The very best on Offensive EPM (into +6 territory) are about twice as high as the best rated on Defensive EPM.
Rookies start out estimated around -2.3 to-3. If they play to -2, they are an average player. If they fade to -4 or worse, they are a bottom 10% player.
There are huge variances in what players ranked near 100th and 200th are paid. Huge agent wins and losses. Huge team wins and losses.
23 at or above +3.
A few rookies contract bargains for now in top 40. Within top 40, getting paid only $25-30 million is a relative bargain compared to some of $45-50+ mil guys.
By EPM, LeBron is 67th, slightly behind I Joe and Aaron Wiggins. Neutral on D.
Scottie Barnes is a very expensive +1. Banchero at +0.9 would be soon.
If you rate better than -1, you are or about a top 200 player. +0.7 or better makes you a top 100. Fractional impact matters.
Cody Williams rates for now as the pretty clear worst performing player here too.
There were a dozen other recent draftees in bottom 35 that I didn't draft rank or barely did.
Observations with EPM
Re: Observations with EPM
Wouldn't an avg player in any PM be zero by definition?
It looks like -3 EPM is where no estimated wins (EW) are created.
https://dunksandthrees.com/epm
Re: Observations with EPM
An average "rating" would be the mean of the population. An average player would be the median. The mean is considerably higher than the median because of better to way better than median players with significantly more minutes.
Scales and what is measured vary by metric.
Scales and what is measured vary by metric.
Re: Observations with EPM
I did not know that avg is the same as median. I don't even know how one could recognize that. Do you count players with just one minute?
The median NBA 3-point FG% (of 569 players) was .328 last year. Median 2fg% was .529, eFG% .533, FT% .763
NBA averages were, respectively .360, .545, .543, and .780
At least, they are called averages everywhere I can find.
I get that there is a temptation to refer to the avg 7-8 player in the rotation as "an average player". But semantically, it just feels consistent to consider "avg NBA play" to be what the "avg NBA player" does.
Does it not make sense to refer to a plus-minus of 0.0 as zero above or below average?
btw, I am enlightened to notice that Durant and Butler are both median rebounders at 9.2 per 100 -- avg is 10.0 by definition.
This presumes they are rebounding against avg NBA rebounders, on avg.
The median NBA 3-point FG% (of 569 players) was .328 last year. Median 2fg% was .529, eFG% .533, FT% .763
NBA averages were, respectively .360, .545, .543, and .780
At least, they are called averages everywhere I can find.
I get that there is a temptation to refer to the avg 7-8 player in the rotation as "an average player". But semantically, it just feels consistent to consider "avg NBA play" to be what the "avg NBA player" does.
Does it not make sense to refer to a plus-minus of 0.0 as zero above or below average?
btw, I am enlightened to notice that Durant and Butler are both median rebounders at 9.2 per 100 -- avg is 10.0 by definition.
This presumes they are rebounding against avg NBA rebounders, on avg.
Re: Observations with EPM
Mean, median, and mode are different measures of center. The group of measures of center are often called loosely "averages" instead because it is slightly simpler and not expecting a debate on the difference.
Here is a example of the BBC calling "them" averages:
"An average is a single ‘typical’ value that is used to represent a set of values. There are three main types of average. They are called the mean, median and
mode."
https://www.bbc.co.uk/bitesize/articles/zj6nb7h
That is a loose definition... but lots of words have loose definitions because words are rough tools to convey varying meanings and are not necesarily absolute straightjackets.
But ok technically different, so call them different measures of "center", defined 3 ways.
"Average performance" is mean. "Average player" might have "meant" median or mean, depending on speaker's intention / strict accuracy in using the word average. It is useful to clarify. I said "An average player would be the median". I should have said it was to me in my usage at that time instead of "would".
Use the one you are interested in / "mean" and hopefully clearly state.
"Do you count players with just one minute?" I guess it does unless you established a qualifier. Without a qualifier, means and medians both include "trivial" players in their calculation. So maybe don't use the loose and debatable value term "average player" and calculate a reasonable / specified qualified / worth including player center of whatever center or representational / 'typical" value most interests you.
I rarely look at medians with regard to basketball, except to observe how it is different from mean (including with salaries).
"Does it not make sense to refer to
a plus-minus of 0.0 as zero above or below average?" It COULD but does not have to be. If you want to reference a player at the mean as an average NBA player, you can go ahead even it that puts them at the 65-75th percentile of players players . The main point is to know if you are measuring / caring about mean or median or both.
