Counting the Baskets

The Benefit of the Doubt: Final Scouting Informed Model

On twitter the other day I noted that statistical analysis is the worst way to select draft picks, except all of the others.  Actually though, best practices would be to combine scouting informaton and analytic analysis.

There are a number of things scouts and video reviewers are able to see that stats can not.  For example, for lack of a better word, character is a tremendous variable in a young player's development.  Some of that shows up through the stat sheet, poor decision makers will have worse shot selection and more turnovers, and hustle stats like boards and steals show hustle. But, many apects of ability and willingness to learn or put in gym time or deal with adversity are less easy to infer.  

It is well known that player stats are less stable when they change teams, with different teammates around them they get different opportunities and fill different rolls.  Taking situation into account can help adjust the numbers players put up, at least at the margins. (It is dangerous over imagine how a player would blossom in a different environment.

On the other hand, as I noted before, it is not difficult to build a model that beats the experts. In and out of sample. 

Essentially, the difficulty, I suspect, is not collecting the information but in weighing it properly and perhaps avoiding information cascades or falling for particular players early in the process. Since I am writing this as the Nets/Kidd meltdown is taking place, I should note that some teams seem to have less developed processes than others, or none, it would seem.

Following a clever idea by Layne Vashro and his Humble model, I tested actual draft place with my model both to see if it added independent value not seen by the stats and to see how it change the effects of some of the stats.  

This is the second run of what I call the Scouting Informed Model, using the data used to train the P-AWS 2.0 model.  It's also it is the first that applies the actual draft order to the 2014 draft class rather than using the Draft Express rankings as a proxy.  

Essentially the model uses the historic value of the pick number balanced against the player's statistical projection to arrive at a new value.   The good and bad of this method is that each front office is assumed to be equal and each draft class is assumed to be average.  

I have also looked at Chad Ford's Draft Tier rankings, which potentially cold be better as they use a consensus of NBA team front offices and allow them to rank prospects on a historical scale, so there is not the same number of say first tier players every year.  Unfortunately, there aren't enough years to properly test that theory yet.

Alternatively access to the major pieces of information historically rated in scouting, whether character issues or pick and roll defense would probably help the model more than an overall rating.  With the scouting information deconstructed the more independently valuable pieces could be emphasized while those that show less independent value would be excluded.

But, even so, draft position in model testing retains some independent ability to improve the model's results.  All things considered, this is likely to be the most accurate of my current models, though after rookie year my research indicates that draft position loses most of its predictive power once rookie playing time and production is considered.

 To get to the actual players, here's how the first round looks.  The table is ordered by draft order, with the Scouting Informed AWS Model (SI-AWS) score shown, SI-AWS rank and the Reach/Steal rating, which is simply the difference between where they were drafted and the model rank (including the knowledge of where they were ranked). Reaches are negative numbers representing players possibly drafted too high and Steals are positive are more likely to perform better than the typical result for their draft spot.

 

Name	Team	LEAGUE	Drafted By	Draft Position	Scouting Informed AWS	Scouting Informed Rank	Reach/Steal
Andrew Wiggins	Kansas	NCAA	CLE	1	6.12	8	-7
Jabari Parker	Duke	NCAA	MIL	2	6.98	1	1
Joel Embiid	Kansas	NCAA	PHI	3	6.39	3	-
Aaron Gordon	Arizona	NCAA	ORL	4	6.59	2	2
Dante Exum			UTA	5		NA	-
Marcus Smart	Oklahoma State	NCAA	BOS	6	6.15	6	-
Julius Randle	Kentucky	NCAA	LAL	7	5.47	12	-5
Nik Stauskas	Michigan	NCAA	SAC	8	4.84	21	-13
Noah Vonleh	Indiana	NCAA	CHA	9	6.27	4	5
Elfrid Payton	La Lafayette	NCAA	ORL	10	5.26	15	-5
Doug McDermott	Creighton	NCAA	CHI	11	5.05	19	-8
Dario Saric	Cibona Zagreb	Adriatic	PHI	12	5.36	14	-2
Zach LaVine	UCLA	NCAA	MIN	13	4.42	24	-11
T.J. Warren	N.C. State	NCAA	PHO	14	4.88	20	-6
Adreian Payne	Michigan State	NCAA	ATL	15	3.35	39	-24
Jusuf Nurkic	Cedevita	Adriatic	DEN	16	5.42	13	3
James Young	Kentucky	NCAA	BOS	17	4.41	25	-8
Tyler Ennis	Syracuse	NCAA	PHO	18	5.79	9	9
Gary Harris	Michigan State	NCAA	DEN	19	5.17	16	3
Bruno Caboclo		Brazil	TOR	20		NA
Mitch McGary	Michigan	NCAA	OKC	21	4.11	27	-6
Jordan Adams	UCLA	NCAA	MEM	22	6.27	5	17
Rodney Hood	Duke	NCAA	UTA	23	3.03	46	-23
Shabazz Napier	Connecticut	NCAA	MIA	24	4.25	26	-2
Clint Capela	Chalon	French	HOU	25	6.12	7	18
P.J. Hairston	North Carolina	NCAA	CHA	26	3.64	33	-7
Bogdan Bogdanovic	Partizan	Adriatic	PHO	27	2.92	50	-23
C.J. Wilcox	Washington	NCAA	LAC	28	2.51	66	-38
Josh Huestis	Stanford	NCAA	OKC	29	3.12	44	-15
Kyle Anderson	UCLA	NCAA	SAS	30	5.74	11	19

