Looking at lineup average offensive and defensive efficiencies would help with the regular season to playoffs comparisons.
Dallas, a high octane offense, had a fairly favorable path to the finals for them in the sense that 2 of the western rounds were against teams with a regular season defensive efficiency above 107 (though barely) and all three western round opponents finished with playoff defensive efficiencies above that standard (2 well above).
From 2001 to 2010 the teams who bounced the Mavericks from the Mavs from the playoffs all had a defensive efficiency of 107.5 or lower in the regular season (most under 107, though a few were slightly above).
The Lakers defense didn't do its job against the Mavs this time.
What if player APM is used as the guide to build lineups?
Re: What if player APM is used as the guide to build lineups
Are good offensive teams more likely to lose to a good defensive team, or to a team with better overall point differential?
I think you've just repeated what the somebody I mentioned had done -- after the fact, the Mavs won, so they had favorable matchups vs opponents who, having lost, weren't very good...
Dallas in the season had been 2.5 ppg better than Portland, and they won that series by 5 ppg.
LA had been 1.6 ppg better in the season, and they'd beaten the Hornets by 7 ppg, before the Mavs dismantled them by 14 ppg.
Oklahoma had had their way with Denver by 5 ppg, after a season that suggested Den by 1 ppg;
favored by 1.3 ppg over Memphis, they won by almost 4 ppg.
The Thunder, just 0.6 ppg behind the Mavs in the season and doing some 3-4 ppg better in the playoffs, nevertheless became just another victim, by 4 ppg.
The Miami Heat, who couldn't seem to beat elite teams all year, had sliced through the East like a hot knife through butter.
Every one of Dallas' opponents was formidable, and in playoffs they'd all outperformed expectations based on the season. Getting thrashed by the Mavs of course trashes their playoff rates.
I think the winner can't be predicted by matchups, season lineup unit performance, offense vs defense, or anything more than by the most general point differential. Beyond that, the team that happens to play better than their usual will gain advantage.
The Mavs played a bit better than their usual in 2 series, and a lot better in the other 2.
No bad series = good chance of winning.
I think you've just repeated what the somebody I mentioned had done -- after the fact, the Mavs won, so they had favorable matchups vs opponents who, having lost, weren't very good...
Dallas in the season had been 2.5 ppg better than Portland, and they won that series by 5 ppg.
LA had been 1.6 ppg better in the season, and they'd beaten the Hornets by 7 ppg, before the Mavs dismantled them by 14 ppg.
Oklahoma had had their way with Denver by 5 ppg, after a season that suggested Den by 1 ppg;
favored by 1.3 ppg over Memphis, they won by almost 4 ppg.
The Thunder, just 0.6 ppg behind the Mavs in the season and doing some 3-4 ppg better in the playoffs, nevertheless became just another victim, by 4 ppg.
The Miami Heat, who couldn't seem to beat elite teams all year, had sliced through the East like a hot knife through butter.
Every one of Dallas' opponents was formidable, and in playoffs they'd all outperformed expectations based on the season. Getting thrashed by the Mavs of course trashes their playoff rates.
I think the winner can't be predicted by matchups, season lineup unit performance, offense vs defense, or anything more than by the most general point differential. Beyond that, the team that happens to play better than their usual will gain advantage.
The Mavs played a bit better than their usual in 2 series, and a lot better in the other 2.
No bad series = good chance of winning.
Re: What if player APM is used as the guide to build lineups
"Are good offensive teams more likely to lose to a good defensive team, or to a team with better overall point differential?"
That would be worth defining fairly and more tightly and checking. Perhaps later.
Not sure who the "somebody" you refer to is. I have a few guesses but if you think it is important to allude to, would you clarify who and what you are specifically referring to?
After the fact playoff data is problematic for evaluating team and opponent performance. With regard to Dallas and quality of defenses faced I looked at both playoff and regular season. I rounded it a bit to make an over-arcing point which I thought interesting. I didn't have the time to find or create the Dallas data series by series but more could be done to refine the story, with the time & interest.
"The Mavs played a bit better than their usual in 2 series, and a lot better in the other 2.
No bad series = good chance of winning."
Yes consistency is important. But so is having one lineup that yielded +115 in 163 minutes when total Dallas +/- for the entire playoffs was +121. One lineup produced 95% of their point differential and about 30 points more than expected based on the regular season. That seems worth noting. The winner can't be predicted with certainty but looking at the detail including lineups and matchups of lineups one probably can do better than without and perhaps do better than just saying the team that plays better will win.
In this case the regular season data for top 5 lineups used in the playoffs for Dallas didn't predict their performances well individually but taken as a group of 5 the regular season performance was strong and the aggregate playoff performance was strong too. Look at top 10 lineups instead of just 5 and there is now a much lower frequency of swings between positive and negative raw +/- ratings from regular season to playoffs. There is noise and things are not entirely clear cut but there is something to gain by somebody in the detail, I think.
