Introducing Stable Player Impact (SPI)

[nbacouchside]

This is the metric I used to make my Team Projections for the Wins Projection Contest. Its DNA is pretty close to that of PIPM with some tweaks. I rolled it out on Twitter a little while back, but it occurred to me that discussion here might also be fruitful.

https://nbacouchside.net/2022/11/05/int ... mpact-spi/

2023 SPI:
https://docs.google.com/spreadsheets/d/ ... edit#gid=0

1997 to 2022 SPI database (RS numbers only):
https://docs.google.com/spreadsheets/d/ ... sp=sharing

I should note that I have developed my own team strength rating that I use for the final team adjustment that is a little bit better than the SRS-style adjustment I was using when I first wrote the introductory piece. All the numbers are updated to reflect the improved team strength ratings.

[Crow]

SPI for 2023 has about 12-13% of players at or above +1, about 7% below -2. So about 80-81% in between.

Do you have an average error estimate?

EPM has almost 19% at or above +1 and 40% below -2 despite only 393 qualifiers.

SPI appears more way regressed to mean than EPM or EPM far less regressed. Not saying one is better or right, but something to note. Difference bigger at bottom than top.

DARKO projection is a different thing but it projects 17% at or over +1 and a bit over 20% at or under -2. Fairly close to other 2 at top, same as EPM. Bottom size is close to middle of the other 2 metrics, with a huge spread. The bottom has some fairly implications for cap management / lower roster construction. Who is / how many are good enough / not too terrible for end of roster varies whichever view you take.

Perhaps the comparison will look different at end of season, but things to note now and watch.

How big a difference does the Games Started squared element make. average and max, in absolute value and how many versions of it did you consider?

[nbacouchside]

1. Do you have an average error estimate?


2. SPI appears more way regressed to mean than EPM or EPM far less regressed. Not saying one is better or right, but something to note. Difference bigger at bottom than top.

Perhaps the comparison will look different at end of season, but things to note now and watch.

3. How big a difference does the Games Started squared element make. average and max, in absolute value and how many versions of it did you consider?

1. If I remember correctly, the MAE on the regression was around 1.0 to 1.1 on both the Offensive and Defensive components. I'd have to re-run the regressions to give more exact numbers.

2. The mean regression is intentionally more aggressive this early in the season due to the smaller sample size. I use stabilization via padding on all components to account for the noise inherent in small samples. The spread should grow as the sample grows. The effect of this is to weight box-score numbers more heavily (as they have less mean regression built in) vis-a-vis on-off components earlier in the year and eventually give relatively more weight to the on-off components as the sample grows.

3. The GS^2 component is meaningful. I only considered it and MPG and GS^2 was meaningfully more significant and they were redundant enough that including both didn't make sense. I don't have the average and max differences handy.

[Crow]

I expected you'd explain item 2 in that fashion but thought it should be said to explain current distribution and lay out expectations for movement as season ages.

In the 26 year database, +1s and better represent 15.1% of total cases. Halfway between season to date SPI rate and the other 2 metrics I mentioned. A modest spreading out of the range but still a little tighter at top than the others.

-2s and worse go to 22.9% of database. Moving to middle of current low SPI rate and current displayed EPM rate for a smaller sample and pretty close to what DARKO has right now. I'd expect the current rate to move toward 20% by season's end and that is probably the rule of thumb- 20% below -2. So try to get guys above that, but accept that some or all of the 3-4 lowest ranking roster guys may be below that.

62% between +1 and -2. Top and bottom groups could be measured alternatively in terms of standard deviations.

Comparisons of distributions could be made to other metrics but the 2 I made were enough for my immediate interests.

If the estimated average errors are that low, SPI looks like it should get attention, get tracked. In this contest of team roll-ups and perhaps in a fuller metric comparison at player level later, removing minute projections from the test.

I might make metric comparisons for specific players later, or perhaps others might.

Do you plan updates here and / or elsewhere?

Thanks for sharing this work.

[nbacouchside]

Do you plan updates here and / or elsewhere?

I am planning on trying to get the process for running this automated and onto a website, so its not a Google sheet that I'm updating daily. I also plan to get playoff data together as soon as I figure out how I want to handle team strength in the playoffs. Ben Taylor's recent work on this has been influencing my thinking, but I also want to still do 4 factor component and opponent adjusted ratings. Ben's piece on it is is behind his Patreon paywall, but here's the link:
https://thinkingbasketball.net/2022/10/ ... -ratings/

I have plans to put together a player tracking data version using the same luck adjusted on-off data, but we'll see how long that takes and I won't have the 26 year RAPM basis for the tracking data since most of the data available for that only goes back to like 2013.

Thanks for sharing this work.

Thanks for the feedback!

[Mike G]

Will you have another metric for guys like Rodman and Artest?

Never mind. It's good to have a Starter% factor! Numbers are more easily racked up against bench players.

Have you considered using Away assist rates, since some scorekeepers are more generous than others, and especially for the home team?
One year I think Blake Griffin had twice the assist rate at home as he did by the standards of the other 29 scorekeepers.

Blocks can also be inflated by the home team. Sometimes they favor individuals, sometimes it's the whole team.
What the league-at-large considers a block or assist, we know by the average of 29 opponents' scorekeepers.

Steals, too, can mean different things with different teams. Last year, Miami opponents had 2.03 turnovers per Miami steal. Memphis' ratio was 1.54. One might infer that a steal by a Heat player was effectively worth 31% more, on avg, than a Grizzly one.

[nbacouchside]

Have you considered using Away assist rates, since some scorekeepers are more generous than others, and especially for the home team?
One year I think Blake Griffin had twice the assist rate at home as he did by the standards of the other 29 scorekeepers.

I did consider using an adjustment, ala the court factors outlined here: https://web.archive.org/web/20161218203 ... 1694917271 and here: https://nbacouchside.net/2022/08/01/nba ... eper-bias/

I may still do in potential future iterations. Would probably improve the correlation with O-RAPM, I'd bet. Might also use an adjustment for teammate field goal percentages to predict potential assists rather than actual assists, which is probably also slightly more indicative of passer quality than actual assists.

Blocks can also be inflated by the home team. Sometimes they favor individuals, sometimes it's the whole team.
What the league-at-large considers a block or assist, we know by the average of 29 opponents' scorekeepers.

Steals, too, can mean different things with different teams. Last year, Miami opponents had 2.03 turnovers per Miami steal. Memphis' ratio was 1.54. One might infer that a steal by a Heat player was effectively worth 31% more, on avg, than a Grizzly one.

I may do a similar thing to the above for these stats. Not sure when I'll have the time to do these sorts of updates, but certainly good thoughts.

[Mike G]

From an article linked in the link:

(these numbers were crunched by a user on the APBRmetrics forum);

yep
(Our friend Alex who used to do the Grizzlies in Vancouver stats)

The assist/court factor formula on your Aug. 1 entry seems to conflate these 2 scorekeeper traits:
- Generosity: giving more or fewer assists to all players
- Bias: giving more to the home team.

A scorekeeper may have one or both or neither of these tendencies. Generosity to the opposition doesn't affect the home team or player factor. Just look at Away rates to get the real (league avg generosity) assist rate:
Ast/FG = aAst/aFG
You can look this up for individuals or for teams. More valid would be for single players, but it's likely more work.

To adjust a player's "official" assist total, divide his home total by (Home Ast/FG)/(Away Ast/FG), and add it to his away total.
Last year, teams ranged from 1.073 "jack factor" (Pacers) down to .977 (Celts)

So if Tyrese Haliburton is in the running for Assist Champion, check those numbers.
I'm pretty sure Chris Paul got a title that Steve Nash should have won.

[Crow]

I compared SPI to EPM on some starters that could be called weak starters.

Looney, I Stewart, Avdija, Dort, Jalen Green, Sochan, S Adams. 5 of the 7 were notably higher on SPI.

Fwiw. Starter factor in formula could be playing a role. Right or wrong. Bigger study would give more information, maybe better perspective.

Inclusion of factor may help more than hurt but any introduced factor could hurt on a portion of the database.

[nbacouchside]

I compared SPI to EPM on some starters that could be called weak starters.

Looney, I Stewart, Avdija, Dort, Jalen Green, Sochan, S Adams. 5 of the 7 were notably higher on SPI.

Fwiw. Starter factor in formula could be playing a role. Right or wrong. Bigger study would give more information, maybe better perspective.

Inclusion of factor may help more than hurt but any introduced factor could hurt on a portion of the database.

I will say that the impact of the games started term will be relatively greater earlier in the season when the stabilization or padding for the on-off components narrows the spread of those terms more significantly than any of the box score components. As the spread becomes greater as the sample grows for the on-off components, games started will have relatively less weight. To the extent that in the early season it is weighed more heavily, I think that is probably more feature than bug, as I think producing against starters is more meaningful than otherwise and coaches also, generally, know what they are doing when making decisions about whom to start and you're likely picking up on some signal from practices and film study that has predictive value going forward for who is better.

[DSMok1]

It is well known that looking at a playing time measure, of which GS^2 is one and MPG is another, has a significant benefit to the accuracy of metrics. As you note, coaches definitely know things that the available statistics do not.

The exclusion of a playing time measure from Box Plus/Minus 2.0 was a conscious design decision, with the knowledge that the overall metric would be less accurate (by a couple %) but that the individual statistic BPM coefficients would be more meaningful and understandable. If a playing time measure is included it muddies the individual statistic coefficients since it is basically measuring how good the player is "a second way".

If the goal is purely accuracy of the resulting single number metric, including a playing time measure makes sense.

[Crow]

There was an update to this metric posted on Twitter by Knarsu3 recently. If interested, find it there

[nbacouchside]

There was an update to this metric posted on Twitter by Knarsu3 recently. If interested, find it there

I don't see it.

[Crow]

https://twitter.com/knarsu3/status/1620 ... jD7QqbBzLA

[nbacouchside]

https://twitter.com/knarsu3/status/1620 ... jD7QqbBzLA

Oh yeah that is a link to my Google Sheet that I try to update every day (but at the very least get done once every third day).