 To look at some of the highlights

• The Scouting Informed Model gives considerable deference to the top picks, otherwise Wiggins might show up as more of a Reach.
• The biggest Steals in the first round go to Memphis for Jordan Adams, Houston for Clint Capela and San Antonio for Kyle Anderson.  It's not a coincidence that these are three of the most analytically inclined front offices around.
• The biggest Reaches were C.J. Wilcox, Rodney Hood,  Adriean Payne, Nik Stauskas and Zach LaVine.
• Rodney Hood is kind of the classic reach candidate because of his scoring and size for position
• Adriean Payne is one of the players the model has consistently disagreed with the scouting.  I sort of suspect his game may translate better than the model projects, but at 15 I think there was more talent on the board than could justify grabbing him there.
• Stauskas and LaVine look too one dimensional to grab there. 

In the second round I think of reaches as being something less of an issue, the model value between te 50th and 70th pick is less dramatic than the difference between the 5th and 25th.  However, there are some potential steals that should be highlighted, basically getting a first round level prospect in the the second.

Not surprisingly the model had the most disagreement with the draft positions in the second round, one because the value of scouting information is so minor given the historic lack of value, especially in the last part of the 2nd.  In addition, with the level of difference between prospects relatively small it is easy for team evaluation and the model's evaluation to diverge.

 That said, there were some real value player on the board who I think have a chance to be the next Chandler Parsons or Paul Milsap type players.

 

Name	Team	LEAGUE	Drafted By	Draft Position	Scouting Informed AWS	Scouting Informed Rank	Reach/Steal
Damien Inglis	Roanne	French	MIL	31	5.11	17	14
K.J. McDaniels	Clemson	NCAA	PHI	32	4.52	23	9
Joe Harris	Virginia	NCAA	CLE	33	2.59	62	-29
Cleanthony Early	Wichita State	NCAA	NYK	34	2.98	49	-15
Jarnell Stokes	Tennessee	NCAA	MEM	35	5.06	18	17
Johnny O'Bryant	LSU	NCAA	MIL	36	1.76	82	-46
DeAndre Daniels	Connecticut	NCAA	TOR	37	2.37	70	-33
Spencer Dinwiddie	Colorado	NCAA	DET	38	3.89	31	7
Jerami Grant	Syracuse	NCAA	PHI	39	4.04	29	10
Glenn Robinson	Michigan	NCAA	MIN	40	4.06	28	12
Nikola Jokic	Mega Vizura	Adriatic	DEN	41	5.76	10	31
Nick Johnson	Arizona	NCAA	HOU	42	3.37	37	5
Walter Tavares	Gran Canaria	ACB	ATL	43	3.95	30	13
Markel Brown	Oklahoma State	NCAA	MIN	44	2.87	52	-8
Dwight Powell	Stanford	NCAA	NOL	45	2.1	75	-30
Jordan Clarkson	Missouri	NCAA	WAS	46	1.98	77	-31
Russ Smith	Louisville	NCAA	PHI	47	2.84	54	-7
Lamar Patterson	Pittsburgh	NCAA	MIL	48	3.46	35	13
Cameron Bairstow	New Mexico	NCAA	CHI	49	2.65	59	-10
Thanasis Antetokounmpo	Delaware	D-League	NYK	51	2.73	55	-4
Vasilije Micic	Mega Vizura	Adriatic	PHI	52	3.64	34	18
Alessandro Gentile	Milano	Italy	HOU	53	2.63	60	-7
Nemanja Dangubic	Adriatic	Mega Vizura	PHI	54	1.92	79	-25
Semaj Christon	Xavier	NCAA	MIA	55	1.82	81	-26
Devyn Marble	Iowa	NCAA	ORL	56	3.02	47	9
Jordan McRae	Tennessee	NCAA	SAS	58	1.49	85	-27
Xavier Thames	San Diego State	NCAA	TOR	59	2.24	74	-15
Cory Jefferson	Baylor	NCAA	SAS	60	2.89	51	9

