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

[Djoker]

Fantastic work OP!

Total Playoff On-Off in 1985 + 1988-1993 + 1996-1998
- On: +8.38
- Off: -8.56 [note: it is -8.83 if using the alternate value for 1990]
- On-Off: +16.94 [note: it is +17.21 if using the alternate value for 1990]

So MJ has roughly +17 ON-OFF for his entire playoff career excluding 1986 (3 games), 1987 (3 games) and 1995 (10 games).

[migya]

Is there a site or place that lists a player's performance against different levels of defense teams? Can someone give links.

Think there was a thread some time ago about players against different defenses.

[Colbinii]

Is there a site or place that lists a player's performance against different levels of defense teams? Can someone give links.

Think there was a thread some time ago about players against different defenses.

There is a thread. You can use the search function to find anything on RealGM.

[lessthanjake]

I think there’s a pretty good argument that the four-year period from 1987-1988 to 1990-1991 is Michael Jordan’s peak individually. After all, it’s the period of time where he peaks in terms of box stats like PER, win shares per 48, and BPM in both regular season and playoffs. Of course, there’s arguments for other time periods for Jordan, especially since the 1987-1988 to 1990-1991 time period wasn’t his most successful in terms of team success. But let’s assume for now that his box-score-metric peak was his peak.

It occurs to me that because the two biggest Squared samples are from that time period, we actually have a lot of on-off data for Jordan for that time period. Indeed, between regular season and playoffs, we have on-off data for 41% of Jordan’s games (158 out of 387 games) in these peak years, and we can combine all that data together. So does the on-off data back up that Jordan was incredible in that time period? Well, in short, yes:

Peak Jordan On-Off per 100 Possessions in Regular Season + Playoffs (1987-1988 to 1990-1991) (sample: 158 out of 387 games)

- On: +7.63
- Off: -13.35
- On-Off: +20.98

In other words, we have data for almost half of peak Jordan’s games, and he’s basically got a +21 on-off per 100 possessions in those games. It’s not all the games, of course, so the actual overall number might differ a bit from that, but this is suggestive of peak Jordan perhaps being the most impactful player ever, as I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value). And of course, we can map this onto the fact that this is also probably the highest four-year peak in terms of advanced box measures. Certainly seems supportive of it being the highest peak ever.

[OhayoKD]

I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value)

RK's magic analysis has him in a similar light and Karl Malone is pulling off that +17 across hundreds of games rather than incomplete samples bookending a 4 year period.

(and that is assuming your calculation isn't inflated...)

[Owly]

I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value)

RK's magic analysis has him in a similar light and Karl Malone is pulling off that +17 across hundreds of games rather than incomplete samples bookending a 4 year period.

(and that is assuming your calculation isn't inflated...)

Robinson has 4 years in that vicinity (high teens) 94-97 (only RS available). I think less probable collinearity issues than Malone.
'97 is mostly just an off sample though (147 minutes on). So perhaps only 3 years.
On the other hand our hand is somewhat forced because we don't know the numbers for '93 (next healthy season ['98] would be worse, and would open up a lot of fuzziness about what's eligible, make it harder to count).

I will say depending somewhat on the source the available games for incomplete data players where the games were almost always going out (i.e. not 60s, more 80s, 90s) might at the margin be tilted towards showpiece games for the superstar ("Do we rebroadcast [and end up with youtube having] the game where X drops 60 or the night he scored 17 at a poor percentage?"). Arguably this is mostly neutralized at some point for someone like Jordan in terms of the sheer volume of stuff that was out there ... if you're getting the majority of games they're probably available just for Jordan being in them (and any "curation" would probably be less influential).

This potential for bias may also be less the case (unless the type of decision making was involved in decisions regarding saving tapes or not) at the NBA archive level.

Regardless I'd be inclined to mentally regress back such an outlier number a little bit (or have an awareness of such) with half the sample still missing.

[Djoker]

I think there’s a pretty good argument that the four-year period from 1987-1988 to 1990-1991 is Michael Jordan’s peak individually. After all, it’s the period of time where he peaks in terms of box stats like PER, win shares per 48, and BPM in both regular season and playoffs. Of course, there’s arguments for other time periods for Jordan, especially since the 1987-1988 to 1990-1991 time period wasn’t his most successful in terms of team success. But let’s assume for now that his box-score-metric peak was his peak.

