Nothing new here. I just thought this looked cool. It's the win probability of an offense with a first down trailing by 3 points. Actually, it is a little new. I'm trying to get smart on multivariate non-parametric kernel smoothing algorithms using R, the open source statistics package. That's just a fancy way of filtering out the noise inherent in the raw data and making a smooth estimate of the true probabilities. This chart is a product of my experimentation.
This axis going from 5-95 represent the midpoint of 10-yard bins of field position. The axis labeled 0-50 represent the game time remaining. The z-axis (vertical) shows the expected win probability. It's a little over-smoothed, at least in terms of field position. It might be under-smoothed in terms of time. Still have a lot learn.
You might find "Numerical Methods that Work" by Acton interesting.
Time and yards come in different units, which is the sort of fact that can become awfully inconvenient when doing kernel methods; the metric is important.
Thanks, Andrew. Right now, I'm relying on the default bandwidth functions. The np documentation is sparse and anything beyond LOESS is fairly varsity stuff for me.
If there is anyone with expertise in R and the np package and is willing to answer some questions drop me a line!
I've been using loess, too - I don't suppose you've seen any way to get an adjusted-R^2? summary gives me an "equivalent number of parameters," so I guess I could tally the residuals myself, but it seems odd to me that I haven't found something like that.
I made my own loess algorithm as an Excel macro, so I don't go to the trouble of doing an r-squared. In R I think it will do it for you. I've been using the np package, which will generate r-squareds.