (unlike APM/RAPM), and it does a pretty good job at approximating the long term APM/RAPM. To do this, AuPM needs to approximate the 'adjustment' part of APM/RAPM, without the play-by-play. This requires knowing the total plus minus data of a player's teammates, but since only Jordan's plus minus data was tracked in these playoffs, we don't have that information. Specifically, we apply an adjustment if a player's teammate has a higher value than them. See original article for details (https://thinkingbasketball.net/2017/09/18/augmented-plus-minus-evaluating-old-pm-data/). We can estimate this (e.g. it's super unlikely that someone would have a better on/off in the 1990 postseason... it's super likely that someone has a better on/off in the 1995 postseason) to get realistic AuPM estimates, so that would inform the estimate. Thus this is a most-likely estimate of 3 year AuPM, until we can get the plus minus data for all Jordan's teammates too.

3-year data:

Spoiler:

3-year On per 48 minutes, 3 year off per 48 minute, 3 year Net on/off per 48 minutes:
1990: +3.4, -19.7, +23.1
1991: +8.2, -13.1, +21.3
1992: +9.0, -3.9, +13
1993: +9.0, +1.0, +8.0
1995: +5.4, +4.1, +1.3
1996: +7.5, +0.8, +6.7
1997: +7.9, -2.5, +10.3
1998: +9.8, -6.2, +16.1
*note: these are calculated by me, but the calculation's fairly simple. Feel free to check if you doubt.

To close, I'll add a little personal analysis here.

Spoiler:

So it looks like the playoffs would rank:
Net/48: 90 > 89 > 97 > 96 > 91 > 98 > 93 > 92 > 88 > 95
Adj On: 96 > 91 > 97 > 98 > 93 > 92 > 90 > 89 > 95 > 88
3-year AuPM: 89–91 > 88–90 > 90–92 > 96–98 > 91–93 > 93–95 > 94–96 > 95–97

-Consistency: Impressive that only 1995 had a negative Net/48. We also have Net/48 estimates for 1985 and 1986, and those are both positive, so all that leaves is 1987 missing (note: I wrote this before Djoker added 1987... it looks like Dojoker found 1987 was positive too!). If ’87 is positive, Jordan would join Curry as the only modern top 10 candidate to have just one negative Net/48 playoff run. LeBron, Shaq (post-97), and Kobe have one negative Net/48 playoff run in their prime, and more pre/post-prime. Duncan and Garnett had multiple negative Net/48 playoff runs in their prime. Many might consider 1995 Jordan to be not fully healthy / in shape, so if you discount unhealthy years, only Jordan and Curry have positive playoff Net/48 in every prime healthy year.

-Peak: 1989–1991 stands out as his peak in AuPM, especially 1990 by Net/48 and 1991 by Adj On & BPM. Surprising dip in 1988, which looks like one of his worst playoffs — I wonder how much is this is noise or signal.

-1st 3-peat vs 2nd 3-peat: it’s slightly surprising 91–93 has a lower AuPM than 96–98. It looks like 1996 and 1997 are right near his peak level, while 1992 (and to a lesser extent 1993) are lower in Net/48 and Adj On. The stats are close enough that noise could explain the difference. But if there’s signal, I wonder whether this says something about 92/93 Jordan’s value, the Bulls’ rotations, or their dependence on Jordan’s role in either of the 3-peats.

If the AuPM estimates here are in per game units, that would give Jordan the 1st, 4th, and 9th best 3-year AuPM runs ever (Duncan’s 2nd at +8.5). Dominant stuff either way!

.

Hey Jake! Feel free to update the OP if you'd like! Using these actual numbers should be more accurate than our by-eye estimates of the plots. Let me know if you have any questions!

This is fantastic info! Thanks for getting it! I will aim to eventually fully update my OP with it. It’d take me a fair bit of time to actually fully do that, because the data is in per-48-minutes terms and I’d tried to convert that to per-100-possessions terms in the OP—which requires pulling some data and making some additional calculations (I unfortunately didn’t keep my original calculations saved in an organized fashion, so that’d make it take longer than it should). And I’m actually not sure anymore that that’s the best way of doing things (since I’ve started to become a bit skeptical of Basketball-Reference’s pace-related data). So if I update the OP with this more precise data, I’m not sure exactly what form it should take (i.e. just leave it in per-48-minute terms or try to convert it to per-100-possession terms). I’m inclined to leave it in per-48-minute terms and just report out the total playoff on-off fully on a per-48-minute basis, in order to avoid potentially serious issues regarding conversion of the data. But that could also be a bit confusing, since some of the other data I’ve cited in the OP for regular season is on a per-100-possession basis, so the post would be using two different types of measuring sticks (not to mention that people are used to per-100-possession on-off values, so it’s not immediately clear exactly how good a number is when reported on a per-48-minute basis). That could be confusing. So I have to think about how best to do it. If you’ve got any particular views on that, I’m certainly happy to hear them! In the meantime, I’ve calculated the per-48-minute data (as per the below) and added that info as a note in my OP (crediting you and Djoker with the extremely helpful info you’ve provided), along with an explanation about how I’m still deciding whether and when to try to convert that to per-100-possession terms. I’ve also added notes in a few places in the OP pointing to the data you reported, as more accurate with regards to the Thinking Basketball data, and noting that the numbers in the rest of the post haven’t actually yet been updated to account for those more accurate numbers.

For now, as alluded to above, we can actually calculate Jordan’s career playoff on-off in per-48-minute terms:

To do this, I’ve taken: (1) the “on” plus-minus, “off” plus-minus, “on” minutes, and “off” minutes you provided above from Thinking Basketball, (2) the “on” and “off” values derived from Squared’s data from the 1984-1985 playoffs (I think he’s taken that data down now—perhaps as a result of arguments on these forums—but I do have record of that, which showed +10 “on” and therefore -32 “off”), (3) the “on” and “off” values for 1986 and 1987 from Djoker’s post (though I know there’s some small disagreement on those numbers, as per the above); and (4) the minutes on and off for 1985, 1986, and 1987 from Basketball-Reference.

And I’ve used that data to compile a full per-48-minute on-off for Jordan in the playoffs:

Michael Jordan Career Playoff On-Off Per 48 Minutes
- On: +6.6573 (+1040 in 7498.58333 minutes)
- Off: -7.2253 (-174 in 1155.936667 minutes)
- On-Off: +13.8826

Very impressive stuff! And, of course, it’s important to recognize that this is in per-48-minute terms. Since there was virtually always fewer than 100 possessions per 48 minutes on average, it is highly likely that the per-100-possession on-off would be a bit higher—without doing the full calculations, a good guess would probably be around a +15 or +16 on-off per 100 possessions (which is consistent with Djoker having calculated a +16.15 per 100 possessions before we got this more precise data).

Glad to help No rush -- I could definitely see it taking a while, and I appreciate your effort to get it in one easily accessible spot!

Since you asked, I tend to agree that having the main post switch units would be too confusing, so I'd stick with either per 48 mins or per 100 pos for the full post.

Per 48 may be slightly more accurate (if Basketball-Reference’s estimation for pace has a small percentage of error), while per 100 may allow for easier comparisons (plus minus numbers in per 100 pos units are far more common, e.g. pbpstats or basketball reference). I'd be fine with either personally.

If you do per 48, people will presumably be looking up the major modern players (probably LeBron, Shaq, Garnett, Duncan, Curry) in per 48 units, which may take a bit of work. I think the only websites that give per48 units are nba stats (which has great data but is pretty inconvenient to search, and doesn't allow you to look at a player's Net on/off per 48 over custom timespans like 3 playoffs or career playoffs) and Thinking Basketball's 3 year on/off (which also doesn't have other timespans). So with per 48, this would be more accurate, but would likely require someone manually calculating for different timespans if they wanted to compare careers or 10-year primes or whatever.

