With March Madness come two things: office pools and upsets. Lots of them.

Edward Kaplan, a Yale School of Management professor in management sciences and public health, has turned these pools into a science of sorts. Kaplan created a statistical model that fills out brackets to generate the maximum number of points for a given pool. This spring marks the second tournament his program has tried to predict.

Kaplan’s model fills out brackets based on one of two basic point structures. In a standard pool like the ESPN.com tournament challenge, the program tries to predict the most accurate results based on each team’s likelihood of winning.

But for “upset pools,” which award more points for predicting upsets correctly, the program bases its predictions on which underdogs are likely to upset, with the goal of maximizing the pool’s more complicated point structure.

In making the predictions to fill out the bracket, the program incorporates a range of outside information — Las Vegas odds, point totals and point spreads, USA Today rankings, team scoring data and strength of schedule. The program compiles that data to make the optimized selections for each type of pool.

If the entry is for an upset pool, which may award 10 points for every No. 16 that defeats a No. 1, nine points for each No. 15 defeating a No. 2 and so forth, Kaplan’s program will take all the outside information and pick the most likely upsets to maximize point totals.

Like so many others, Kaplan used to enter an office pool and lose every year. So he decided to do something about it.

“[A former faculty member] ran the pool with a bizarre structure for awarding points,” Kaplan said. “I knew there had to be a mathematical way to take that all into account, so I finally decided to try and figure it out.”

And he’s done a pretty good job, at least in his upset model. In pools that award more points for underdog victories, Kaplan’s models have produced solid results thus far in the tournament. All his Final Four teams are still in the running — Duke, Stanford, Arizona and Gonzaga.

Yes, Gonzaga.

But there is a method to Kaplan’s madness.

“In the pools that award lots of points for upsets, you gamble a lot,” Kaplan said. “Because even if only a few come in, you still earn a lot of points. The model helps you to pick the right upsets.”

In the ESPN.com tournament challenge contest, Kaplan’s models are around the median. He is doing slightly better than if he had just picked the higher seeds.

“Sometimes they make a few mistakes with the seedings,” Kaplan said. “To figure out the real ability of the teams, look at the odds or the Sagarin rankings. Then you see that sometimes lower-seeded teams actually are better. This gives you seed reversals. The model optimizes this information telling you which ones to pick.”

This is why one of his models has No.12 seed Gonzaga in the Final Four but not similarly seeded Hawaii, which the program accurately predicted to lose in the first round.

Kaplan’s model is enjoying even more success in the women’s tournament. He is currently ahead of 99.7 percent of those entered in the ESPN.com tournament challenge for the women’s tournament. The amazing part of this success is that it has come in a standard pool, not an upset pool.

Why has this model been so successful?

“A good piece of luck,” Kaplan said. “That, and the fact that the model takes into account the structure of the tournament.”

His women’s Final Four produces no surprises. The program picked all four No. 1 seeds — Tennessee, Connecticut, Duke and Notre Dame.

Kaplan is undecided as to whether he will continue the models in future tournaments. But, he has a much more important use for the mathematical formulas used in creating the models. His main area of work is in HIV prevention, and he says that much of the math used in creating the tournament models can be used in this area as well.

“The success of the models in predicting March Madness shows people that these kinds of mathematical models can be successful,” Kaplan said. “It’s a tremendous confidence builder, and gives people confidence that the models will work in HIV prevention, which is far more important.”

[ydn-legacy-photo-inline id=”20665″ ]