It occurs to me that because the two biggest Squared samples are from that time period, we actually have a lot of on-off data for Jordan for that time period. Indeed, between regular season and playoffs, we have on-off data for 41% of Jordan’s games (158 out of 387 games) in these peak years, and we can combine all that data together. So does the on-off data back up that Jordan was incredible in that time period? Well, in short, yes:

Peak Jordan On-Off per 100 Possessions in Regular Season + Playoffs (1987-1988 to 1990-1991) (sample: 158 out of 387 games)

- On: +7.63
- Off: -13.35
- On-Off: +20.98

In other words, we have data for almost half of peak Jordan’s games, and he’s basically got a +21 on-off per 100 possessions in those games. It’s not all the games, of course, so the actual overall number might differ a bit from that, but this is suggestive of peak Jordan perhaps being the most impactful player ever, as I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value). And of course, we can map this onto the fact that this is also probably the highest four-year peak in terms of advanced box measures. Certainly seems supportive of it being the highest peak ever.

I just quickly checked the W-L % as well.

In the 158 games sampled, the Bulls went 100-58 (63.3%).
In the 229 games not sampled, the Bulls went 151-78 (65.9%).

So it's not inconceivable at least based on team results that the ON-OFF would hold.

A more detailed analysis would look at the MOV in games sampled vs. not sampled.

[lessthanjake]

I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value)

RK's magic analysis has him in a similar light and Karl Malone is pulling off that +17 across hundreds of games rather than incomplete samples bookending a 4 year period.

(and that is assuming your calculation isn't inflated...)

Robinson has 4 years in that vicinity (high teens) 94-97 (only RS available). I think less probable collinearity issues than Malone.
'97 is mostly just an off sample though (147 minutes on). So perhaps only 3 years.
On the other hand our hand is somewhat forced because we don't know the numbers for '93 (next healthy season ['98] would be worse, and would open up a lot of fuzziness about what's eligible, make it harder to count).

I will say depending somewhat on the source the available games for incomplete data players where the games were almost always going out (i.e. not 60s, more 80s, 90s) might at the margin be tilted towards showpiece games for the superstar ("Do we rebroadcast [and end up with youtube having] the game where X drops 60 or the night he scored 17 at a poor percentage?"). Arguably this is mostly neutralized at some point for someone like Jordan in terms of the sheer volume of stuff that was out there ... if you're getting the majority of games they're probably available just for Jordan being in them (and any "curation" would probably be less influential).

This potential for bias may also be less the case (unless the type of decision making was involved in decisions regarding saving tapes or not) at the NBA archive level.

Regardless I'd be inclined to mentally regress back such an outlier number a little bit (or have an awareness of such) with half the sample still missing.

I don’t really think that we have an issue with the games being tilted towards showpiece games. For one thing, the Bulls actually did worse in the sample of games we have data for than they did in the games we don’t have (see Djoker’s post above). Second of all, the games weren’t really chosen from stuff put on YouTube I don’t think. The data includes all playoff games (so no selection bias there, as it’s just all of them). The rest of the data we have is from Squared, and my understanding is that Squared has a massive amount of game footage that he has gotten from the NBA and is just slowly going through it randomly, so I’m not sure there’s any reason to think it’s biased by showpiece matches—it’s not just someone looking through things that happened to be posted on YouTube.

All that said, that doesn’t mean that there’s no sampling error. It’s only like 41% of the games from Jordan’s peak years, so it’s certainly possible that the full data wouldn’t be as good. I just don’t think that this particular issue would be tilting things.

I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value)

RK's magic analysis has him in a similar light and Karl Malone is pulling off that +17 across hundreds of games rather than incomplete samples bookending a 4 year period.

Yes, Magic looks great in what we have. We don’t have as much concentrated stuff for him (so we can’t exactly make as confident a claim about a 4-year span for him, due to sample size issues), but I think what we do have overall actually looks pretty similar in terms of on-off. It’s a bit lower sample size, but it’s enough that my baseline assumption is that Magic’s on-off was sky high too.

[rk2023]

I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value)

RK's magic analysis has him in a similar light and Karl Malone is pulling off that +17 across hundreds of games rather than incomplete samples bookending a 4 year period.

(and that is assuming your calculation isn't inflated...)

FWIW -
1988 (47 Games):
10.85 +/- Per-48 with Magic, -10.29 +/- Per-48 without

1991 (48 Games):
6.62 +/- Per-48 with Magic, -9.03 +/- Per-48 without

1988 (42 Games):
5.65 +/- Per-48 with Mike, -17.14 +/- Per-48 without

1991 (56 Games):
9.84 +/- Per-48 with Mike, -7.63 +/- Per-48 without

Take this however y'all will

[homecourtloss]

I’m not aware of anyone with that kind of on-off over a four-year span (the closest I’m aware of is Steph from 2014-2017, who had about a +17 RS+PS on-off, and with a higher “on” value)

RK's magic analysis has him in a similar light and Karl Malone is pulling off that +17 across hundreds of games rather than incomplete samples bookending a 4 year period.

