stop making college football preseason rankings

Just saw the 2026 preseason AP Top 25. I really wish we would stop ranking teams before ~week 8.

What are preseason rankings even based on?

past performance? No guarentee of future returns.
brand? Talk to Cig.
roster/recruiting? See Texas A&M.

You’re just a whiner. Rankings are fun

Okay fair enough, but I contend that this actually matters. Playoff rankings are quite subjective and the people voting have a limited context window with heavy recency bias. I have seen a good deal of anecdotal evidence over the years that rankings are path dependent ^[1]. That is: the previous week’s rankings provide a sort of psychological inertia for generating the next.

A thought experiment

Consider this all-too plausible CFB scenario:

Week 4: Team A, B and C (all 2-1) sit at ranks 7-9 respectively. Team C beats team A. Team B plays an unranked team. Naturally, team C moves up relative to its neighbor B since they have a “quality win” ^[2] against a ranked team.

Week 7: Team B and C continue to win, but there is no change in ranking relative to one-another. Surprise, the highly speculative preseason rankings were total BS and Team A keeps losing. They fall out of the Top 25. Do we reverse Team C’s week 4 jump over team B now that we know more about A? No, that hardly seems fair since team C hasn’t lost to merit the drop.

Week 8: Team B plays and beats Team A, but that’s not a “quality win” against a ranked oponent anymore. Once again, there is no relative change in ranking.

Week 12: Teams B and C both end 9-3 with no h2h matchup or other common oponents. Team C makes it into the CFP, Team B is first out

All seems reasonable here. Just one problem: Imagine if we just switched the schedule so that the Week 8 A vs B game hapenned in Week 4 instead. Now B stays ahead of C, and all the rest of the logic holds to maintain their relative positions for the rest of the season. The CFP outcome was entirely contigent on the order of the schedule and the low signal-to-noise ratio rankings as they stood in week 4.

Conclusions

One would actually expect that if rankings were not path dependent, the relative positioning of teams within the T25 would change seemingly arbitrarily every single week as new information about all teams (including those outside T25) is aquired and all its historical consequences factored in and transitive implications propogated through. In practice this rarely ever happens. Rankings are remarkably stable week-to-week, even though there are only 12 games in a season ^[3].

The earlier in the season a particular ranking is made, the lower it’s signal-to-noise ratio and the more potential for undue effect on the season’s outcome. Ideally, we probably wouldn’t have any rankings at all until week 12, where the rankings would be determined in one shot by a comprehensive overview of each team’s resume with the complete season of information.

But at a minimum can we just eliminate preseason rankings?

I think it would be interesting and fairly straightforward to flesh out this path dependence from real data. Train a MLP that takes in a (permutation-invariant) set of each week’s h2h scores and predicts rankings. Train another model that takes only the current week’s scores but is also autoregressive on the previous week ranking. I hypothesize that the autoregressive model would have comparable or higher predictive accuracy by week 12 ↩︎
Don’t even get me started about those “quality losses” they have down south. Fun fact: an SEC team hasn’t appeared in a national championship game since 2022. Coincidentally, ever since NIL kicked in. ↩︎
With so few games, the variance of any kind of “holisitic rankings” should be so high that large weekly fluctuations would be completely expected ↩︎