Written by: John McCool (Desertrose28)
- Melo Trimble and Dennis Smith Jr. are the highest rated draft prospects, based on the model.
- The number of steals and free throw attempts per game in college are predictive of future NBA success.
- The model tends to value guards and small forwards over power forwards and centers.
This year’s NBA draft is one of the deepest in a long time. This means that teams drafting in the mid to late first round have a decent shot of landing a future starter.
Of course, players like Markelle Fultz, Lonzo Ball, and Josh Jackson will be swooped up early in the draft, but other players such as Lauri Markannen, Kadeem Allen, and Ike Anigbogu could also be high value selections later in the first round.
Aside from the first two picks in the draft, there is very little distinction between the value of picks number three through ten according to the data journalism website Five Thirty Eight. For example, Ben Simmons, the top pick in 2016, hasn’t even suited up for the Philadelphia 76er’s due to injuries. Finding underrated prospects is a difficult task in the draft and requires a healthy dose of luck.
Using analytics can be useful in projecting the future value of draft prospects. While most players are evaluated primarily through scouting, modeling can offer useful insights into a prospective draft pick’s future ability in the NBA.
Data and Methods
The college and NBA data used in this model comes from four sources: Basketball Reference,Sports Reference,, and Draft Express. The primary goal is to project NBA prospects’ value above replacement (VORP) over their first three seasons in the league.
According to Basketball Reference, VORP is a box score estimate of the points per 100 team possessions that a player contributed above a replacement level player (ex: -2.0), translated to an average team and prorated to an 82-game season. In other words, VORP measures the value of an NBA player against a G-League level player who could be signed as a free agent. This makes VORP very useful for player evaluations and comparisons across the NBA.
The model is trained on 269 current NBA players drafted from 2000 through 2016. It is important to note that a small subset of players in the model were drafted in either 2015 or 2016. In turn, their first or summed second year VORP (for players drafted in 2015) was used to approximate their three-year VORP. International players were not included in the model since the level of competition and data availability varies in different leagues.
In addition, a small group of former draft picks did not partake in the NBA combine. We approximated these players’ physical features such as wingspan and vertical jump by taking a positional mean (i.e. point guard, small forward). Obviously, this allows for an element of bias in the model.
We tested several predictive models including random forest and linear regression, but found that the gradient boosting model had the lowest mean squared error (1.02 MSE compared to 1.32 MSE in the random forest model). Thirty-three variables were used to project the 2017 draft classes’ three-year VORP based on college shooting, rebounding, defense, and NBA combine metrics.
According to the model, the most predictive variables for early NBA success (in terms of VORP) are steals, free throw attempts, assists, weight, and two-point field goal attempts per game. The player’s college position is also an important predictor of early career VORP. The least predictive variables in the model are the player’s wingspan, field goals per game, three pointers made per game, and whether he played in a power conference or not.
Figure 1: The relative influence of variables in the gradient boosting model.
As noted, steals and free throws attempted per game in college are the most important indicators for players’ three-year VORP after entering the NBA. Below is a smooth spline fit (Figure 2) that shows a positive relationship between steals per game (x-axis) and VORP (y-axis). The grey shading represents 95% error bars on the graph.
Figure 2: College steals per game against player’s VORP over his first three seasons in the NBA.
A 0.37 correlation exists between players’ college steals per game and VORP over their first three seasons in the NBA. This trend is even more evident for point guards, shooting guards, and small forwards (0.44 correlation). For reference, ten of 17 former draft picks that averaged at least two steals game in college, such as Stephen Curry, Marcus Smart, and Chris Paul, had at least a 2.0 VORP during their first three seasons in the NBA.
The predictive power of steals slightly shrinks for power forwards and centers (0.33 correlation between VORP and steals per game). While Nerlens Noel’s averaged 2.1 steals per game at Kentucky and a 3.3 VORP during his early NBA career, other college stars including Kelly Olynyk, Nikola Vucevic, and Myles Turner averaged 0.5 steals per game or less in college but were still productive early in their careers.
Figure 3: College free throw attempts per game against player’s VORP over his first three seasons in the NBA.
Free throw attempts are also predictive of future VORP but there is significantly more noise surrounding this metric. The above graph shows a steady rise in players’ VORP until around 3.50 free throw attempts per game. Past this point, it is difficult to assess whether free throw attempts are still predictive of VORP or if there is a point of diminishing returns.
The number of free throw attempts per game are much more predictive for point guards, shooting guards, and small forwards (0.24 correlation between free throw attempts and VORP). The ability of guards and small forwards to get to the line by driving to the rim and creating contact is a valuable asset at the NBA level, particularly as teams move towards “small-ball” lineups.
However, free throw attempts per game are poor predictors of power forwards and centers’ future VORP (0.07 correlation). Players filling these positions tend to have bigger bodies and are less mobile, making drawing contact around the rim less of an acquired skill than more mobile guards and small forwards. Among the current NBA power forwards and centers in the model, Michael Beasley, Derrick Williams, and Kris Humphries made the most trips to the line in college but averaged just -1.33 VORP over their first three seasons.
Figure 4: Top 20 NBA prospects with the highest three-year projected VORP.
It’s a bit surprising that Melo Trimble, a former point guard at Maryland, is projected to be the best player of the 2017 draft class with 6.21 VORP over his first three seasons in the NBA. In fact, the former Maryland point guard is a fringe NBA prospect according to the basketball website Draft Express, meaning that he may not be drafted!
Based on Trimble’s 15.9 points, 3.9 assists, and 5.8 free throw attempts per game college averages, it appears that Trimble has the potential carve out a career in the NBA. The Washington Wizards recently expressed interest in him as a late second round pick and other teams will likely target him as well.