(and that is assuming your calculation isn't inflated...)

FWIW -
1988 (47 Games):
10.85 +/- Per-48 with Magic, -10.29 +/- Per-48 without

1991 (48 Games):
6.62 +/- Per-48 with Magic, -9.03 +/- Per-48 without

1988 (42 Games):
5.65 +/- Per-48 with Mike, -17.14 +/- Per-48 without

1991 (56 Games):
9.84 +/- Per-48 with Mike, -7.63 +/- Per-48 without

Take this however y'all will

1990-1991

Regular Season (source: Squared data; 56 games)
- On: +9.85
- Off: -9.77
- On-Off: +19.62

Thank you for your calculations here and then other places, RK.

If LessThanJake is doing a per 100 calculation and you have done a per 48 calculation, how is the ON value pretty much the same but the off value so much higher in ltj’s here? In any case, +9.85 on court in the consensus peak year? That’s…underwhelming.

Also, the 1991 Bulls were a +9.5 NRtg (i.e., per 100) team. IF Jordan wound up with something like +9.85 ON for the season, the Bulls could NOT be a -9.77 with him off and still wind up a +9.5 team.

If he was +9.85 over an entire 82 games (to extend the 56 game sample) then for the Bulls to wind up a +9.5 team, the Bulls would have to be like a +8 team in the minutes he was off. If he was +9.85 for 82 games and the Bulls were -9.77 in the minutes he was off in these 82 games, Bulls would wind up like a +5.5 team. So the other 26 games must have some wildly different numbers. If anything, this should make us wary of partial samples, and we should caution against it it.

Look at the Nuggets this year. If we take the +12 ON of BKREF for the minutes Jokic was on and the -10 when he was off, we get the +22 ON-OFF and the Nuggs go from the +12 with Jokic on to the +3.4 overall (Jokic had more off minutes than Jordan).

Also, if Jordan does wind up with something around +9.85 on court in his consensus peak year (or about +10.9 for the RS+PS), well…that doesn’t seem like a GOAT peak when you have players with multiple year runs of RS+PS with on court numbers better than +9.85 or +10.9 and especially so if we’re talking single season. In a previous discussion on another thread, it was mentioned that a large on-off isn’t as impressive because someone is taking a bad team and making them good and that what was more impressive was the on court rating.

Peak Jordan On-Off per 100 Possessions in Regular Season + Playoffs (1987-1988 to 1990-1991) (sample: 158 out of 387 games)

- On: +7.63
- Off: -13.35
- On-Off: +20.98
.

So Peak Jordan was +7.63 RS+PS on court over a 158 game sample. That’s…nice I guess, but not the type of astonishing high on court number that one would expect of a consensus peak stretch for a goat candidate. I recall in a different thread, KG’s on-off wasn’t all *that* impressive as it was mostly a high on-off raising a bad team, etc.

And even leaving aside raw achievement, I’ll repeat again that there’s diminishing marginal returns on things like on-off. It is definitely easier to raise an awful team a lot in terms of on-off than it is to raise a good team. And that also means it’s easier to accrue high RAPM on a bad team—because ultimately RAPM is aiming to isolate the effect you had on your team’s net rating and you can have a bigger effect on your team's net rating if the baseline for the team is bad. So, to me, I don't really find Garnett's on-off numbers or RAPM with the Timberwolves all *that* impressive. It’s mostly just high on-off (and high RAPM) as a result of being on a team that was awful without him. Which I think inflates the numbers due to being far on the no-diminishing-returns end of things.

Btw,

KG 2002 through 2004 (272 games): +6.21 on, +17.75 on-off
KG 2003 and 2004 (188 games): +6.73 on, +21.45 on-off

[lessthanjake]

RK's magic analysis has him in a similar light and Karl Malone is pulling off that +17 across hundreds of games rather than incomplete samples bookending a 4 year period.

(and that is assuming your calculation isn't inflated...)

FWIW -
1988 (47 Games):
10.85 +/- Per-48 with Magic, -10.29 +/- Per-48 without

1991 (48 Games):
6.62 +/- Per-48 with Magic, -9.03 +/- Per-48 without

1988 (42 Games):
5.65 +/- Per-48 with Mike, -17.14 +/- Per-48 without

1991 (56 Games):
9.84 +/- Per-48 with Mike, -7.63 +/- Per-48 without

Take this however y'all will

1990-1991

Regular Season (source: Squared data; 56 games)
- On: +9.85
- Off: -9.77
- On-Off: +19.62

Thank you for your calculations here and then other places, RK.

If LessThanJake is doing a per 100 calculation and you have done a per 48 calculation, how is the ON value pretty much the same but the off value so much higher in ltj’s here?

I can’t speak to anyone else’s calculation. It’s possible I made a data entry error (or I guess perhaps RK did). I might eventually go back and see if I made a mistake. But of course, one possible explanation here is that the Bulls played a substantially lower pace with Jordan off the floor, which could cause the “off” value to be worse in per-100-possessions terms than it looks in per-48-minutes terms. (And, by the way, if that were the case, then it’d imply that the data I derived assuming equal pace “on” and “off” the court might be underestimating Jordan’s on-off).

Anyways, as an aside, I’ll note that, unlike the method using Thinking Basketball’s video, my method for the Squared stuff did not include any per-48-minute calculations, because we actually have possession numbers from Squared. I did derive a total possessions estimate (and therefore an “off” possession estimate after subtracting out Squared’s “on” possession numbers) using BBREF’s pace numbers on a game-by-game level. If anything, that should make the data worse for Jordan, though, since apparently Basketball Reference counts more possessions than Squared does (Squared has a post about it), which means that this method would estimate a higher number of “off” possessions than Squared would’ve. But that shouldn’t really make a big difference, since the differences in BBREF’s and Squared’s possession calculations weren’t large.

Also, the 1991 Bulls were a +9.5 NRtg (i.e., per 100) team. IF Jordan wound up with something like +9.85 ON for the season, the Bulls could NOT be a -9.77 with him off and still wind up a +9.5 team.

As I noted in my original post, the Bulls games in the Squared sample for 1991 were actually substantially less good for the Bulls than the full season. The Bulls only lost one of the unsampled games and had a bunch of blowout victories in the unsampled games. So this is not surprising at all. And it relates to something that I noted from the very beginning in my OP. As I mentioned in my OP, I’m not sure how it cuts in terms of bias, but my general inclination is to think that Jordan would probably look better in a better sample of games. We don’t know though, of course, and sampling error is definitely an issue with the Squared stuff and I think is a particular concern with the 1990-1991 season data since we know it’s not really a representative sample of games.

Also, if he does wind up with something around +9.85 on court in his consensus peak year (or about +10.9 for the RS+PS), well…that doesn’t seem like a GOAT peak when you have players with multiple year runs of RS+PS with on court numbers better than +9.85 or +10.9 and especially so if we’re talking single season. In a previous discussion on another thread, it was mentioned that a large on-off isn’t as impressive because someone is taking a bad team and making them good and that what was more impressive was the on court rating.

Jordan’s “on” number is surely substantially higher for the entire season, because, as noted above, the unsampled games were a lot better for the Bulls than the sampled games. So this isn't really a particularly valid point. I do agree, though, that the super high on-offs in Jordan's earlier years aren't as impressive as they would be if they were with a higher "on" value. Which is something I actually flagged in another thread when I noted that Jordan's four-year peak on-off in the data we have is higher than any four-year period I'm aware of for others but that Steph is fairly close but with a higher "on." In any event, Jordan does have some fantastic “on” values in the full data we have for the 1996 and 1997 regular seasons, so he’s not lacking in those sorts of seasons.

Peak Jordan On-Off per 100 Possessions in Regular Season + Playoffs (1987-1988 to 1990-1991) (sample: 158 out of 387 games)

- On: +7.63
- Off: -13.35
- On-Off: +20.98
.

So Peak Jordan was +7.63 RS+PS over a 158 game sample. That’s…nice I guess, but not the type of astonishing high on court number that one would expect of a consensus peak stretch for a goat candidate. I recall in a different thread, KG’s on-off wasn’t all *that* impressive as it was mostly a high on-off raising a bad team, etc.

Btw,

KG 2002 through 2004 (272 games): +6.21 on, +17.75 on-off
KG 2003 and 2004 (188 games): +6.73 on, +21.45 on-off

You’re not quite using four-year spans there (though I understand your point is similar number of games), but yeah, that’s a very high on-off span for Garnett—basically the highest in the play-by-play era.

And the fact is that if Jordan was rocking a +21 on-off in a four-year span, it would be unmatched by anyone in the play-by-play era, including Garnett (with the caveat again that Steph’s higher “on” value makes his best four-year span arguably more impressive in this regard). We don’t know for sure if Jordan really did have that high an on-off over four years, since we only have 41% of the data from those years, meaning there’s potentially significant sampling error. But what we do have looks undeniably incredible. And I’m a little confused by you saying this “is not the type of astonishing high on court number that one would expect of a consensus peak stretch for a goat candidate” since the player voted #1 by this board has no such stretch. You seem to often create an artificial bar for Jordan that you don’t apply to anyone else (and that no one else would meet), and then say Jordan doesn’t meet it.

[Squared2020]

.

[MrVorp]

Also, the 1991 Bulls were a +9.5 NRtg (i.e., per 100) team. IF Jordan wound up with something like +9.85 ON for the season, the Bulls could NOT be a -9.77 with him off and still wind up a +9.5 team.

If he was +9.85 over an entire 82 games (to extend the 56 game sample) then for the Bulls to wind up a +9.5 team, the Bulls would have to be like a +8 team in the minutes he was off. If he was +9.85 for 82 games and the Bulls were -9.77 in the minutes he was off in these 82 games, Bulls would wind up like a +5.5 team. So the other 26 games must have some wildly different numbers. If anything, this should make us wary of partial samples, and we should caution against it it.

Look at the Nuggets this year. If we take the +12 ON of BKREF for the minutes Jokic was on and the -10 when he was off, we get the +22 ON-OFF and the Nuggs go from the +12 with Jokic on to the +3.4 overall (Jokic had more off minutes than Jordan).

Also, if Jordan does wind up with something around +9.85 on court in his consensus peak year (or about +10.9 for the RS+PS), well…that doesn’t seem like a GOAT peak when you have players with multiple year runs of RS+PS with on court numbers better than +9.85 or +10.9 and especially so if we’re talking single season. In a previous discussion on another thread, it was mentioned that a large on-off isn’t as impressive because someone is taking a bad team and making them good and that what was more impressive was the on court rating.

Except in that sample the Bulls were a +6 team when they finished as a +9 team. So the games not sampled in that season would have the bulls playing as a roughly +15 team.

[Djoker]

Here's some of the Jordan playoff game data I have. I didn't have which games were which tagged as this data is older.









Edit: I do recall Ben Taylor noticing a timing error in one of those files a year ago. The plus-minus is good; but I am sure there's one playing time mistake in there. -- found it: Game 2 vs. New York; entered game with 12:17 remaining. I think it was actually 11:17 remaining.

You're doing god's work.

If you have it, can you post the plus minus for all MJ games for the 1995 playoffs?

[WestGOAT]

If anything, that should make the data worse for Jordan, though, since apparently Basketball Reference counts more possessions than Squared does (Squared has a post about it), which means that this method would estimate a higher number of “off” possessions than Squared would’ve. But that shouldn’t really make a big difference, since the differences in BBREF’s and Squared’s possession calculations weren’t large.

On here or on his site? Could you provide a link? I'm trying to estimate game-level possessions for games prior to 1985:

Total number of NBA games missing possessions by season:

Spoiler:

Code: Select all

1955     618
1956     622
1957     614
1958     618
1959     620
1960     650
1961     682
1962     778
1963     778
1964     774
1965     772
1966     774
1967     872
1968    1064
1969    1226
1970    1228
1971    1474
1972    1468
1973    1476
1974     870
1975     876
1976     804
1977     978
1978     770
1979     584
1980     632
1981     832
1982     892
1983     342
1984       2

Percentage of games missing possessions by season:

Spoiler:

Code: Select all

1955    100.000000
1956    100.000000
1957    100.000000
1958    100.000000
1959    100.000000
1960    100.000000
1961    100.000000
1962    100.000000
1963    100.000000
1964    100.000000
1965    100.000000
1966    100.000000
1967    100.000000
1968    100.000000
1969    100.000000
1970    100.000000
1971    100.000000
1972    100.000000
1973    100.000000
1974     58.943089
1975     55.796178
1976     51.015228
1977     51.204188
1978     40.398741
1979     30.575916
1980     33.263158
1981     41.767068
1982     45.050505
1983     17.342799
1984      0.097847
1985           NaN
1986           NaN

Would be great to eventually try to calculate Offensive-WOWY and Defensive-WOWY, similar to what Moonbeam did.

[lessthanjake]

If anything, that should make the data worse for Jordan, though, since apparently Basketball Reference counts more possessions than Squared does (Squared has a post about it), which means that this method would estimate a higher number of “off” possessions than Squared would’ve. But that shouldn’t really make a big difference, since the differences in BBREF’s and Squared’s possession calculations weren’t large.

On here or on his site? Could you provide a link? I'm trying to estimate game-level possessions for games prior to 1985:

Total number of NBA games missing possessions by season:

Spoiler:

Code: Select all

1955     618
1956     622
1957     614
1958     618
1959     620
1960     650
1961     682
1962     778
1963     778
1964     774
1965     772
1966     774
1967     872
1968    1064
1969    1226
1970    1228
1971    1474
1972    1468
1973    1476
1974     870
1975     876
1976     804
1977     978
1978     770
1979     584
1980     632
1981     832
1982     892
1983     342
1984       2

Percentage of games missing possessions by season:

Spoiler:

Code: Select all

1955    100.000000
1956    100.000000
1957    100.000000
1958    100.000000
1959    100.000000
1960    100.000000
1961    100.000000
1962    100.000000
1963    100.000000
1964    100.000000
1965    100.000000
1966    100.000000
1967    100.000000
1968    100.000000
1969    100.000000
1970    100.000000
1971    100.000000
1972    100.000000
1973    100.000000
1974     58.943089
1975     55.796178
1976     51.015228
1977     51.204188
1978     40.398741
1979     30.575916
1980     33.263158
1981     41.767068
1982     45.050505
1983     17.342799
1984      0.097847
1985           NaN
1986           NaN

Would be great to eventually try to calculate Offensive-WOWY and Defensive-WOWY, similar to what Moonbeam did.

MrVORP linked to it here: https://forums.realgm.com/boards/viewtopic.php?p=107992860#p107992860

[lessthanjake]

I mentioned this a bit above, but one think I want to note is that comparing per-100-possession data from Squared’s 1990-1991 sample with RK’s per-48-minute analysis of that data does actually suggest that the Bulls played a substantially lower pace with Jordan off the court.

Specifically, RK’s analysis indicates that the Bulls’ net rating per 48 minutes with Jordan on the court was essentially identical to the per-100-possession number I got using Squared’s specific count of Jordan’s possessions on the court (+9.84 vs. +9.85). Definitionally indicates that the Bulls were playing with about a 100 pace with Jordan on the court. After all, pace is a measure of number of possessions per 48 minutes, and if the net rating per 100 possessions and the net rating per 48 minutes are the same, then the team must be playing 100 possessions per 48 minutes. But we also know from Basketball Reference data that the Bulls overall pace that year was only 95.6. And in the specific 56 games Squared sampled, the Bulls’ pace was 95.0. (Squared tends to estimate slightly lower numbers of possessions, so Squared’s estimate of total possessions would presumably actually be slightly lower than that). If the Bulls’ pace with Jordan was about 100, and the Bulls’ overall pace was just 95, then that’d suggest the Bulls played at a much slower pace without Jordan. That doesn’t really matter much for the purposes of this particular estimate, since I didn’t make any equal-pace assumption there (because we had “on” possessions for Jordan from Squared). But if this were true more generally, then the equal-pace assumption I made would actually underestimate Jordan’s on-off, and it might underestimate it by a good deal if the difference were as big as it looks like it was in 1990-1991.

[AEnigma]

Now you are finally hovering around what I have been gesturing at with all these assumptions about how pace “should” translate to per 48 measures.

Say we take it as a given that the Bulls played at a 95 pace across those games. And for the sake of ease, say we assume Jordan was matching his season average in minutes (this does not really matter to make the point). If Jordan was playing at a 100 pace on the court in 37 minutes a game, for the team’s total average to be 95 by playing 11 minutes a game, the Bulls would need to play at a pace of 78 without Jordan. Maybe you will decide that seems like a reasonable assumption, but the problem with playing around with minutes adjustments and unofficial pace adjustments is this phenomenon where the per 48 numbers end up strangely similar to “official” pace numbers, even on supposedly slow-paced teams.

Quick exercise since I have done this one a few times now: 2021 postseason Lebron is +10 in 224 minutes (perhaps technically 223.6 minutes), so logically that should be +2.14 per 48 minutes. Yet again, basketball-reference says +1 per 100. And it estimates the pace at either 92.9 (by the series page) or at 93.8 (postseason summary page). Well, logically, for Lebron to go down from +2.14 per 48 minutes to +1 per 100 possessions in a series that was somewhere in the vicinity of a 93 pace, then Lebron must have been booking it… right?