... semantically, it just feels consistent to consider "avg NBA play" to be what the "avg NBA player" does." One could be rigid and correct. Semantically I'd suggest be outspoken and consistent in that rigidity or specify, specify as you wish. "average player" is not universally defined and is not simply "average". It is the intention of the speaker or author. Middle or center player would be clearer. Maybe I'll use them to clarify that usage. But I specifically stated that the two were different and that is a main, meaningful, if obvious point.
"avg is 10.0 by definition" for mean rebounding. And median level is also good to know. Focus of understanding the data.
This response could be shorter and / or reordered or nothing at all, but that is where it stands for now.
EWins sets 0 at the very bottom / no
value or nearly so. EPM, RAPM, LeBron and many other metrics set 0 rating at average / mean. With EWins per se, "average" is some value, probably not zero and may or may not be clearly stated. Find or define a replacement level and calculate the difference if you want.
This scale difference has been discussed before. You go your way, they are staying their way.
Here is a example of the BBC calling "them" averages:
"An average is a single ‘typical’ value that is used to represent a set of values. There are three main types of average. They are called the mean, median and
mode."
https://www.bbc.co.uk/bitesize/articles/zj6nb7h
That is a loose definition... but lots of words have loose definitions because words are rough tools to convey varying meanings and are not necesarily absolute straightjackets.
But ok technically different, so call them different measures of "center", defined 3 ways.
"Average performance" is mean. "Average player" might have "meant" median or mean, depending on speaker's intention / strict accuracy in using the word average. It is useful to clarify. I said "An average player would be the median". I should have said it was to me in my usage at that time instead of "would".
Use the one you are interested in / "mean" and hopefully clearly state.
"Do you count players with just one minute?" I guess it does unless you established a qualifier. Without a qualifier, means and medians both include "trivial" players in their calculation. So maybe don't use the loose and debatable value term "average player" and calculate a reasonable / specified qualified / worth including player center of whatever center or representational / 'typical" value most interests you.
I rarely look at medians with regard to basketball, except to observe how it is different from mean (including with salaries).
"Does it not make sense to refer to
a plus-minus of 0.0 as zero above or below average?" It COULD but does not have to be. If you want to reference a player at the mean as an average NBA player, you can go ahead even it that puts them at the 65-75th percentile of players players . The main point is to know if you are measuring / caring about mean or median or both.
... semantically, it just feels consistent to consider "avg NBA play" to be what the "avg NBA player" does." One could be rigid and correct. Semantically I'd suggest be outspoken and consistent in that rigidity or specify, specify as you wish. "average player" is not universally defined and is not simply "average". It is the intention of the speaker or author. Middle or center player would be clearer. Maybe I'll use them to clarify that usage. But I specifically stated that the two were different and that is a main, meaningful, if obvious point.
"avg is 10.0 by definition" for mean rebounding. And median level is also good to know. Focus of understanding the data.
This response could be shorter and / or reordered or nothing at all, but that is where it stands for now.
EWins sets 0 at the very bottom / no
value or nearly so. EPM, RAPM, LeBron and many other metrics set 0 rating at average / mean. With EWins per se, "average" is some value, probably not zero and may or may not be clearly stated. Find or define a replacement level and calculate the difference if you want.
This scale difference has been discussed before. You go your way, they are staying their way.
Re: Observations with EPM
In eWins, avg = 1.00
A 1.50 player is 50% better than that, and .50 is half as good; so it's straightforward to say how much a player has improved or declined -- defined as how many more wins he's given his team.
Below zero is an experiment that didn't or hasn't worked out.
I notice the EPM has a small, faint number under the value, and I think it's the Percentile. So if you scroll down and find 50, you've got the median of the sample. For EPM it's -1.1, Usg is 17.1
https://dunksandthrees.com/epm/actual
The lowest 3fg% is .156, and the median is .354 -- quite a bit higher than including all players.
The sample is only those who attempted >30.
A 1.50 player is 50% better than that, and .50 is half as good; so it's straightforward to say how much a player has improved or declined -- defined as how many more wins he's given his team.
Below zero is an experiment that didn't or hasn't worked out.
I notice the EPM has a small, faint number under the value, and I think it's the Percentile. So if you scroll down and find 50, you've got the median of the sample. For EPM it's -1.1, Usg is 17.1
https://dunksandthrees.com/epm/actual
The lowest 3fg% is .156, and the median is .354 -- quite a bit higher than including all players.
The sample is only those who attempted >30.
Re: Observations with EPM
Yes the lower EPM values are percentiles but you are right to ask about what the population(s) is (are). Qualifiers are involved in estimated skill calculations but not immediately clear for all stat displays.
