Pushing Buttons

Plus an interview with a vorps and shnorps acolyte

This is The F5, a twice-weekly newsletter that helps readers be less wrong about the NBA. If you enjoy it, think its interesting, or want to learn how to make the visualizations you find in this newsletter (below), please consider a paid subscription!

---

One of the early season surprises so far has been the hot start of the Cleveland Cavaliers. The Cavaliers have jumped out to an 8-0 record backed by the league’s 2nd ranked offense and 5th ranked defense. What’s new and interesting about these Cavaliers is that after five straight years of finishing in the bottom ten in pace they are now ranked 10th, which I believe can be directly attributed to the hiring of Kenny Atkinson over the summer.

As the head coach in Brooklyn, Atkinson’s teams finished 1st, 6th, 10th, and 11th in pace. With hardly any personnel changes, the Cavaliers are suddenly playing at one of the league’s fastest paces. That holds true whether you use the NBA’s definition of pace (a team’s number of possessions per 48 minutes) or something like Inpredictable’s “average time to shot”, where the Cavaliers rank 13th overall, but 8th after opponent made shots. To put it simply, Atkinson has the Cavaliers taking the ball out of the net faster than ever before.

The table above details the team-level rankings of every Kenny Atkinson-coached squad. Looking at the table, some patterns emerge.

As we’ve discussed, Atkinson’s teams are known for playing fast. But they’re also known for limiting opponent three-point attempts. All of Atkinson’s teams have finished in the top 10 in opponent three-point attempt rate (3PAr)¹. I don’t think these patterns are coincidences. Let me explain.

In my job with the Knicks I spent a lot of time thinking about what was in a coach’s control and which buttons they had available to push. When you’re coming up with a game plan for an upcoming opponent it’s not a good idea to recommend, “Score more points.” Instead, you focus on the things that can be influenced by giving them extra emphasis in practices and walk-throughs. “Push the ball up the court” is a lot more realistic for players to execute than “Put the ball in the basket.”

So what is in a coach’s control? Which buttons can they reliably push independent of player personnel? To find out, I looked at the rankings of every team that was in the first season with a new head coach and compared it to the rankings of the teams that had previously employed the same head coach. I wanted to see if a coach that likes to play a certain way with one team, will those same stylistic preferences show up in their first season with a new team?

The table below shows the correlation between team-level rankings in a coach’s first year with a new team vs their last year with their old team. I’ve ordered the stats below in order of the strength of their correlation. A correlation of 1 (or -1) indicates a perfect positive (or negative) correlation, while a correlation of 0 means there is no relationship at all.

Pace is by far the strongest correlation of the bunch, suggesting that coaches that like to play fast with one team are likely to play fast with their next team. This should come as no surprise to anyone that’s followed Mike D’Antoni’s career. His first seasons with Denver (1998-99), Phoenix (2003-04), New York (2008-09), Los Angeles (2012-13), and Houston (2016-17), resulted in pace rankings of 6th, 6th, 2nd, 7th, and 3rd, respectively.

On the other side of the table is opponent three-point percentage, giving support to the analytics-backed belief that teams — and more specifically coaches— have little-to-no control over how well their opponents shoot from three.

However, at the same time, opponent three-point attempt rate is one of the stronger correlations. So even if a team can’t control how many threes their opponent makes, they can control how many threes their opponent takes.

I was surprised to see how weak the correlation was on team three-point attempt rate (0.09). In a previous edition of this newsletter I pointed out that every Mike Budenholzer-coached team has finished in the top half of the league in three-point attempt rate. Sure enough, as of this writing, the 6-1 Budenholzer Suns rank 5th in three-point attempt rate. That correlation ranking could be the result of small samples (we’re working with a sample size of 139 coaches that have led more than one team since 1996-97) or it could be something else entirely, like the fact that coaches who like their teams to shoot threes are less likely to switch teams. But that’s just a guess. To be honest, it’s something I’d need to investigate further.

Back to Cleveland. The Cavaliers are 10th in pace and 4th in opponent three-point attempt rate, which we would expect from an Atkinson-coached team. The question is whether these changes will result in anything more than an early round exit in the playoffs.

