2022 Playoff lineup analysis (and more)

[Crow]

2022 playoff lineup analysis

There are different levels of lineup data. Usually I focus on the most used. This time my focus will begin with small minute but impactful lineups. At the moment, I will call them "lineup grenades" for their ground gaining impact.

The initial criteria is 20 plus minutes at +20 pts / 100 possessions. Could be fairly small minutes but not very small. At minimum they had a +8 actual impact on the series. That is big enough to potentially matter a lot.

I will be looking at pair, trio and 5 man data. Might come back to look at quads compared to trios and 5man lineups later. The comments will be brief and generalized. Plenty of reasons to potentially go deeper, if there is more interest.

Pair data:

Celtics- Pritchard was in most of the very few pairs to qualify by this criteria. More than anyone else.

Mavs, Bulls, Jazz, Hawks, Suns and Nets had no qualifying pairs. If it is championship or bust, there should be change seeking such strong pairs imo.

All the qualifying pairs the Warriors had left with Otto Porter.

Adams was in the 2 Griz pairs that remain after Anderson left. So I'd keep him.

For Heat, the hot name surprisingly was D Robinson.

For Bucks, it was J Carter.

For Timberwolves, the only such pair was D Russell - McDaniels.

Trey Murphy was in most or all the strong Pelican pairs.

For Raptors, the only such pair was Flynn - Boucher. Says something about "needing" small samples to "pop" big but also maybe something about the "stars".

This data made me think about role player value… and the value of disrupting your competitors if you can.

Trios

D Wright is out of Atlanta and so are all the strong qualifying pairs.

None left for Brooklyn.

Mavs, had trio of Doncic Bertans Kleber… and nothing else.

Warriors… ONLY 1 OF 13 SUCH STRONG TRIO REMAINS AVAILABLE.

ANOTHER TEAM HAD 19. WHO WAS IT?

That and 2/3rds of the remaining very first cut findings are being held. More cuts of the data identified, pending motivation.

[Crow]

Now some small sample lineups will be noise or more noise than signal. But they warrant study. The +20 / 100p standard for consideration is not that hard to achieve randomly but it is not nothing.

[Crow]

The Grizzlies had 19 of the 97 qualifying strong trios. But… only 2 of 17 qualifying 5 man lineups. Seems Coach Jenkins has plenty of work to do refining 5 man lineups. Both had Bane Brooks Jackson. One had Morant, one didn't.

Heat had 10 strong trios. 3 or 4 each for Butler, DRob, Oladipo and Vincent. Just 2 for Adebayo.I think it was 1 or none for Herro.

Pelicans had 7 strong trios, all with Murphy. 3 For McCallum and Jones, 1 for JVal.

14 for Bucks.

76ers, just 3. 1 Embid, all Niang.

Suns 4 trios, none with Ayton. Though he was in both of the strong 5 mans. Maybe not so vital?

4 for Timberwolves. Just 1 with Towns.

Just 1 for Raptors and an obscure one not involving core. They probably need lots of changes to change and advance.

1 for Jazz. With Conley and not Gobert. Blow it up.


5 man lineups

Celtics with 3 qualifiers. 2 with D White. Plenty of lineups with D White don't work, so trying greater concentration on the positive is probably the right action item to test.

Both of Mavs 2 strong 5 mans had Dinwiddie Bullock DFS. One with Luka, one without.

As mentioned at trio level, Warriors offseason departures have major impacts. Only strong 5 mans gone with Porter.

Heat had 2 best 5 mans. In the +40 to +53 / 100 p range. Butler Strus Oladipo Adebayo. Best with Lowry, still great with Herro.

TWolves best gone with Beverley.

76ers with 2. Neither with Embid.

97 divided by 16 playoff teams equals an average of about 6. But of course it will vary depending how many rounds played. Still you can see who is low and who is high.

[Crow]

Warriors, Grizzlies, Bucks and Heat look better from this partial perspective than others. Each have issues though. Celtics probably in tier 2. 76ers, Suns, Raptors and Mavs in tier 2 at best and way behind.

[DarkStar48]

FYI, as per the work of the creator of DARKO player stats — one of the most accurate metrics at the aggregate level — Kostya Medvedovsky (@kmedved), “raw” 5-man lineup offensive rating takes ~550 possessions (about 264 minutes played) to become predictive, while defensive rating takes ~850 possessions (approx. 408 min.). As a result, you can’t rely on net rating until both values stabilize with at least almost 408 minutes of playing time.

So, for 5-man combinations, you probably shouldn’t be that confident in the stats for any lineups that haven’t played for that long.

Alternatively, BBall Index has a tool that projects actual and custom 5-man ratings that utilizes LEBRON stats as well as Bayesian priors as a part of its estimates.

Link: https://t.co/Zz1omWODAu

[Crow]

Thanks for the read and for the other lineup thoughts.

While teams with a plan and discipline could / should have multiple (2-5) 400+ minute regular season lineups, there were only 9 in the league. The Knicks were the only team with 2.

But this was about playoffs. And not statistical significance, because that would be almost impossible in playoffs.

But the lineups and sub-lineups had notable actual impact. As such, worth knowing about them. Either "guess" with them or guess on less impressive actual performers. Neither will be very reliable but got to guess on some.

A weighted combination of regular season and playoffs would probably be better than either alone, but I started with playoffs only because I had yet to see it anywhere else.

I haven't used the Bballindex tool. While it is probably somewhat helpful, it is based on individuals. How closely do lineups behave as a simple sum or adjusted blend of individuals? I dunno. Maybe fairly close on average, but there will be important variances in the real performances compared to the estimated. Which to go with more? I am not sure but I want to at least know the actual.

If advising a team, it might be useful to try to merge regular season and playoff actuals with individual based lineup projections. I wasn't doing projections but that might be the way to go there.

For knowing what actually happened, the actuals should be the focus and the most impactful per minute should be a major focus and perhaps the greatest focus.

[Crow]

If player interaction terms are notable, sustained on average, "true", then an individual based lineup model, which lacks those specific interaction terms, might not be the best model because of it or may not even be a good proxy. Individual ratings will be influenced by interactions but by all interactions as weighted by specific lineups choices and other coaching choices.

I haven't seen any RAPM pair data recently. I would of course appreciate such if publicly provided or a possible private share. Regular season, playoffs, combined, possibly multi-year.

If player interaction terms were minor or mostly noise, there would be a lot less reason to pursue lineup insights. I don't believe lineups to be largely devoid of true and notable interactions. I believe the lack of clarity about the impact of interactions is largely due to very chaotic / poor testing and generally inadequate analytic focus on interaction terms. In public and probably inside many teams or at least Coaching brains.

[Crow]

Only 34 regular season lineups got 200+ minutes of test. Only 95 got 100 plus. So statistically significant tests are very rarely approached.

No playoff lineup reached 200 minutes. Only 11 exceeded 100 minutes. Only 25 total reached 50 minutes.

The desired standards will rarely be reached and so are not that relevant if you still want to look at actual data.

I agree with anyone thinking that is a waste of time looking at super super dink lineups. They make up most of the approximately 20,000 lineups NBA Coaches use and think they "need" or are smart to use.

20 minute standard for playoffs was as low as I'd go. 40-50 is another choice but that would mean looking at even fewer notable lineups. 20 minute criteria seemed like a good starting point. I've looked at 50 and 100 minute minimum data but I may not write it up without team interest. They are in the small 20 minute plus group already.

[Crow]

Lineup data gets very little public interest and pretty limited media attention.

I can't know what teams are analyzing and producing. I can only see what they "do". And many, many teams have highly questionable numbers of lineups, concentration levels and priorities that are not panning out.

I'd think it would be wise / beneficial to have more folks doing lineup analysis and commenting on it and having weight in decisions. I'd think it would be wise to see lineup management as at least shared territory and not "Coach's prerogative". How shared it is now, I dunno. I would not tolerate Coaches who do not at least listen and debate and try some suggestions. Their actual lineup management records have so many blatant issues to yield to them complete carte blanche to repeat those patterns.

[Crow]

Knowing that lineup grenades exist in actuality is a start. How real they are is an important research project. How actively to search for them from the sideline is debatable but they should be recognized in lineup data history.

