Bracket Voodoo: Beating the Experts -- or just your friends

PLAY: Bracket Manager | Bracket Challenge | Round-by- Round | Experts Challenge

Every year I like to gather a few expert or celebrity brackets and gauge how they would stack up in various types of pools. This year, I decided to take a look at the battery of Expert brackets and take a peek inside the office pool. (Note: I don’t think this pool really exists, but I wanted to look at these brackets in that context to illustrate where the edge is derived in a typical March Madness pool.)

First we analyzed this “pool” by simulating the entire March Madness tournament a few hundred thousand times with the prediction engine and looking at which expert came out on top each time:

So in this pool, Jeff Borzello and Gregg Doyel are co-favorites, with Jerry Palm not far behind, while Steve Lappas is on the outside looking in. Sorry Steve.

In order to better understand where Borzello and Doyel derive their edge, let’s add a couple of metrics to the table -- each bracket’s expected score (the average number of points we think it would receive) and a similarity score ranging from 0 to 1 for each bracket (where a 0 would represent a completely unique bracket and 1 represents a bracket that is exactly like every other bracket).

Yesterday we discussed the importance of trying to differentiate your bracket in order to give yourself a chance in an extremely large pool like the Billion Dollar Bracket Challenge, but the same holds true for a smaller pool like this one. You want to differentiate your bracket as much as possible while sacrificing as little expected value. And sure enough, the two best-positioned brackets in this pool are the one with the highest expected value and the one that is the most unique.

We can see this laid out visually in the graph below. Each bubble represents one of the expert brackets and the size of the bubble represents that bracket’s probability of winning this pool (larger is better). The x-axis represents a bracket’s expected number of points; the y-axis is the similarity score. The large bubble at the right is Borzello (high expected score), and the large bubble at the bottom is Doyel. The almost dot-like bubble at the top is Lappas. Despite having an expected score that is actually greater than Doyel’s, Lappas has a much less unique bracket, and as a result ends up with a bracket that is 1/6th as likely to win this “pool”.

So what exactly is “wrong” with Lappas’s bracket. It’s not that the picks are bad, it’s just that they are too similar to those of everyone else in the pool. Not only has Lappas picked a champion (Michigan State) that coincides with two competitors, but every one of his Final Four and national finalist picks is duplicated by about half of his competitors or more. When you get into situations like this you start to need elaborate combinations of events to play out in order to win your pool. Have you ever had a bracket that was doing poorly and you thought to yourself, “OK, now I just need Michigan State to win it all and Arizona to make the Final Four but then lose to Louisville…” That’s not a good place to be. And that’s the sort of place Lappas’s bracket would be in from the start of this “pool”.

This is a common pitfall that people fall into when filling out their March Madness brackets. Sometimes it isn’t avoidable -- since you usually don’t know what everyone else is going to pick. But sometimes it is -- since you should know that Michigan State is over-picked. Steve Lappas didn’t fill this bracket out in the context of an actual pool, much less this one, so I don’t want to criticize him, but I do want to point out that it is a mistake for people to submit brackets that are too aligned with the masses in just about any sort of March Madness pool.

But enough with this hypothetical office pool -- what does this mean for your pool? It means you need to take after Doyel and Borzello and differentiate your bracket, but without sacrificing expected value. And in fact there is opportunity to go much further. For example, if you were in a pool like this and happened to know that the other folks in your pool were high on Michigan State and Louisville, you might consider the following bracket.

Riding on both Arizona (our favorite to win the tournament) and Virginia, two teams the panel of experts isn’t very high on, this bracket would have a similarity score of 28%, 11 points lower than the lowest among the experts, and a 27% chance of winning this hypothetical pool -- almost double that of any of the experts. And oh yeah, since is actually running an expert challenge where you can compete against some of these brackets maybe you should consider a similar strategy for that pool.


Brad Null is the founder of, the world's most advanced NCAA Tournament bracket analysis and optimization engine. Try it out now at

Show Comments Hide Comments
Our Latest Stories