If you do per 100, that would allow easier comparison with pbpstats or basketball reference (which do custom timespans!), but I would definitely include a qualifier in your post that there may be a few percent error bars from the pace estimation and include link to the the raw data without error bars somewhere in your post.

Otherwise, your choice! And just for fun, If we went with my plus minus for 1986 Game 1 (which I'm pretty confident in), Jordan's total on gets +4 and his off gets -4. So that total would be:
Michael Jordan Career Playoff On-Off Per 48 Minutes (my calculation for 1986 Game 1)
- On: +6.68 (+1044 in 7498.58333 minutes)
- Off: -7.39 (-178 in 1155.936667 minutes)
- On-Off: +14.07

It's mainly the off sample that changes (A 4 point swing is a bigger percentage change when we were dealing with a smaller number to begin with). A good reminder that there's still some uncertainty in hand-tracked plus minus stats -- the first or second decimal we have may not be gospel if a person made an error by 1 or 2 points. But if your argument depends on the first or second decimal of a raw plus minus stat, you should probably make a better statistical argument lol. Regardless, pretty dominant stuff from Jordan, and it's cool to see the full career playoff on/off value (even with uncertainties)!

[scrabbarista]

Third quarter is the Quarter of Champions!

[Djoker]

This is incredible work! Thank you very much! Would you be okay with me editing my OP to take account of this (crediting you with it of course)? I don’t actually have time to add it at the moment anyways, but also want to be sure you are okay with that if/when I have time to update.

On the substance of this, those overall numbers are extremely impressive. I am not aware of anyone who has playoff on-off numbers that are this high! So this data definitely is extremely good for Jordan, and compares favorably to any other player in history that we have playoff on-off data for.

As you say, even if we want to leave aside the debate about whether the per-100-possession numbers calculated in this thread can validly be compared to Basketball-Reference’s per-100-possession numbers, we can instead just look at per-48-minute numbers and see that this looks like unprecedented playoff impact. A +14.33 on-off per 48 minutes is absolutely massive. It would be a huge number even as a per-100-possession number, but is even more incredible on a per-48-minutes basis (since there’s typically fewer than 100 possessions per 48 minutes, a player’s per-48-minutes on-off would typically be smaller than a per-100-possession on-off).

Of course it's ok to update the OP with the data. That's why I posted it here and tagged you!

Great stuff Djoker! I actually get a slightly different value for ~1986 Game 1~, which is now on YouTube:

Watch on YouTube

Specifically, I get:
Margin: -19
Raw On Net: -2 in 43 minutes (not - 6)
Raw Off Net: -17 in 5 minutes (not -13)

Here are my tracking notes.

Spoiler:

First Jordan rest:
Jordan out at 10:30 in the video above, 2nd quarter
Bulls up 43-33 (+10)

4 point game by 32:42, bulls timeout
At 34:25, 3 point game and 6:50 remaining in 2nd

34:40 Parish fouled, Jordan comes in, Parish makes both foul shots, Bulls up +1 (at 35:34: “now they lead by 1”). But Jordan came in after the foul, so the free throws count for when he was off
--> So during 1st rest, off Net: -9. (from Bulls +10 to +1)

Second Jordan rest:
41:50 Jordan’s on, teammate fouls
- does Jordan come out of the game here? Yes.
McHale makes 2/2 that would count with Jordan on (since foul was committed when Jordan was on), so Bulls up +2 (46:39)

Celtics tie without Jordan.
Celtics score two more by the end of the half without Jordan, so Bulls -2 (Bulls 59-61 at 49:59)
--> So during 2nd rest, Off Net: -4 (from up 2 to down 2)

Third Jordan rest:
1:31:10 Jordan off during timeout
Bird scores 2 free throws
Bulls fast break to score 2
Bird lovely turnaround 2
Celtics score 2 more
So in this stretch, Celtics go 6-2, -4 Net Bulls. Score is 99-116
--> So during 3rd rest, Off Net: -4

Total off Net in game 1: -9 + -4 + -4 = -17.
Margin: -19
Total On net: -19 - -17 = -2

.
So why might we disagree?
Option 1) It's possible one of us missed one of the times Jordan rested. During Jordan's 2nd and 3rd rest, the non-Jordan bulls were -4, so if you missed one of these rests, that would entirely explain the difference. Likewise, if there was a 4th rest that I didn't catch where the non-Jordan bulls were +4, that could explain the difference.
Side note: this plus minus I tracked is actually better than homecourtloss' "most optimistic view"! Where does the discrepancy come? Using the plays available in a highlight video, homecourtloss estimated reasonable On and Off net for Jordan, but he assumed Jordan didn't take any rests during the 4th quarter. That's actually not true (e.g. third rest was in fourth quarter), so that's where the discrepancy comes from.

Option 2) Difference in tracking methodology. For example, earlier in this thread, I accidentally counted free throw points based on when a player was subbed in, rather than when the foul was committed. Squared2020 said people usually add the free throw points based on the lineup on the court when the foul was committed, even if the lineup changes during the free throws. A difference here might contribute too.

Let me know if you figure out why we differ! Anyways, if we go by my plus minus value for 1986 game 1, the Total 1986 Playoff Plus Minus Data for Jordan would be:
Total series:
Off Net: -17 -2 -1 = -20
Off minutes: 5 + 5 + 9 mins = 19 mins

On Net: -2 -2 -17 = -21
On minutes: 43 + 53 + 39 mins = 135 mins

Off per 48: -50.53
On per 48: -7.47

On/off per 48: +43.06

The standard practice is to attribute the free throws to the lineup that was on the floor when the foul occurred not one that was substituted after the whistle whether before or between free throws. I also did that.

Here is my log of 1986 Game 1:

Spoiler:

0-0 Jordan IN -- Game Starts
43-33 Jordan OUT (+10)
43-42 Jordan IN
97-110 Jordan OUT (-14)
99-116 Jordan IN
104-123 Jordan OUT (-2) - Game Ends

Total: -6

My data completely agrees for the 1st and 3rd rest on your records. Turns out I missed the second rest you logged by MJ with 52 seconds left at the half. I'm re-watching the tape now. The announcer talks about Corzine leaving the game as McHale shoots the free throws with no mention of MJ (maybe the announcers missed it too LOL) but he definitely wasn't on the floor for those last 52 seconds in which the Celtics made a 4-0 run.

It was such a baffling substitution too. 52 seconds at the half and Celtics were already making a run. MJ did have two fouls so maybe the coach doesn't want him picking up a third but they had the ball to go on offense after McHale shoots the freebies so timing also quite weird.

Here is the updated table with 1986 Game 1 fixed.



I applaud you for double-checking the data. I urge other people to double check as well. I'm sure I have another couple of mistakes somewhere in other games as well.

[lessthanjake]

This is incredible work! Thank you very much! Would you be okay with me editing my OP to take account of this (crediting you with it of course)? I don’t actually have time to add it at the moment anyways, but also want to be sure you are okay with that if/when I have time to update.

On the substance of this, those overall numbers are extremely impressive. I am not aware of anyone who has playoff on-off numbers that are this high! So this data definitely is extremely good for Jordan, and compares favorably to any other player in history that we have playoff on-off data for.

As you say, even if we want to leave aside the debate about whether the per-100-possession numbers calculated in this thread can validly be compared to Basketball-Reference’s per-100-possession numbers, we can instead just look at per-48-minute numbers and see that this looks like unprecedented playoff impact. A +14.33 on-off per 48 minutes is absolutely massive. It would be a huge number even as a per-100-possession number, but is even more incredible on a per-48-minutes basis (since there’s typically fewer than 100 possessions per 48 minutes, a player’s per-48-minutes on-off would typically be smaller than a per-100-possession on-off).

Of course it's ok to update the OP with the data. That's why I posted it here and tagged you!