Dink or near dink lineups are a large share of lineup experience. The main rationale of that, if there is one, is to find and more fully take advantage of / further test lineup grenades and similars.

I've focused on net impacts; but, of course, study of why / how would be a next level. So much could be done. It starts with degree of interest.

[DarkStar48]

Also, a little bit of a pro-tip: you can do your own back-of-the-envelope calculations to project the “true” 5-man lineup ratings from small samples (e.g., the playoffs) using the padding method.

That is,
expected_lineup_pm = 100 * {obs_lineup_pm + (pad_poss * [lg_avg_lineup_pm/100])} / (obs_lineup_poss + pad_poss)

where,
obs_lineup_pm is the observed total lineup, opponents and net points.
lg_avg_lineup_pm is the league average lineup offensive, defensive or net rating
pad_poss is the padding value for lineup possessions
obs_lineup_poss Is the number of observed lineup possessions

pad_poss(offense) = 550 poss.
pad_poss (defense) = 850 poss.
pad_poss (net) = 850 poss.

For example, in the 2022 playoffs, the Maxey-Harden-Green-Harris-Embiid lineup for the 76ers had totals of 410 points, about 367 for opponents and approx. 43 on net in 345 possessions resulting in an offensive rating of 118.8 points per 100 poss., 106.4 on defense, with 12.5 net rating. Assuming a league average lineup rating of 112.0/100poss. on both offense and defense (0 net rating), therefore:

expected_lineup_ortg = 100 * {410 + (550 * [112.0/100])} / (345 + 550)
= ~114.6

expected_lineup_drtg = 100 * {367 + (850 * [112.0/100])} / (345 + 850)
= ~110.4

expected_lineup_nrtg = 100 * {43 + (850 * [0/100])} / (345 + 850)
= ~3.6

[Crow]

Thanks for the formula and explanation. It would be a useful next step if making a presentation to a team. For quick work, knowing the actual performance, small size and the principle of regression to the mean can give rough in the head estimate, knowing that use of the formula also produces a rough estimate. (Using this method moves you from actual to "expected"... on average with "true" values having a range around that expected.)

More accuracy is good but it is often not that hard to decide whether a lineup should be tried more. Less might be a tougher, slower call, depending on degree of attachment for other reasons.

Also isn't the padding method making some assumptions of similar nature of distributions across league vs team and that specific lineup? They aren't necessarily exactly the same.

[Crow]

Warriors used almost twice as many lineups in playoffs run as Celtics.

Celtics gave their 50 most used lineups about 97% of all playoff minutes. Grizzlies only about 85%.

This isn't much concentration but better than regular season. I'd like to see 85-90% of regular season minutes go to 50 lineups. I'd actually like way more concentration than that but I'd shoot for that much first. That is enough to try multiple subs and multiple player positions. Not all of them, but main ones, "enough" of them for realistic / "responsible" maximum test set.

Grizzlies gave 50 most used lineups in regular season about 55% of minutes. Celtics about 59%. Jazz, probably the leader at about 68%.

Jazz used 383 lineups in regular season. 333 beyond the 50. Celtics 453, Grizzlies 533, Thunder 753. Lakers 727. Lakers more ridiculous because trying to win. Not with super high chaos. Thunder were near tops on regular season lineup chaos. Lever pulling fanatics Rick Carlisle and Steve Nash were even a bit higher on the chaos. Popovich also in the 700 club.

[Crow]

More on regular season lineup distribution:

The median most used lineup by team is about 300 minutes or less than 4 minutes per game across the full season. 2nd most used median is under 2 minutes. 3rd about 1.5 Minutes, 4th and 5th about a minute, 6th and 7th closer to 40 seconds per game. 7 most used on average will get less than 25% of total minutes.


In playoffs,Celtics 7 most used get a bit more than 50% of total minutes, Warriors a bit less than 50%.

Great teams can do what they are doing in regular season. Teams striving for more than they are achieving in regular season might want to act more like they do in the playoffs and show more lineup concentration for generally better results. Or not and settle for lower seeds and outcomes.

[Mike G]

Playoff rotations are more fixed and less experimental than regular season.
Also nobody is getting traded during PO. During RS, teams may have very different rosters available at various points.