Different ways of doing things. Explanations are useful but not always complete or read.
Not sure if I had read the full explanation of the whole / current EPM world but saw it this morning: https://dunksandthrees.com/about/epm#e-skills
There are other pages I have not previously explored. Might have questions or comments after I reread fully and more closely. Commentary from others about EPM methodology could be interesting / useful potentially, if any are inclined to wade in deep enough and engage in such. Or just understand, accept and use it as one might.
I could spend more time comparing actual to predictive EPM. That could be very important, especially for team decision making (or gambling). Anyone heavily engaged in that and wiling to say?
There could be many hours or a lifetime of comparison and consideration of all the different metrics out there. I do some, could do even more. Not clear how many others do, how far. Teams or outsiders. Many in media or on X appear to do little or no metric use or comparison. Simply blends are available and useful but more rigorous multi-model constructions might be even better if any have done or wanted to try & share.
Yes EWins scale has a rationale and utility.
(Dean Oliver's offensive and defensive ratings used 100 as average for similar reason and effect.) I know you have explained parts or all in past. It is useful to inform new readers (to the extent there are such) and refresh.
Consuming and understanding and explaining data all take time when done to best practice. Sometimes they are more rushed. Slow down when one wants to / can.
Different ways of doing things. Explanations are useful but not always complete or read.
Not sure if I had read the full explanation of the whole / current EPM world but saw it this morning: https://dunksandthrees.com/about/epm#e-skills
There are other pages I have not previously explored. Might have questions or comments after I reread fully and more closely. Commentary from others about EPM methodology could be interesting / useful potentially, if any are inclined to wade in deep enough and engage in such. Or just understand, accept and use it as one might.
I could spend more time comparing actual to predictive EPM. That could be very important, especially for team decision making (or gambling). Anyone heavily engaged in that and wiling to say?
There could be many hours or a lifetime of comparison and consideration of all the different metrics out there. I do some, could do even more. Not clear how many others do, how far. Teams or outsiders. Many in media or on X appear to do little or no metric use or comparison. Simply blends are available and useful but more rigorous multi-model constructions might be even better if any have done or wanted to try & share.
Yes EWins scale has a rationale and utility.
(Dean Oliver's offensive and defensive ratings used 100 as average for similar reason and effect.) I know you have explained parts or all in past. It is useful to inform new readers (to the extent there are such) and refresh.
Consuming and understanding and explaining data all take time when done to best practice. Sometimes they are more rushed. Slow down when one wants to / can.
Re: Observations with EPM
Is there any interest in comparing EPM with other metrics (PER, WS, BPM, eWIns,...) in their correlation with minutes per game?
In other words, which metrics do NBA coaches tend to "agree" with?
I've been doing this more in playoff series, where decisions have to be made sooner rather than later.
Unsurprisingly, winners tend to have higher correlations.
We could make blends of metrics based on this correlation level.
[I don't see an easy way to grab the EPM from the website.]
In other words, which metrics do NBA coaches tend to "agree" with?
I've been doing this more in playoff series, where decisions have to be made sooner rather than later.
Unsurprisingly, winners tend to have higher correlations.
We could make blends of metrics based on this correlation level.
[I don't see an easy way to grab the EPM from the website.]
Re: Observations with EPM
That is a good question. I have interest. Might do something more later.
Several blends are out there: CraftedPM, ValuePM, a simple EPM-RAPM blend...
Should the test be minutes weighted metric values to minutes given correlation? Probably?
Found this statement from 9 years ago.
"For now all I can say, BPM, WS, RAPM and RPM are all worse than MPG at assigning individual player value."
viewtopic.php?t=9207
Fwiw.
MPG mix might predict lineup results decently but generally better?
PER was said to be better. Better would suggest closer than the others, more sinilar to coach given minutes. Closer for an offensive and usage biased metric? Closer is not surprising. "Better" is.
Some new and some perhaps improved metrics since this statement.
Metric value vs. minutes is one comparison. But lineup management is key with both. We only know what the Coach did, not all the other options which would likely produce different player metric values.
I don't think this test is the end of the story. But would be a good step. I looked quickly and did not find such a comparison... on first look.
On second look I found this:
https://www.unc.edu/wp-content/uploads/ ... search.pdf
It considers coaching performance in comparison to expected BPM and RAPM. It is focused at coach level and there is variance. Results with coaching may be a little better than expected on average but that is a guess. Not immediately sure how BPM did compared to coaching RAPM for correlation to MPG. Maybe I'll re-read more carefully later.