Great stuff Djoker! I actually get a slightly different value for ~1986 Game 1~, which is now on YouTube:

Watch on YouTube

Specifically, I get:
Margin: -19
Raw On Net: -2 in 43 minutes (not - 6)
Raw Off Net: -17 in 5 minutes (not -13)

Here are my tracking notes.

Spoiler:

First Jordan rest:
Jordan out at 10:30 in the video above, 2nd quarter
Bulls up 43-33 (+10)

4 point game by 32:42, bulls timeout
At 34:25, 3 point game and 6:50 remaining in 2nd

34:40 Parish fouled, Jordan comes in, Parish makes both foul shots, Bulls up +1 (at 35:34: “now they lead by 1”). But Jordan came in after the foul, so the free throws count for when he was off
--> So during 1st rest, off Net: -9. (from Bulls +10 to +1)

Second Jordan rest:
41:50 Jordan’s on, teammate fouls
- does Jordan come out of the game here? Yes.
McHale makes 2/2 that would count with Jordan on (since foul was committed when Jordan was on), so Bulls up +2 (46:39)

Celtics tie without Jordan.
Celtics score two more by the end of the half without Jordan, so Bulls -2 (Bulls 59-61 at 49:59)
--> So during 2nd rest, Off Net: -4 (from up 2 to down 2)

Third Jordan rest:
1:31:10 Jordan off during timeout
Bird scores 2 free throws
Bulls fast break to score 2
Bird lovely turnaround 2
Celtics score 2 more
So in this stretch, Celtics go 6-2, -4 Net Bulls. Score is 99-116
--> So during 3rd rest, Off Net: -4

Total off Net in game 1: -9 + -4 + -4 = -17.
Margin: -19
Total On net: -19 - -17 = -2

.
So why might we disagree?
Option 1) It's possible one of us missed one of the times Jordan rested. During Jordan's 2nd and 3rd rest, the non-Jordan bulls were -4, so if you missed one of these rests, that would entirely explain the difference. Likewise, if there was a 4th rest that I didn't catch where the non-Jordan bulls were +4, that could explain the difference.
Side note: this plus minus I tracked is actually better than homecourtloss' "most optimistic view"! Where does the discrepancy come? Using the plays available in a highlight video, homecourtloss estimated reasonable On and Off net for Jordan, but he assumed Jordan didn't take any rests during the 4th quarter. That's actually not true (e.g. third rest was in fourth quarter), so that's where the discrepancy comes from.

Option 2) Difference in tracking methodology. For example, earlier in this thread, I accidentally counted free throw points based on when a player was subbed in, rather than when the foul was committed. Squared2020 said people usually add the free throw points based on the lineup on the court when the foul was committed, even if the lineup changes during the free throws. A difference here might contribute too.

Let me know if you figure out why we differ! Anyways, if we go by my plus minus value for 1986 game 1, the Total 1986 Playoff Plus Minus Data for Jordan would be:
Total series:
Off Net: -17 -2 -1 = -20
Off minutes: 5 + 5 + 9 mins = 19 mins

On Net: -2 -2 -17 = -21
On minutes: 43 + 53 + 39 mins = 135 mins

Off per 48: -50.53
On per 48: -7.47

On/off per 48: +43.06

The standard practice is to attribute the free throws to the lineup that was on the floor when the foul occurred not one that was substituted after the whistle whether before or between free throws. I also did that.

Here is my log of 1986 Game 1:

Spoiler:

0-0 Jordan IN -- Game Starts
43-33 Jordan OUT (+10)
43-42 Jordan IN
97-110 Jordan OUT (-14)
99-116 Jordan IN
104-123 Jordan OUT (-2) - Game Ends

Total: -6

My data completely agrees for the 1st and 3rd rest on your records. Turns out I missed the second rest you logged by MJ with 52 seconds left at the half. I'm re-watching the tape now. The announcer talks about Corzine leaving the game as McHale shoots the free throws with no mention of MJ (maybe the announcers missed it too LOL) but he definitely wasn't on the floor for those last 52 seconds in which the Celtics made a 4-0 run.

It was such a baffling substitution too. 52 seconds at the half and Celtics were already making a run. MJ did have two fouls so maybe the coach doesn't want him picking up a third but they had the ball to go on offense after McHale shoots the freebies so timing also quite weird.

Here is the updated table with 1986 Game 1 fixed.



I applaud you for double-checking the data. I urge other people to double check as well. I'm sure I have another couple of mistakes somewhere in other games as well.

Thanks! So as I mentioned in a prior post to DraymondGold, I’ve semi-updated the OP with this data to make note of the full-playoff per-48-minute data and to give credit to you and DraymondGold for the info you provided. I’ve not yet used all the more precise Thinking Basketball data or your data to provide specific year-by-year estimates (which means the OP’s year-by-year playoff estimates based on Thinking Basketball’s data are the same as they were originally, and I don’t have the year-by-year data for the 1986 or 1987 playoffs in there yet).

Part of my reason for not fully updating it yet is just time (particularly to adjust to per-100-possessions terms). But one is that I’m not *entirely* sure what to do with the new data. Specifically, the precise Thinking Basketball data and your data for the 1986 and 1987 playoffs are both in per-48-minute terms. In my OP, I did a calculation to adjust things to per-100-possession terms, based on using Basketball-Reference’s pace estimations and assuming equal pace with Jordan on and off the floor. That adjustment inherently causes some error (since it is based on a pace *estimation* and the equal-pace assumption is not going to be *exactly* right). I’d assumed that the error caused would be small, though. That seemed like a reasonable assumption IMO. However, it has been pointed out in this thread that if you made that adjustment to players’ plus-minus data that BBREF has plus-minus and on-off data for, the result generally is actually pretty far from BBREF’s reported on-off (having checked a lot of random players for this, I can say that the result using this adjustment usually, but not always, is a higher on-off than the on-off BBREF actually reports). I don’t know what is causing this discrepancy, and I actually think BBREF has some significant issues with its on-off (including that they sometimes have things like a positive “off” value for a player where BBREF’s own plus-minus says the team was outscored when the player was off the court). But regardless of what is causing that, I guess I’m hesitant to take precise data and then layer on an adjustment that may be inaccurate and controversial. So I’m inclined to just leave this new data in per-48-minute terms. At the same time, though, there’s downside to doing that. People are used to seeing plus-minus data in per-100-possession terms, so it won’t naturally be so clear to people exactly how good a given per-48-minute on-off is. Furthermore, some of the data in the OP is in per-100-possession terms, without any real ability to put it in per-48-minute terms. Most pertinently, I believe the 1996 regular season data from Pollack is in per-100-possession terms, but with no data on how many possessions and therefore no way to convert to per-48-minute terms. So if I edited the OP to report most things out in per-48-minute terms, it’d be confusing since at least one data point would be using a different measuring stick. That’s not ideal. Do you have any particular thoughts on how it would be best to handle this? There doesn’t seem to be an ideal solution either way.

[lessthanjake]

~Thinking Basketball's Actual Jordan Playoff Plus/Minus Data~

Over the holidays, I was able to contact Thinking Basketball, who was kindly willing to share the actual plus minus data from his 'I tracked every Michael Jordan playoff game for a decade' video. This is Jordan's Plus/Minus data for every playoffs from 1988–1998. Of course, full credit goes to Thinking Basketball and the rest of the team that tracked this data. It's due to the ambitious work of basketball historians like TB and Squared2020 that we're able to get this entirely new insight into past players. And plus, learning about the old stats is just fun! So big thanks to the team that tracked this

Here's the raw data copied directly from Ben (I've added commas between column headers for readability) . First the year by year data:

Spoiler:

PS, On Net, On Min, Off Net (at check in time), Off Min, On/48, Off/48, Net/48, Year, Opp RS Net, Adj On
1988 -33 431.96 -8 48.04 -3.7 -8.0 4.3 1988 3.0 -0.7
1989 63 726.76 -41 99.24 4.2 -19.8 24.0 1989 5.7 9.8
1990 100 678.66 -47 86.67 7.1 -26.0 33.1 1990 4.2 11.3
1991 196 694.623333 3 126.376667 13.5 1.1 12.4 1991 2.8 16.4
1992 135 917.37 15 143.63 7.1 5.0 2.1 1992 4.3 11.3
1993 120 786.29 -9 140.71 7.3 -3.1 10.4 1993 5.4 12.7
1995 -15 421.76 24 63.24 -1.7 18.2 -19.9 1995 5.8 4.1
1996 199 731.16 -9 137.84 13.1 -3.1 16.2 1996 5.0 18.1
1997 136 804 -31 108 8.1 -13.8 21.9 1997 6.6 14.7
1998 158 872 -11 146 8.7 -3.6 12.3 1998 5.1 13.8
That's by year. Let me get by series...