Okay, so say you agree that basketball-reference did a garbage job in that instance. What if we used a “better” source. This time we can look at the 2010 postseason (because we want “slower team when off”). Per 48 (+44 in 460 minutes), Lebron is +4.59 on. Basketball-reference says +4.6 per 100, but we are ignoring them. However, NBA.com says +4.5 per 100, so was Lebron playing fast? To go from 4.59 per 48 minutes to +4.5 per 100 possessions — and to maximise the benefit of the doubt, how about we pretend the real number is +4.549 or something — Lebron would need to be playing at least at a pace of 101. And NBA.com says the pace for the team was 93.82. Since Lebron played 460 minutes on and only 68 minutes off, that would mean his team would have needed to play at a 45 pace without him!

And then we check Lebron’s on/off for pace and see that NBA.com says Lebron played at a 94.29 pace and the team played at a 93.77 pace without him.

These numbers do not extrapolate the way you assume they do. If I were doing the same blind Jordan process for Lebron, I would see basketball-reference list a 92.6 postseason pace and say he should actually be +4.96 per 100 on. Or if I think there is better data on NBA.com, I could see that 93.82 team pace and conclude he was +4.89 per 100 on. Or maybe I would see the 94.29 on-court pace number and say he should be +4.87 per 100.

However, that is not what happens with any of these sites. So why would you conclude that is what would happen with Jordan? Your process inflates numbers relative to what is listed on all these “official” sites; I am not criticising that in itself, but it is why I insist on you applying the process evenly. Numbers are not meaningful without equal comparison. And this is not isolated to Lebron: this applies to everyone you are comfortable declaring to have fallen short of Jordan’s values based on completely disparate means of measurement.

Here, try it out with Curry if you do not care about running numbers for Lebron. Surely you must be at least a little curious about what would happen if you went through the same process you used on Jordan? I get +10.1 per 48 on for 2015-19 Curry, but pbp only has him at +9.74 even though it also says he played at a 97.893 pace. Do you not think that is something worth considering so long as you continue to advance this methodology?

[Djoker]

@Enigma

The issue you're describing is particularly bad with OFF samples which tend to be very small. Even for Lebron's entire playoff career he probably set on the bench for like 3000 minutes which is equivalent to one full regular season of minutes. With most players it's even much less than that so the OFF sample even for an entire career isn't very large.

OFF samples can also be affected by how we count minutes. Pace corrections can make it worse.

For instance say we have a hypothetical game where a team is +10 in 45:19 that player X is ON and -2 in 2:41 where player X is OFF. Per 48 that translates to +10.592 ON and -35.776 OFF.

However when rounding the minutes which is done often, that game becomes 45 minutes ON and 3 minutes OFF. Recalculating the per 48 we now get +10.667 ON and -32.000 OFF. The OFF value is quite different..

Of course since these are random errors, you would hope that they even out over large samples. In some games the ratings are overestimated and in some underestimated.

I think pace estimates are quite inconsistent for a variety of reasons (how pace is counted, how minutes are counted) but over large samples this is unlikely to skew the results just IMO. Because random errors happen in every which direction. Kind of who the wind blows in different directions every day.

We see that with PBP ON-OFF for Lebron. The data is different than BRef but some years it makes Lebron look better and some years worse. Overall it makes him look a bit better maybe +11 as opposed to +10.2 for his career but it's still in the same ballpark.

It's fair to give the numbers with error bars of some sort though. So for instance +17 ON-OFF for Jordan in the OP can be stated as let's +16 to +18. Or if larger error then like +15 to +19.

That's also why all serious analysts such as Ben Taylor give their low range and high range estimates. But that's beyond the scope of this thread as I doubt folks here are either willing or capable to do high level statistical analysis.

[lessthanjake]

Now you are finally hovering around what I have been gesturing at with all these assumptions about how pace “should” translate to per 48 measures.

Say we take it as a given that the Bulls played at a 95 pace across those games. And for the sake of ease, say we assume Jordan was matching his season average in minutes (this does not really matter to make the point). If Jordan was playing at a 100 pace on the court in 37 minutes a game, for the team’s total average to be 95 by playing 11 minutes a game, the Bulls would need to play at a pace of 78 without Jordan. Maybe you will decide that seems like a reasonable assumption, but the problem with playing around with minutes adjustments and unofficial pace adjustments is that this phenomenon where the per 48 numbers end up strangely similar to “official” pace numbers, even on supposedly slow-paced teams.

