APBRmetrics

Mike G · Posted: Thu Mar 05, 2009 11:46 am Post subject:

Good general points, Dean.

I have an old pickup truck that I never take out of the county, and never take out in bad weather. It's worth the insurance and licensing expense to use it for minor hauling, but barely. It's on its last legs.
Manute on his last legs was barely viable, and only on defense. To get any value at all out of him, and not incur the liability of having him clog the middle on offense, Nelson left him far away, where he might scare the opponent with the possible 3. This also allowed him to be back on D.

Since this was tried and tested over multiple seasons, it was apparently not a key to success. If 10 of us get together and try it, we will figure it out in a few minutes. You stay 30 feet away, and I can neutralize your offense by playing 20' out. I'll get every long rebound and still get out on the break. It would be 4-on-4.5 .

Ben reaches something I've been thinking in general terms. If 'conventional wisdom' isn't all-knowing, it's still a reference point. Maybe a rating system is only as good as the correlation between the rating and the minutes a player gets. Since everyone (coaches, analysts, players) is striving for that competitive edge, they should be moving toward possible strategies. The successes are self-reinforcing; failures are self-defeating.
_________________
`
36% of all statistics are wrong

**HoopStudies** · Posted: Thu Mar 05, 2009 3:35 pm Post subject:

THWilson · Joined: 19 Jul 2005 Posts: 164 Location: phoenix

What I've done, not for basketball, is create a separate model for different input ranges. You can then make a sort of piece-wise function that is a good fit both in the middle and on the edges.

Say you wanted to model the relationship between height and FT%, you might have a model each for <6'6", 6'6"-7'3", 7'3"<.

The two major problems with this approach: 1) it is more complicated, 2) you have less data. The first means that you have a much more difficult task trying to explain your model to others, especially those less number savvy. The second means that you are more likely to create an over-specified model. In the above example, the highest group might show a positive correlation between height and FT% because there are only a few players and Yao Ming is un-repeatably excellent at FT.

Anyway, the more fundamental question should be: is it desirable for player interaction models to give good results for extreme outliers? I imagine that it is desirable, but probably not very important.

gabefarkas · Joined: 31 Dec 2004 Posts: 1313 Location: Durham, NC

Have you thought about splines? Or is that basically what you meant?

THWilson · Joined: 19 Jul 2005 Posts: 164 Location: phoenix

Ryan J. Parker · Joined: 23 Mar 2007 Posts: 708 Location: Raleigh, NC

I believe Gabe was referring to fitting a spline to the data since it allows for the type of piecewise stuff you were referring to.
_________________
I am a basketball geek.

gabefarkas · Joined: 31 Dec 2004 Posts: 1313 Location: Durham, NC