By John Ezekowitz and Chris Bruce
Last week, we introduced our methodology for predicting the college football AP Poll before it came out. This week, we continue our prediction series and add the Coaches’ Poll, which is calculated as part of the BCS. Interestingly, whereas the AP Poll model had an R^2 of .57, the same model used to predict the Coaches’ Poll yields an R^2 of .68. The inputs from the previous week post predict a significantly larger portion of the variation of the Coaches’ Poll than they do in the AP Poll.
Interestingly, the direction and magnitude of the significant coefficients are similar in both models (their 95% confidence intervals overlap quite a bit). That implies that both groups of poll voters react to results very similarly. This raises some interesting questions about the validity of using polls. If two distinct sets of humans are both acting similarly and predictably in response to college football results, what makes one poll better than the other?
Additionally, the success of these regressions (which we plan to fine tune as the season goes along. Please keep adding your suggestions in the comments) allows us to think about human polling as a mechanistic process. If almost 70% of the variation in the movement in the polls can be predicted through a formula, is the Coaches’ Poll any more than just another computer ranking with a dose of random noise?
Without further ado, here are our predicted rankings for the AP and Coaches’ Poll this week: