RAPM factors (and more)

[Crow]

Anybody capable and willing to produce public RAPM factors again? As a public service or as a stepping stone to team work?

RAPM pairs (and other lineup levels)?

We had RAPM factors at least 3 or 4 times but not continuously and not currently.

I consider such data highly desirable over not, though of course subject to imperfect estimate / error.





Somewhat related addition:

Fwiw, re-thinking my approach for "counterpart RAPM" (simple concept from years ago, but flawed and now removed) it should be RAPM or RAPM factors compared (subtracted one from other to find net) by best possible matching up of main opponents. The matching up will be imperfect but it might be worthwhile.

Could simply match 1 to 1 or do some math to include estimated crossmatch percentages resulting from team defense.
I.e. PG against PG at end of possession for say estimated 55% of time, against SG 20%, SF 15%, PF 5%, C 5% or some other rough breakdown from general eye test experience (unless there is available play level video matchup determinations by paid video scouts or AI). And similar matchup estimates for the other positions (careful to sum to 100% for each at team level).

The matchup process could also be attempted alternatively with various hybrid boxscore / adjusted plus-minus metrics.

I'd probably assume starter vs. starter and first sub vs. first sub in initial cut but could hybridized this too.

Could do a full matchup for a game, compare to actual results, fine-tune process.

Doing at factors level (pure RAPM or otherwise) would be the most desirable to try to learn about factor level interactions in matchups. That is where things grind from expected to actual.

If you compared expected results to actual for every play, every game you could potentially build a probabilistic model / simulation that went beyond simple, always standard match-up rules to ones based on patterns observed in match-ups between players of different quality levels by factor.

Dean's "Roboscout" approach to factor data was not revealed but I had assumed it was team level not player level. Wonder if any of that will ever be revealed.

[Crow]

CraftedNBA has a few adjusted factors but the future is cloudy.

[TeemoTeejay]

Anybody capable and willing to produce public RAPM factors again? As a public service or as a stepping stone to team work?

RAPM pairs (and other lineup levels)?

We had RAPM factors at least 3 or 4 times but not continuously and not currently.

I consider such data highly desirable over not, though of course subject to imperfect estimate / error.





Somewhat related addition:

Fwiw, re-thinking my approach for "counterpart RAPM" (simple concept from years ago, but flawed and now removed) it should be RAPM or RAPM factors compared (subtracted one from other to find net) by best possible matching up of main opponents. The matching up will be imperfect but it might be worthwhile.

Could simply match 1 to 1 or do some math to include estimated crossmatch percentages resulting from team defense.
I.e. PG against PG at end of possession for say estimated 55% of time, against SG 20%, SF 15%, PF 5%, C 5% or some other rough breakdown from general eye test experience (unless there is available play level video matchup determinations by paid video scouts or AI). And similar matchup estimates for the other positions (careful to sum to 100% for each at team level).

The matchup process could also be attempted alternatively with various hybrid boxscore / adjusted plus-minus metrics.

I'd probably assume starter vs. starter and first sub vs. first sub in initial cut but could hybridized this too.

Could do a full matchup for a game, compare to actual results, fine-tune process.

Doing at factors level (pure RAPM or otherwise) would be the most desirable to try to learn about factor level interactions in matchups. That is where things grind from expected to actual.

If you compared expected results to actual for every play, every game you could potentially build a probabilistic model / simulation that went beyond simple, always standard match-up rules to ones based on patterns observed in match-ups between players of different quality levels by factor.

Dean's "Roboscout" approach to factor data was not revealed but I had assumed it was team level not player level. Wonder if any of that will ever be revealed.

Is RAPM Factors just getting "RAPM" but instead of points it would be like, the 4 factors?

I think I might have had code for that awhile back I cant remember if it was any good but I can try it if that sounds like a cool thing

[J.E.]

Adjusted FtR is kind of pointless, so it's a good idea to combine eFG and FTs (including And1s) into a single factor

[TeemoTeejay]

Adjusted FtR is kind of pointless, so it's a good idea to combine eFG and FTs (including And1s) into a single factor

That makes sense

So it would basically be RAPM but instead of points it would be effect on OREB%, TOV%, and that then right?

I remember awhile back I feel I saw it on that nba rapm site before it all changed and everything, they did it in a way to get those things all add up to the RAPM number for the offensive or defense side but I think the way they did that part was just RAPM - OREB RAPM - TOV RAPM, if I’m not tripping.

I think getting RAPM on those 3 factors separately seems straightforward it’s just the, getting them to equal the number part I’m a bit confused by unless it just comes out to that

[Crow]

Adjusted FtR is information. Better to see than not. Might reveal something in some cases.

rTS% is bigger and more meaningful because of size but good to see proportions of impact of the 2 factors.


Anyone who is able and willing to provide well done RAPM factors for public consideration, great.


Whether to do RAPM factors in way they are forced to add up exactly or not, I am open-minded. Ideally you'd have both and compare.

[J.E.]

So it would basically be RAPM but instead of points it would be effect on OREB%, TOV%, and that then right?

