Seeding & Expert Fallibility

In the first round, the higher-seeded team wins 72.9% of the time. You might expect that number to climb in later rounds — after all, the surviving teams have proven themselves, and the committee's best-seeded teams should separate from the pack. Instead, the upset rate stays remarkably flat:

This is a key economic insight: the seedings contain meaningful information, but roughly a quarter of it is wrong in any given game. If seeding were a financial forecast, you'd say the model has a consistent ~27% error rate that it can't seem to reduce, regardless of the sample or the round.

If anything, the rate has drifted slightly upward — from 23.9% in the first decade to 25.2% in the most recent. The committee has access to more data, more analytics, and more computing power than ever before. Yet their hit rate hasn't budged. This is consistent with a well-known finding in forecasting research: expert judgment tends to plateau, and more information doesn't automatically translate into better predictions.

This is the single most striking pattern in the data and the clearest evidence of expert fallibility. If seeding were perfectly calibrated, performance should decline monotonically from seed 1 to seed 16. It doesn't. The 11-seed line is out of order:

The numbers get more dramatic the deeper you go:

11-seeds have made 6 Final Fours. 9-seeds and 10-seeds have made 1 combined. This isn't a fluke — it's a 40-year pattern across 174 11-seed entries. Several possible explanations:

The 5-vs-12 matchup is famous in bracket circles as the most common upset pick, and the data validates it:

From 2005-2014, 12-seeds won 45% of the time — almost a coin flip. The rate has pulled back recently, but the structural issue remains: the committee consistently assigns 5-seeds to teams that aren't meaningfully better than 12-seeds. The gap between a "good but not great" major-conference team (typical 5-seed) and a "best team in a mid-major conference" (typical 12-seed) is smaller than the committee prices it.

1-seeds are supposed to be the committee's four most confident selections. They get the easiest path — facing a 16-seed in round one, then an 8 or 9 seed. Yet:

The fact that both 1-vs-16 upsets happened in the last 6 years is noteworthy. Either the committee's top selections have gotten weaker, the 16-seed play-in game provides a warm-up advantage (similar to our 11-seed hypothesis), or random variance is clustering in a small sample.

If we compare each team's actual tournament wins to the expected wins for their seed, we can measure whether a conference's teams are systematically over- or under-seeded:

The ACC consistently outperforms its seeding (+0.19 wins per entry), meaning the committee gives ACC teams slightly worse seeds than they deserve. Conversely, the Mountain West is the most "overseeded" conference — their teams win a third of a game less than their seeds would predict.

The coefficient of variation (standard deviation divided by mean) tells us how "noisy" each seed's performance is relative to expectations:

Seeds 1-2 are reasonably well-calibrated (CV around 50-63%). By seed 5, the CV crosses 100% — meaning the standard deviation is larger than the mean. The committee's predictions for seeds 5-12 contain more noise than signal. In economic terms, the "price" the committee assigns to these middle seeds is barely more informative than random assignment.

Champions the committee underrated:

1-seeds that flamed out earliest: