Aggregating Expert Ratings Using Preference-Neutral Weights: The Case of the College Football Polls

I use two principal college football polls to illustrate a preference-neutral linear programming procedure for determining optimal weights for aggregating expert ratings. I compute the weights for 84 weekly polls released during the 1999 through 2003 college football seasons. The weights vary from week to week (sometimes considerably over a year), the range of weights over which the aggregate ranking holds also tends to vary and to be quite small (with a range of zero almost half the time), and nothing is systematic about the week-to-week changes in either the weights or their ranges. The results suggest that predisposition to a particular set of weights is a bad idea, not just for the purpose of aggregating the football polls, but in any situation in which one wants to aggregate ratings provided by multiple sources of complementary expertise.