vjl110's Draft Data: Comparisons, Expected Peak Values, Star-Starter-Bust Likelihoods

JUNE 10th EDIT: Moved this back to the cover because of the new chart at the bottom of the article, that compiles all of the data and shows a few changes in the model. Go C.J. McCollum!-- EiM.

2013 NBA Draft Summary:

This post is a collection of my different draft projection models/tools and their results for the 2013 draft class. I wanted to put them all in one place where I can reference and update them in the future. Below we have two "Expected Win-Peak" models, one using a deep dataset but lacking in combine data, appreciation for the distinction between offensive and defensive rebounds, and subject to the faults of Win Shares as a catch-all player evaluation metric, and another model that includes more detailed information and uses an RAPM-based evaluation metric but only draws from players in this millennium. There is also another projection model that focuses on 'likelihood of becoming a bust/bench/starter/star" rather than projecting exact win performance. Finally, there are two different comparison model results. One looks only at the physical profile of prospects and compares them to well-known players, while the other draws on all of the available information to identify the best player-comparisons in the past 30 years.

If you want to jump right into the numbers, here is a link to the google spreadsheet with results for NBA draft prospects from 1981 to 2014.

Expected Win-Peak Models:

These models use historical data to predict how many wins a prospect will be responsible for during his best season. The methodology is based on multilinear regression. These models identify how things like college production, size, age, and level of competition helps explain the actual NBA performance of past prospects, they then use this knowledge to generate predictions for future NBA players based on how they look across those measures.

Winshare model (1981 to 2012)
The first model uses individual collegiate box-score statistics (points, rebounds, steals...), height, weight, age, and competitive context (team's rating and strength of schedule as well as era in which players played) to predict NBA "win shares" during a player's season of peak performance. This dataset draws information from players going all the way back to 1981.

RAPM model (2002 to 2012)
The second model uses individual collegiate box-score statistics (points, rebounds, steals...), measurements collected at the NBA pre-draft combine (standing reach, vertical jump...), age, and competitive context (team's rating and strength of schedule) to predict NBA "RAPM wins above replacement" during a player's season of peak performance. This dataset includes information starting with players selected in the 2002 draft.

Bust or Star? Model

The "Bust or Star? model" is constructed from the same 1981 to 2012 dataset as the "Win Share model" above, but rather than trying to estimate a precise "peak win production" it estimates the likelihood that a player will become either a "bust", a 'bench-warmer", a "starter", or a "star".

Player Comparisons

These models use statistical information to find which NBA players make the best comparisons for each prospect. I recommend you use these to help form a picture of the player in your head and maybe help develop expectations for what the prospect needs to do in order to suceed and what paths may lead towards failure.

Statistical comparison:
This model uses the same 1981 to 2012 dataset employed in the models above. It finds the players who's collegiate box score stats, size, competitive context, and expected NBA performance were on average the most similar to each prospect. Comparisons are listed with the most similar first and the eighth most similar last.

Physical Comparisons

These comparisons use only draft combine measurements to find the NBA players with the most similar physical profile to each prospect. These comparisons do not care how the prospect actually puts those physical tools to use relative to possible comparisons, but only how his size, length, speed, agility, and jumping ability compare to NBA players. The chart below lists the closest physical comparison for each prospect and then gives two additional measures "SIZE" and "ATHL" to help expand on that comparison. The numbers in the "SIZE" and "ATHL" give the average standard deviation difference between the prospect and the comparison in terms of size (height, wingspan, reach, weight, and bench) and athleticism (agility, sprint, and vertical) respectively. This allows you to say for instance "Ben McLemore looks physically similar to what Brandon Roy did coming out of college, but he isn't quite as big (-0.28 of a standard deviation smaller) or athletic (-0.35 of a standard deviation less athletic)".

[Note: Players who did not participate in the pre-draft combine do not have the necessary data to compute physical comparisons so there comps are left blank until those measurements become available]

Full Data:

Same numbers as above, but lumped together in one friendly image file:

To help contextualize the model results here is a link to the google document with numbers for all players in my data.