Now for the series by series data:

Spoiler:

SERIES, On Net, On Min, Off Net (at check in time), Off Min, On/48, Off/48, Net/48
CLE88 16 217.68 -8 22.32 3.5 -17.2 20.7
DET88 -49 214.28 0 25.72 -11.0 0.0 -11.0
CLE89 22 216.36 -18 28.64 4.9 -30.2 35.0
DET89 -19 258.61 12 29.39 -3.5 19.6 -23.1
NYK89 65 240.29 -30 38.38 13.0 -37.5 50.5
DET90 -13 297.17 -8 38.08 -2.1 -10.1 8.0
MIL90 68 169.62 -30 22.38 19.2 -64.3 83.6
PHI90 45 211.87 -9 26.21 10.2 -16.5 26.7
PHI91 29 172.58 0 35.75 8.1 0.0 8.1
NYK91 48 113.643333 12 30.356667 20.3 19.0 1.3
DET91 47 160.32 -1 31.68 14.1 -1.5 15.6
LAL91 58 224.08 -9 20.92 12.4 -20.7 33.1
MIA92 56 118.77 -2 25.23 22.6 -3.8 26.4
NYK92 39 302.06 2 33.94 6.2 2.8 3.4
CLE92 20 245.97 -9 42.03 3.9 -10.3 14.2
POR92 20 250.57 24 42.43 3.8 27.2 -23.3
ATL93 55 103.39 -6 40.61 25.5 -7.1 32.6
CLE93 31 154.27 3 37.73 9.6 3.8 5.8
NYK93 25 251.09 3 36.91 4.8 3.9 0.9
PHO93 9 277.54 -9 25.46 1.6 -17.0 18.5
CHA95 22 165.42 -8 31.58 6.4 -12.2 18.5
ORL95 -37 256.34 32 31.66 -6.9 48.5 -55.4
MIA96 72 99.45 -3 44.55 34.8 -3.2 38.0
NYK96 30 216.24 1 28.76 6.7 1.7 5.0
ORL96 53 163.52 14 28.48 15.6 23.6 -8.0
SEA96 44 251.95 -21 36.05 8.4 -28.0 36.3

And finally the 3-year AuPM approximation:

Spoiler:

We can't technically know his AuPM, but here are 3-year estimates per game:

1990 +7.7
1991 +8.7
1992 +7.1
1993 +6.3
1995 +5.5
1996 +5.5
1997 +5.6
1998 +6.7

Now just for readability, here's a summary that's in a slightly more legible format:
Playoffs: Net On/off per 48 minutes (On per 48 minutes, Adjusted On relative to opponent's average MoV per 48 minutes)
1988: +4.3 (-3.7, -0.7)
1989: +24 (+4.2, +9.8)
1990: +33.1 (+7.1, +11.3)
1991: +12.4 (+13.5, +16.4)
1992: +2.1 (+7.1, +11.3)
1993: +10.4 (+7.3, +12.7)
1995: -19.9 (-1.7, +4.1)
1996: +16.2 (+13.1, +18.1)
1997: +21.9 (+8.1, +14.7)
1998: +12.3 (+8.7, +13.8)

3 year AuPM (estimate)
1990: +7.7
1991: +8.7
1992: +7.1
1993: +6.3
1995: +5.5
1996: +5.5
1997: +5.6
1998: +6.7
*note: year listed is the final year, so 1990 means 1988–1990. Why are these an estimate and not the true value? AuPM takes a player's net plus minus data over a timespan (e.g. a playoff run) and estimate their APM/RAPM, without the full play-by-play data needed to actually calculate APM/RAPM. Pros: it's much stabler in small samples, like a playoff run (unlike APM/RAPM), and it does a pretty good job at approximating the long term APM/RAPM. To do this, AuPM needs to approximate the 'adjustment' part of APM/RAPM, without the play-by-play. This requires knowing the total plus minus data of a player's teammates, but since only Jordan's plus minus data was tracked in these playoffs, we don't have that information. Specifically, we apply an adjustment if a player's teammate has a higher value than them. See original article for details (https://thinkingbasketball.net/2017/09/18/augmented-plus-minus-evaluating-old-pm-data/). We can estimate this (e.g. it's super unlikely that someone would have a better on/off in the 1990 postseason... it's super likely that someone has a better on/off in the 1995 postseason) to get realistic AuPM estimates, so that would inform the estimate. Thus this is a most-likely estimate of 3 year AuPM, until we can get the plus minus data for all Jordan's teammates too.

3-year data:

Spoiler:

3-year On per 48 minutes, 3 year off per 48 minute, 3 year Net on/off per 48 minutes:
1990: +3.4, -19.7, +23.1
1991: +8.2, -13.1, +21.3
1992: +9.0, -3.9, +13
1993: +9.0, +1.0, +8.0
1995: +5.4, +4.1, +1.3
1996: +7.5, +0.8, +6.7
1997: +7.9, -2.5, +10.3
1998: +9.8, -6.2, +16.1
*note: these are calculated by me, but the calculation's fairly simple. Feel free to check if you doubt.

To close, I'll add a little personal analysis here.

Spoiler:

So it looks like the playoffs would rank:
Net/48: 90 > 89 > 97 > 96 > 91 > 98 > 93 > 92 > 88 > 95
Adj On: 96 > 91 > 97 > 98 > 93 > 92 > 90 > 89 > 95 > 88
3-year AuPM: 89–91 > 88–90 > 90–92 > 96–98 > 91–93 > 93–95 > 94–96 > 95–97

-Consistency: Impressive that only 1995 had a negative Net/48. We also have Net/48 estimates for 1985 and 1986, and those are both positive, so all that leaves is 1987 missing (note: I wrote this before Djoker added 1987... it looks like Dojoker found 1987 was positive too!). If ’87 is positive, Jordan would join Curry as the only modern top 10 candidate to have just one negative Net/48 playoff run. LeBron, Shaq (post-97), and Kobe have one negative Net/48 playoff run in their prime, and more pre/post-prime. Duncan and Garnett had multiple negative Net/48 playoff runs in their prime. Many might consider 1995 Jordan to be not fully healthy / in shape, so if you discount unhealthy years, only Jordan and Curry have positive playoff Net/48 in every prime healthy year.

-Peak: 1989–1991 stands out as his peak in AuPM, especially 1990 by Net/48 and 1991 by Adj On & BPM. Surprising dip in 1988, which looks like one of his worst playoffs — I wonder how much is this is noise or signal.

-1st 3-peat vs 2nd 3-peat: it’s slightly surprising 91–93 has a lower AuPM than 96–98. It looks like 1996 and 1997 are right near his peak level, while 1992 (and to a lesser extent 1993) are lower in Net/48 and Adj On. The stats are close enough that noise could explain the difference. But if there’s signal, I wonder whether this says something about 92/93 Jordan’s value, the Bulls’ rotations, or their dependence on Jordan’s role in either of the 3-peats.

If the AuPM estimates here are in per game units, that would give Jordan the 1st, 4th, and 9th best 3-year AuPM runs ever (Duncan’s 2nd at +8.5). Dominant stuff either way!

