Nylon Calculus: Three models to project the best offenses for next season
Most analytic-based team win projections are built on one or more of the one-number overall player metrics, weighted by expected minutes played. The most successful over the last few years have been used some combination of Regularized Adjusted Plus Minus (RAPM) and box score metrics like Box Plus Minus or my Player-Tracking Plus-Minus, or ESPN’s Real Plus Minus (RPM), which effectively combines the RAPM and a box score prior.
Even though it is widely held that having the proper fit among players’ skills is beneficial there are almost too many aspects of player fit for us to reliably quantify good- or poor-fitting talents so far, and how roster fit factors into a team’s win projection. One thing we have a good handle on is the value of spacing in the form of 3-point shooting. Another is that the five cardinal positions, while still useful as a rough sorting of roles, are not adequate to describe the varieties of roles players fulfill or the subtleties of player fit.
When I saw Abhijit Brahme release his research defining ten new roles using historical data from 1998, I decided to build on his work to explore adding more elements of roster fit to some projections.
Read More: Ranking the best and worst scorers in every offensive role
As Brahme’s player type clusters had some of the most interesting categories on offensive player types, I decided to start by using the clusters to project a team’s offensive efficiency, as captured by offensive rating (points per 100 possessions: ORTG). To do this, I modeled on the percent of minutes played by incoming or returning players for each player type. For example, if the players for the Indiana Pacers played 18,000 minutes last year and 2,700 minutes were played by “Typical Bigs” and 2,200 by “Offensive Playmakers” those categories were modeled at 15 percent and 12.2 percent respectively against the out of sample next year’s ORTG.
Given the changing modern team roster make-up Brahme identified, I thought it was best to concentrate on more recent years. I trained three models using performance over the last three years — one using only the ten player types and total minutes played the prior year by players on the roster, another that included Offensive BPM and one with my Offensive PT-PM using one year only.
The Player Type model performed fairly well, explaining about 60 percent of the team variance on ORTG, with a penalty to the adjusted R^2 for using so many variables. The key variable in different model selection methods was percent of minutes played by Gods vs Not Gods in the Player Types. I also built a Bayesian regression tree model, which gave similar results. Below, I am using a model that excludes the Gods, giving most everyone else a negative impact on offensive performance and a small bonus for the total number of minutes logged in the NBA the prior year.
Percent of minutes given to Rough Riders, which includes players like Noah Vonleh and Omer Asik, are the most detrimental, followed by Average Point Guards, Average Wings and Typical Bigs. Other than the Gods, Offensive Playmakers and All Around Bigs are the most valuable offensive players to have on the court. At the very bottom I have some examples of each Player Type from last year.
The modeling including Offensive BPM with the Player Types tended to give low value to the Player Types and was somewhat unstable in the variable selection. That is not too surprising, given that the BPM metric and player types were built using the same advanced box score data. The only other metric included with BPM was the total minutes in the prior year.
The Offensive PT-PM had a somewhat more complementary relationship with the Player Types. Essentially the Typical Bigs, and to a lesser degree PnR Bigs, were negative when paired with PT-PM, while the other Player Types did not have a consistent effect. This has to be interpreted within the relationship of PT-PM to offensive efficiency, namely that the Player Types are picking up something PT-PM is missing. In addition, average usage weighted by minutes played the prior year was a consistent positive with PT-PM in predicting.
.
Using those three models and the projected rosters as of today, I was able to make preliminary prediction on the most efficient offenses for next year. Sit down folks; I think the Warriors and the Rockets are going to have pretty good offenses this year.
A couple of notes on the projections:
The Warriors have the second-highest offensive efficiency projection in two of the three models and first in the PT-PM based model, for first overall.
Houston is rated as the best in the BPM based model, third in PT-PM and fourth via Player Types, with about 24 percent of their minutes played by ‘Gods’ the Rockets are a little lower in the Play Type model.
Boston and Minnesota might be surprises to many people being this high. One aspect is that the Player Type ‘Gods’ is binary, so that Gordon Hayward counts as much as Steph Curry, even though not all Basketball Gods are actually equal. Another is that Minnesota’s fit doesn’t look as bad as some might think based on these rankings, avoiding ‘Typical Bigs’ and ‘Ave Point Guards’ that most penalize an offense.
Portland’s offensive projection is weighted down by Jusuf Nurkic’s numbers from his Denver stint, Noah Vonleh and because CJ McCollum is considered an offensive playmaker, rather than a true basketball God.
Chicago looks bad.
There are other issues that need to be explored here. Both Brahme and Senthil Natarajan have done some interesting work on lineups and fit issues recently. Both roster analysis and lineup analysis have issues. Rosters only indirectly look at fit, since there are many different lineups that can be combined in one roster over the course of a year. Lineups in the NBA suffer from the opposite issue of having small sample sizes, only one five man roster played over 1,000 minutes together in the NBA last year per Basketball-Reference, for example. Looking at data from both directions is helpful, I think.
Of course, there is always that other pesky defensive side of the ball, where the bigs might have their revenge.
Player Cluster, description and examples:
Cluster 9 — Average PG: Rajon Rondo, TJ McConnell
Cluster 8 — Gods: LeBron James, Giannis, Nikola Jokic
Cluster 7 — Rough Riders: Noah Vonleh, Omer Asik
Cluster 6 — Average Wings: Stanley Johnson, Brandon Ingram
Cluster 5 — Typical Bigs: Trevor Booker, Robin Lopez
Cluster 4 — Offensive Playmakers: Goran Dragic, Carmelo Anthony
Cluster 3 — Defense/PnR Bigs: Steven Adams, Tristan Thompson
Cluster 2 — All-Around Bigs: Marc Gasol, Myles Turner
Cluster 1 — All-3, No-D: Harrison Barnes, Allen Crabbe
Cluster 0 — Defensive Guards: KCP, Kris Middleton, Jae Crowder