Wow. I expected a lot more variation.

Performance against the spread is also pretty tight (top 40 between 45-52% accurate).

I suspect if we filtered out all the "obvious" games (e.g. games where all rating systems correctly predicted the same winner), and low-interest games (e.g. games between two teams with losing records) we would see a lot more variation in their ability to predict. Plus, we would care more about the result.

Other interesting filters:
- Non-conference games, since the outcome is less obvious due to few common opponents
- Late-season games between teams with high winning percentages, since the outcome is more consequential