APBRmetrics

For lineups under 400 possessions there will be plenty of noise but comparisons of lineups can still suggest with high probability which is better.

But my priority is testing lineups over to way over 400 possessions and taking the likely better.

I am not really interested in debating stability of lineups of 100 possessions or less. I concede that small samples will be volatile. I hope others will concede that bigger tests increase knowledge and bigger minute positives are the best choices to guess on.

Here is a chart showing the point differential between 2 identical teams, with identical odds of scoring 0, 1, 2, 3, and 4 points on a possession, after 250 possessions have been played.



Remember, these teams have identical true talent. 5% of the time, one or the other team is at least 50 points better after 250 possessions! This is assuming no added source of noise (rest differential, injuries, hot streaks, etc.)

(Let me know if the chart doesn't come through.)

250 possessions is still a pretty small sample. The game could be volatile.


More importantly imo, show a matchup of lineups of equal current performance but one tested to 500 or 1000 possessions and the other to only 100 or 250.

Fundamentally the error is going to be of the order of 1/sqrt(n) where n is the number of possessions. Unfortunately this means quadrupling the sample to halve the error. Such is life, and such is why I take lineup data with a large grain of salt regardless of the sample.

v-zero wrote: ↑Fri Jul 21, 2023 3:14 pm Fundamentally the error is going to be of the order of 1/sqrt(n) where n is the number of possessions. Unfortunately this means quadrupling the sample to halve the error. Such is life, and such is why I take lineup data with a large grain of salt regardless of the sample.

Agreed. The goal here was to clearly parameterize just how much random noise is inherent in basketball.

Take lineup data with a grain of salt or more.

What should teams do? Primarily trust coaching instincts (with huge documented misses, many of which are probably errors in judgment) or use lineup analytics to try to blend with and improve coaching? I am not for accepting the status quo.

I don't know how successful internal team analytics efforts for lineup management "are" or are regarded typically or best case but it should be possible to add some value and / or subtract some harm. If people think *analytics" can help with shot selection, player selection, salaries, etc., why decline to support analytic efforts on lineup management?

From my observation most of the strong performing, big minute lineups stay the best bets to me (even if they are not always at their best) and the dogs of moderate size tend to stay that way. Sure it is not totally clean and simple. That's why the efforts should be far greater to gain what can be gained. Don't oversell but try.


Apparently knocking teams that use 800-1000 lineups with low levels of concentration or showing that they have a majority or supermajority of negative lineups in most used 5, 10 or 20 and entirety is not enough to get the attention of many or support for change.

Big change, not small change. The game is pretty random and a good part of it is very sloppy lineup management. I find the apparent lack of interest in lineup management in the "analytics community" as incredible as with teams. But whatever, keep playing that coach-diva-magician driven mega hundreds of sloppy, mostly unknown dink lineups style ball. I don't and won't be buying that this is the best that can done.

Now if a team tried massively more concentrated lineups derived with heavy analytic input and decision weight and failed, then walk way. But who will really try?

I generally focus on 5 man data and mostly the very top of that data but if a team can work on sub-lineup issues (with much bigger sample sizes) and have greater success than currently, do that too or even primarily. Can teams learn more and better manage their top 20 pairs? I think so.

If actual data is not enough and never will be even with heightened concentration and greater analytical efforts, build the best damn lineup simulator and add that to the mix of considered content.

If lineup data showed anything close to a reasonable pattern of testing and improvement, I wouldn't be as loud. The failure to be reasonable is so egregious imo, I have been loud about it. The playoff are fairly different and serve as part of the indictment of regular practices. But playoff ball is not independent. Failure to reasonably test in regular season or change from it leads to many playoff missed opportunities and failures.

One way to better understand coach thinking on lineup would be, as I have suggested before, to require him to file a report with management explaining the thought behind every substitution in every game in writing. Does any team require it? I doubt it is done, but would want to hear if anyone knows it is done.

Management should give feedback / suggestions where it has any. If you want more concentration for better testing and hopefully better results, say so repeatedly, every game or even every case for it.

I wish some ex-coach would do that in public, probably for a playoff game rewind. Or two opposing coaches side by side talking thru it. But they probably never will beyond a few antedotes or at all. If you are playing 3D, 6D, 9D chess all the time, every beat, necessarily for the win, show it, prove it.

In this thread from 2018:

"Only team stylistic choices showed good correlation between the first half of the season and the second--things like pace, 3PA, and DRB."

Are the correlations between the first half of the season and the second higher for the top 10 teams? For coaches with more experience? How much?

Do good teams / more experienced coaches tend to be stable across all of pace, 3PA, and DRB simultaneously or often here and there? Are pace, 3PA, and DRB low, moderately or highly correlated among themselves in general or for good teams? They could be. I have guesses for the others but not pace and DRB. Which correlations are highest and with whom?

Any interest in presenting / summarizing this data and analysis for last season?

What are the correlations between stat stabilities and winning? In general and for standout teams? Better to have a set way to win or do it many ways? Starter heavy stability vs. beach heavy stability? Home / road? By quality of competition?

What is the average variance in points per game in league and offensive and defensive efficiency?

You could just copy lineup plus-minus a few times -- monthly, perhaps -- and see if there is month to month correlation, or trend.

I was talking in immediately above post about at team level and league level, not lineup level.

Lineup level does have quite a bit of variance in most dink cases, though much less so in the very few cases where a lineup has decent minutes size.

I occasionally look at lineup result trends, usually dividing data into 2 segments to get the simplest level of progression while trying to maintain sample sizes. Could graph monthly or compute rolling trends if interest warrants. Given team lineup management practices, it is virtually useless to look at trends beyond for the most used 1-5 llineups. The average performance of the total sample is going to be closer to long-run "true" than most time slices.

NBA average efg% set an all-time high again for 8th straight season.

Average NBA FT/FGA per game ends up last season as 11th lowest in 77 years but up a bit from 5 straight immediately prior years of being even lower.

3ptas as a percent of all fgas ends season at 38.7%. Temporary lull in 3pt shooting or was the peak reached?

3ptas per game comes in as 3rd highest. 3ptrs made also comes in as 3rd highest but 2 straight down years.

The coaching experience with small samples is likely to:

create anguish and despair of simple / ready knowledge

encourage reaches for special insight / hubris to overcome the common experience of failure and try to deny it's imminent and frequent re-arrival for them.

The response to small sample experience is to further heighten small sample experience thru skidish, almost mystical response to it, to try to beat or dance ahead of chance, of at least the last negative experience in short term memory, never or almost never having explored life under much larger sample size campaigns (into the land of greater stability / somewhat more useful knowledge).

So the grieving repeat the fight in a cage of their own making and never invest in a possible way out.

And the shamen will be quick to rebuke questions and challenges. Knowing the tenuous nature of their knowledge and power. However hard and earnest their struggle, mostly in darkness, for it.

Favor closer to clinical trials over ad-hoc witch-doctor hit and miss.

DSMok1:

"

DSMok1:

"my feeling is that lineups are generally best evaluated as the sum of their 5 players"

Looked at RAPM factors for Nuggets starting five. Sums to +11.8 / 100p. Actual performance was +13.7. Half from own efg%. 6 positive factors, 2 small negatives. 30 individual positive factors, 10 negative (only 1 over -0.5). That's a good lineup. A lineup to believe in.

KCP was negative on 6 of 8 factors immediately before acquisition. But 5 of the 6 were tiny. Gordon in his last season with Magic was 5 positive, 3 negative. A sure bet. Both went to 8 non-negative with Nuggets (KCP with 2 neutrals, Gordon all positive).

Looks like some positive synergies based on team and individual data. But it could be some degree of random variances. We'll see what happens this season. Probably still great even without synergy / random variance.