This is my submission to the kaggle march mania competition. It was made with the encouragement of the Columbia Data Science Society.

The final model I ended up using was gbm. It aggregates a bunch of information, much of which ended up being pretty irrelevant. In particular I create some aggregate statistics about each team over the regular season, like win percent, and margins for winning/losing (surprisingly winning margin is super predcitive).

I use the PlayerRatings library to calculate a glicko rating for each team, and then calculate a prediction from that (along with raw difference in rating).

I merge in some tournament information, including seed number (obviously very predictive). Finally I added some of the ranking systems that Nate Silver identifies as important, Pomeroy, Moore, and LRMC. I couldn't get easy access to Sagarin's predictor ratings, so I used Sagarin's regular ratings.

The gbm's relative influence information is:

              var   rel.inf
       seedn.diff 45.272478
       glickopred 14.791104
      winpct.diff  6.241056
 wmargin_avg.diff  5.389482
      glicko.diff  5.316646
         POM.diff  4.761056
         SAG.diff  4.173324
    wmargin_avg.1  4.018286
         MOR.diff  3.604845
         LMC.diff  2.514890
  margin_avg.diff  2.133449
    wmargin_avg.2  1.783383
         seedpred  0.000000