We have a long ways to go until the postseason, but I think these early season indicators are positive signs for their chances of matching up against Boston, specifically. One of the reasons I think Joe Mazzulla said the Indiana Pacers gave the Celtics the most trouble last year is because Indiana sold out to take away the three last season. Boston ranked first in three-point attempt rate in 2023-24, but Indiana ranked first in opponent three-point attempt rate. To put it another way, Indiana was — and now Cleveland is — built to take away the thing that Boston wants to do the most.

---

Coaching Profiles

Something I learned while writing this edition of the newsletter is that there is not anywhere online where you can quickly look up the rankings for every team led by a specific coach. To find out where every Mike D’Antoni-led team finished in pace, you’d have to go to 16 different pages on nba.com/stats. And then you’d have to look at 16 different pages to find out where those teams ranked in three-point attempt rate. No thanks.

To solve this problem I wrote some code to pull the team-level rankings for every coach season since 1996-97. This was a somewhat arduous undertaking since many coaches are fired (or hired) midseason and to find out where their teams ranked at the time of their firing (or hiring) requires a lot of careful filtering of start and end dates.

I will be publishing an interactive/searchable table with the rankings of every coach season since 1996-97 on Friday for paid subscribers, but wanted to show a preview of some of the more interesting coach profiles I made.

---

Short Leash, Long Bench

The biggest button that head coaches can push is the minutes button. The chart below summarizes each team’s standard rotation size by looking at the average number of players that receive at least ten minutes of playing time in a given game.

In a typical game, Golden State is giving real minutes to up to 11 players. Meanwhile, the Knicks are keeping a tight eight-man rotation on average which sometimes results in games like Monday’s matchup in Houston where just seven Knicks players received ten or more minutes.

I’m not sure either strategy is “good” or “bad” so much as they are different approaches to different goals. The Warriors clearest path to a championship this season is by making a consolidation trade for a star that puts them over the top. But to make that trade they have to show they have multiple players that are worth trading for. The Knicks on the other hand are limited by the number of NBA-level players they have on their roster after making a pair of consolidation trades of their own this offseason.

---

A Refreshing Interview with Taylor Snarr, creator of Estimated Plus Minus (EPM)

If I could only look at one stat for the rest of the NBA season it would be Estimated Plus Minus (EPM). Not only is EPM one of the most trusted advanced statistics by NBA executives, the site (dunksandthrees.com) that hosts it is one of the most aesthetically pleasing basketball sites around.

Like it’s all-in-one player value metric brethren (BPM, RAPTOR, LEBRON, and others), EPM attempts to figure out how impactful every player in the NBA is. Its creator, Taylor Snarr, developed the idea of EPM after working as the Utah Jazz’s first dedicated analytics staffer.

Recently, EPM underwent a big shift.

Snarr detailed the changes on Twitter, but the gist of it is that EPM is now “forward-looking” instead of “backward-looking.” In practice, what this means is that EPM is now better suited to answer questions like “Who is the most impactful player in the NBA right now” and less useful for answering questions like “Who has been the most impactful player this season.”

I spoke to Snarr about the changes to EPM, handling public criticism, and Franz vs. Paolo.

---

F5: What inspired you initially to develop EPM?

TS: I've always been drawn to all-in-one metrics because, while not perfect, I feel they provide the most objective approach to identifying how good players are, and I love tracking and watching the greatest players in the world.

Some of my favorite metrics include Daniel Myer's Box Plus-Minus (BPM), and Jeremias Engelmann / Steve Illardi's Real Plus-Minus (RPM), which both served as inspiration for EPM, among others.

I wanted to build an NBA analytics site because I enjoy the possibilities and creativity that coding allows, and I thought it would be interesting to have a metric of my own. I really enjoy pursuing models that test well and also generate output that feels like it's capturing reality on some level. I also thought it could be interesting to include newly published player-tracking data into the metric.

Why did you make the decision to change EPM from being a backward-looking metric to a forward-looking one. EPM is already highly respected by people in and around the NBA. So if it ain’t broke, why fix it?

The change from backward-looking, season EPM to forward-looking EPM was largely about using a lot more data and handling sample size issues in a better way. For season EPM², you really need to wait fairly deep into the season for player values to stabilize, and I wouldn't even publish values until 15 to 20 games into the season. But there's an entire ocean of data from previous seasons that can help provide a realistic estimate of player impact for the beginning of seasons or any point in time.

But the true power of the new version comes from how data from previous games/seasons are used. With the old EPM, the stats were entered as-is, without regard to how much sample is needed to trust them. It was often the case with 3-point shooting, which can be the source of a lot of random point-variance, that might lead you to believe a player is better or worse than they actually are.