Memphis and Denver probably have the steals of the draft, according to the model.  Memphis has become one of the most analytical teams under John Hollinger, so their inclusion is no surprise.  Their pick at number 35, Jarnell Stokes put up great numbers and is a physical load, which makes it surprising to me he was still around.

Denver picked up Nikola Jokic, a very young Center out of Mega Vizura.  Jokic was specifically made for the P-AWS model as a passing big man with three point range.  There are issues with lateral speed on defense, as well as a lot of development.  But, Jokic is a promising young player, in fact the model likes him more than Denver's frist round pick Jusuf Nurkic, though Jokic is further away from having an impact.  The influence of the Nugget's head European scout, Rafal Juc, can probably be seen in their selection of two promising young Euro centers.

Milwuakee gets a good overall grade.  Parker was a good choice and I love the Damien Inglis pick, even if the fit behind Parker, the Greek Freak and Middleton isn't ideal. Johny O'Bryant isnt rated as a good prospect by the model, especially with players like Dinwiddle, Jerami Grant and Glenn Robinson still on the board.  

Speaking of Robinson, I put it at essentially 50/50 as to whether he or Minnesota first round pick Zach LaVine turn into better players. 

Statistical Meet and Greet for the 2014 Celtic Rookies

As a follow up to the models I have been creating for the draft I made a data visualization of the prospects posted at Hardwood Paroxysm before draft night.  Now that the teams have made their selection I have a new one focusing on the Celtics new rookies.  

The visualization uses Tableau's "Story" format, which is basically an interactive data focused Power Point.  You can either click through the slides with my filters applied to see some of the data that jumps out for me from their seasons last year, or change the filters around to see their stats compared to different groups of players.

The first group of data slides use standardized data for different classes of data like steals and scoring so that the scale will be the same.   For example, you'll see that Marcus Smart's steals and blocks rating was 2.3 standard deviations over the average prospect, that's elite.  A zero is exactly average, while +1 is good in that category and -1 is not good overall.  However, some categories are fairly position dependent, like assists or rebounds, so I have the vizualizations grouped by position. Age is another occassionally underappreciated factor, so the main vizuations compare to players in the rough cohort of the Celticis draftees.

But, feel free to mix and match.

 

Learn About Tableau

 

Overall, I am more excited by Marcus Smart than James Young, when I look at these numbers.  It is unfortunate that his one obvious flaw jump shooting, overlaps with the team's current best player Rajon Rondo's  weakest point.  But his all around games is really strong, if he can improve the jumper he'll have a chance to be elite.

I rated Young as an end of the first round prospect.  But Young is really young, and he showed definite improvement as the year went on.  On offense, his catch and shoot jumper is solid (almost 90% of his threes were assisted acoording to Hoop-Math, but not fantastic at this point, but he pulls up for too many mid-range twos when driving (17.6 percent of non-transition two point jumpers were unassisted).  On the other hand his turnovers were low for his age and assists were decent for a wing player, especially a freshman.  But, the steals and blocks tell the same story as the scouting reports, his defense will need to improve to actually stay on the court.

Starter Odds Model 2.0

Last wek I revealed the final version of my base draft model, the P-AWS 2.0 model.  I am finally ready to put out the companion model the Starter Odds model, which is an ordinal logit model. Simply, the target for the model is a Y/N variable keyed on whether the player put up 'starter quality' numbers as measured by their alt. win score (AWS) efficiency.  

The determination of 'starter' level numbers is a bit subjective, but I determined it by looking at the numbers put up by solid starters in the NBA over the last two years.  The measure, with a replacement level prior applied, came out to a 5.4 AWS.  This is slightly higher than my previous estimate, so fewer prospects rate as future starters, which is more realistic given the number of slots available. 

