## Dave Meyers’ genius solution for automating MWS48

As I stated in a previous post, at the moment all of my “blog time” is being devoted to calculating and finishing Win Charts for every 2009-10 NBA team.  I’m at the “16 teams to go” mark.  You cannot believe how much mind-numbing work is involved in those things, mainly because few NBA players play one position, so I have to calculate multi-position players proportionally.

But the results so far have been interesting and even more accurate than past seasons.  I’m very encouraged.  This summer myself and hopefully some readers of this blog need to really put Marginal Win Score to the test.  I’m beginning to believe it may be the real deal.  But more on that when I complete the project.  I’m efforting that as fast as I can.

#### Dave Meyers genious idea

I have not yet followed up with reader Dave Meyers on his absolutely ingenious solution to the automation of Marginal Win Score.  But I’d like to briefly share it with you to see if anyone else has more to add.   Since the WoW Journal has recently announced that win production based on the metric Win Score will be fully automated next season, there is pressure to get this project done or the alternative Marginal Win Score win production argument will possibly drift out of the conversation.

But automating MWS is going to be a huge project.  Step one was determining how in the world we could automate positional matchups.  Into the breech stepped reader Dave Meyers.

Here was Meyers’ creative solution.  The problem, again, was automating players “defensive” assignments for every single moment of NBA action.  I felt I could, under most circumstances, accurately predict who would “check” whom based on a balancing test I developed that considers the player’s “recognized” position, his height, and his weight (along with knowledge based on casual observation), with the first element being given the least weight in the test.  As I stated in the past, the test I have developed is nothing more than an application of the “pairing” process that takes place prior to every pick-up basketball game in every gymnasium.  (If you notice, and I’m not trying to be facetious, but only when a female is on the court — because of the “no-win” nature of the matchup — or when players are too redundant on the 1-5 spectrum is the matchup process ever allowed to delay the commencement of action.  That’s how universal it is).

So I felt like I could tell a person at all times who is probably counterpart to whom.  The problem was, how could I tell the computer?

Meyers suggested the following.  Rank every player on the roster according to their positioning on the left-right, Position 1 to Position 5, player spectrum.  The computer could simply apply the ranking to game information provided by the NBA’s play-by-play transcripts and the program would be able to immediately decipher who was matched against whom at any moment on the court.  Then it could simply calculate the Win Score of any player, along with the cumulative Win Score of every counterpart that player was matched against.

To put it in practical terms, here’s an example of the Meyers solution at work, using a pair of Milwaukee Bucks players as illustration.

Under every circumstance last season, Brandon Jennings defended the oppo point guard, no matter who else was on the court.  Therefore, under every circumstance his assigned number must have the lowest value of anyone who is on the court at the time.  Thus Jennings would be assigned the number (1).

Similarly but with a key distinction, under nearly every circumstance, Luke Ridnour would also defend the oppo point.  But when Jennings was on the court simultaneously, Ridnour defended the shooting guard.  Therefore Ridnour’s number must be lower than every player’s except for Jennings, and thus Ridnour would be assigned the number (2).

You could apply that same logic to every player relationship and all the program would have to do is choose player matchups according to assigned numbers in ascending order.

Brilliant.  But now the ball is in my court, and I am turnover prone.  I must follow through with every effort to get this thing done.  Marginal Win Score is backed by solid logic, intuition, and increasingly by numerical data.  In short, I think I’m on to something, but if I don’t make this automation happen whatever I’m on to might simply disappear.  82games.com is no longer reliable, and besides I would like to have more control over matchup data.  I don’t think you will see a big difference, but nevertheless that kind of control at some point becomes necessary to the legitimacy of the whole enterprise.

So I’ve got to stay on this project.  It cannot just become another one of my “big ideas” that dissipate like smoke from my Uncle’s pipe.

If anyone has any interest in chipping in with computer knowledge, basketball know-how, ideas of any sort, any help would be greatly appreciated.

Thanks!

### 3 Responses to “Dave Meyers’ genius solution for automating MWS48”

1. Abe Jaroszewski Says:

Not sure what I could do, but I would be willing to help you out if needed.

2. tywill33 Says:

Thanks Abe! Could you shoot me an email at bucks_diary@yahoo.com when you get a chance?

3. palamida Says:

Ty, not to take anything away from Mr. Meyers but this is exactly what 82games have been doing all along… They rank all the players on a roster in a logical order from “most PG’ To “most center”, and then let the computer do all the sorting, in every lineup that was employed.

While it’s certainly better then nothing, and thus I see the data (when the sample size is sufficient) generated using their counterpart data as – useful, at the very least, it’s still far from ideal.
The most glaring problems with this method are probably these:
1.) When players check a certain player on one end but are not checked by him at the other. This is especially evident with players whom are perceived (usually correctly) as being “extreme” on either end of the offense-defense scale.
For instance Thabo sefolosha. Thabo is seen as a “defensive specialist”. Calling him a SF or a SG will simply be misleading. More often that not Thabo will guard the opposing wing player who is the more “threatening” offensively, regardless of position. Similarly his opponent will usually be “based” not so much on size alone – If a wing player is seen as superior defensively, he will usually check Durant\Green etc. and not Thabo.
Perimeter oriented big men create similar problems. think… Bosh\Bargs or Okur\Boozer.
2.) Certain coaches use systems that promote prompt rotations and switches, others – not so much. In the aforementioned system players spend a considarable amount of PT guarding… some other player’s man. Mind you this is not a testament of their ability! For instance: your common pick n’ roll.
Some systems call for players to “stay (or at least attempt to) with their original man” at all costs. In those systems, the counterpart data, derived in the manners you describe still holds plenty of value. But what do you do, when players are encouraged to rotate? and this is not just a zone vs. man thing, it’s much more widespread than that.
Naturally, other then charting – there is no way to address these issues, but with Synergy’s data been made available to the general public (at a fee), I wonder if such a feat can’t be performed.

This is not meant btw, to dismiss the data derived from such (82games etc.) methods or it’s validity – As you well know i’ve long been a supporter of your work and the premises (and promises :)) of using counterpart data in general. I’m merely saying that this “new” proposition is really more of the same, providing roughly the same level of data “reliability”, that is, if I’m reading it correctly.
Sadly I can’t offer any assistance in the computer dpt., If you have an idea\question you want to run by me as far as the theoretical goes, please do – either through the comment section or via E-mail.