Imagine two teams play the same schedule.
One team goes undefeated but keeps on winning squeakers (think the Chiefs pre-SB this year).
One team obliterates 95% of its opponents but has one or two bad games.
Obviously the undefeated team deserves a regular season championship, a higher seed, and being lauded for its accomplishments. But which team is worse in the computer model...?
The team with a few losses. On average, they've played at a higher level even with the bumps.
Vegas would have them favored over the undefeated team, too.
This way is a more realistic way to look at sports like football and basketball where margins of victory tell you so much more than mere win/loss. But under an Elo rating system, the undefeated team is going to keep going up and up and up while the other team is going to take a few dings in its ratings.
Vegas would laugh and try to exploit the arbitrage presented.
Elo has its merits. It was developed decades ago, it is easy to calculate, and it works very well for any sort of contest where readymade and objective score data doesn't really exist outside of win/loss splits (mostly strategy games, including videogames), and it works well with very large sample sizes.
College basketball just doesn't work the same way.
Elo throws out a lot of data (scores) and college basketball only gives you a short season (31 games plus postseason) with 364 teams with very few of the 66,066 theoretical matchups ever being played. You need a system that can worth with smaller sample sizes and make more background inferences.