Hey Jake! Feel free to update the OP if you'd like! Using these actual numbers should be more accurate than our by-eye estimates of the plots. Let me know if you have any questions!

This is fantastic info! Thanks for getting it! I will aim to eventually fully update my OP with it. It’d take me a fair bit of time to actually fully do that, because the data is in per-48-minutes terms and I’d tried to convert that to per-100-possessions terms in the OP—which requires pulling some data and making some additional calculations (I unfortunately didn’t keep my original calculations saved in an organized fashion, so that’d make it take longer than it should). And I’m actually not sure anymore that that’s the best way of doing things (since I’ve started to become a bit skeptical of Basketball-Reference’s pace-related data). So if I update the OP with this more precise data, I’m not sure exactly what form it should take (i.e. just leave it in per-48-minute terms or try to convert it to per-100-possession terms). I’m inclined to leave it in per-48-minute terms and just report out the total playoff on-off fully on a per-48-minute basis, in order to avoid potentially serious issues regarding conversion of the data. But that could also be a bit confusing, since some of the other data I’ve cited in the OP for regular season is on a per-100-possession basis, so the post would be using two different types of measuring sticks (not to mention that people are used to per-100-possession on-off values, so it’s not immediately clear exactly how good a number is when reported on a per-48-minute basis). That could be confusing. So I have to think about how best to do it. If you’ve got any particular views on that, I’m certainly happy to hear them! In the meantime, I’ve calculated the per-48-minute data (as per the below) and added that info as a note in my OP (crediting you and Djoker with the extremely helpful info you’ve provided), along with an explanation about how I’m still deciding whether and when to try to convert that to per-100-possession terms. I’ve also added notes in a few places in the OP pointing to the data you reported, as more accurate with regards to the Thinking Basketball data, and noting that the numbers in the rest of the post haven’t actually yet been updated to account for those more accurate numbers.

For now, as alluded to above, we can actually calculate Jordan’s career playoff on-off in per-48-minute terms:

To do this, I’ve taken: (1) the “on” plus-minus, “off” plus-minus, “on” minutes, and “off” minutes you provided above from Thinking Basketball, (2) the “on” and “off” values derived from Squared’s data from the 1984-1985 playoffs (I think he’s taken that data down now—perhaps as a result of arguments on these forums—but I do have record of that, which showed +10 “on” and therefore -32 “off”), (3) the “on” and “off” values for 1986 and 1987 from Djoker’s post (though I know there’s some small disagreement on those numbers, as per the above); and (4) the minutes on and off for 1985, 1986, and 1987 from Basketball-Reference.

And I’ve used that data to compile a full per-48-minute on-off for Jordan in the playoffs:

Michael Jordan Career Playoff On-Off Per 48 Minutes
- On: +6.6573 (+1040 in 7498.58333 minutes)
- Off: -7.2253 (-174 in 1155.936667 minutes)
- On-Off: +13.8826

Very impressive stuff! And, of course, it’s important to recognize that this is in per-48-minute terms. Since there was virtually always fewer than 100 possessions per 48 minutes on average, it is highly likely that the per-100-possession on-off would be a bit higher—without doing the full calculations, a good guess would probably be around a +15 or +16 on-off per 100 possessions (which is consistent with Djoker having calculated a +16.15 per 100 possessions before we got this more precise data).

Glad to help No rush -- I could definitely see it taking a while, and I appreciate your effort to get it in one easily accessible spot!

Since you asked, I tend to agree that having the main post switch units would be too confusing, so I'd stick with either per 48 mins or per 100 pos for the full post.

Per 48 may be slightly more accurate (if Basketball-Reference’s estimation for pace has a small percentage of error), while per 100 may allow for easier comparisons (plus minus numbers in per 100 pos units are far more common, e.g. pbpstats or basketball reference). I'd be fine with either personally.

If you do per 48, people will presumably be looking up the major modern players (probably LeBron, Shaq, Garnett, Duncan, Curry) in per 48 units, which may take a bit of work. I think the only websites that give per48 units are nba stats (which has great data but is pretty inconvenient to search, and doesn't allow you to look at a player's Net on/off per 48 over custom timespans like 3 playoffs or career playoffs) and Thinking Basketball's 3 year on/off (which also doesn't have other timespans). So with per 48, this would be more accurate, but would likely require someone manually calculating for different timespans if they wanted to compare careers or 10-year primes or whatever.

If you do per 100, that would allow easier comparison with pbpstats or basketball reference (which do custom timespans!), but I would definitely include a qualifier in your post that there may be a few percent error bars from the pace estimation and include link to the the raw data without error bars somewhere in your post.

Otherwise, your choice! And just for fun, If we went with my plus minus for 1986 Game 1 (which I'm pretty confident in), Jordan's total on gets +4 and his off gets -4. So that total would be:
Michael Jordan Career Playoff On-Off Per 48 Minutes (my calculation for 1986 Game 1)
- On: +6.68 (+1044 in 7498.58333 minutes)
- Off: -7.39 (-178 in 1155.936667 minutes)
- On-Off: +14.07

It's mainly the off sample that changes (A 4 point swing is a bigger percentage change when we were dealing with a smaller number to begin with). A good reminder that there's still some uncertainty in hand-tracked plus minus stats -- the first or second decimal we have may not be gospel if a person made an error by 1 or 2 points. But if your argument depends on the first or second decimal of a raw plus minus stat, you should probably make a better statistical argument lol. Regardless, pretty dominant stuff from Jordan, and it's cool to see the full career playoff on/off value (even with uncertainties)!

So I think it’d probably be an easier decision if there was a way to do per-48-minute data for the full post. But the trouble is that, for at least one of the data points, there isn’t. Specifically, I believe Pollack’s on-off data is in per-100-possession terms. But there’s actually no information I’m aware of that we could use to convert that to per-48-minute terms, because it doesn’t have the raw plus-minus or the number of possessions. To be honest, I guess it’s *possible* that Pollack’s data is in per-48-minute terms to begin with—there’s not really any context to tell. But my assumption has been that it’s in per-100-possession terms. If that’s the case, then I don’t see a way to make the full post be all in per-48-minute terms. And keeping the whole post in per-100-possession terms basically layers on an adjustment that has proven to be controversial (and I think correctly so actually, since, for reasons I can’t quite figure out, replicating that adjustment to plus-minus data for more recent players generally doesn’t match the actual on-off data for those players, and almost always is too high). So that leaves me unsure of what to do when I fully update the OP. I guess I’m probably inclined to convert the OP to all per-48-minute terms and then just leave a note in the post flagging that I believe the Pollack data is in per-100-possession terms. But there’s real downsides to that approach.

[eminence]

So I think it’d probably be an easier decision if there was a way to do per-48-minute data for the full post. But the trouble is that, for at least one of the data points, there isn’t. Specifically, I believe Pollack’s on-off data is in per-100-possession terms. But there’s actually no information I’m aware of that we could use to convert that to per-48-minute terms, because it doesn’t have the raw plus-minus or the number of possessions. To be honest, I guess it’s *possible* that Pollack’s data is in per-48-minute terms to begin with—there’s not really any context to tell. But my assumption has been that it’s in per-100-possession terms. If that’s the case, then I don’t see a way to make the full post be all in per-48-minute terms. And keeping the whole post in per-100-possession terms basically layers on an adjustment that has proven to be controversial (and I think correctly so actually, since, for reasons I can’t quite figure out, replicating that adjustment to plus-minus data for more recent players generally doesn’t match the actual on-off data for those players, and almost always is too high). So that leaves me unsure of what to do when I fully update the OP. I guess I’m probably inclined to convert the OP to all per-48-minute terms and then just leave a note in the post flagging that I believe the Pollack data is in per-100-possession terms. But there’s real downsides to that approach.

I believe Pollack's data is just raw cumulative +/- (so like +900 for the season or whatever), somebody else (I forget who, sorry friend) converted it into a /100 estimate.

