2015 NBA Draft

[nrestifo]

Haven't seen a 2015 draft thread yet (did i miss one?) so I figured I'd start one. Been thinking a lot about the draft with the tourney and all happening, and last night finished some work I've been doing on my draft model. You can check out the results here. https://docs.google.com/spreadsheets/d/ ... =561660716 Thoughts? Criticisms? Questions? Would love to start a dialogue and see some of your models' results.

[Crow]

File isn't stable on my phone so just caught short glimpses. Low rankings for some bigger names- Oubre, Lyles, Selden, LaVert, etc. Will look more later. Is retrodiction file available generally? Not to my click.

Biggest differences between your rankings, Layne Vashro's, Pelton, Ford, hoopsnerd, etc.? At some point a table would be helpful.

[Chris Hoffman]

Hey nice predictions on the 2015, what kind of analysis did you do to produce those numbers and what do they mean?

-chris

[nrestifo]

Crow,

Let me know if you can't see the retrodictions tab. Works fine for me. There is the occasional wonky high pick, but other than that, retrodictions look OK to me (theyre only back to 2002). Kyrie, KD, Beasley, Oden, and AD are 1-5.

Chris,

The numbers are based on a neural network and regression prediction of a normalized RAPM/WS blend. The RAPM/WS blend uses the 2 year peak of RAPM and WS for each player, just as in Layne's model. The numbers on that doc do not represent Wins or Plus Minus, rather they represent units of variance. KD has an Eval. of 6.534, so that can be interpreted that KD projected to be 6.534 StDeviations better than the average NCAA player since 2002. For this model I evaluated every prospect, not just the ones that were drafted or played in the NBA. For features I looked at Age, Height, HSRank, SOS, TS%, 3ptPts%, Ptsper40, Rbsper40, Astsper40, Stlsper40, Blksper40, and TOsper40. To chose the players that are in the 2015 Draft tab I made a chaid decision tree and neural network model that predicts whether a player will be drafted. If either of these model's predicted they would, I included them in the 2015 draft tab. I also included all players in the current 2015 Mock draft on DX, in case any of my models didn't think one of those players wouldn't be drafted. They only predicted a handful on the DX board wouldn't be. Personally I think the model overvalues things like PGs that play at big schools and undervalues uber-athletes (while stl rate considered very important here, doesn't seem to be as impt. as with other models i've seen) and Euros/Internationals in general. I'll continue to tinker with the methods right up to the draft, because as with any method, there are some flaws.

Based on what I see from this model and Layne's model and others, I do think people should be taking D'Angelo more seriously. This model is far from gospel, but it has D'Angelo at a Greg Oden level. Layne's has him right between Kyrie and Jason Kidd.

[Crow]

The retrodictions page partially loaded today (perhaps gave it a bit longer to load) but wouldn't scroll past row 62 (in the time I waited). But the issue must be with my phone.

[Statman]

Haven't seen a 2015 draft thread yet (did i miss one?) so I figured I'd start one. Been thinking a lot about the draft with the tourney and all happening, and last night finished some work I've been doing on my draft model. You can check out the results here. https://docs.google.com/spreadsheets/d/ ... =561660716 Thoughts? Criticisms? Questions? Would love to start a dialogue and see some of your models' results.

Just a heads up, you left your sorting formula on the rank (forgot to copy/paste data only?) - so if you sort by player name the rank has #REF!.

Of course, I love this stuff - I have purposely not allowed myself to run mine yet. Going to add this season's college to pro data when the NBA season ends to adjust the factors a little.

I'll probably send out stuff to all NBA & DL teams - see if I can't land a gig first - I'm done with retirement. If I don't, I expect I'll post all current projections and all retrodiction projections back to '97.

I don't believe my rankings will have Kaminsky so low - his overall college rating is just too good (across the board), despite not getting those big 19-22 yo NBA career curve upswings on his stats, he'll almost certainly project a bit above average NBA player next season (production wise) with 3/4 more upward trajectory years.

Few questions - you are using pace adjustment on your college data - right? That would partly explain Kaminsky so low if not. Obviously you are using SoS. Are you using weighted 2 seasons for non frosh?

[Statman]

File isn't stable on my phone so just caught short glimpses. Low rankings for some bigger names- Oubre, Lyles, Selden, LaVert, etc. Will look more later. Is retrodiction file available generally? Not to my click.

Biggest differences between your rankings, Layne Vashro's, Pelton, Ford, hoopsnerd, etc.? At some point a table would be helpful.

I doubt any projection model will like LaVert much, his college production just hasn't ever looked NBAish. Similar to Zach Lavine - except he's a little more productive, but older & coming off injury.

I'm certain someone will throw together a consensus this season - I saved one that Nick threw together last minute last year: http://hoopsnerd.com/?p=626 We were very twitter active right before the draft, and he was nice enough to try to pull it all together last minute.

I expect we'll have more draft models to add to the mix this year.

I'm also attempting to improve on mine - since I project every rating subset, I decided I will try to run similarity scores on the yearly PROJECTIONS of every past college player (NBA year by year based on age - so I can compare everyone at similar ages) - to see if that helps distinguish certain player types even more. I'll run the similarity score based on projections - but compile the actual career WAR results from those scores (including playoffs, we want guys that help teams win).

If I were working for a team - I'd add combine stuff (height, length, etc) to my work & similarity scores & lessen the noise of my outliers even more (my top rated none draftees are almost always extremely all around productive 6'7" PFs from mid majors). But I stubbornly won't do that yet, because I want to show how very viable my college ratings are, and how helpful they alone can be (looking at how they break down & including age) in better tabbing future productive NBA players, particularly in later picks.

