Four Factor RAPM
Something that I thought would be fun to do was Four Factor RAPM, it ended up being pretty straightfoward so I did it last night after dinner and I ran it overnight. RAPM can be seen as a fancy way to parse a player’s impact from teammates, but some consistent weird results can pose more questions than answers. For the full breakdown of RAPM and some concerns I have with it, you can find that on this the first few sections of this post here. https://www.teemohoop.com/mamba/Blog%20Post%20Title%20One-mm8gk
If you go there or are visiting for the first time, maybe check out the All in One I made, the updated (non crazy long) write up is available at the top of the link I sent, or you can go to it directly here https://www.teemohoop.com/mamba/Blog%20Post%20Title%20One-mm8gk-cy9wh
Most of this is copy and pasted from Linkedin, but a few extra examples at the end:
Why does Jokic’s defensive RAPM always end up being so good?
Why are 2023 and 2024 Embiid and post-GSW Durant's ORAPM not quite as high as the top offensive players in the NBA?
Why does Caruso’s defensive RAPM always look like he's the greatest bald guard defender in NBA history?
A flaw of RAPM, (beyond noise and ignoring context) , is it doesn't explain the “Why.”
So, many treat it as a signal to see if something is worth analyzing or looking into. While I this opinion over this kind of data, I thought making it point out the “why” when its not noise, might be possible, to an extent.
Last night I thought it would be interesting to make “Factor” RAPM, instead of getting impact on point differential, you get impact on TS%, Rebounds, and TOs. (Think the four factors, combining FTR and EFG and FT% into one). You end up:
OREB impact
OTOV impact
OTS impact
DREB impact
DTOV impact
DTS impact
You then scale it in a way so it generally adds up to Offensive and Defensive RAPM. Which I did just running linear regression using those variables on the component, and getting the coefficients and multiplying. I did this in a way so a positive number always was good, and a negative number was bad. (Positive DTOV = force more turnovers).They don't add up perfectly, but generally add up pretty close (0.1 away on average).
Practically, how does this help with questions from earlier?
With Jokic: As much as his hands/positioning/general Defensive IQ is praised, the entirety of his positive defensive impact pretty consistently is from his insane impact on the defensive glass.
With Embiid and KD: I’ve heard arguments its from a lack of playmaking, but there’s isn't much evidence here suggesting this is what’s behind their lower than expected offensive RAPM. Durant’s impact on TS% was better than Jokic from 2021-2023, and Embiid was 3rd his MVP year and 2nd this year, despite his offensive RAPM being outside of the top 10 both years. Its mainly from negative impact on the Offensive Glass, especially for Embiid
With Caruso: He led the league in “Defensive Turnover Impact” for the last 3 years
I caution making any sweeping conclusions off this or raw RAPM in general, especially off two year samples, but I believe there are interesting practical takeaways you can get from this type of analysis, at least when its not just noise. Sometimes you can even see if a player's ranking might be dubious for something out of their control, if a perimeter player who doesn't crash the glass has an exceptionally high Offensive RAPM one year because OREB impact, and its a one year thing that isn't consistent year o year, maybe its noise or luck.
It wouldn’t be difficult to add more years or change the bounds of years depending on if the NBA PBP Possession format from the API is the same
data from 2016 to 2024 (2 year stints): https://timotaij.github.io/FactorRAPMScaled/
Raw version (unscaled to RAPM) from 2016 to 2024 (2 year stints): https://timotaij.github.io/FourFactorRAPMRaw/
Running a WNBA version of this is INCREDIBLY noisy because, well RAPM is incredibly noisy for those samples, but some interesting practical takeaways from that, thats private though lol.
I haven’t done anything on MAMBA since I made that blogpost, but might take out the offensive and defensive team * minutes effect that was originally there.