The key is that the model is only able to tell that the player did or did not hit that level, not how much they missed the mark or exceeded it.  The model instead calculated the odds that the player will or will not hit that mark.  Of course, players with star types numbers end up being much more likely to at least hit that level, so the result is fairly linear with the P-AWS numbers.

The factors driving the Starter Odds model are very similar to the linear P-AWS model.  The biggest difference is that the age factor is less weighted in the Started Odds model relative to production efficiency.

That gives a slightly different rankings, but overall the models are giving well aligned signals on players.

• Clint Capela and Jordan Adams switch places at one and two.
• The model doesn't know, of course, about Joel Embiid's injuries outside of their effect on his minutes played.  He rises to fifth in this model.

Three Point Shooting Prospects: Expect Regression

The law of small numbers is well neigh inescapable, in any small sample expect regression to the long term mean.  In basketball social media circles this gets referenced as 'small sample size theater,' as a winking acknowledgement of the overwhelming urge to make too large conclusions from a couple of games even while doing exactly that.

It's most obvious in basketball in three point shooting, that most variable offensive statistic.  

For draft purposes that means that anytime we reference a shooter's three point percentage we have to keep in mind their number of attempts as well.  Simply, the more attempts they take the more sure we can be of their true talent shooting the three.

In addition, looking at the NBA I have that consistently more attempts at the three is a sign, in and of itself, that the player can make threes.  I also, by the way, found the same thing for mid-range shots.

With that in mind I looked at the attempts and efficiency of draft prospects, and plotted a trend line that gives a very rough idea of the regression to expect. The more attempts a player has the closer their long term mean is to their current percentage and the higher the expected mean they're regressing towards. (1)

The visualization below is the number of three point attempts (approximately, actually, as it's using pace adjusted numbers) and the accuracy of prospects in this year's draft for any player with at least 10 attempts.

• Expect regression from Noah Vonleh and Kyle Anderson.
• Doug McDermott, Nik Stauskas and Shabazz Napier can shoot.  The issues with their games is in other areas though.
• Don't pay any attention to Bogdan Bogdanovich's field goal percentage, or anyone else's, he took alot of threes.  Many of them as pull ups due to his ball handling duties with Partizan.
• I included age as a color coding because in addition to regressing, young players develop and the jump shot is one of the more developable skills through gym work and reps. (Though not easily and not always).

Learn About Tableau

 

 

 

 

 

1. Regression towards the mean does not imply regression to the population mean. Based on the variance of three point shooting in the NBA I estimated that a decent formula for estimating a player's 'true' three point shooting prowess is their current average times actual attempts plus their expected mean times 150 attempts, then divided by actual attempts plus 150.  So a player shooting 40% on 150 attempts with an expected regression to 34% would, have an expected long term mean of 37%.

Draft Model Comparisons- Acronym Convergence

I finally broke down and got ESPN Insider, just for the month, because I wanted to see if I could use Chad Ford's tier system as something of a better approximation of scouting information than the actual draft order.  I had hoped that the tier system would leverage the 'wisdom of crowds' a bit and take advantage of the ability to rate players in a longer term perspective (there's a number one pick every year no matter how strong or weak the class), would make it more informative.  

Unfortunately, the tier system doesn't go back as far as I had hoped, and what rankings there were appeared to be no more or less informative than actual pick selection. Oh, well...

But, it did give me a chance to look at Kevin Pelton's draft rankings and compare them to mine. Pelton's system is based on his own WARP player rankings. This was particularly interesting because Pelton also rate international players.   With a quick copy and paste I grabbed his rankings and compared them with P-AWS.  They turned out to be pretty highly correlated, with an almost 90% correlation for an R² of .79.  

 

The two biggest outliers were PJ Hairston, who was 1.27 standard deviations higher in WARP than in my ranking and Jerami Grant, who was 1.07 standard deviations lower in WARP.  As far as Hairston goes, Thanasis Antetokoumpo, another D-League player in the draft, is also ranked higher by WARP.  So, there may be a systematic difference in how we are weighting D-League experience.  

I also decided to look at my rankings compared to Layne Vashro's rankings based on win shares, Expected Win Peak (EWP).  There are a slightly higher, though not in any meaningfully so, correlation between P-AWS and EWP, with a correlation of .90 for an R² of .81.  There the biggest divergence was on Embiid, with EWP much higher on him and Doug McDermott, who P-AWS ranked higher.

