Dean Oliver at ESPN introduces Net Points

[Mike G]

https://www.espn.com/nba/story/_/id/440 ... y-findings

more detail:
https://espnanalytics.com/nba-net-pts
toggle to net pts per 100 poss. or to WAR

[Crow]

Yes, it has been trickling out.

Some use, some caveats.

There if inclined to use.

I expect to check occasionally. More in playoffs.

Need to check closer how shot defense is handled.

[v-zero]

Far as I can tell there is no reinvention of the wheel here. Reading between the lines, because that is all we can do given that Oliver has made none of the finer detail public, this appears to be essentially a statistic built from extending the counting stat concept of the box score to observations available either by parsing the play by play (and I don't mean plus-minus data), or by using additional data sources such as tracking. Then, some regression has been done of observed counting stats vs results, and the weights derived therein.

This will, as they have noted, produce a metric which is inherently descriptive of the average value of those counting stats happening, but says nothing about the propensity for a player to repeat that action. This is the crux of where predictive value is derived. It is essentially useless to observe that the value of an action is X, if the repeatability of that action is essentially unknown. That is, the extent to which a counted event occurs due to chance rather than skill, and is thus not entirely repeatable.

They have noted on ESPN that they intend next to create a predictive model based on the same principles. I believe it is likely to be very good, given that I know from (several years now) of experience that this method is of significant value above and beyond what can be achieved by merely spinning up another iteration of box score enriched RAPM.

I notice too that EPM has recently received a (long overdue) refit. It appears now to be following the same principles of the best available public metric, DARKO. That's pleasing, as I see EPM referenced a lot, and people have lauded it because it does well in (completely useless) retrodiction contests. Fact is it's not been a good predictor of game outcomes since I have been feeding it as a feature into my models, in fact in general it doesn't even pass the variable selection phase during the daily refit of said models, swamped by other AIO metrics within the feature set which simply provide better value for predicting outcomes.

A case of convergence of a sort, it's nice to see. I still don't think anybody has publicly quite figured out that they don't need RAPM or anything like it to point a linear model at in order to create a box score (or extended box score) metric and have it be predictive, but behind closed doors I have no doubt many have found that step to be a blunder, and have also located the better path.

[Mike G]

It was just a bit unnerving to read the article without much detail into the process. And that it strongly seems to parallel Win Shares -- guys who don't miss many shots nor turn the ball over seemed to be favored over those who control the ball and create the shots.
Correlations between the per minute (or per possession) Net Points and a few other summary stats; and those stats to minutes per game:

Code: Select all

correl. to NP/100       correl. to MPG
PER   WS/48  BPM     NP/100  PER  WS/48  BPM
.799  .853  .840      .291  .470  .225  .443

WS always ranks very low against what coaches think is important -- minutes being a proxy measure for that. NP may be a modest improvement.
When Win Shares first appeared at b-r.com, it was accompanied by Loss Shares. The "worst" players in the league were whoever got the most minutes for the worst teams. We had intense debate here, and then it quietly went away.

The "worst players" in the league according to NP look like the same subset: members of the weakest teams with lots of minutes.
Luke Kornet is the best player for Boston (according to NP, WS)
Chi: Jalen Smith
Cle: Jarrett Allen
GSW: Kevon Looney and Butler
Hou: Steven Adams up there with Sengun
LAL: Vanderbilt and LeBron (and Luka)
Min: Gobert
etc.
Centers should do what they're good for, and most should not try to create all their shots. This metric seems to suck credit away from the players who are setting up their good shots.

[Crow]

V-zero, say or not say what you want, but I'll ask:

Does your model build entirely on individuals or does it explicitly consider lineups as a building block at all?

Can you meaningfully model coaching lineup management in abstract or in matchups? Is the lineup chaos such that it preferable to ignore lineup affects?

Coaching lineup management behavior is baked into individual performances so maybe it doesn't need separation at one level but I'd think understanding the two separately might have some advantages.
Past lineup usage doesn't necessarily maych future lineup usage. Lineup usage on average is not necessarily the same in a time decay fashion.

Lineup management probably can't be understood at granular level but perhaps in a simplified form. Starters and 1 change variations as a group, then 2-3 changes and then high to exclusive bench units.

[DSMok1]

When you divide credit on an individual possession level, it is difficult to account for the benefits of marginal shot creation. In other words, even with assistedness being the same, a player taking more shots has more value. When you look at things in larger chunks, that is much easier to include.

Notably, BPM is trying to do the exact same thing as this new metric. Simply to divide up the credit for the team's success. BPM is also explicitly not predictive. Just trying to divide up the credit for what has actually happened. Clearly, BPM has ended up a very different place than this metric.

From Dean:
"It basically divides credit on every shot, turnover, rebound, free throw. It uses some info not in the box score to do so - layups have a higher OR% than jumpers, for example. The quality of the shot generated by an assist. The play-by-play captures that better than the box. A general explanation of how it does the division of credit/blame is here: https://ldeano.substack.com/p/piece-of-harmony "

The concept reminds me of Evan Zamir's work back 15 years ago.

http://www.d3coder.com/thecity/2010/12/ ... valuation/

It's not a bad concept, I like it in theory, but something seems a little bit off here.

[Crow]

Statistical plus minus usually does not account for quality of teammates and opponents on the court. That is a big factor to ignore. It is a big reason for considering RAPM based metrics.

