Layne Vashro’s Draft Projections Tool

Nov 26, 2014; Durham, NC, USA; Duke Blue Devils forward Amile Jefferson (21) and center Jahlil Okafor (15) push the ball up the court against the Furman Paladins in their game at Cameron Indoor Stadium. Mandatory Credit: Mark Dolejs-USA TODAY Sports

I build models to predict the future NBA impact of current NCAA or international draft prospects. This tool allows you to navigate the draft projections of these models, both for the current 2015 draft class as well as historical players going back to 1990.

I have three datasets to choose from. The “Draft” dataset looks only at actually drafted players, or players in the current Draftexpress Mock Draft and gives a projection that aggregates across their pre-draft career.  The “NCAA” and “Euro” datasets include a more complete list of prospects and lists each season separately.

Once you have selected a dataset use the season slider to highlight the season(s) you are interested in.  You can then use the “At least X minutes played” box to filter noisy cases.  You can then click the separate columns to sort the list, or simply type the name of a specific player/team/league into the search box.

The two models currently included in this data are my “Estimated Wins Peak” (EWP) and “Humble” (HUM) models (more will be added later).

“Expected Wins Peak” is a quasi-poisson regression model used to project how well a player will perform at the peak of his NBA career. NCAA and international prospects are projected using two similar but distinct models. The NCAA model uses age, experience, MPG, SOS, Height, Weight, and basic per-possession box-score statistics fit from all players to play in both the NCAA and NBA since 1990. I have outlined the international model in detail in an article series (here, here, and here).

The values in the “EWP” column give the number of “wins” each player is projected to be worth at his peak. Wins are defined by a combined two-year average of Win Shares and RAPM-wins.

The “Humble” model has the same ultimate goal as the EWP model. However, it differs in two significant ways. First, the Humble model uses a randomForest machine-learning process to fit the value of each variable instead of classic linear regression. Second, the Humble model includes a transformed variable that captures consensus scouting opinion of each player using actual draft order and popular draft big-boards. The Humble model is currently only calculated for NCAA prospects.

The values in the “HUM” column should be interpreted in the same way as the EWP column.

Obviously these projections are not 100% accurate, but both models have proven a very useful tools for evaluating prospects. In fact, out-of-sample retrodictions independently do a better job of identifying talent than actual NBA front offices over the past decade. That is great, but I still recommend using these to inform a more complete prospect evaluation rather than treating outputs as a single deterministic value.

To aid in that process I recommend several other approaches and sources of information:

1) Identify different player types and how they translate between leagues. Simple statistical models are likely to overlook the fact that some player-types typically do not work in the NBA even if they put up aggregate numbers that look good. I recommend using my prospect comparison tool, and NCAA-to-NBA correlation tool to help identify potential problems of this type.

2) Look for system effects. Offensive and defensive systems vary dramatically across NCAA teams. The models above assume a steal/block/rebound in Bob Knight’s system is the same as one in Jim Boeheim’s system, but it seems obvious that players in a strict man-to-man and 2-3 zone defense are facing at least a slightly different challenge.

It is important to try to account for these types of system issues and I offer some tools to help with that. Use the “By-coach NCAA-to-NBA correlation tool” to highlight players under different coaches in order to assess how different systems impact their translation between leagues.  Furthermore, tracking system effects may require moving beyond individual coaches. I also offer a coaching network tool that will make it easier to identify related coaches who may support the patterns you find.

3)  Read/watch scouting reports.  My models do an excellent job of optimally using basic information, but they are necessarily limited in the resolution they work from.  Stripping away context and detail benefits from aggregation, but it loses much in the process. There are also tons of subtleties like defensive rotations and screens that are completely ignored or at best indirectly captured in my models. Use available scouting tools like Draftexpress.com’s excellent video breakdowns to flesh out your prospect evaluation.

4) Watch games.