And, for good measure, I compared WARP and EWP.  Again there was a fairly high correlation, though a little lower than either had with P-AWS, .83 for an R² of .695.  There the biggest outliers were Kyle Anderson, with EWP being much higher and, again, McDermott was an outlier with WARP being much higher on him than EWP.

 One additional note is that none of the three models ranked exactly the same players, so I used whichever players were in any two result sets.

Update:

Because much of the value in the draft is at the top of the draft, I looked at the number of top ten prospects shared by each projection model combo.

• EWP and WARP shared 7 of the top 10
• P-AWS and WARP shared 8 of the top 10
• P-AWS and EWP also shared 8 of the top 10 prospects in common, both having Aaron Gordon, Jarari Parker, Joel Embiid, Jordan Adams, Kyle Anderson, Marcus Smart, Noah Vonleh and Tyler Ennis.  Though, my total top ten includes a few international players, namely, Clint Capela and Nikola Jokic.  

Then I ran one more comparison between the three metrics, that paints a slightly different picture of the amount of agreement between the metrics.  I looked at the average absolute difference (ie regardless of plus or minus) in rank of the shared prospects with each metric.

• Once again WARP and P-AWS were the closest, the average difference in rank was 7.5 and the median was only 5 places.  Because both Pelton's WARP and my P-AWS model rank international and D-League prospects they rated 79 of the same potential draftees.  Meaning that the average difference was 9.6% of the list and the median difference was 6.3% of the list.  The average can also be calculated as a coefficient of dispersion as 19.3%.
• Between P-AWS and EWP the average difference in rank was also 7.5 and the median was 6, but since they only ranked 61 of the same prospects, the average difference was 12.3% while the median was 9.8% of the list.   
• Between EWP and WARP the average absolute difference in rank was 8 and the median difference was also 6 places.  With 60 prospects ranked in common the average difference was 13.4% and the median difference was 10% of the list.

 Related articles

Re-Drafting 2012

Pelton: Projecting top 30 NBA prospects

P-AWS Draft Model Two Point O

Related articles

P-AWS Draft Model Two Point O

This should be my final update of my base draft model, the Projected Alternate Win Score (AWS) model (P-AWS) before this year's draft.   This one represents a modest change to model, which I will detail, as well as a completion of nearly all of the player's data.

In terms of possible draft prospects, I think Alessandro Gentile is the only one still playing as Olimpia Milano plays for the Italian championship.  (Gentile is on the bubble of being draftable, according to my model moving up from 77th to 63rd with a good run at the end of the season in Euro League and the Italian Playoffs. So these are big minutes for him!)

To review, breifly, my draft model is built off of a statistical analysis using AWS as the measure of efficiency at the end of their rookie contract, with a replacement level reversion included in order to stabilize the measure and account for the fact that better players tend to get played more. However, the value of playing time is not linear, a player putting in 30 minutes a night shouldn't be assumed to be twice as good as some one playing 15 minutes.  Something I looked at in a post called Time is a Flat-ish Curve.  It helps avoid what I call the Sebastian Telfair fallacy, over crediting an inefficient player who happens to get a good deal of time on the court.

I used pace adjusted data from Draft Express, in both the training set and applying it the current class.  I also use Draft Express's Top 100 as a filter for applying the model to this class, since that more closely mimics the training set built off of actually drafted players.

There were a couple of changes applied to the P-AWS 2.0 model from the last model.  The biggest is the inclusion of the prior year's data, which is something that had been dropped from the original model during development.  However, I decided to look at it again during some other data cleaning.

The informational value of data tends to depreciate over time, how fast depending on the source and rapidity of change.  In basketball, for example, looking at Kevin Garnett's efficiency in 2007 is obviously much less predictive of his prowess now than last year.  For developing players there is also a depreciation in the predictive ability of prior years as the player develops.

In making predictions about team records last year, I found that performance one year removed was less predictive of performance than the immediate prior year, even if the player remained on the same team and an age factor was applied.  In a sub-study for the P-AWS 2.0 model I found that the difference was even bigger in pre-NBA stats.  But the earlier year does have some independent value, and including both years weighted towards the most recent year does improve the model's predictive ability in a robust number of filters tested.

One of the follow on changes resulting from the inclusion of prior year data is that age is slightly less weighted, as players in the development stage typically have worse numbers some of the skill development is captured more directly.