That would be worth defining fairly and more tightly and checking. Perhaps later.
Not sure who the "somebody" you refer to is. I have a few guesses but if you think it is important to allude to, would you clarify who and what you are specifically referring to?
After the fact playoff data is problematic for evaluating team and opponent performance. With regard to Dallas and quality of defenses faced I looked at both playoff and regular season. I rounded it a bit to make an over-arcing point which I thought interesting. I didn't have the time to find or create the Dallas data series by series but more could be done to refine the story, with the time & interest.
"The Mavs played a bit better than their usual in 2 series, and a lot better in the other 2.
No bad series = good chance of winning."
Yes consistency is important. But so is having one lineup that yielded +115 in 163 minutes when total Dallas +/- for the entire playoffs was +121. One lineup produced 95% of their point differential and about 30 points more than expected based on the regular season. That seems worth noting. The winner can't be predicted with certainty but looking at the detail including lineups and matchups of lineups one probably can do better than without and perhaps do better than just saying the team that plays better will win.
In this case the regular season data for top 5 lineups used in the playoffs for Dallas didn't predict their performances well individually but taken as a group of 5 the regular season performance was strong and the aggregate playoff performance was strong too. Look at top 10 lineups instead of just 5 and there is now a much lower frequency of swings between positive and negative raw +/- ratings from regular season to playoffs. There is noise and things are not entirely clear cut but there is something to gain by somebody in the detail, I think.
Re: What if player APM is used as the guide to build lineups
I defined a good team as one with a regular season SRS of 4 or better and a good offensive team as one with a regular season offensive efficiency of 109.5 or better. I looked at the teams who knocked good teams and good offensive teams by these criteria out of the playoffs the last 3 seasons.
For good teams the opponent who knocked them out of the playoffs averaged a regular season offensive efficiency of 109.9 and a defensive efficiency of 104.6. The weighted league average efficiency for these teams was 107.6. So the team which knocked them out average 2.3 points better on offense than weighted league average efficiency and 3.0 points better on defensive efficiency.
For good offensive teams the opponent who knocked them out of the playoffs averaged a regular season offensive efficiency of 110.0 and a defensive efficiency of 104.6. The weighted league average efficiency for these teams was 107.6. So the team which knocked them out average 2.3 points better on offense than weighted league average efficiency and 3.1 points better on defensive efficiency.
To answer the question posed above, yes, good offensive teams were more likely on average for this timespan to lose in the playoffs to a team who is better on defense than offense compared to the league average. But for this timespan than was no difference in the offensive / defense bias of what knocked good teams or good offensive teams out of the playoffs.
If you have a regular season offensive efficiency near 110.0 and a defensive efficiency near 104.6 you are in pretty good shape to take your opponents out in the playoffs, based on this data. Which teams this playoff were closest to these marks in terms of total shortfall from these goals?
Miami (exceeded both goals)
Lakers (exceeded both goals)
Dallas (0.3 short on offense, 0.4 short on defense)
San Antonio (1 short on defense)
Chicago (1.7 short on offense)
Why didn't Miami win? Probably lineup and match-up management against the Mavs. For more on that story see Wayne Winston's blog.
Why didn't the Lakers beat the Mavs? You already know some of the possible answers and either find the answers acceptable or not.
For good teams the opponent who knocked them out of the playoffs averaged a regular season offensive efficiency of 109.9 and a defensive efficiency of 104.6. The weighted league average efficiency for these teams was 107.6. So the team which knocked them out average 2.3 points better on offense than weighted league average efficiency and 3.0 points better on defensive efficiency.
For good offensive teams the opponent who knocked them out of the playoffs averaged a regular season offensive efficiency of 110.0 and a defensive efficiency of 104.6. The weighted league average efficiency for these teams was 107.6. So the team which knocked them out average 2.3 points better on offense than weighted league average efficiency and 3.1 points better on defensive efficiency.
To answer the question posed above, yes, good offensive teams were more likely on average for this timespan to lose in the playoffs to a team who is better on defense than offense compared to the league average. But for this timespan than was no difference in the offensive / defense bias of what knocked good teams or good offensive teams out of the playoffs.
If you have a regular season offensive efficiency near 110.0 and a defensive efficiency near 104.6 you are in pretty good shape to take your opponents out in the playoffs, based on this data. Which teams this playoff were closest to these marks in terms of total shortfall from these goals?
Miami (exceeded both goals)
Lakers (exceeded both goals)
Dallas (0.3 short on offense, 0.4 short on defense)
San Antonio (1 short on defense)
Chicago (1.7 short on offense)
Why didn't Miami win? Probably lineup and match-up management against the Mavs. For more on that story see Wayne Winston's blog.