It is surface logical that players will improve if the coach is bringing any real value from one season to next or from considering usage elsewhere. The age curve might have a general positive effect with likelihood of more players before peak playing more minutes than after peak. Reading performance in a specific game is a different thing thing than average previous performance. Easier in ways, maybe harder at times.
Those are quick comments. I am not really into deep study of this at the moment.
Maybe others will read, analyze & comment?
Several blends are out there: CraftedPM, ValuePM, a simple EPM-RAPM blend...
Should the test be minutes weighted metric values to minutes given correlation? Probably?
Found this statement from 9 years ago.
"For now all I can say, BPM, WS, RAPM and RPM are all worse than MPG at assigning individual player value."
viewtopic.php?t=9207
Fwiw.
MPG mix might predict lineup results decently but generally better?
PER was said to be better. Better would suggest closer than the others, more sinilar to coach given minutes. Closer for an offensive and usage biased metric? Closer is not surprising. "Better" is.
Some new and some perhaps improved metrics since this statement.
Metric value vs. minutes is one comparison. But lineup management is key with both. We only know what the Coach did, not all the other options which would likely produce different player metric values.
I don't think this test is the end of the story. But would be a good step. I looked quickly and did not find such a comparison... on first look.
On second look I found this:
https://www.unc.edu/wp-content/uploads/ ... search.pdf
It considers coaching performance in comparison to expected BPM and RAPM. It is focused at coach level and there is variance. Results with coaching may be a little better than expected on average but that is a guess. Not immediately sure how BPM did compared to coaching RAPM for correlation to MPG. Maybe I'll re-read more carefully later.
It is surface logical that players will improve if the coach is bringing any real value from one season to next or from considering usage elsewhere. The age curve might have a general positive effect with likelihood of more players before peak playing more minutes than after peak. Reading performance in a specific game is a different thing thing than average previous performance. Easier in ways, maybe harder at times.
Those are quick comments. I am not really into deep study of this at the moment.
Maybe others will read, analyze & comment?
Re: Observations with EPM
In the browse for relevant information on this question, I came across a "Thibodeau" formula or metric. I didn't stop and focus on it. I will try to find it again.
Here it is, or at least one reference.
viewtopic.php?t=9207
This is Thib's simplistic formula, not what he actually did. The metric was found to be middle of the pack and super similar in prediction to RAPM, surprisingly. That is a good result, with BPM only slightly better. MPG, PER and Usage performed relatively poorly here (but not terrible) despite the above statement about MPG being best.
The data is from a long time ago. A new test would be good, if any wanted to do it.
I'll suggest that one could also look at what coaching got from team as intentional and perhaps derive a coach specific metric of value for any of them.
Another article talk about Thib's defensive metrics. He said he didnt use what others were (overall defensive efficiency / rating) and had his own. They were lower level and incomplete, focusing on shot defense at several levels and defensive rebounding. Opponent turnovers and fouls given went completely unmentioned.
This adds to the cautionary tale of allowing coaches doing their own analytics, things their own way, imo.
He cherry picks his 2 better defensive factors and ignored the lesser ones and the 19th overall rating.
https://www.nydailynews.com/2023/05/02/ ... analytics/
"I dont-go-by-'new-age'-defensive-analytics". New age metrics. A clever attempt at deflection of criticism but a crock.
Here it is, or at least one reference.
viewtopic.php?t=9207
This is Thib's simplistic formula, not what he actually did. The metric was found to be middle of the pack and super similar in prediction to RAPM, surprisingly. That is a good result, with BPM only slightly better. MPG, PER and Usage performed relatively poorly here (but not terrible) despite the above statement about MPG being best.
The data is from a long time ago. A new test would be good, if any wanted to do it.
I'll suggest that one could also look at what coaching got from team as intentional and perhaps derive a coach specific metric of value for any of them.
Another article talk about Thib's defensive metrics. He said he didnt use what others were (overall defensive efficiency / rating) and had his own. They were lower level and incomplete, focusing on shot defense at several levels and defensive rebounding. Opponent turnovers and fouls given went completely unmentioned.
This adds to the cautionary tale of allowing coaches doing their own analytics, things their own way, imo.
He cherry picks his 2 better defensive factors and ignored the lesser ones and the 19th overall rating.
https://www.nydailynews.com/2023/05/02/ ... analytics/
"I dont-go-by-'new-age'-defensive-analytics". New age metrics. A clever attempt at deflection of criticism but a crock.