Hearing the phrase, "but it's a small sample size", is super frustrating because the next natural thought is "well how much do I need?", which is hardly ever answered. So the new EPM is optimized to handle the sample size question for you by weighting past data appropriately for each stat individually, producing stats you could more easily trust at any point in time. This approach was inspired by how Kostya Medvedovsky's DARKO model stabilizes player stats, which I thought would be a powerful pairing with the EPM model. 

I think people often used EPM to support their arguments, like why a player deserved MVP. But I’m not so sure the new version of EPM is well suited to answer those questions anymore. Are you worried at all that people will unintenitonally misuse the new version of EPM? 

I can understand that it will be an adjustment to go from season-scoped values to forward looking values for any date. After-all, things of consequence in the NBA happen in individual seasons, so it is very natural, myself included, to want to know impact for seasons. Beyond it being a natural inclination, I think it is also useful to look at past impact for seasons to see how we got to where we are.

The common misinterpretation of past impact is assuming it will continue as is, that it represents how good players are now. So the shift to forward-looking EPM takes what has happened and gives you a better answer moving forward.

The good news is, I will be bringing back season EPM so both will be available to help answer different types of questions.

Something that stands out about EPM is the presentation. Its clear a lot of thought went into how information should be displayed. A small example is the beeswarm plot that’s on the right side of the EPM page that gives the viewer has a sense of the distribution of values, which can get lost when just looking at values in a table. How much of your time is spent fine tuning and developing your metrics vs. designing the website, dunksandthrees.com

Creating the new EPM and underlying projection system took most of my time over the last while, but I'd say in general it's about 60% tuning models and 40% developing the site. I tend to compartmentalize as well, so I might go months without developing any new stuff on the site while I'm immersed in building/tuning models.

One thing I've noticed about you is that don’t appear to take things personally. How do you deal with all the negative feedback when EPM is for lack of a better word, your baby

I admire people who put their work out into the public domain, and it's something that hasn't come very easily for me. I do take things personally from time to time, but ultimately I want to do things the right way. I love the scientific approach of dispassionately pursuing reality, and I value a similar approach in my life in general, which oftentimes, for me, just means trying to observe my inner experience and give it space for what it is.

Are there examples about “dispassionately pursuing reality” in your day-to-day life that come to mind?

When there's something triggering or distressing, I will sometimes take the time to slow down and try to get clear on the stories that I'm believing, and then gently challenge those stories. I think it's human nature to assume that what we are feeling represents reality, but it seems to me that beliefs/stories often just happen to us, and so I have found it helpful to try to process my inner experience like this to be more aligned and to make decisions from a more grounded place.

Changing directions a bit, are there times where you see people cite EPM in bad faith or in ways you don’t think are appropriate?

I don't see a lot of that, but there are limitations to EPM and impact metrics in general. I call it "Estimated" to convey that true impact is unknown. EPM relies on player stats fairly heavily, and these capture a decent amount of impact on offense but less on defense.

EPM uses some player tracking data on defense which helps, but there are very limited player tracking data available to the public.

Regularized Adjusted Plus Minus (RAPM) captures impact beyond player stats, and I've found that it really helps improve EPM, but RAPM is largely a black box and doesn't consider roles players play.

So a common mistake is assuming a player's impact will be the same on a different team which could involve a different role or system. I think corroborating EPM with film is important.

Couple of rapid fire questions to wrap this up. Franz or Paolo, who ya got? Franz is an EPM darling while Paolo is not as much, which I think runs contrary to what most people would assume.

I'll take Franz. Franz's True Shooting % skill is much higher than Paolo's right now, meaning it believes his scoring efficiency is more proven in spite of a great start for Paolo this season. Paolo is tracking well, though, and I think they're both great players right now and will both be even better.

Any interest in working for a team again or is that part of your life behind you now?

If a team doesn't mind that DunksAndThrees stays public then I could be open to exploring something.

What’s one thing you can’t live without during the NBA season?

You mean other than the F5? I think the F5 is fantastic and I'm honored to do this Q&A. Also, I seem to be obsessed with Couilably's usage. And I love watching Steph Curry play.

1. meaning each of his teams finished the season with one of the ten lowest opponent three-point attempt rates

2. the backward-looking version