APBRmetrics

Paging the RAPM experts (J.E., vzero, etc.)

How does RAPM account for the fact that lineup fit is important? I'll give 2 examples.

Let's say you have a lineup of 5 of +0 defenders.

Guard
Guard
Guard
Small Forward
Power Forward

Thats a small lineup that will play below average on defense. I assume that RAPM would punish those guys even though they were put in the wrong position.

Another hypothetical. Let's say LeBron (+5 on offense) plays with different lineups

LeBron + 4 floor spacing 3 point shooters (assume they are true +0 guys)

RAPM would say they would be expected to put up +5, but instead they put up +9. Do those guys get extra rewards for that?

Another example. Lets call them the 2011 Heat.

scrub PG (-2)
Wade (+5)
LeBron (+5)
Bosh (+4)
Scrub center (-2)

RAPM would say that lineup would produce +10, but the 2011 Heat didn't because of diminishing returns and lack of fit. How does the RAPM credit get distributed in the scenario when there are diminishing returns for playing 3 great players together.

RAPM do not directly addresses those issues but,

1. Ridge regression limits unusual values comes from variety of reasons yours' included.
2. RAPM data size is huge. Due to nature of statistical data, all of the examples you give will be valid for most of the players eventually, and the effect of it will be minimial if we consider the size and variety of the data along with the way regression especially one big ridge regression solves the problem.

These bring their own problems too such as bias but they are considerably minor compared to others.

The point of RAPM (and virtually all one-number player evaluation metrics) is to isolate a player from his teammates and give some manner of 'average' value - what he's worth over the course of all his minutes in a season, or whatever timeframe the data covers. So if you took five RAPM +0 defenders, who happened to be small guys, I don't see how you could use RAPM to predict anything other than that they would be a +0 defensive line-up.

An exception I can think of would be if a player spent most/all of his time in the particular arrangement you're talking about. If one of those +0 defenders spent all his time playing in smallball line-ups, then you have to assume he's actually an +0 smallball defender. If you then decide to say that means he's probably an above-average defender with a 'standard' line-up, that's a leap you have to make.

I do think there's been some work on 2-, 3-, and 4- (maybe all the way up to 5?) player combinations. I suppose someone could try to classify those groups as certain types (smallball, spacing, twin towers, etc) and see if they give systematically different ratings from the individual players' ratings.

It doesn't. It merely tries to find the average impact a player had on the court within the sample in question. However it would be possible to add in dummy variables to try to factor in some element of fit.

Yes, dummy variables for unusual lineups would be the way to go.

Regular RAPM can't help those players that are put into bad situations by their coaches.

I've ran into a similar problem a few years back when I computed adjusted rebounding: many of the Don Nelson coached players were (probably) unfairly showing up as horrible defensive rebounders - that's what happens when you're in lineups that feature Al Harrington as center

As Mark Cuban said in regards to the original APM system developed by Wayne Winston.

APM measures not necessarily the best players in the league, but the players put in the best position to succeed.