In absolute terms, Trimble, Smith Jr., and Jordan Bell hold the loftiest projections averaging 5.20 VORP over their first three seasons according to the model. Other highly touted prospects include Ball, Fultz, and Jonathan Isaac, a balanced scorer and rebounder, who declared for the draft after his freshman season at Florida State.
Jordan Bell, a 6’7 power forward with a 38” max vertical jump, made a big leap in the model from 27th (late first round) to third overall. Other standouts in the model feature freshman point guards Derrick White (ranked 7th), who averaged 4.4 steals and 18.1 points per game during his college career, and Jawun Evans (9th), who recorded 5.5 assists per game. White and Evans are projected to land in the late first round of the draft per Draft Express.
The model overvalues shooting guard Sindarius Thornwell (48th to 11th overall), point guard Thomas Bryant (28th to 14th), and small forward Kyle Kuzma (36th to 17th). The trio is projected to land somewhere in the mid to late second round of the draft. Power forward Nigel Hayes (6thoverall in the model) and center Thomas Welsh 2.65 VORP (10th) are also overvalued. Both players are unranked by Draft Express but still might be drafted in the late second round. In addition, Malcolm Hill and Isaiah Briscoe, both borderline NBA picks, fall into the top twenty according to this model. On the flip side, Jayson Tatum, Josh Jackson, and Lauri Markkanen’s draft stock is downgraded from potential lottery picks to mid-first round selections.
That being said, the model projects Dennis Smith Jr., Lonzo Ball, Markelle Fultz, and Jonathan Isaac as lottery picks, which falls in agreement with most NBA forecasts across the Internet.
Figure 5: NBA prospect’s projected model ranking compared to his actual ranking according to Draft Express.
Dennis Smith Jr. (ranked 2nd) is perhaps the most dynamic player in the draft. The freshman point guard averaged 18.1 points per game and 6.2 assists at North Carolina State. The model projects Smith to accumulate 5.05 VORP over his first three seasons in the NBA. Lonzo Ball (4th) is another intriguing point guard in the draft. He averaged 14.8 points, 7.6 assists, and 6 rebounds per game as a freshman at UCLA. As the projected top pick, Markelle Fultz averaged the most points per game (23.2) of any player in the draft.
Figure 6: The 20th through 40th best NBA prospects in the model (three-year projected VORP).
Among the mid-tier NBA prospects, P.J. Dozier (39th to 21st), Edmond Sumner (35th to 22nd), Frank Mason (40th to 25th), Nigel Williams-Goss (49th to 29th), and Wesley Iwundu (47th to 31st) were significantly propped-up by the model. Dozier, a shooting guard, averaged 10.4 points and 1.3 steals in 70 games with South Carolina and Mason, a senior point guard, averaged four assists, 1.1 steals, and 1.2 blocks per game at Kansas and boasted a 41” maximum vertical jump (the eighth most predictive variable), which was third in the draft class trailing only Devin Robinson and Frank Jackson.
On the other hand, De’ Aaron Fox, T.J. Leaf, Zach Collins, and OG Anunoby’s draft stock fell by an average of 21.5 spots from their Draft Express rankings. For example, Fox, ranked 28th in the model, will likely be a lottery pick.
Some mid-tier prospects made the jump from fringe draft picks to late first or early second round picks in the model. Take Omer Yurtseven and Kennedy Meeks, who are ranked 23rd and 24threspectively in the model and aren’t even projected to be draft picks! The same goes for Chris Boucher (34th), Andrew Jones (35th), Rawle Alkins (40th), and Luke Kornet (41st).
Late selections in the draft
The model projects Monte Morris, a senior point guard from Iowa State, to be the best “late-selection” in the draft. Morris was an elite ball handler in college averaging 5.5 assists and 1.6 steals per game, which can make him a valuable asset for any NBA team.
Figure 7: The 41st through 60th best NBA prospects in the model (three-year projected VORP).
Among the late round selections that were snubbed in the model includes Donovan Mitchell (11thto 46th), Luke Kennard (15th to 49th), John Collins (17th to 52nd), Malik Monk (6th to 56th), Tony Bradley (34th to 57th), Caleb Swanigan (32nd to 58th), Jonathan Motley (29th to 59th) and Harry Giles (22nd to 60th) are undervalued by at least 20 picks. Meanwhile, Davon Reed, Damyean Dotson, Jamel Artis, and Isaiah Hicks, fringe draft prospects, landed among the 40th to 60th best NBA prospects in the model.
Falling outside of the top sixty college prospects are Devin Robinson, Justin Jackson, Cameron Oliver, Dwayne Bacon, Tyler Dorsey, and D.J. Wilson. This group’s draft position fell by an average of 29.5 spots. They are projected to average -0.73 VORP over their first three seasons in the league.
The model overall outperforms the top fifteen prospects in our dataset in 2012 and 2013 drafts (in terms of VORP) by +15.2 and +8.0 VORP respectively. As noted, however, this model doesnot include every college prospect drafted between 2000 through 2016.
Other factors including injury history and even player psychology can influence future success in the NBA. In addition, the model gives an incomplete picture of a player’s defensive value in college. VORP tends to overvalue steals and blocks and under-rates defensive specialists that do not generate a ton of steals and blocks. Measuring the number of contested shots and even loose ball recoveries might give us more accurate defensive projections. Nonetheless, the model shows that analytics are an important component of talent evaluation and should continue to play a larger role in the NBA draft in the coming years.
 Trimble sat out his senior year at Maryland and decided to work out for the draft on his own.