Crow wrote:
Saw you include high school recruiting rankings. How big a weight? (Don't have to give exact weight I'd you don't want to but would you call it light or moderate?) Any study of indicator value of what college / coach they picked / which & which "level" (college performance or future pro prep) picked them (as an add on to HS recruit rank)?
I would say it is a light weight in PNSP but moderate weight in the NBA Role Prob Model. We have looked at college / conference but didn't find anything meaningful.
Hoe do you team stats? Do you use strength of opponent? At game or season level? Any consideration of extra weight on games with / actual counterpart minutes against NBA prospects?
Right now we are simply using strength of schedule. We have explored incorporating Team wins and performance against top level teams but haven't found a solution that provides anymore predictive power.
Shooting weighted heavily. Is usage level considered?
Usage is not used in our models.
Correlations with mainline metrics checked. Closest to PER. Ok with that?
This is interesting. So intuitively it makes sense that our scale and aggregate method of NBA stat lines closely aligns with PER because both are built off basic box score statistics. While PER is not my favorite all encompassing metric, I feel better because our scale and aggregate of predictions on NBA Stat lines predicted "better" than just predicting on any of the all encompassing metrics RPM, SPM, VORP, PER, etc.
Physical measurables included. How many? Any use of Combine athletic tests?
Wingspan, Weight and Max Vert. Obviously, top prospects do not go to the combine these days so we grab Draft Express Measurements (if not too out dated) and impute the remaining players.
By position. By college position or projected NBA position? Any strategy for dealing with NBA "playing small"... at PF or C?
Position is difficult given the position-less NBA. But we do incorporate to account for differences in statistics across positions. We are using their projected NBA position and for the most part use Draft Express's position assignment of players although we have gone back historically and adjusted players to fit into the PG/SG,..,PF/C roles. So we have 9 position groups PG, PG/SG, SG, etc... I will say that shuffling Bigs through PF, PF/C, and C usually gives a similar PNSP score for a given player (within maybe 5-7 points) and similarly for PG, PG/SG, but big position jumps will have massive changes.
Machine learning. Applied to individuals or also "player types"?
Machine learning applied to individuals.
Has the model been run before other drafts or is it built / run retrospectively to date?
We ran it prior to the 2016 draft and those results are posted on our website, but retrospectively if we run them now they are slightly different obviously.
Plan on offering projections to Draft Express for anticipated draft analytics model? Blending your models? Using what principle? Any use of subjective / scouting for adjustments of player scores or draft ranks? How did your models do on projecting the young TWolves? Does how teams weight all-star vs starter weights depend on what they have, what the intermediate goals are? In general how much should teams shot for stars vs. get realistic value? LaVar Ball with light yard ahead highest all -star potential in draft class? Ever or by year 4? Second highest ever probability in your dataset to Durant?
We would be open to offering projections to Draft Express if that is opportunity (although we do not have a model for international players yet). We could incorporate Draft Express rankings into our Model but opt for separate models. I wouldn't be opposed to incorporting subjective rankings with someone that I liked their scouting approach. Models really liked Karl Towns, saw Andrew Wiggins as a solid starter, did not particular like Zach LaVine nor Kris Dunn, but did like Tyus Jones as a solid NBA player. Definitely apart of team construction in terms of NBA Role Probs and Risk / Reward on prospects. The outliers are interesting - here are the top 5 All Star probs in our retrodictions:
1. Kevin Durant 85%
2. Ben Simmons 84%
3. DeMarcus Cousins 80%
4. Michael Beasley 79%
5. Joel Embiid 70%
Basically set with your methodologies or wrestling with additions / changes? What is on your to do or someday wish lists? Any interest in RPM or using what exists / new stuff to yr to project RPM or RPM similars within existing player similars? Any consideration for projection adjustments for college coach impacts on player stats?
We do feel strongly about our methodology but we do know its limitations and have a good sense for some of the "gaps" in the model but we will always wrestle with new ideas, additions and changes. Remember, "All Models are wrong, some are useful." We are working on our version of RAPM that should be primed for next season as well team projections, draft pick values related to our draft model, lineup construction, yr to yr NBA projections and a whole host of other NBA projects on our list.
Light on defense due to boxscore limitations. Any interest in trying to address this with strategies to guessestimate / incorporate more D?
Definitely, searching for ways to encompass a prediction on a defensive NBA measure to incorporate in the calculation of PNSP.
Have you considered running role probability model customized to selecting team? Who got more / less projected value FOR THEM?
Interesting, haven't explored.
How would you describe your 2 principal components in the similarity model? Is that only model adjusted for SOS or just the only one where that is mentioned? Is there anything you learn about the college to NBA transition in general from your simulated annealing step?
So we are using K means clustering to group players prior to running similarity scores. This groups players fairly nicely into Ball Handlers, Wings and Bigs (obviously not perfect). We just use the first two PC to visualize the data. More interestingly, the simulated annealing finds different weights for statistics by clustered groups. Most heavily weighted by position measured by correlation to NBA similarity:
Ball Handlers - Points, Steals, Assists
Wings- 2P Percent, Rebounds, Blocks, Wingspan
Bigs - 3P Percent, FT Percent, Assists
Application of models to international players? Playing overseas... or in the NCAA environment? Is NCAA environment / resulting the same for internationals or do they need a third way translation path?
In the works - currently work on something similar to what Layne Vashro put together a couple of years ago comparing leagues.
Significance of your site name?
Comes from a room at school where my buddy Sam and I first worked on sports analytic projects many years ago (oh, and where he ran the March Madness Model that successfully predicted the Shabazz Napier led 7 seeded UCONN Huskies to win it all).
Seeking NBA positions or direct consulting opportunities? Your website has a team... uses a team approach? Would you consider adding scouting talent or feedback / critique talent or coaching perspective or psychological assessment talent or anything else?
Most of the work we do in pairs or groups. I cannot speak for all of us but for now we will continue to publicly post our research while trying to add some humor for fun. Definitely, open to adding scouting, critiques, and definitely psychological assessment talent. We will get a podcast running in the near future but that will cover multiple sports.
NBA betting model appears to be entirely at team level with no consideration of player or lineup matchups? True or not?
True - plenty of room to explore in this area.
Ideas for types of math or math techniques that could be applied to NBA but not yet seen, to your eyes? Any techniques that could be taken from re-insurance brokerage business, say related to performance "risk"? Does volatility of player performance game to game (or by other filters, yr to yr, opponent type, home / road, tournament, etc.) interest you for projections? Is it actively considered right now?
Lots and lots of opportunities. Volatility is probably the most common term used in both avenues of work for us (especially in the betting world). We do already incorporate differing levels of volatility measures.
I've been told by a certain figure in the world of basketball analytics that "question guys" aren't that rare or special. How did I do for 80 minutes of immediate feedback?
Keeping me on my toes, love it.
Last caveat I will add, is that while predicting actual NBA box score statistics is difficult, we have found when scaling predictions by position they have been predictive of where a player lies in comparison to other players at a similar position in the NBA. Breaking down the models and identify those areas of under/over performance in a given statistic is very useful in prospect evaluation, IMO. We plan to breakdown some of the top prospects and pull out interesting parts of the Models for each prospect over the next couple of weeks.