Not sure where the original is sitting around, but I'm sure someone has it. Owly? Doc?

[lessthanjake]

So I think it’d probably be an easier decision if there was a way to do per-48-minute data for the full post. But the trouble is that, for at least one of the data points, there isn’t. Specifically, I believe Pollack’s on-off data is in per-100-possession terms. But there’s actually no information I’m aware of that we could use to convert that to per-48-minute terms, because it doesn’t have the raw plus-minus or the number of possessions. To be honest, I guess it’s *possible* that Pollack’s data is in per-48-minute terms to begin with—there’s not really any context to tell. But my assumption has been that it’s in per-100-possession terms. If that’s the case, then I don’t see a way to make the full post be all in per-48-minute terms. And keeping the whole post in per-100-possession terms basically layers on an adjustment that has proven to be controversial (and I think correctly so actually, since, for reasons I can’t quite figure out, replicating that adjustment to plus-minus data for more recent players generally doesn’t match the actual on-off data for those players, and almost always is too high). So that leaves me unsure of what to do when I fully update the OP. I guess I’m probably inclined to convert the OP to all per-48-minute terms and then just leave a note in the post flagging that I believe the Pollack data is in per-100-possession terms. But there’s real downsides to that approach.

I believe Pollack's data is just raw cumulative +/- (so like +900 for the season or whatever), somebody else (I forget who, sorry friend) converted it into a /100 estimate.

Not sure where the original is sitting around, but I'm sure someone has it. Owly? Doc?

Oh! Well, if the original is sitting around somewhere with raw plus-minus, then I could definitely convert that to per-48-minute terms.

Also, if the original was just raw plus-minus, then that kind of makes me think the data we have is in per-48-minute terms, since I doubt anyone else had the ability to go back and log the number of possessions in every single game in the 1996 regular season. But perhaps someone basically converted it to per-100-possession terms using BBREF pace estimates. If that’s the case, I could try to derive the raw plus-minus by undoing that, but I wouldn’t be confident in that since BBREF’s pace estimates might’ve changed in the meantime.

[eminence]

So I think it’d probably be an easier decision if there was a way to do per-48-minute data for the full post. But the trouble is that, for at least one of the data points, there isn’t. Specifically, I believe Pollack’s on-off data is in per-100-possession terms. But there’s actually no information I’m aware of that we could use to convert that to per-48-minute terms, because it doesn’t have the raw plus-minus or the number of possessions. To be honest, I guess it’s *possible* that Pollack’s data is in per-48-minute terms to begin with—there’s not really any context to tell. But my assumption has been that it’s in per-100-possession terms. If that’s the case, then I don’t see a way to make the full post be all in per-48-minute terms. And keeping the whole post in per-100-possession terms basically layers on an adjustment that has proven to be controversial (and I think correctly so actually, since, for reasons I can’t quite figure out, replicating that adjustment to plus-minus data for more recent players generally doesn’t match the actual on-off data for those players, and almost always is too high). So that leaves me unsure of what to do when I fully update the OP. I guess I’m probably inclined to convert the OP to all per-48-minute terms and then just leave a note in the post flagging that I believe the Pollack data is in per-100-possession terms. But there’s real downsides to that approach.

I believe Pollack's data is just raw cumulative +/- (so like +900 for the season or whatever), somebody else (I forget who, sorry friend) converted it into a /100 estimate.

Not sure where the original is sitting around, but I'm sure someone has it. Owly? Doc?

Oh! Well, if the original is sitting around somewhere with raw plus-minus, then I could definitely convert that to per-48-minute terms.

Also, if the original was just raw plus-minus, then that kind of makes me think the data we have is in per-48-minute terms, since I doubt anyone else had the ability to go back and log the number of possessions in every single game in the 1996 regular season. But perhaps someone basically converted it to per-100-possession terms using BBREF pace estimates. If that’s the case, I could try to derive the raw plus-minus by undoing that, but I wouldn’t be confident in that since BBREF’s pace estimates might’ve changed in the meantime.

Here is fpliii's original thread on it on this forum, viewtopic.php?f=344&t=1343246

Don't have time to read through it all now, but it seems Colts or Elgee put together the estimate you're thinking of

[lessthanjake]

I believe Pollack's data is just raw cumulative +/- (so like +900 for the season or whatever), somebody else (I forget who, sorry friend) converted it into a /100 estimate.

Not sure where the original is sitting around, but I'm sure someone has it. Owly? Doc?

Oh! Well, if the original is sitting around somewhere with raw plus-minus, then I could definitely convert that to per-48-minute terms.

Also, if the original was just raw plus-minus, then that kind of makes me think the data we have is in per-48-minute terms, since I doubt anyone else had the ability to go back and log the number of possessions in every single game in the 1996 regular season. But perhaps someone basically converted it to per-100-possession terms using BBREF pace estimates. If that’s the case, I could try to derive the raw plus-minus by undoing that, but I wouldn’t be confident in that since BBREF’s pace estimates might’ve changed in the meantime.

Here is fpliii's original thread on it on this forum, viewtopic.php?f=344&t=1343246

Don't have time to read through it all now, but it seems Colts or Elgee put together the estimate you're thinking of

Thanks! It looks to me like the links to the original data being discussed there don’t work anymore. However, I was able to find in that thread people saying the data says Jordan had a +980 plus-minus in the 1996 regular season—which is all the info I need to get per-48-minute calculations for that season.

[Owly]

I actually think BBREF has some significant issues with its on-off (including that they sometimes have things like a positive “off” value for a player where BBREF’s own plus-minus says the team was outscored when the player was off the court).

I only skimmed so maybe I missed something ... and maybe it's so obvious but also so marginal as to not be worth raising ... and this certainly doesn't preclude other issues but ...

at the margins I think (as I understand it) this kind of thing could reasonably happen. A defensive specialist playing more defensive possessions than offensive ... could be a positive defender and net player for on-off per 100 possessions. But his team could be outscored with them on by virtue of playing more defensive possessions. Conversely on offensive specialist could see the team outscore opponents when they were on the court ... in part because he gets more offensive possessions and the other lineups more defensive possessions ... yet have a negative "off". Game management at the end of quarters can net (or lose) teams a couple of extra possessions a game too.

Equally this will be the case for the off samples. The team with the defensive player could outscore opponents in absolute terms with him off and yet per possession basis be negative with him off. And the team with the offensive player could be outscored with him off in absolute terms, but after accounting for possessions be reckoned better than neutral with him off (such as your example).

[lessthanjake]

I actually think BBREF has some significant issues with its on-off (including that they sometimes have things like a positive “off” value for a player where BBREF’s own plus-minus says the team was outscored when the player was off the court).

I only skimmed so maybe I missed something ... and maybe it's so obvious but also so marginal as to not be worth raising ... and this certainly doesn't preclude other issues but ...

at the margins I think (as I understand it) this kind of thing could reasonably happen. A defensive specialist playing more defensive possessions than offensive ... could be a positive defender and net player for on-off per 100 possessions. But his team could be outscored with them on by virtue of playing more defensive possessions. Conversely on offensive specialist could see the team outscore opponents when they were on the court ... in part because he gets more offensive possessions and the other lineups more defensive possessions ... yet have a negative "off". Game management at the end of quarters can net (or lose) teams a couple of extra possessions a game too.

Equally this will be the case for the off samples. The team with the defensive player could outscore opponents in absolute terms with him off and yet per possession basis be negative with him off. And the team with the offensive player could be outscored with him off in absolute terms, but after accounting for possessions be reckoned better than neutral with him off (such as your example).