Why didn't the Lakers beat the Mavs? You already know some of the possible answers and either find the answers acceptable or not.
Re: What if player APM is used as the guide to build lineups
The Lakers starting lineup had a raw +/- of over +14 per 100 possessions last season and an estimated lineup Adjusted +/- of just under +10. The sum of 1 year traditional Adjusted +/- measurements for those players for that season was about +12.5. For 2 yr. traditional APM the sum was about +12.5. For 09-10 1 year APM it was about +10, it you want to try to use that as a projection. The lineup Adjusted +/- and the sum of player APM was pretty close.
2010-11 RAPM measured the sum of the players at about +11. 2009-10 RAPM summed to +15.5. LambaPM summed to +16. The average of the 6 measures was about +13. There was variation but they all suggest it was (or was going to be) a great lineup. I am not surprised that the sum of the individual measures almost always overestimated the lineup performance when all of the best options for each position are out there together.
The RAPM lineup estimate was +12.
The starting lineup of the Spurs was +10 on raw +/- and about +6.5 on traditional APM. The sum of 1 year traditional Adjusted +/- measurements for those players for that season was about +8.5. For 2 yr. traditional APM the sum was about +10. For 09-10 1 year APM it was about +12.5, it you want to try to use that as a projection. 2010-11 RAPM measured the sum of the players at about +9.5. 2009-10 RAPM summed to about +11. LambaPM summed to +15.
The RAPM lineup estimate was a bit over +8.
Pretty similar story. The sum of the current season APMs for the players was always higher than the traditional APM for the lineup, even with less than the 5 best out there. The variance between the lineup estimates and the various sums of the players was larger but were no really weird, chaotic estimates. All sums suggested it was (or would be) a good lineup. Not surprisingly 1 year player APM was the most consistent with 1 year lineup APM.
The player sums fit tighter around the RAPM lineup estimate in these cases, with a few cases where the player sum was under the lineup estimate but the large majority still being over.
Of course would want to check more lineups, including some not as strong, near neutral and negative, before locking in any assumptions. Near neutral lineups are probably more likely to have player sums above and below the lineup estimates. Players sums might be more likely too low on average for poor performing lineups. Takes a bit of time to analyze the data. Shame that the numbers don't reveal everything immediately, all by themselves.
2010-11 RAPM measured the sum of the players at about +11. 2009-10 RAPM summed to +15.5. LambaPM summed to +16. The average of the 6 measures was about +13. There was variation but they all suggest it was (or was going to be) a great lineup. I am not surprised that the sum of the individual measures almost always overestimated the lineup performance when all of the best options for each position are out there together.
The RAPM lineup estimate was +12.
The starting lineup of the Spurs was +10 on raw +/- and about +6.5 on traditional APM. The sum of 1 year traditional Adjusted +/- measurements for those players for that season was about +8.5. For 2 yr. traditional APM the sum was about +10. For 09-10 1 year APM it was about +12.5, it you want to try to use that as a projection. 2010-11 RAPM measured the sum of the players at about +9.5. 2009-10 RAPM summed to about +11. LambaPM summed to +15.
The RAPM lineup estimate was a bit over +8.
Pretty similar story. The sum of the current season APMs for the players was always higher than the traditional APM for the lineup, even with less than the 5 best out there. The variance between the lineup estimates and the various sums of the players was larger but were no really weird, chaotic estimates. All sums suggested it was (or would be) a good lineup. Not surprisingly 1 year player APM was the most consistent with 1 year lineup APM.
The player sums fit tighter around the RAPM lineup estimate in these cases, with a few cases where the player sum was under the lineup estimate but the large majority still being over.
Of course would want to check more lineups, including some not as strong, near neutral and negative, before locking in any assumptions. Near neutral lineups are probably more likely to have player sums above and below the lineup estimates. Players sums might be more likely too low on average for poor performing lineups. Takes a bit of time to analyze the data. Shame that the numbers don't reveal everything immediately, all by themselves.
Re: What if player APM is used as the guide to build lineups
It appears that there were at least 12,000 possible 5 man combinations in the league during the 2010-11 season. Of those used and qualified on Jerry's 2010-11 lineup RAPM (with playoffs), 59 lineups were used 500+ total possessions (offense + defense) and were estimated at +3 or better during 100 possessions for each team. Playoff teams had about 85% of these lineups. Using individual Adjusted +/- the average rank of these lineups on expected performance among all possible 5 man combinations was 865th out of about 12,000, or about top 7%. But the RAPM estimate for their performance together had an average rank of 65th or top 0.5%.