Oh, yeah, I still have to work in international players. I really probably need an NBA gig for that, my methodology is GREAT for foreign leagues (my work adjusts for league statistical nuances) - but I refuse to take the time to try to compile ALL the data I need with so many other things I need to finish first. I'd expect an NBA team could get that data easily, save me loads of time.

I also like the idea of working in (weighted of course, smaller data) any international competition results I can between counties. Also, probably any high school summits & all star weeks in which I can get complete player stats. AAU is getting better at stat keeping, I expect I'll be doing work there too.

But, all in good time.

I think the Sixers (& their 6 picks) should send plane tickets & per diem to every draft model guy that is active & has results that aren't too absurd - and have us round table ideas and talk about the outliers the day before the draft. It could be our own little analytics conference (pseudo job interview), & the Sixers might find someone(s) they want to lock up for future work & their upcoming D league draft. I good far reaching draft model can be very fluid - and could really help pinpoint where to best send resources (scouts).

Again, I love this stuff. Thanks Nick for getting the discussion started.

[Statman]

Based on what I see from this model and Layne's model and others, I do think people should be taking D'Angelo more seriously. This model is far from gospel, but it has D'Angelo at a Greg Oden level. Layne's has him right between Kyrie and Jason Kidd.

I expect my model will love Russell as well, maybe #1.

I think Towns will be limited somewhat in my model by his big foul rate - he won't be projecting big minutes per game at any point in the future. If it wasn't for that foul rate, I'd think he'd be #1. I expect he'll be top 3 nonetheless with Okafor & Russell.

I'm very curious where Kaminsky lands in mine. Top 5 maybe?

[nrestifo]

Stat,

Thanks for the heads up on the live formula issue, just fixed that up and made those values. Would be happy to set up the consensus again this year too. And plane tickets to go discuss draft in Philly, that would be fun huh?

[nrestifo]

Your questions:

I am using a weighted average of SOS based on GP. For college data, it's per 40, but not pace adjusted. Should incorporate that.

[Statman]

Your questions:

I am using a weighted average of SOS based on GP. For college data, it's per 40, but not pace adjusted. Should incorporate that.

Yeah, pace adjusted is pretty important from my perspective. Plus, it really helps filter out every non prospect without purposely limiting your prospect pool to do it. You'd like your model to be viable no matter the player you throw at it.

For example - if you ran your projection on Iona players, you'd possibly have 4 guys rating higher than many of the prospects in your spreadsheet. Possibly 4 guys that rate higher than any Virginia guy?

BYU, Arkansas, LSU, UNC, & Iowa State guys are gonna look much better in your model than maybe they should. With no pace adjustment, Virginia & Wisconsin, two of the very top teams in the nation, will have like 1 NBA prospect (Kaminsky) between them - and he's not rated very high.

[ampersand5]

im confused when you say you are using a blend consisting of RAPM.

Theres RAPM for college now?
or did you make your own SPM that you are playing to college players?

on a somewhat unrelated note, Karl Towns' BPM is so incredibly high that I'm surprised he isn't a consensus first.

[nrestifo]

When I say I'm using a RAPM/WS blend in my draft model, what I mean is I'm predicting that number. For each NBA player I am taking their WS and RAPM of their peak two-season span, normalizing them, and averaging them together to form the blend. Then, I'm using neural networks and regression models to predict that blend from a player's per40 college stats.

Towns is phenomenal. Most models will see him as the best or one of the best projected players. The biggest reason Okafor sneaks ahead of Town in my model is because Okafor shot 66.4% and had 23pts per40 this year, which is insane. But Towns shot 56.6% from 2, averaged 12.7 rbs per 40, decent steal rate for a big, has a top 15 block rate amoung centers that played more than 15 minutes per game, 81% from the line, 19 years old out of Kentucky, there's a lot to Towns that's just as insane.

[ampersand5]

ok, that makes sense.

some observations:
why not use a RAPM-BPM blend rather than WS?
How are you accounting for age in college stats?
If you are including height, why not include wingspan as well? - seems to be a very desirable feature in the nba draft

Have you thought about running separate neural network/regression models for each position? It seems as if what scouts are looking for is very position dependant.

[nrestifo]

I havent yet looked into a RAPM-BPM blend, but I could rather easily. I experimented with varying numbers/blends but used RAPM-WS because I liked how that blend ranked the NBA's players. While we know that things like RAPM are more predictive that WS, there are some plus-minus darlings that I don't want a model to consider as an ultimate NBA success story (Amir Johnson, etc.). RAPM-WS provides a nice balance and also produces sell-able results.
My age variable is based on the birthdays listed on DraftExpress.com and the February 1st before the draft a player was drafted in. I used February 1st because that is the cutoff-basketball reference uses for their age variable. Not that I used basketball reference's data for age, but it's just nice to have mental consistency. The age variable is then normalized with z score normalization and included in the models like any other variable. It's not treated in any special way.
I did not include wingspan because, like weight, it correlates significantly with height. Although I have used PCA in draft models before, this draft model does not utilize any kind of PCA or factor analysis to account for multicollinearity, so although many of my included features correlate sizeably with each other, adding wingspan under the current methodology would result in an especially bad case of double-counting in theory. I have been looking into wingspan further and may adjust this approach further down the line.

I have experimented with position dependent models with mixed results. The problem with that approach is you considerably cut your sample, giving the models less examples to tease out the relationship. But I will continue to play with this. The good news is neural networks can, to some extent, tease out some of these position-dependent trends without being trained on guards/wings/bigs separately. The bad news about neural networks is that they overfit. But giving regression models a word in the final say helps to balance that out.