jalengreen wrote:Doctor MJ wrote:jalengreen wrote:Assuming you two are alluding to multiplying D-LEBRON by Minutes? This would not give you the "AR" part of WAR because replacement level is not 0, it is -2.7.
So if you add 2.7 to each player's D-LEBRON and then multiply that by their minutes, then I have the ordering as Zubac > Gobert > Draymond.
Suppose one could then question why we're going off of replacement level, but that's just what the site is doing for WAR so I stuck with it.
(Let me know if I incorrectly assumed)
I love that you're thinking about the true replacement level.
Question: How was the replacement level defined?
They say the replacement level of -2.7 is the "estimated LEBRON for a G-League replacement added to the roster", nothing I see beyond that
As for what you said earlier about role adjustments, it's not entirely clear to me what role the player classifications play in the calculation. Definitely a good thought and now I wish I knew more about the methodology.
They mention using offensive roles to help w/ stabilizing box score priors:Stabilization
To deal with smaller minute players from previous year and better process data for future seasons to help determine if high performance on a small sample is noise or real, we’ll be stabilizing/padding data by combining a technique outlined by Kostya Medvedovsky here with our Offensive Archetypes, which label players based on their jobs on offense.
What this approach does is determine the volume at which each box score statistic stabilizes (and becomes a good indicator of performance rather than noise). A tiny sample of outlier performance won’t get the math’s full buy-in, but sustained performance over a higher sample will be respected by the math.
Incorporating role allows us to treat the expected values component of that math with a bit more common sense. An Off-Screen Shooter, operating via pin downs and flare screens often for 3-point looks, won’t use the same average value in their calculations as a Roll & Cut Big, who does their work at the rim and rarely (if ever) takes 3-point shots.
Through these techniques, we’ll end up with a stabilized and role-adjusted version of boxPIPM as our box score prior.
It only mentions offense so unclear if it's doing what you referenced, by applying the defensive roles to defensive box score stats. I imagine it wouldn't affect much in this discussion specifically because we're not talking about any lower minute players, but regardless it's definitely something to consider with the metric anyway.
Appreciate the response and the details, some thoughts on Replacement Level?
1. The question of course is how they did the study to conclude that this is what G-league players do when they come to the NBA. Are they taking guys who had been in the G-league in the past year, grouping them together as a single player, and regressing that to a single value? Are they including a guy like Scoot Henderson who last year as a rookie played WAY more minutes than a typical called up G-leaguer and played worse than a typical G-leaguer?
2. Is 2.7 is the value they got for overall, shouldn't the offensive & defensive be less than that and probably different from each other?
Regarding Stabilization, I totally get it for Prediction but would just emphasize against that I'm not sure I'd want to use this factor when considering past achievement. Not that I think it literally makes the general Past results worse on average, but when considering individuals, I think it's going to overrate and underrate certain guys in predictable ways...but not to predictable degrees.
Last note:
I have to acknowledge I didn't do Replacement at all for my table and so I think I probably should. I think I'd be incline to treat half of 2.7 as the replacement level for each side of the ball. That's obviously imperfect, but definitely folks should let me know if it seems particularly wrongheaded.