An example from this year shows the scale of this effect and the direction.  Kyle Anderson, is a favorite of the P-AWS 1.0 model, putting up fantastic numbers last year at UCLA in his sophmore year.  Anderson's numbers, in fact, jumped considerably from his freshman year from an 7.9 AWS to 10.8.  The freshman number was still a good one, a little above Andrew Wiggins number this year, but the jump is well above the  typical developmental improvement.  Consequently, Anderson drops from 2nd in the class projection to 7th.  

On the other hand, Adeian Payne moves up from 72nd to 56th, his numbers last year were good also, though not great, but his consistency improves his overall rating.

An interactive variable between rebounding and three point shooting was also added,which had only a limited effect but tested as significant and fit other studies I have done. 

Also, previously I had posted the P-AWS results alongside a Logit model version that estimated the odds of a player becoming a 'quality starter' player.  Right now, I am working through that model and I think it deserves its own post in any case, which I hope to do later in the week.

Finally, a couple of notes on the top prospects according to the model.

• Jordan Adams is the darling of a number of analytics models, in fact he's the third highest projected prospect in the last three years.  He's significantly behinf Nerlens Noel and Anthony Davis, but ahead of everyone else. But he is not a favorite of any of the scouts, due in part to limited athleticism.  So he will be an interesting case both in the draft and n his career.  If the Spurs snag him in the late first, you have been warned.
• Clint Capela's rating is right behind Adams, and the difference isn't really significant.  Capela also topped Kevin Pelton's international rankings at ESPN.  In contrast with Adams, Capela has athleticism and length to spare.  His strengths are just the kind of thing GM's seem to like, and his weaknesses, physical strength and mid-range touch, are fairly fixable as the 19 year old develops. I really can't see how he falls out of the lottery.
• Jabari Parker is the first of the traditional top three in the model.  Definitely looks like an "NBA ready game."
• Nikola Jokic is a bit of a surprise in the four spot, but he was almost made for the P-AWS 2.0 model, a young big that passes and shoots the three (though not as well as one would like).  There are concerns about his foot speed and defense, but he could be a sleeper high upside kind a kid.
• Noah Vonleh and Aaron Gordon are right around the spots they're expected to go in the draft.  Both very young players with good numbers.  Of the two I prefer Gordon because of his defensive potential.
• Kyle Anderson, as discussed above, drops not because of the 'who does he guard' question, rather because the model doesn't 'believe' that all of his increase in production is real.  Anderson's increase in three point range is most obvious case for skepticism, going from 21.1% to 48.3%.
• Skipping down to Joel Embiid, he comes in tenth in part because of his lack of minutes and because he is actually a fairly old freshman.  There is no factor on his late basketball start, which may or may not indicate he has more development potential than a typical player of his age.

Here's an image of the model results applied to this year's draft class, and the results in Google Docs:

 

This Summer's Worst NBA Mock Draft

So far, I have created two draft models, a linear model and logit model, an Update Model for rookies, combining the draft model and their rookie stats, and a Scouting Informed Draft model.  Now in a sure sign that I have been looking at this data set for too long, I have a statistics based Pick Prediction model.

The idea for this one is to simply use the same playing statistic data used in the draft models to predict where a player will land in the draft.  The model does fairly well, explaining 39% of the variance in selection place using a curved draft pick scale (reflecting the more significant differences between the top picks compared to the lower picks).   

While the end result is probably this year's worst NBA mock draft, the more interesting result is the differences between what the GM's appear to value and what P-AWS model indicates enhances performance in the NBA.  

Things GM's have historically 'Over Valued:'

• Scoring, this the least surprising finding ever, is about 15% more valuable in  Pick Prediction Model.
• Youth, in the Pick Prediction model age is an even larger factor than in the P-AWS model. There is some dual causation here since players that are highly thought of by GM's are more likely to enter the draft, and less sought after players are more likey to stay in college or European leagues.
• Blocks, slight surprise here, but the P-AWS model is less of a fan of blocks than many other draft models.  Also, when the Pick Predict model is applied to this year's top one hundred prospects, it is clear that big men have tended to valued more highly than their overall statistics may have warranted otherwise, particularly the 'rim protector' type bigs.

