[grndmstr_c]

Okay, was playing around with this playoff odds issue a bit more, and decided do a little modeling using Hollinger's playoff odds and power rankings to try and generate some guesses about what Dallas' actual probability of making the playoffs is right now.

The approach was pretty simple, take Hollinger's posted playoff probabilities and model them as a function of each team's: 1) Rating, from Hollinger's espn power rankings, 2) win percentage (technically odds or log odds) to date, and 3) conference, west vs east (conceived as a simple way to get at likely strength of schedule and quality of competition). The goal was to generate a model for approximating Hollinger's playoff odds that could be used to generate new guesses for different (novel) combinations of predictors. For Joshi and anyone else who cares, I used generalized additive models with a beta distribution for the response variable.

Once the models had been fit (and the fits were pretty good; e.g., fitted playoff probabilities for the Mavs under the models hovered ~ 10%, whereas Hollinger's actual prediction was 11.4%), I generated predictions for Dallas given their current win odds and conference, but with ratings set equal to the ratings of each of the other 29 teams. The idea being to generate guesses that could be interpreted against the backdrop of comparative assessments of team strength.

In general terms, the results were as follows:

- If Dallas actual quality is on par with teams like GSW, Indiana and Chicago (minus Rose, of course), then Dallas' probability of making the playoffs as of right now as predicted by the meta-models is something in the neighborhood of 35%-45%.

- If Dallas actual quality is on par with Brooklyn and Denver, then Dallas' probability of making the playoffs is in the 45%-55% range.

- If Dallas actual quality is on par with Memphis or NYK, then Dallas probability of making the playoffs is in the 55%-65% range.

There's an obvious criticism of those projections, and that is that they're based on combinations of predictors that are bivariate outliers with respect to the training data set. But since there's not really anything that can be done about that...