2018 NBA Draft Model 1.0

An exercise in forecasting the NBA stars of the future

As the calendar turns to March, NBA fans and executives focus more and more of their attention on what college heroes of March could find their names on NBA rosters throughout the league in six months time. Building an NBA draft model, to project the top collegiate prospects, is not only a good exercise in modeling, but a good thought exercise in determining what matters in the making of an NBA success story.

Of the top three players on the last 10 NBA champions, 17 (57%) were acquired via the draft, while only 7 were acquired via trade and only 6 due to free agency. Thus, of the three paths to landing a star, the draft has become the most enviable way to construct a roster.

The Model

The biggest difficulty in draft modeling is that any one, top-down statistic is going to be built on some underlying assumption that doesn’t account for large aspects of the game. Thus, this model — Player Market Value Predictor (PMVP) — uses a player’s market value as its response variable. What is a player’s true market value? Well, PMVP predicts the percentile of a player’s second contract value (first year value), as a percentage of the salary cap (to adjust for league inflation), amongst players of similar positioning (PGs, Wings, & Bigs — to normalize for scarcity amongst some positions). This response, though far from perfect, doesn’t have the typical bias against defensive ability, but is more reflective of what the league thinks of a player as they enter restricted free agency after their rookie-scale contract. But how do you interpret this? Well, for example, Stephen Curry received his second-contract prior to the 2013–14 season for $9.9 million which was 16.9% of the salary cap at the time. Yet, that 16.9% of the salary cap ranks in the 76.7th percentile of NBA point guards. Typically, those at about the 75th percentile are fringe all-star players (remember, Curry had major injury concerns when he signed that second deal) and those near the 90th percentile are usually the top-10 or so players in the league.

PMVP is built using random forest regression, a machine learning technique, and Monte Carlo simulation on players drafted between 2003 to 2013 (you can find more on the methodology at the end of this post). The table below shows the mean of results for players within the top 50+ of ESPN/DraftExpress’s Top 100 players in the draft class of 2018. You can also see the ceiling(max) and floor for these players (the top and bottom 5% of possible careers). Remember this list won’t contain the likes of Luka Doncic, Michael Porter Jr., and Mitchell Robinson, but will paint a picture of the players with significant collegiate seasons vying for an NBA contract next season.

Results

Analysis

The model makes a lot of statements in its projections that you may or may not agree with. But, I figured I would highlight a few of the intriguing prospects and what the model says about them.

Troy Brown, Oregon Fr.

Troy Brown is not talked about a lot amongst the top few tiers of player’s in the draft, but PMVP forecasts him as the 4th best collegiate prospect and the best non-big in the class. His numbers do not jump of the page initially, but is steady across the board. Being one of the youngest players in the draft (which the model values highly), and having the athletic ability, at 6'7" and 215 pounds, to guard multiple positions should make him one of the most enviable talents entering a league that values wings with this talent highly. His shooting is what prevents him from being in the discussion for an early lottery pick, but at 31% from three this season, his mechanics and shot selection do not show any signs of him not being able to improve this in pre-draft stages or with a solid player development program.

Jaren Jackson Jr., Michigan State Fr.

Jackson has been a player in the discussion amongst the top 5–10 prospects, yet PMVP rates him as just the 21st best collegiate prospect. His youth and defensive instincts provide a lot of NBA scouts with hopes of him becoming a solid starting NBA center. But, it was his slower start and relegation to a second and many times third option in Tom Izzo’s offense that limits the sample size of his potential playmaking ability offensively. It will be intriguing to see if his stock changes with a productive, long run in March.

Keita Bates-Diop, Ohio State Sr.

Bates-Diop was a player many thought could be a late second round pickup or two-way contract candidate coming into this season, but has emerged into a quality, efficient scoring threat from multiple levels with the versatility to guard across multiple positions. PMVP does rightly show he has a lower floor, but if he can translate into the coveted 3-and-D player in the NBA, a late first round listing would become very understandable.

Methodology & Future Adjustments

All the data for this project came from web scraping of kenpom.com or draftexpress.com. Data and code can be found on my github page. The model is built on collegiate entries to the draft from 2003–2013 using a random forest model with a number of predictors from the players’s college statistics, biometrics, and how good their college team was. Using a normal distribution, the fitted response for each player in the 2018 draft, and the residuals of the model within each position group (PG, Wing, Big) the 2018 players’s careers were then simulated 10,000 times using Monte Carlo simulation to account for the variability in potential outcomes. Of these results the mean provides the expected market value for a given player, while the top 5th percentile of the simulations provides the “ceiling” of a player, and the bottom 5th percentile provides the “floor” of a player’s career.

I have already begun working on future versions of the model that can decrease variability and provide a better estimate of a player’s NBA career. A few of those concepts are using more than simply the last collegiate season of a player and weighting more recent seasons more heavily. Secondly, instead of dividing into the three position groups of point guards, wings, and bigs, I plan to use an expectation-maximization (EM) algorithm to cluster players into more unique positions. This will provide a better estimate within the random forest, but particularly the simulation because we know that while Chimezie Metu and Deandre Ayton are both classified as “Bigs”, they are world’s apart in terms of their style. Thirdly, I hope to tinker with a player’s college team’s statistics to determine how the style of team might positively or negatively effect a player. For instance, Kyle Kuzma and Donovan Mitchell are already looking much better than they did in college. Maybe their role in college hid their true abilities, but could be parsed out if we factor in more the unique styles and makeups of college programs.

Expect version 2 of PMVP to be presented within the next few weeks as well as further draft analysis, but in the meantime checkout my other work, email me, or reach out to me on twitter for further discussion/questions.