Things GM's have historically 'Under Valued:'

• Offensive Rebounds, the P-AWS model treats offensive rebounds as approximately twice as valuable as defensive rebounds, in line with the Alt Win Score model, as well as a good deal of other research.  The Pick Prediction model shows offensive rebounds as only about as valuable as defensive rebounds.  That's driven, in part, by examples of players like Paul Milsap, DeJuan Blair and Kenneth Faried falling further than their ultimate production indicates they should have.
• Turn Overs, basically the Pick Prediction model indicates that GM's have not cared a lick about turn overs.  Part of this is likely to be the preference for offensive stars that handle the ball all the time and part is the belief that turnovers are one of the more correctable elements of a young player's game.
• Assists, assists were consistently indicated as a positive sign for future success, particularly in front court players in all of the development of the P-AWS model.  In the Pick Prediction model, however, assists were seen as much less valuable.

In essence the Pick Prediction model is the reverse exercise of the Scouting Informed Model.  In the SI model I was trying use selection rank as a way to glean what independent information non stat sources could give me, though with a bit of a blunt instrument.  Here I am using that same blunt instrument to gain some inferences about how the stats I have been using have been weighed by NBA teams.  

Finally, the part where I put up a pretty conditionally formatted table for your viewing pleasure. The prediction model is likely no closer to what this year's draft order will look like than either the P-AWS model or Player Odds model, so this is probably the least accurate 'Mock Daft' you'll see this summer. 

This one is just for fun, so please, no wagering.

 

 

I guarantee that this is not how draft night will go.  But, don't be surprised to see some of the young Centers move up on draft night.

Scouting Informed Draft Model

For a while I have been playing around with a version of my draft model that includes the position the player was taken in the draft.  The idea is completely derivative of Layne Vashro's "Humble" version of hs Expected Win Peak model.  But it's a good idea, certainly as a test of the model and exploratory exercise.  

The are two advantages to this approach, looking at the change in the coefficients we can infer something about what scouts and teams are picking up that raw player stats are not seeing and we can get an estimate of the relative independent percentage contribution of the expert opinion in addition to the stats.

The big mover in terms of relative coefficient value between the P-AWS  model and the Scouting Informed model is the age adjustment.  That was somewhat expected given that the age adjustment in the P-AWS model consistently came in 10%-15% higher than the aging curve I developed looking at NCAA players.  In effect, because highly scouted players tend to enter the draft earlier, a portion of the age coefficient acts as a proxy for some of the things team scouting values.  In the Scouting Informed version, age falls to 10% below the age curve, indicating that there maybe a slight overvaluation of youth in actual drafts.  

The downside is that we don't actually know who will be drafted so can not actually adopt the same methodology with our projected draft class, but only approximate it.  Also, draft order itself only represents a blunt approximate of the scouting consensus with heavy doses of GM idiosyncrasies, 'fit' and need considerations and the like thrown into the mix.  (cough Anthony Bennett cough). Ideally one would want to have specific draft contemporary scouting analysis to feed into the model that could be tested independently and weighted in relavent pieces, for example, defensive ratings, character ratings, shooting mechanics, and so on, to maximize the information in the model.

But, absent that level of detail the lump sum rating still provides some information. One of the more valuable insights is the scale of difference between the pre stats P-AWS model and the Scouting Informed version.  

Here is an image of the table using Draft Express as the draft order proxy:

And here's the table in Google Doc's:  Table Link.

The biggest mover is Adreian Payne, who I like, but the P-AWS model doesn't.  In the last two out of sample drafts, older bigs have been the largest outliers, and when the model is separated into front and back court players the age curve is less strongly correlated to back court players.  So, I can imagine that Payne is a player the stats model underestimates.

Another big mover is Wiggins, though he still rates as only the eighth best player.

Stats model favorite Jordan Adams has a large relative fall, but still ranks in the top ten.

Draft Rank Update 6-3

I should have a good deal of, hopefully, substantive analysis on the draft coming up soon.  But, I realized that I don't think I have put up a full ranking using my two base models, P-AWS and the related Player Odds model, that gives the odds of the player developing into a 'quality starter' caliber player.  (One of the substantive things I intend to do is to write up an omnibus explanation of the models, but, for now if one is interested, this is the best post, or following along with the Draft Model tag).

Below are my rankings updated with the full season of data for all but a one or two European prospects still playing their league's play offs.  

 

The three biggest surprises to me, were how much the model liked Sim Bhullar and how much is did not like Adreian Paine.  With Paine, age is a big part of the issue, but I have to say that the biggest misses for the model out of sample have come with older centers, both of the Plumlees, Withey and Gorgui Dieng.  In fact, looking at that issue is one of the 'substantive' issues I hope to get to soon.