I remember awhile back I feel I saw it on that nba rapm site before it all changed and everything, they did it in a way to get those things all add up to the RAPM number for the offensive or defense side but I think the way they did that part was just RAPM - OREB RAPM - TOV RAPM, if I’m not tripping.

I think getting RAPM on those 3 factors separately seems straightforward it’s just the, getting them to equal the number part I’m a bit confused by unless it just comes out to that

It's not a "requirement" for the 3 factors to perfectly add up to standard RAPM, although if you're very far off there's probably something wrong

You can simply determine a total of 6 factors where
oRAPM = f1 * oPPS + f2 * OReb% + f3 * oTOV%
dRAPM = f4 * dPPS + f5 * DReb% + f6 * dTOV%
where PPS is the combination of eFG and FtR

[TeemoTeejay]

Anybody capable and willing to produce public RAPM factors again? As a public service or as a stepping stone to team work?

RAPM pairs (and other lineup levels)?

We had RAPM factors at least 3 or 4 times but not continuously and not currently.

I consider such data highly desirable over not, though of course subject to imperfect estimate / error.





Somewhat related addition:

Fwiw, re-thinking my approach for "counterpart RAPM" (simple concept from years ago, but flawed and now removed) it should be RAPM or RAPM factors compared (subtracted one from other to find net) by best possible matching up of main opponents. The matching up will be imperfect but it might be worthwhile.

Could simply match 1 to 1 or do some math to include estimated crossmatch percentages resulting from team defense.
I.e. PG against PG at end of possession for say estimated 55% of time, against SG 20%, SF 15%, PF 5%, C 5% or some other rough breakdown from general eye test experience (unless there is available play level video matchup determinations by paid video scouts or AI). And similar matchup estimates for the other positions (careful to sum to 100% for each at team level).

The matchup process could also be attempted alternatively with various hybrid boxscore / adjusted plus-minus metrics.

I'd probably assume starter vs. starter and first sub vs. first sub in initial cut but could hybridized this too.

Could do a full matchup for a game, compare to actual results, fine-tune process.

Doing at factors level (pure RAPM or otherwise) would be the most desirable to try to learn about factor level interactions in matchups. That is where things grind from expected to actual.

If you compared expected results to actual for every play, every game you could potentially build a probabilistic model / simulation that went beyond simple, always standard match-up rules to ones based on patterns observed in match-ups between players of different quality levels by factor.

Dean's "Roboscout" approach to factor data was not revealed but I had assumed it was team level not player level. Wonder if any of that will ever be revealed.

Finally had time last night to try this out, the NBA API I ended up combining FTr and EFG%, could be willing to try seperating the two, although that would miss the component of ft%. there's at least one WNBA example of a player who really heavily effected one and it was offset to an extend by the other iirc, not sure in the NBA though (On Defense).

I did put a Raw version and one that attempted to get things added up. it wasnt perfect, but the residual between the added numbers and the final result was around 0.1 iirc, so I think its fine, passes the smell test.

Jokic defensive impact on defensive rebounding was very funny to see lol

Numbers formatted in a way that positive = good.

Its all 2 year RAPM, so its 2014-2015 to 2015-2016 for example, up to this past year.
No Edit:
https://www.dropbox.com/scl/fi/5ke8asht ... brfsx&dl=0

Editted to roughly add up:
https://www.dropbox.com/scl/fi/awlckzcm ... 1y24h&dl=0

[Crow]

Thanks very much.

Look forward to using.

Any chance you'd also offer an additional file trimmed to just lastest timeperiod? The whole history file is quite a challenge for my phone.

[TeemoTeejay]

Thanks very much.

Look forward to using.

Any chance you'd also offer a file trimmed to just lastest timeperiod? The whole history file appears to a great challenge for my phone.

Heres a sortable github version where it will only load the selected year, I basically just copy and pasted lol. The Raw numbers tell a story but wont add up to the RAPM number, the scaled one does.

Raw table
https://timotaij.github.io/FourFactorRAPMRaw/

Scaled table
https://timotaij.github.io/FactorRAPMScaled/

[DSMok1]

This is wonderful work!

[Crow]

Thanks again.

Gained insights on one team and few other stars on first use.

More to come.

Filters by team, position, salary would be even more friendly but I am quite pleased to have this.

[TeemoTeejay]

Thanks again.

Gained insights on one team and few other stars on first use.

More to come.

Filters by team, position, salary would be even more friendly but I am quite pleased to have this.

I can try that out, maybe can add more to it too, if I have some free time there

Could look into adding more years too if the raw PBP data pre 2015 is the same format that I can just pass it through

[TeemoTeejay]

This is wonderful work!

Thanks! it was really interesting to do for sure some interesting stuff.

Jokic defensive rebound impact is not suprising but its hilarious lol

[J.E.]

In the 28-year rebounding version of this he's 2nd out of ~2,700 players in DReb% impact
https://docs.google.com/spreadsheets/d/ ... sp=sharing
Most players in his DReb% impact range are retired