APBRmetrics

Crow wrote: ↑Tue Mar 11, 2025 1:56 pm "And I have such a model, it is very edifying, it's a strong predictor of team performance."

You are most interested in predicting team performance. I am most interested in improving it.

We have different priorities so we may not overlap that much or that well in conversation.

But others can read and use what is made available according to their own interests and perspectives.

I would posit that the most beneficial instrument you could have in order to improve team performance is one which allows you to answer the question of predicting how a novel combination of players will work as a team.

I've been very rude and sharp elbowed here, and the dullest of wits is often combined with the sharpest of elbows, so it is not a flattering look. I do regret being quite so forthright.

My personal failings aside, you quoted me form another thread regarding how lineup analysis is useful-ish to me, but what the data has said to me, and what it I believe would say to others who look at it, is that lineup analysis lends itself to playoff basketball, in which you are trying to find the key (lineup) for the lock (opponent lineup). This is why a seven game series ends up feeling so game theoretical, at least when you are coached competently, which I do agree is not always.

In the regular season I would suggest most coaches optimise who is on the court to match a general play style, with little concern as to the opponent. I would also suggest this is a very logical approach, and that obsession over lineup performance is less so. Granted you can disregard that statement out of hand, as I have not corroborated it. You can do that with every word I say, if you like.

On that point, of why I don't share my numbers? Not only because they are proprietary to me and valuable to me, but because I think the cultivation of progress in any discipline is in the work of doing it for yourself, with a little encouragement and education along the way. Sometimes that encouragement gives you a direction to take, sometimes it brings about in a person the desire to prove that encouragement foolish, malformed, misdirecting.

I do not find it edifying, and do not think it is edifying, to sit in front of a set of numbers in the abstract, without having a very strong understanding of how those numbers have arrived.

Lastly, you asked about what I mean by teammate-marginal production. It's a simple thing. Within each game, a player's teammate marginal production of each counting stat (and indeed of raw +/-) is their count of that statistic, minus the average production of their teammates in that same statistic, within that same game, over an equal number of possessions.

What this fulfils is to ensure that teams who have a glut of production in one area do not overly bias the situation. If every one of your players is the same player, who produces the same box score on average, and whose plus minus matches the others, then there is no way to prefer any of them, beyond perhaps age or injury history, but let's imagine we do simply have twelve of the same player - we want the model to trivially be able to say they all get the same number of minutes, and that because each has zero marginal production compared to the other, they each tell the model nothing (this is a good thing), and get allocated by the model 5/12ths of the available court time merely by that being the case.

Another way to think about it, is if you are small squad, then a mediocre center is likely to find more court time on your squad than on one with other, potentially superior centers. This calculation allows for that implicitly, and the fact that two statistically similar players on different teams might be allocated very different minutes does not then inherently cause the model any great problems.

To put it another way, if you are the only player on TerribleTeam who ever scores any significant points, you are going to get minutes, even if you are dreadful at everything else.

Thanks for responding to the requested clarification.



A beneficial instrument you could have in order to improve team performance can be one which allows you to answer the question of predicting how a novel combination of players will work as a team.

And also careful review of what has been tried and use of what actually has worked, more often and what has worked less well.

Do both if you can.

DSMok1 wrote: ↑Tue Mar 11, 2025 1:53 pm In my years investigating these things, I've found essentially zero value in looking at point differential at the individual lineup level... The sources of noise and other variance swamp any signal. It takes multiple years of data, thousands of minutes, to start to get that noise under control... And at that point the long time period adds another source of noise that removes any value.

There can be some value in player pair or trio data, but you really have to control for player and opponents to have any chance at it being meaningful. J.E. did this at one point using a RAPM approach, and very few player combinations had enough signal to trust there were synergies that could be identified purely from point differential.

This is where coaches can recognize good and bad patterns of play and synergy or the reverse through their own pattern recognition and judgement far, far faster (multiple orders of magnitude faster) than it can be identified statistically/mathematically.

This is more or less precisely how I see it.

Crow, are you recording any of your results with lineup analysis? You cited the Celtics a couple of years ago, but that seems to just be anecdotal? Unless you have framework within which you are actually statistically testing the power of your hypotheses, you are essentially not doing anything which I would call rigorous analysis.

"I think the cultivation of progress in any discipline is in the work of doing it for yourself, with a little encouragement and education along the way. Sometimes that encouragement gives you a direction to take, sometimes it brings about in a person the desire to prove that encouragement foolish, malformed, misdirecting."

I could possibly undertake some bigger projects undone by others or in similar spaces as others available. Sometimes that is about time allocation and respect for the skill required to do certain things well or better than others. Sometimes it is about getting quicker payoff from simpler things not beind done by others. Know your skills / time / preferences for work.

You suggest "a desire to prove that encouragement foolish, malformed, misdirecting." Not the case.



"... dullest of wits is often combined with the sharpest of elbows, so it is not a flattering look."

Assessment of wits can be wrong and are often a defensive reaction. And not a flattering look.

I was speaking of myself, having reflected, and fed some ducks.

I explained that proposing alternative lineups that have never have tried or barely have been tried will limit followup evaluation. I have set about followup evaluation on a number of occasions and have confronted this difficulty repeatedly. I could persevere and perhaps present in detail what I can find & conclude more often. And might if there is genuine interest. I have received so little interest in my lineup analysis that I sometimes leave followup to myself or undone, awaiting possible interest of a team that wanted more analysis, recommendations and followup. More possible with more time and more incentive and the potential for utilization.

"I was speaking of myself, having reflected, and fed some ducks "


OK, then I won't take it personal unless you say it directly about me.

So the answer to my question, as I understand it, is no. You haven't devised a framework within which you can test your hypotheses. I cannot in good conscience take them seriously, in that case.

Perhaps they have value, but if it cannot be proven statistically, then it is hard to make that argument and seem credible. If you had such strong evidence, I imagine you'd find your way into a front office. Perhaps you can devise some sort of natural experiment and statistical tests which will begin to build a body of evidence in support of your analysis? I will admit I have no suggestions of how to take what you do and make it mathematically rigorous, as I understand it.

Crow wrote: ↑Tue Mar 11, 2025 3:10 pm "I was speaking of myself, having reflected, and fed some ducks "


OK, then I won't take it personal unless you say it directly about me.

I was saying it about myself. My sharp elbows don't do me any good, only make me seem a bull in a china shop. Make me seem dull but hoping belligerence will cover for that.

Ok.

"So the answer to my question, as I understand it, is no. You haven't devised a framework within which you can test your hypotheses. I cannot in good conscience take them seriously, in that case.

Perhaps they have value, but if it cannot be proven statistically, then it is had to make that argument and seem credible"


Fair enough.

Could do more.

Might.

I will consider critique and decide what I can and will do.

There are people who build sophisticated models, better than I probably could without major new effort. They can interpret them with greater depth of understanding.

I offered the OP 20-30-40 comments on twitter baaed on my level of knowledge ... and then suggested he talk to the experts.

He did come here seeking such dialogue.

There have been a few brief comments but not a sustained dialogue.

Any want to get back to that?

Seems unlikely, especially with OP stopping after first post.

But there might be a forthcoming more formal article, at least of the possibility of such mentioned.

Perhaps such an article would benefit from further discussion, and maybe beforehand than after.

Perhaps not. But the possibility is available if any have further comments directly and in detail on that topic.

I'd be happy to explain to the op all of the flaws I can see in their critique of RAPM, if they ever return to this thread and are curious.

To reiterate I don't really value RAPM.

They seemed quite zealous whilst being inaccurate, however, so I am not sure they'll be returning.

Yeah, it is indeterminate. But thanks for the generous offer. As I said at the beginning, I would probably mainly listen from here on that topic, if re-started here with expert participation.

I don't know whether this is within or beyond what you are willing to say about methods v-zero, but does your model using performance to predict minutes tend to over-predict minutes given to better performers, roughly match or under-predict? Under-predict weaker performers minutes given? Not expecting perfection or detailed television but is there a tendency?

Other people's experiences?

Has every team done this? Any or many teams? By position? Tracking over time and discussing with coaches? If not, why not?