Quick exercise since I have done this one a few times now: 2021 postseason Lebron is +10 in 224 minutes (perhaps technically 223.6 minutes), so logically that should be +2.14 per 48 minutes. Yet again, basketball-reference says +1 per 100. And it estimates the pace at either 92.9 (by the series page) or at 93.8 (postseason summary page). Well, logically, for Lebron to go down from +2.14 per 48 minutes to +1 per 100 possessions in a series that was somewhere in the vicinity of a 93 pace, then Lebron must have been booking it… right?

Okay, so say you agree that basketball-reference did a garbage job in that instance. What if we used a “better” source. This time we can look at the 2010 postseason (because we want “slower team when off”). Per 48 (+44 in 460 minutes), Lebron is +4.59 on. Basketball-reference says +4.6 per 100, but we are ignoring them. However, NBA.com says +4.5 per 100, so was Lebron playing fast? To go from 4.59 per 48 minutes to +4.5 per 100 possessions — and to maximise the benefit of the doubt, how about we pretend the real number is +4.549 or something — Lebron would need to be playing at least at a pace of 101. And NBA.com says the pace for the team was 93.82. Since Lebron played 460 minutes on and only 68 minutes off, that would mean his team would have needed to play at a 45 pace without him!

And then we check Lebron’s on/off for pace and see that NBA.com says Lebron played at a 94.29 pace and the team played at a 93.77 pace without him.

These numbers do not extrapolate the way you assume they do. If I were doing the same blind Jordan process for Lebron, I would see basketball-reference list a 92.6 postseason pace and say he should actually be +4.96 per 100 on. Or if I think there is better data on NBA.com, I could see that 93.82 team pace and conclude he was +4.89 per 100 on. Or maybe I would see the 94.29 on-court pace number and say he should be +4.87 per 100.

However, that is not what happens with any of these sites. So why would you conclude that is what would happen with Jordan? Your process inflates numbers relative to what is listed on all these “official” sites; I am not criticising that in itself, but it is why I insist on you applying the process evenly. Numbers are not meaningful without equal comparison. And this is not isolated to Lebron: this applies to everyone you are comfortable declaring to have fallen short of Jordan’s values based on completely disparate means of measurement.

Here, try it out with Curry if you do not care about running numbers for Lebron. Surely you must be at least a little curious about what would happen if you went through the same process you used on Jordan? I get +10.1 per 48 on for 2015-19 Curry, but pbp only has him at +9.74 even though it also says he played at a 97.893 pace. Do you not think that is something worth considering so long as you continue to advance this methodology?

I think this provides good reason to not be particularly certain about extrapolations about exact on and off pace differences based on per-48-minute data. Which, I’ll note, is why my last post did not provide precise numbers and was not worded in any definitive way (I talked about what the data “suggests” and said “if this were true” and that it “might underestimate” Jordan’s on-off “ if the difference were as big as it looks like”). The truth is that we don’t really know for sure what, if any, pace differences on and off the floor there were for Jordan but, as I’ve said, we have no reason whatsoever to believe that difference would be the same direction and magnitude for Jordan as it is for LeBron, and so using data for LeBron that we know would incorporate this error (when we have data without the error) does not make sense since there’s no reason to believe that would cancel out any error from pace difference for Jordan (indeed, the equal-pace assumption might well *hurt* Jordan, rather than help like it would for the LeBron).

As for the rest of it, you’re largely just pointing out that on-off models can be a bit wonky (especially at smaller sample sizes), likely due to possession estimations sometimes being inaccurate at small sample sizes.

At a fundamental level, on-off models are absolutely just trying to take +/- when “on” and divide by an estimate of possessions “on” and taking +/- when “off” and dividing by an estimate of possessions “off.” And that is exactly what I did, except with an equal-pace assumption to determine the breakdown between “on” and “off” possessions. So I don’t know what other error/bias you’re specifically asserting there is. Is your assertion that the possession estimate formula BBREF uses is different when calculating pace than it is when calculating on-off? That seems…quite unlikely, but if you have evidence of that then please show it. Otherwise I don’t really see what source of error/bias you’re identifying beyond error caused by the equal-pace assumption, and we have no idea the direction or magnitude of that for Jordan (and have at least some indication it might be *hurting* his numbers). The bottom line is that if you think my method is biasing the data in a specific way beyond the equal-pace assumption, you need to explain exactly how rather than just declaring that it must be different without identifying the specific mechanism through which that’d be the case.