Reweighting the Bowl Championship Series

Abstract

The majority of statistical work on college football's Bowl Championship Series (BCS) has involved proposing or categorizing computer ratings of teams. Computer algorithms, a coaches' poll, and a media poll make up the three ratings systems that are currently equally weighted to produce an overall BCS rating, which ultimately determines which schools will compete in lucrative post-season BCS bowls. We focus on investigating the performance of the BCS as implemented for the 2004, 2005, and 2006 seasons to determine whether equal weights are appropriate. Our Bayesian analysis shows that while the posterior mode places more than half the weight on the media poll, the 95% HPD credible interval contains the equally-weighted scheme. We relate our work to the ongoing controversies over the BCS.

Minimum violations and predictive meta‐rankings for college football

“…The problem with this dichotomous performance is that there is also a dichotomous perspective on ranking fairness. Some observers believe it fairest that a ranking match past game results, while others believe it fairest that a ranking predicts future games well [6] and otherwise view the ranking as unreasonable [5] or "fundamentally deficient" [17]. This quandary between "prediction" and "performance" (i.e., Correspondence to: B.J.…”

“…This quandary between "prediction" and "performance" (i.e., Correspondence to: B.J. Coleman (jcoleman@unf.edu) retrodictive fit) [5] is acknowledged by Stern [24]: "to obtain a uniquely best [ranking] method would first require that we all agree on the objective. There is clearly a tradeoff between rewarding performance [valuing wins]...and trying to build a model that would effectively predict the outcome of future games."…”

“…Some observers believe it fairest that a ranking match past game results, while others believe it fairest that a ranking predicts future games well and otherwise view the ranking as unreasonable or “fundamentally deficient” . This quandary between “prediction” and “performance” (i.e., retrodictive fit) is acknowledged by Stern : “to obtain a uniquely best [ranking] method would first require that we all agree on the objective. There is clearly a tradeoff between rewarding performance [valuing wins]…and trying to build a model that would effectively predict the outcome of future games.”…”