WarriorGM wrote:CDM_Stats wrote:WarriorGM wrote:
If you're looking at them in isolation but we're looking at multiple similar lineups and applying a prior framework based on prior years to see consistency in expectations.
That's actually worse. There's no consistency to looking at prior years and comparing multiple tiny datasets is exponentially worse
Looking at 5 man units like this is most effective with consistency. N=1 and all that. If Kerr continues to use a wide rotation with a ton of changing lineups, the value of comparing 5 man units shrinks a ton because that consistency is shattered to the wind. If Kerr rides hot hands or plays lineups based on matchups to the degree he has in the first 7 games, there's not going to be much value to gleam there. Value would have to be deferred to individual metrics, individual successes, and applying THOSE variables
For example - the Celtics are a perimeter based team. JK is not a perimter-based player. However, if JK is only given minutes against non-perimeter based teams, his perimeter defense might look stronger, indicating a false positive of him being a good option against a perimeter team. That's at an individual level. It is 4-5x less likely to be an accurate representation when stacking more and more players on top of it
Thats what analytics is.. not just reading metrics, but understanding how and when to apply them
Bottom line - if you are using metrics based on single digit minutes in single digit games right now to evaluate 5 man lineups and their efficacy, I love that motivation, but its not going to give you anything meaningful to use. Its way too reactive. Unless you're evaluating every 5 man lineup the Warriors have put out there while factoring in every 5 man unit the opposition has put out there. And still, there's going to be massive variance game to game because of how basketball works
From a strict standpoint yes. Maybe this is all just association and no causality is involved. That's why randomized controlled trials are the gold standard in a scientific setting. But absent lab conditions you're always going to get some kind of confounding variables in the mix.
Happily just because association cannot prove causality it doesn't mean either that it isn't there. Moreover what we are studying isn't completely 100% known. There are various factors or interactions that might be escaping the common understanding. Just because doctors of the past didn't know exactly why willow bark might help certain maladies just knowing it did help was still useful.
The datasets are what they are but that doesn't preclude us from suspecting some lineups might have a chance of being great while others won't work. If we're doing that at a greater than random rate then we probably know something.
When I say n=1, there never really is a 1, just that results have more confidence as variables change less
The metric I showed in the offseason had a confidence level in the 40s.. with a full season of data and a lot of similar rotations
There is nothing that can be gleamed from 8 games of data. Is Hield going to shoot 50% from deep on high volume? Is Moody going to shoot around 50% from deep? Of course not.. all its doing is narrowing possibilities at this stage. In the tracking models I've seen, 5 man units don't even count until there's 30 minutes. Not a strict standpoint either - that's actually extremely generous
In terms of ideas, models would still be in the brainstorming stage. No conclusions should be set, other than conclusions based on basketball acumen outside of the model. Example - TJD's showing to be a meh defender and poor team rebounder. A lineup with him and SloMo (another poor team rebounder) is likely going to force the team to be extremely efficient offensively. No metrics needed, but a credible hypothesis
Even ESPN didnt show that RPM number until there was like 15 games in, because measuring on small datasets not only is inaccurate, it erodes confidence in the metric altogether