Ah, that’s an interesting point! It’s possible that the discrepancies are caused by something like this. That said, I’m not *sure* this comes into play, since it seems like Basketball-Reference’s possession estimate doesn’t really distinguish between a team’s possessions and the opponent’s possessions (see the formula here: https://www.basketball-reference.com/about/glossary.html). Perhaps I’m missing something, but it seems like Basketball-Reference’s estimated number of possessions is just calculated using both team and opponent data and averaging those two. So if a player is on the court for more opponent possessions than possessions for their team, then it seems like Basketball-Reference wouldn’t be adjusting for that. Again, though, I may be missing something—especially as I’m not privy to how BBREF calculates on-off specifically. If their calculation does try to adjust for this (i.e. for instance, by calculating the “on” value based on looking at points per team possession minus points per opponent possession, rather than just dividing plus-minus by number of possessions), then this is a potential explanation for the issue I’ve noted, since it is a potential way that a negative plus-minus could turn into a positive “on” or “off” value. It’s definitely the best potential explanation I’ve seen! As you say, it seems so marginal that I’m not sure it can fully explain some of the examples of this I’ve seen, but this at least gives a potential logical mechanism through which the sign could change.

[eminence]

Also, I said +900 or whatever, lol. That's an extreme extreme number. Guys we have on record who hit +900 on the season: '96 MJ, '15-'17 Curry, and '16 Green. Poor '96 Pippen one point short.

And given overall team quality/number of games in a season, it's entirely possible those are the only players to ever do it.

[Owly]

Also, I said +900 or whatever, lol. That's an extreme extreme number. Guys we have on record who hit +900 on the season: '96 MJ, '15-'17 Curry, and '16 Green. Poor '96 Pippen one point short.

And given overall team quality/number of games in a season, it's entirely possible those are the only players to ever do it.

Teams with a points dif of better than 800 through 2019 were.

1971-72 Los Angeles Lakers* 1007
1970-71 Milwaukee Bucks* 1005
1995-96 Chicago Bulls* 1004
2016-17 Golden State Warriors* 954
1971-72 Milwaukee Bucks* 915
1996-97 Chicago Bulls* 886
2015-16 Golden State Warriors* 882
2015-16 San Antonio Spurs* 872
1991-92 Chicago Bulls* 856
2007-08 Boston Celtics* 841
2014-15 Golden State Warriors* 828

If (hypothetically) those teams maintained a constant points margin regardless of lineup (or did so with regard to the following players) the minute leaders and their "expected" pro-rata plus-minus would be

Wilt Chamberlain 883.0341254
Kareem Abdul-Jabbar 835.0442366
Michael Jordan 786.2037506
Klay Thompson 638.8134479
Kareem Abdul-Jabbar 828.7272497
Michael Jordan 697.3938165
Draymond Green 622.9014085
Kawhi Leonard 526.60746
Scottie Pippen 682.9006556
Paul Pierce 611.7524677
Stephen Curry 548.2929549

[eminence]

Also, I said +900 or whatever, lol. That's an extreme extreme number. Guys we have on record who hit +900 on the season: '96 MJ, '15-'17 Curry, and '16 Green. Poor '96 Pippen one point short.

And given overall team quality/number of games in a season, it's entirely possible those are the only players to ever do it.

Teams with a points dif of better than 800 through 2019 were.

1971-72 Los Angeles Lakers* 1007
1970-71 Milwaukee Bucks* 1005
1995-96 Chicago Bulls* 1004
2016-17 Golden State Warriors* 954
1971-72 Milwaukee Bucks* 915
1996-97 Chicago Bulls* 886
2015-16 Golden State Warriors* 882
2015-16 San Antonio Spurs* 872
1991-92 Chicago Bulls* 856
2007-08 Boston Celtics* 841
2014-15 Golden State Warriors* 828

If (hypothetically) those teams maintained a constant points margin regardless of lineup (or did so with regard to the following players) the minute leaders and their "expected" pro-rata plus-minus would be

Wilt Chamberlain 883.0341254
Kareem Abdul-Jabbar 835.0442366
Michael Jordan 786.2037506
Klay Thompson 638.8134479
Kareem Abdul-Jabbar 828.7272497
Michael Jordan 697.3938165
Draymond Green 622.9014085
Kawhi Leonard 526.60746
Scottie Pippen 682.9006556
Paul Pierce 611.7524677
Stephen Curry 548.2929549

So only the '72 Lakers, '71 Bucks, and '92 Bulls survive as prior to '94. Seems likely someone else did it prior to '94, but I'd put it at likely we can count the total number of players to do it on our fingers.

[Djoker]

Part of my reason for not fully updating it yet is just time (particularly to adjust to per-100-possessions terms). But one is that I’m not *entirely* sure what to do with the new data. Specifically, the precise Thinking Basketball data and your data for the 1986 and 1987 playoffs are both in per-48-minute terms. In my OP, I did a calculation to adjust things to per-100-possession terms, based on using Basketball-Reference’s pace estimations and assuming equal pace with Jordan on and off the floor. That adjustment inherently causes some error (since it is based on a pace *estimation* and the equal-pace assumption is not going to be *exactly* right). I’d assumed that the error caused would be small, though. That seemed like a reasonable assumption IMO. However, it has been pointed out in this thread that if you made that adjustment to players’ plus-minus data that BBREF has plus-minus and on-off data for, the result generally is actually pretty far from BBREF’s reported on-off (having checked a lot of random players for this, I can say that the result using this adjustment usually, but not always, is a higher on-off than the on-off BBREF actually reports). I don’t know what is causing this discrepancy, and I actually think BBREF has some significant issues with its on-off (including that they sometimes have things like a positive “off” value for a player where BBREF’s own plus-minus says the team was outscored when the player was off the court). But regardless of what is causing that, I guess I’m hesitant to take precise data and then layer on an adjustment that may be inaccurate and controversial. So I’m inclined to just leave this new data in per-48-minute terms. At the same time, though, there’s downside to doing that. People are used to seeing plus-minus data in per-100-possession terms, so it won’t naturally be so clear to people exactly how good a given per-48-minute on-off is. Furthermore, some of the data in the OP is in per-100-possession terms, without any real ability to put it in per-48-minute terms. Most pertinently, I believe the 1996 regular season data from Pollack is in per-100-possession terms, but with no data on how many possessions and therefore no way to convert to per-48-minute terms. So if I edited the OP to report most things out in per-48-minute terms, it’d be confusing since at least one data point would be using a different measuring stick. That’s not ideal. Do you have any particular thoughts on how it would be best to handle this? There doesn’t seem to be an ideal solution either way.

I have Pollack's raw ON for the 95-96 season.

ON: +980
OFF: +24 (because 1004 total team margin)

Convert that to per/48 and then the whole career data is per/48. That's what I would do. B-Ref pace estimates seem weird and I haven't been able to reproduce their numbers. Even MJ's 1997 and 1998 numbers are a bit different for ON samples and very different for OFF sample when I manually calculate them using B-Ref's own data compared to what they have on there.

...

I hate to be a party pooper but I see some inconsistencies in Thinking Basketball data too.

In the 1989, 1990, 1992 and 1995 postseasons, the raw ON and raw OFF numbers don't add up to the total team margin.

1989: +63 ON, -41 OFF (+4 total margin)
1990: +100 ON, -47 OFF (+55 total margin)
1992: +135 ON, +15 OFF (+136 total margin)
1995: -15 ON, +24 OFF (+10 total margin)

1990 and 1995 are really close so no major issue there. But 1989 and 1992 seem significantly off.

When PHILA tracked the 1992 playoffs, he got +128 ON and +8 OFF. Not saying he's correct and Thinking Basketball isn't but his numbers add up to the total margin.

Maybe these discrepancies are a result of missing footage? Maybe Ben is simply not considering anything that doesn't exist on footage. For example if two minutes are missing in the 2nd quarter of one of the games, his data simply excludes that segment.

In the 1989 and 1991 playoffs, the sum of individual series totals in ON and OFF don't match the postseason totals. For 1991 the postseason totals match the team margin but the sum of the individual series doesn't. The Philly series seems to be the culprit where +29 ON 0 OFF doesn't match the +44 total series margin.

