I have admired Dalt’s models for years and wanted to wait until after his was posted to post mine. As with all models, there are certain biases. But I believe this is interesting not just for the rankings, but for looking at what statistics the model felt were influential and furthering the conversation.
For some 101 on what these models are, please see part 1 here:
In part 3 of the series, I will rank the prospects in tiers and wrap everything up.
Without further ado, let me present the rankings. Note these are for college players only. I am keeping this write up very non-technical. I want to present the rankings and talk a bit about what statistics were meaningful and how it might apply to the Kings. I am happy to answer any more technical questions in the comments.
To make the ratings parallel with Dalt’s in terms of interpretability, here’s the scale that generally reflects the players ceilings in that range.
11.5 and up: Superstar potential. (Note: the actual number doesn’t seem to matter, but appears to be the computer putting players into the same distribution as win scores were in the training set)
7.7-11.4: Solid starter with all star potential
5.8-7.1: Mostly 6th man / solid starter potential
5.0-5.8: Solid bench contributor
Below 5.0: Lucky to be in the league
For reference here are the rankings for the past few years:
In general the system does well. Like any model, it has it’s fair share of misses. Some which seem unique to this model (e.g. John Wall rates a tier lower than he should) and some players that seem to present problems for almost any model I have seen (e.g. Lillard and Noah). Likewise it may seem odd to have Irving as the top rated player, but he only played 11 games at Duke, most of which were non-conference games early in the season. Had he played a full NCAA season, including conference games against tougher competition, he would likely still be highly ranked, but would be much lower than he is here.
So what makes this model tick? How does it relate to the Kings and add to the conversation? Here’s the heart of the beast.
One note is that I was actually torn between two models. I went with the simpler one because when in doubt, you want to pick the simpler one unless there is a compelling reason not to. While there were a couple of differences between them, what was most compelling was the similarities.
One thing that makes my model unique and where I want to start the conversation is that I tried to add scouting insights to the statistics. I combed through Chad Ford’s ESPN draft profiles, which date back to 2005, and coded in a number of scouting variables to try to capture information such as does the player have character red flags, does he have a high basketball IQ, etc. Two variables ended up mattering.
First and perhaps the most important observation I can make in this whole post is that the most important scouting variable, which the computer consistently picked as one of the most important variables was whether the scouts said the player was a "high IQ basketball player." Now let this sink in for a moment. Ideally, something we find in a model fits with our intuition about how the world works. Obviously, we know the number one thing a basketball team needs is a superstar. If you can get Shaq, Duncan, James, etc. you are going to be a great team. However, thing about the teams we have seen compete who did not have a true superstar (or have had them past their prime). The Spurs (they do not have a prime superstar anymore, but continue to win), Warriors (perhaps Curry now, but bear with me), Hawks, Rockets, etc. Not every player is a rocket scientist, but they have specifically targeted high IQ players. The Warriors have drafted Curry, Klay, Barnes and Green – all guys notes as high IQ. It’s one reason teams prefer veteran players. Conversely look at teams like us and the Pistons, who have amassed great talent, but have looked lost on the court. Now you still need to consider draft tiers when selecting a pick. Clearly, a team would want to roll the dice on Wiggins over Kyle Anderson. But don’t be surprised when Anderson is a key cog for the Spurs, Warriors or Hawks in a few years. When the high iq stat was combined with steals per 40, it became an even better predictor and drives a lot of what you see above.
The other important scouting variable was "elite defender," but only when it was combined with a more heady variable like A/T ratio. Basically, great defenders with some court awareness were much more likely to become good NBA players. This certainly helped Aaron Gordon, who is widely recognized as a great defensive player and is a very good passer, but who doesn’t rack of the steals and blocks in Arizona’s more conservative defense. One weakness of this system is it does rely on ESPN’s scouting. They did not call Gordon a high IQ player, if they did, I ran an alternate version of him and he would be at the top of the class.
Speaking of Gordon, another statistic that helps him is his age. Age is important in any model, because players tend to get exponentially better (or at least the good ones do) in their late teens and early 20s. This shouldn’t be a surprise to anyone, but it really helps Gordon, who is the youngest player in the draft and who doesn’t turn 19 until September.
I am happy to answer any questions in the comments. But to briefly run through the other important statistics, they were Assists per 40 combined with points per 40. This explains why guys like Ennis are lower in this projection. It rewards great passers who can also score. Although, going back to the scouting note from earlier, Ford did not call Ennis a high IQ player. This surprised me. But part of having a system is that I don’t get to randomly inject my opinion. However, an alternate Ennis I tested with the high IQ variable shot back up to the top 5.
There were also a couple of meaningful statistics that were controlled for position. That may sound complicated, but essentially a metric like offensive rebounds per 40 is important regardless of what position a player plays (e.g. oboards are important for PGs and PFs). However, you can’t compare the number of oboards a PG gets to the number a PF gets. So you compare the player to his peer group. Once you are comparing apples to apples, then the statistic becomes very meaningful and leads to better results than if you are just comparing oboards across all players. There were also two important combination statistics that were controlled by position. The first was a combination of blocks, steals and points. And the second looked at turnovers by ended possessions (e.g. shots + assists + turnovers). So basically defensive ability combined with the ability to score on offense and how well a player took care of the ball.
Finally, in the alternate, model that was narrowly cut in favor of the more simple one, true shooting percentage, strength of schedule free throws attempted per field goal attempted and minutes player compared to others at that position (important as guards tend to play more minutes than bigs) were all important.
As a final comparison, here is my player ranking compared to a few other online models:
(Note: My rankings and VJL’s only have college prospects, the others include international)
I hope all of this is useful to the community, I look forward to the discussion below.
Update 1: Added in pictures since tables did not paste cleanly and the preview function wasn't working.
Update 2: Noticed my '10-'13 projections had gotten truncated so there was no Anthony Davis and some players who withdrew from the '13 draft were still listed. Both have been fixed.