Justin Jacobs

George Clooney hanging out at a Lakers #NBA game in April 1989. Chick Hearn makes no notice...

Just a random day -- back in the day -- hanging out with Trae Young.

JE demonstrates how to incorporate priors for RAPM. One of the tricks I've used during my years on pro teams has been to do this similarly using tracking data metrics. Check out this post to see how to build these types of priors carefully -- even when there's no closed form.

@yolo_pinyato That translated across two teams!

@Squared2020 yes, I'm glad to finally confirm my suspicions about Horace having a Draymond like impact with more of an offensive tilt.

Updated my #NBA RAPM and RAPM supplemental pages! Up to well over 2,000 games and 400,000 possessions. #Analytics

Supplemental Page: squared2020.com/2025/06/01/sup…

RAPM Page: squared2020.com/2025/01/26/his…

The Pistons and their private jet for the 1989 #NBA season.

@SilentSpy1 @mikegrib8 @DQuinn1575 100%. DRAPM is great for context, but usage, role, and matchup patterns matter a ton for interpretation.

@Squared2020 @mikegrib8 @DQuinn1575 drapm is a pretty good stat, but it’s probably not 100% fair to compare drapm among heavy usage offensive stars to defensive specialists or lower minute bench players. Historically speaking, there aren’t many usage offensive stars who also are > 1-2 drapm (especially guards)

What no one else will tell you so I will.. THere is a decent likelihood that Toni Kukoc was a more impactful defensive player than Pippen and Jordan on the left DRAPM from possessions tracked by Justin Jacob 85-96 and on right 30 year RAPM since 96

@DQuinn1575 @mikegrib8 No problem! I do this for the love of the game. I enjoy your guys' debates and always hope to help add context. I know sometimes my data and presentation can seem confusing, so asking questions is definitely encouraged.

@Squared2020 @mikegrib8 Thank you for the explanation. I appreciate all the work you did on this

@mikegrib8 @DQuinn1575 So I think there's some merit to your discussion, but I'd warn against gross overgeneralization and focus on the nuance on what makes each player great.

@mikegrib8 @DQuinn1575 Where Kukoc would look inflated is that RAPM treats Jordan against fresh starters the same as Kukoc coming in fresh against players that are 5-6-7 minutes in. That said, Kukoc was a really good defender -- even probably a better on-ball defender than Jordan.

@DQuinn1575 @mikegrib8 The pre-97 DRAPM is a full RAPM-style model, not simple on/off. It’s built from possession-level lineup data, and the regression adjusts for teammates, opponents, and scoring context using a ridge-regularized framework (the same structure used in modern RAPM).

@Squared2020 @mikegrib8 Thanks, question is your DRAPM data pre 97 more than simple on-off? Do you go thru the math of adjusting for teammates & opponents?

@DQuinn1575 @mikegrib8 So the DRAPM sample for Kukoc is structurally incomplete, and the modern RAPM dataset starts in 1997. The two datasets are disjoint and not meant to be combined into a single ‘career defensive value’ comparison.

@DQuinn1575 @mikegrib8 Just to clarify on the Kukoc numbers: he had a very solid defensive season in 1996, but my dataset does not include the bulk of 1994 and 1995... the years where Kukoc played without Jordan and logged his highest on-ball usage.

@AliKnowsBall @JerryEngelmann I’m happy that helps! 15 years ago, that plot use to frustrate me because RAPM is just a Gaussian-Gaussian model, but the sampling distribution across all dimensions (2*players+1) is not Gaussian, nor should the sub-distributions center on zero.

@Squared2020 @JerryEngelmann Interesting. Thanks for explaining it to me! And looking at that graph I do understand now why percentiles make more sense to use 🤣

I was asked recently how does #NBA RAPM data compare the 1980's players. So I ran it on the partial data I have. Some of the players makes sense. But there's a long way to go. Data is only from games during the 1980 through 1989 seasons.

@AliKnowsBall @JerryEngelmann You’re right. Just a victim of my Top-26 cutoff 😅 Here’s the full ORAPM vs DRAPM distribution from the sample. It aligns with what you see from JE. Also, the median needs no be zero. It's actually that the associated +/- needs to average out appropriately. IE: 2 pts is 2 pts

@Squared2020 Interesting. What's notable to me though is how they're negative instead of positive and near zero. Shouldn't about half of the players be below zero and half above zero? For example in the one-year RAPM for last year by @JerryEngelmann, there's lots of positive DRAPMs.

@inpredict The looks I get when I say "Matrix multiplication is just composition of functions." So beautiful. (The algebra, not the looks.)

Counterpoint: I remember my college linear algebra professor began her course with the statement: "Matrices are the greatest mathematical notation convention ever created." Although I'm guessing this part of the article was somewhat tongue-in-cheek.

@AliKnowsBall Through this lens, an ORAPM of ~+6.0 can be roughly equivalent to a DRAPM of ~+2.5. That said, elite bigs with ~+2.0 DRAPM often still post positive ORAPM. Their defense naturally boosts team offense, or they’re offensive forces themselves (Ewing, Robinson, Hakeem).

@AliKnowsBall This is common in traditional O/D RAPM. Since the game is heliocentric on offense, but not on defense, ORAPM picks up player-defined signals much quicker than player-defined signals in DRAPM. So we tend to look at the percentiles of ORAPM and DRAPM to infer "player value."