Of course these discrepancies are unlikely to shift the data much in either direction. ON sample is particular is rock solid but it's still annoying.

[lessthanjake]

Part of my reason for not fully updating it yet is just time (particularly to adjust to per-100-possessions terms). But one is that I’m not *entirely* sure what to do with the new data. Specifically, the precise Thinking Basketball data and your data for the 1986 and 1987 playoffs are both in per-48-minute terms. In my OP, I did a calculation to adjust things to per-100-possession terms, based on using Basketball-Reference’s pace estimations and assuming equal pace with Jordan on and off the floor. That adjustment inherently causes some error (since it is based on a pace *estimation* and the equal-pace assumption is not going to be *exactly* right). I’d assumed that the error caused would be small, though. That seemed like a reasonable assumption IMO. However, it has been pointed out in this thread that if you made that adjustment to players’ plus-minus data that BBREF has plus-minus and on-off data for, the result generally is actually pretty far from BBREF’s reported on-off (having checked a lot of random players for this, I can say that the result using this adjustment usually, but not always, is a higher on-off than the on-off BBREF actually reports). I don’t know what is causing this discrepancy, and I actually think BBREF has some significant issues with its on-off (including that they sometimes have things like a positive “off” value for a player where BBREF’s own plus-minus says the team was outscored when the player was off the court). But regardless of what is causing that, I guess I’m hesitant to take precise data and then layer on an adjustment that may be inaccurate and controversial. So I’m inclined to just leave this new data in per-48-minute terms. At the same time, though, there’s downside to doing that. People are used to seeing plus-minus data in per-100-possession terms, so it won’t naturally be so clear to people exactly how good a given per-48-minute on-off is. Furthermore, some of the data in the OP is in per-100-possession terms, without any real ability to put it in per-48-minute terms. Most pertinently, I believe the 1996 regular season data from Pollack is in per-100-possession terms, but with no data on how many possessions and therefore no way to convert to per-48-minute terms. So if I edited the OP to report most things out in per-48-minute terms, it’d be confusing since at least one data point would be using a different measuring stick. That’s not ideal. Do you have any particular thoughts on how it would be best to handle this? There doesn’t seem to be an ideal solution either way.

I have Pollack's raw ON for the 95-96 season.

ON: +980
OFF: +24 (because 1004 total team margin)

Convert that to per/48 and then the whole career data is per/48. That's what I would do. B-Ref pace estimates seem weird and I haven't been able to reproduce their numbers. Even MJ's 1997 and 1998 numbers are a bit different for ON samples and very different for OFF sample when I manually calculate them using B-Ref's own data compared to what they have on there.

...

I hate to be a party pooper but I see some inconsistencies in Thinking Basketball data too.

In the 1989, 1990, 1992 and 1995 postseasons, the raw ON and raw OFF numbers don't add up to the total team margin.

1989: +63 ON, -41 OFF (+4 total margin)
1990: +100 ON, -47 OFF (+55 total margin)
1992: +135 ON, +15 OFF (+136 total margin)
1995: -15 ON, +24 OFF (+10 total margin)

1990 and 1995 are really close so no major issue there. But 1989 and 1992 seem significantly off.

When PHILA tracked the 1992 playoffs, he got +128 ON and +8 OFF. Not saying he's correct and Thinking Basketball isn't but his numbers add up to the total margin.

Maybe these discrepancies are a result of missing footage? Maybe Ben is simply not considering anything that doesn't exist on footage. For example if two minutes are missing in the 2nd quarter of one of the games, his data simply excludes that segment.

In the 1989 and 1991 playoffs, the sum of individual series totals in ON and OFF don't match the postseason totals. For 1991 the postseason totals match the team margin but the sum of the individual series doesn't. The Philly series seems to be the culprit where +29 ON 0 OFF doesn't match the +44 total series margin.

Of course these discrepancies are unlikely to shift the data much in either direction. ON sample is particular is rock solid but it's still annoying.

Yeah, I’m in the middle of drafting up a revamp of the OP to update with new data and make it all in per-48-minute terms, and while doing it I noticed those issues with the Thinking Basketball tracking data as well. I’m inclined to just update the post with that data, but make note of the issue. The original post essentially had this same issue as well anyways—we just didn’t know about it back then!

[Owly]

Also, I said +900 or whatever, lol. That's an extreme extreme number. Guys we have on record who hit +900 on the season: '96 MJ, '15-'17 Curry, and '16 Green. Poor '96 Pippen one point short.

And given overall team quality/number of games in a season, it's entirely possible those are the only players to ever do it.

Teams with a points dif of better than 800 through 2019 were.

1971-72 Los Angeles Lakers* 1007
1970-71 Milwaukee Bucks* 1005
1995-96 Chicago Bulls* 1004
2016-17 Golden State Warriors* 954
1971-72 Milwaukee Bucks* 915
1996-97 Chicago Bulls* 886
2015-16 Golden State Warriors* 882
2015-16 San Antonio Spurs* 872
1991-92 Chicago Bulls* 856
2007-08 Boston Celtics* 841
2014-15 Golden State Warriors* 828

If (hypothetically) those teams maintained a constant points margin regardless of lineup (or did so with regard to the following players) the minute leaders and their "expected" pro-rata plus-minus would be

Wilt Chamberlain 883.0341254
Kareem Abdul-Jabbar 835.0442366
Michael Jordan 786.2037506
Klay Thompson 638.8134479
Kareem Abdul-Jabbar 828.7272497
Michael Jordan 697.3938165
Draymond Green 622.9014085
Kawhi Leonard 526.60746
Scottie Pippen 682.9006556
Paul Pierce 611.7524677
Stephen Curry 548.2929549

So only the '72 Lakers, '71 Bucks, and '92 Bulls survive as prior to '94. Seems likely someone else did it prior to '94, but I'd put it at likely we can count the total number of players to do it on our fingers.

Ad hoc thoughts ... seems sensible enough.

Now Curry doing it in limited minutes by substantially exceeding pro-rata expectations (in addition, indeed, to exceeding team on-off) made me want to give a few more teams just in case. And players sometimes can sometimes quickly build up a difference between their plus minus and overall points dif (thinking Embiid 2019 versus Toronto). So I'll give the next band of teams showing a few more pre-plus-minus teams in case anyone's curious

1985-86 NBA Boston Celtics* 772
1966-67 NBA Philadelphia 76ers* 765
1986-87 NBA Los Angeles Lakers* 763
2012-13 NBA Oklahoma City Thunder* 755
1971-72 ABA Kentucky Colonels* 754
1990-91 NBA Chicago Bulls* 746
1969-70 NBA New York Knicks* 745
1993-94 NBA Seattle SuperSonics* 745
1985-86 NBA Milwaukee Bucks* 741
1961-62 NBA Boston Celtics* 739
2008-09 NBA Cleveland Cavaliers* 732
2018-19 NBA Milwaukee Bucks* 727
1996-97 NBA Utah Jazz* 721
1990-91 NBA Portland Trail Blazers* 712
1967-68 NBA Philadelphia 76ers* 705
1999-00 NBA Los Angeles Lakers* 701
1972-73 NBA Los Angeles Lakers* 700


Whilst I felt happy enough with the 800 arbitrary cutoff as a multiple of 100, and because GS clearly cleared it (828) and the next team clearly didn't (772) for instance another top heavy team with a high impact superstar and limited depth would have some shot. That next team is the '86 Celtics. Bird is no longer a complete minutes monster by this point but he still has an advantage in this regard over Curry ... then again you need some minutes for a garbage bench unit to be actively harmful once the team is below +900. Ultimately it gets really tricky for this group: '97 Karl Malone has +768 plus-minus (from Statmuse) via a +21.9 on/off on a 7.97 SRS, 721 points diff team ... he is high impact on a really good team with a thin bench to pull the team (in the off) down ... and his still only a little over 5/6ths of the way there.