Is this a sign that lineup construction really does matter, beyond individual ratings? To pursue it a bit further I also looked at the 2008-11 file. There were 205 lineups that qualified on the 2 criteria. Using individual Adjusted +/-, the average rank of these lineups on expected performance among all possible 5 man combinations was 4441th out of at least 43,000 possibilities, or just outside top 10%. But the RAPM estimate for their performance together had an average rank of 264th or top 0.6%. So it is not a one year thing.
How much of the improvement is random noise and how much is lineup synergy I am not sure. What were the size of the standard errors?
More research would be warranted but I would not agree with those who dismiss the importance of lineups as distinct operating units and think of them as mere sums of individual ratings, or only casually consider lineup data for a few lineups and generally run their rotation by coaching judgment (strategic and match-ups for the immediate situation) and individual stats & other considerations (seniority, pay, past usage) and largely independent of the lineup performance data, limited in quality though it may be usually due to low sample sizes in all but a small set of cases.
Minnesota was the only team who had no qualifying lineups quality at +3 or better for the 3 year period. 5 teams only had 1 or 2 "good" lineups used over 500 possessions. Some more analysis and targeted experimentation seems in order especially for them, but also for anyone trying to get better actual team performance.
The qualifying lineups average about 1300 total possessions of use (offense + defense). Given about 15,000 total possessions in a full season for a team, you could conceivably run about 12 lineups for this average length, exclusively (approach A). Or more practically say 6 lineups averaging 1300 total possessions, 10 averaging 500 total possessions and give 2200 total possessions (or about 15%) to the dink lineups which might be needed or appropriate for specific matchups and special situations (approach B). Instead teams appear to use just a couple lineups at most over 1000 total possessions and then about a dozen more somewhere over 200 total possessions (200 total possessions= a bit over 50 minutes use) with perhaps half to 2/3rds of all time going to lineups below this cutoff level (approach C). Even if approach B is too rigorous and too difficult for many teams to achieve, they could try to move some distance from the typical approach C levels towards what was suggested in approach B or what the better / probably usually more concentrated lineup management teams are actually doing.
Of course you don't just want bigger usage lineups, you want bigger usage lineups that work. League-wide about 75% of lineups used over 230 minutes or about 900 total possessions were positive and only about 5-10% were more than modestly bad. I'd think that would be encouraging to find and play more of the better remaining candidates currently dividing time with hundreds of other often lower performing lineups.
Is this a sign that lineup construction really does matter, beyond individual ratings? To pursue it a bit further I also looked at the 2008-11 file. There were 205 lineups that qualified on the 2 criteria. Using individual Adjusted +/-, the average rank of these lineups on expected performance among all possible 5 man combinations was 4441th out of at least 43,000 possibilities, or just outside top 10%. But the RAPM estimate for their performance together had an average rank of 264th or top 0.6%. So it is not a one year thing.
How much of the improvement is random noise and how much is lineup synergy I am not sure. What were the size of the standard errors?
More research would be warranted but I would not agree with those who dismiss the importance of lineups as distinct operating units and think of them as mere sums of individual ratings, or only casually consider lineup data for a few lineups and generally run their rotation by coaching judgment (strategic and match-ups for the immediate situation) and individual stats & other considerations (seniority, pay, past usage) and largely independent of the lineup performance data, limited in quality though it may be usually due to low sample sizes in all but a small set of cases.
Minnesota was the only team who had no qualifying lineups quality at +3 or better for the 3 year period. 5 teams only had 1 or 2 "good" lineups used over 500 possessions. Some more analysis and targeted experimentation seems in order especially for them, but also for anyone trying to get better actual team performance.
The qualifying lineups average about 1300 total possessions of use (offense + defense). Given about 15,000 total possessions in a full season for a team, you could conceivably run about 12 lineups for this average length, exclusively (approach A). Or more practically say 6 lineups averaging 1300 total possessions, 10 averaging 500 total possessions and give 2200 total possessions (or about 15%) to the dink lineups which might be needed or appropriate for specific matchups and special situations (approach B). Instead teams appear to use just a couple lineups at most over 1000 total possessions and then about a dozen more somewhere over 200 total possessions (200 total possessions= a bit over 50 minutes use) with perhaps half to 2/3rds of all time going to lineups below this cutoff level (approach C). Even if approach B is too rigorous and too difficult for many teams to achieve, they could try to move some distance from the typical approach C levels towards what was suggested in approach B or what the better / probably usually more concentrated lineup management teams are actually doing.
Of course you don't just want bigger usage lineups, you want bigger usage lineups that work. League-wide about 75% of lineups used over 230 minutes or about 900 total possessions were positive and only about 5-10% were more than modestly bad. I'd think that would be encouraging to find and play more of the better remaining candidates currently dividing time with hundreds of other often lower performing lineups.