NCAA Tournament Simulations

[mehmattski]

I didn't want to bury this at the back of the "Dork Polls" thread, but feel free to move this if necessary...

One of the advantages of the various computer rankings is that they spit out "expected winning percentages." This is useful because if you have the "odds" of two events independently, but can only choose one, there is a pretty straightforward formula for figuring it out. See: Odds Ratio, or in the stat-head world, the log5 method. I suppose one could use raw winning percentages, but I doubt that method has much prediction ability.

Given the expected winning percentages, it is possible to generate a large contingency table, with the odds of each team progressing to each round. One approach is to give the discrete odds, while others like to do independent simulations of the bracket a bunch of times and sum the results. I wanted to use this thread to start collecting the various "Dork Poll" prognostications (there will certainly be many more). Here are three:

1. Sagarin (Predictor) based, 5000 simulations. waynewinston.com

2. Pomeroy based: click and create a "random bracket," with winners chosen weighted by their Pomeroy ratings! Click over and over and over until you get Lehigh beating Kansas! billmill.org

3. My method is similar to the randomizer (#2), except I tabulate the results of one million runs of the tournament. Duke fans will enjoy the results of this one, although if you've been following the Dork Polls thread, you already know who's favored to win the National Championship... immaculateinning.com

This year will be a great test of how feasible it is to use pre-tournament Pomeroy rankings to predict tournament results. There are a lot of teams ranked way off their seedings when compared to the kenpom method. Kentucky and Syracuse, in particular, could be very sad to meet Wisconsin and FSU, respectively.

[Troublemaker]

This is outstanding stuff. Very cool. Hopefully Basketball Prospectus releases their log5 projections tomorrow.