It would be possible to adjust statistical plus minus by various methods. Any examples of this being done?

[v-zero]

V-zero, say or not say what you want, but I'll ask:

Does your model build entirely on individuals or does it explicitly consider lineups as a building block at all?

Can you meaningfully model coaching lineup management in abstract or in matchups? Is the lineup chaos such that it preferable to ignore lineup affects?

Lineup analysis does have some value to me, but I have not found a good use for it in AIO metric construction, which is the backbone (though perhaps this beast has two backs, 'a' backbone might be more appropriate) of regular season prediction. In the chess that is playoff basketball lineups can make or break a series.

Statistical plus minus usually does not account for quality of teammates and opponents on the court. That is a big factor to ignore. It is a big reason for considering RAPM based metrics.

It would be possible to adjust statistical plus minus by various methods. Any examples of this being done?

Yes, my AIO metrics are all opponent aware. You will benefit more from the same performance against a better set of opponents (and indeed amongst a better set of your own teammates) than against a lottery team.

From Dean:
"It basically divides credit on every shot, turnover, rebound, free throw. It uses some info not in the box score to do so - layups have a higher OR% than jumpers, for example. The quality of the shot generated by an assist. The play-by-play captures that better than the box. A general explanation of how it does the division of credit/blame is here: https://ldeano.substack.com/p/piece-of-harmony "

Glad you were there to point me to the right place, googling had failed to find me that, and Oliver hasn't been active enough for me to have ever sought out his substack.

It's an interesting line of thinking. I like the metric less than when I assumed he was using statistical methods to find the appropriate weights for a set of extended counting stats, his arguments have merit but I am not sure they hold water.

[TeemoTeejay]

closed doors I have no doubt many have found that step to be a blunder, and have also located the better path.

Out of curiousity, within AIO metrics which ones did you use and between EPM and LEBRON which one did you prefer?

[v-zero]

closed doors I have no doubt many have found that step to be a blunder, and have also located the better path.

Out of curiousity, within AIO metrics which ones did you use and between EPM and LEBRON which one did you prefer?

EPM now, to me, seems better than LEBRON. Previously I thought they were about even, or at least they seemed equally useful in predicting the future. DARKO is the only publicly available metric that generally finds itself in my final model for a given day, once feature selection is performed. The others are of my own invention.

[Crow]

Would "Net Points" be useful for finding marginal player points (at least by factor) to predict minutes and / or to try optimize minutes and lineups?

[Crow]

For adjusting for opponents and teammates without using raw +/- or RAPM, one set of choices would seem to be any of various player impact metrics. And / or minutes. And / or salary. Or maybe number of starters. Or some combination of the above. There may be other options I am not thinking of at the moment. Anybody else have other choices or guesses they are willing to name?


I should check back with what ESPN is doing with Net Points more often. Are any other ESPN writers using it?

[Crow]

Current Net Points ratings by position:

"Pure" Centers (over 500 minutes)

10 over +3
17 total over +1
24 above neutral
Only 7 more above -2
Then big dropoff to Missi (-2.99)
and Sarr (-4.11).
70+% positive.

Go cheap here and you probably suffer.

If you don't have better than neutral, probably better do something. Draft, free agent, trade. 2 would be desirable. Few have that. Cavs, Rockets, Thunder if you go to C/Fs.

All 13 listed as C/PF are rated positive. Almost same for 22 F/Cs but Markannen comes in 20th at -0.73. Who is buying him and for what price? Would have to believe in unfavorable context impacts and / or rehabilitation.


102 "Fs". 40 positive, 62 negative. 6 over +3.
Buzelis 89th, Saluan 92nd, Risacher 95th, Cody Williams dead last.

Just 9 F/Gs. Doncic and Tatum and then not much else.

26 G/Fs head by Ty Jerome and Jalen Williams. 50/50 positive / negative.

136 Guards. 93 negative. Bottom 50 includes in descending order: S Sharpe, Beal, Henderson, Black, M Christie, Westbrook, Jalen Green, Castle, Coulibaly, Carrington and Collier. I was way lower than consensus on all these recently drafted guards (except Coulibaly). SGA 2 to 1 over next best guards.


Will continue to do more metric comparisons and proximate blends as time and interest dictates.

Who are the cases with widest / most significant metric differences. That would take more time to compile. What if anything do teams do with that information? Other outside analysts / fans?

[Mike G]

When I see "negative" player, I mentally translate it as "below average" or "not in the top 100".
Only 3 non-negative players (with serious minutes) per team. If the league contracts to 15 teams, will we see mostly positive players?
Nope. Same ratio, since "average" would be even more exclusive.

This being a thread on DeanO, I realize he creates similar language. I just get weary of the translating.

[Crow]

Without a minutes filter, Net Points has 214 positives. Darko about 170. BPM 193. 6-7 per team.

Stretch down to -0.5 estimates, it is Net Points 226, Darko 224, BPM 242. 7-8 per team.

Pretty tight range and then tighter.


Median obviously stretches lower. To -1 or a bit lower thru the on average 9th man.


The very best teams have 9-12 metric positives.

Looking by factor would tell some things about specialization and depth of quality.


A starter who is negative on many or all metrics has a story worth pursuing. Slight negatives would not be alarming but larger ones should prompt a review of alternative players, systems and roster & lineup construction.