APBRmetrics

I looked at BLI's recent article again. http://bigleagueinsights.com/top-ten-nb ... ince-2000/

156,716 different lineups used by coaches since 2000. Only about 1.5% used over 100 minutes total ever.

Coaches try to coach the situation but this just boggles my mind and smells of potentially tremendous inefficiency.

On average a lineup got 10.6 minutes of use ever. The top 1.5% averaged a bit over 250 minutes ever.

Very few coaches have shown a considerably tighter lineup distribution. It would be interesting to look at the relative degree of success of the top ten current coaches with the highest distribution of minutes to top 3-5-10 lineups vs. the middle and bottom 10. There may not be a correlation given variation in team talent and coaching talent outside this trait. But I can't help but wonder if a significant edge could be obtained for regular season and playoffs by more disciplined testing of the expected strongest lineups and more heavy use of those that show strong performance.

How many teams have a lineup distribution plan for the season? How many have an on-going detailed review of the results of that distribution? How many coaches take it seriously? How many are made to take a serious, disciplined, strategic in addition to tactical approach to lineup minute allocation?

I know at least a few teams have analysts with lineup simulators. How common is it overall? How good are they? How rigorously are lineup actuals analyzed to check model performance and improve them? Will any of the big public analytic sites develop such a model and detail their methodology, results and evolution? Will there eventually be a public access lineup model application?

Only 6 lineups got used 100 plus playoff minutes and had a net rtg of plus 4.9 or better. 3 of the conference finalists had one, Spurs 2 (and ATL and LAC). Spurs only got slightly higher rate of pts edge from these 2 lineups than rest of lineup time but Miami and Indiana got all their edge essentially from their single best big minute lineup (for other two teams, the top lineup helped them get where they got but wasn't enough to overcome deficit from rest of lineups). 3 teams can field the same lineups next time. Will any do so at significantly higher min. / game and will that make a critical difference?

Crow wrote: I know at least a few teams have analysts with lineup simulators. How common is it overall? How good are they? How rigorously are lineup actuals analyzed to check model performance and improve them? Will any of the big public analytic sites develop such a model and detail their methodology, results and evolution? Will there eventually be a public access lineup model application?

I really want to create a lineup simulator based on all my ratings across the statistical spectrum, since the ratings are adjusted for pace & such. Something that I can adjust the weights of any subset (say block rate, assist rate, any combinations of stats) and have the program spit out a projected lineup W% based on historical team data. Lineup optimization so to speak.

I think this is where the player ratings debate could end up being more illuminating. You could create a set lineup of 4 players - and then have the simulation rank all the players (either on the same team or even league) that would make these set 4 achieve the most. You could set 3, & have the simulation find the best pairs. Player rankings could vary quite wildly depending on the teammate constants.

Using player projections (&/or D to NBA conversion), it'd make things quite a bit easier for a team to figure out who to pick up when a roster spot opens, who to trade for, who to sign.

Like many other facets of this analytical world that rattle around in my brain - I do not nearly have the practical knowledge to create such a thing. I grind out all kinds of stuff in my spreadsheets, but machine learning (is that what this would be called?) or creating complex simulations is well beyond my grasp for now.

I've been wondering the same thing Crow.

156,716 different lineups used by coaches since 2000. Only about 1.5% used over 100 minutes total ever.

Coaches try to coach the situation but this just boggles my mind and smells of potentially tremendous inefficiency.

Last season the Spurs were the first team in NBA history without a single player to average at least 30 minutes per game. You'd think with so much experimentation with Lineups, this approach might have been tried before.

There is an interesting experiment going on with the Kings' D-League team this season, the Reno Bighorns. The Kings hired a coach from Division III Grinnell college, who attempted to get 94 shots per game:

"We try to get a shot off every 12 seconds and get the ball back every 10 seconds," Arseneault Jr. said. "We sub the first whistle after 35 seconds of game time. We use different subbing combinations every game, but 15 different guys get on the court within the first three minutes."


Check out their stats:
http://www.nba.com/dleague/reno/statistics/

In any case, this hockey-type of approach to subbing has never been tried before.

I'm open to doing some work to learn more about how to improve efficiency in the lineup area, just not sure exactly what the questions are.

Simulators based on 4 factors straight up would be one thing but ideally you'd want the model to account for four factor interactions in an "accurate" but not overly rigid way.

xohoops was / is a place where a lineup simulator operates with results in member view, though not the backend mechanics to my past dated knowledge. I'd go back there if I found the time. The original creator works for Philly and previously Portland. I've wondered how much the simulator advanced, was used.

With accurate sportVU player data on touches, passes, shot types / fg%s, defensive behavior, etc. I'd that play by play level simulators could be far more realistic these days. I understand some gamblers / syndicates have had simulators for some time too.

Very cool, thanks Crow.

We're working on some related simulation type stuff, will keep you posted.

Alright.

Statman, you and could probably figure out most or all the steps and considerations needed but to string altogether with appropriate random value generators and have everything be automated for an entire game, season or 1000 seasons, does take technical skills I have not previously attempted. (I did produce a really simple simulator long ago) So I mention it with hopes that some with those skills (like it appears BLI) go for it and we can get some of the benefits.

statman, the same4 player data at jerry's site has the advantages of being real not simulated and just sitting there for consideration. They are probably an underutilized resource, in part because they a click away from surface view.

A simple way to build a good rotation is to find a very good trio or quad and then use as many of the lineups with that as a core which are good for as many minutes as possible. You often don't have to find 10 separate diamonds, maybe just 2 or 3 and the variations.

Are there really 500 game situations requiring 500 truly different lineups by teams over a season? I doubt it. Sure use 100 micro-lineups if you can really justify it but I'd think try to keep it down to under 20-30% of total time.

Crow wrote:statman, the same4 player data at jerry's site has the advantages of being real not simulated and just sitting there for consideration. They are probably an underutilized resource, in part because they a click away from surface view.

A simple way to build a good rotation is to find a very good trio or quad and then use as many of the lineups with that as a core which are good for as many minutes as possible. You often don't have to find 10 separate diamonds, maybe just 2 or 3 and the variations.

Are there really 500 game situations requiring 500 truly different lineups by teams over a season? I doubt it. Sure use 100 micro-lineups if you can really justify it but I'd think try to keep it down to under 20-30% of total time.

You stated yourself that the "real" data sets are so small, so how much can we learn from past lineups?

A properly done simulation could take every possible lineup in the NBA, and rank them by projected strength. That simulated info could be used not just in decided what lineups to play most, but what players to try to get to fill out rosters around the stars.

With proper player projections (which I believe I do well right now), it could help a team decide which players to draft who compliment each others skillsets & would develop best together.

I pretty much agree with you about lineups in general, it seems like coaches might not be figuring out their best 5 & allowing their best 5 to play together nearly long enough as a season progresses.

Yes, I agree a massive set of hypothetical lineup testing and for specific situations would dwarf the real dataset and perhaps learning from it even the small sample size of the vast majority of lineups. I was just noting the low effort cop out option available for use, the bare minimum effort.

I once read that Rich Cho had gotten some Microsoft programmers to develop application that showed every legal trade possible with set of team assets and that they were scored with a one number value. So maybe somewhere out there are team execs willing to study and score every possible lineup against confronting at least 30 / 300 or whatever team / lineup / game situations.

I am getting the sense that there are analysts out there capable of doing super huge data projects and at least a few teams doing them but I wonder how much is getting used by coaches with lots of functional autonomy and zealous guardianship of it.

Yes, I agree teams should evaluate both current and future players with clear understanding of their player roles, the requirements and potentials of each spot / role to maximize the team contribution from putting players in specific contexts / roles / opportunities. Some fols have fought specific metric weighting for positions, but I could see value in very specific position or role scoring guides for all in one metrics or even the parts of box score or rpm metrics. You could have teams discover that the best realistic signing for them is a guy with a adj off PPS of plus 1.5, an adj own turnover of plus 1, a 3pt frequency of x, etc. Outright or given market pricing.

Cavs under Blatt have used 79 lineups so far. Most used at a bit over 10 minutes a game, K. Irving | L. James | K. Love | S. Marion | A. Varejao is +4 per 100 possessions. Ok, though nothing special. But 3 of the top 5 most used and 10 of the top 16 are negative. Top 5 only account for about 45% of total minutes.

On average teams have one lineup used over 50 minutes. 12 are better than +5, 10 worse than -5 per 100 possessions. No team has more than two.

Some teams have already used over 100 lineups.

Good article on Spurs lineup management, including cool graphs with BPM:
https://medium.com/the-cauldron/gaming- ... a2a7fd6ca7

Wow, great link. Second graph, devastating support for win with bench strategy as practiced (designed) by Spurs. Could be pivotal step in league minute allocation awareness. If read, thought and acted upon. Good quote of Steven Shea.

Is win with bench in abstract a good small market strategy, last hurrah of great team strategy? Already shown it is not for most, not for most not built that way. Too much focus on starters? Open big can of worms...

Talkingpractice indicates on twitter that he plans to look at performance by length of time on court. Checking performance level trends could shed light on wisdom of coaching practices, etc.

I think when the salary cap goes up, there will be several small market teams trying to employ a similar strategy.

It would be interesting to see that same BPM graph but with minutes distribution by salary rather than BPM. My guess is that the strength of the Spurs bench is due to a more even distribution of salary and efficient spending by the front office (obviously it helps when your stars are giving up some of that cap money to the bench players).