Tom Federico met his mathematician and computer scientist partners as a freshman engineering student at Stanford. TeamRankings.com, the statistically rigorous Web site the trio founded in 2005, offers bracketeers paid access to a world of predictive data through its BracketBrains.com spin-off.
Speaking with amNY this week, Federico advised pickers to focus on the NCAA tournament matchups whose outcomes the office-pool competition may be guessing wrong.
How does your site work? We run our predictive models and ... see where there might be inefficiencies in how the public is predicting teams versus what the real data-driven odds are saying.
What is an ‘inefficiency’? The NCAA tournament — sort of a financial analogy — is a marketplace, and your job is to find inefficiencies in the market. Our job is to help you know when is a favorite really a favorite, and when is there an underdog that the public is not giving enough credit to in terms of their actual odds to win.
Where could the public go wrong? You might see that 25 percent of the public is picking Kentucky to win the NCAA tournament, but their actual algorithmic win odds are only 10 percent. ... It gives you the data you need to make a decision that says, hey, I can optimize my win odds to win if I don’t pick Kentucky to win because it’s a negative-value pick, if you will.
Which team would be a positive-value pick two win the national title? We’re still getting the public data in, but right now I would say for the national title, Kansas is definitely the odds-on favorite. We still think they’re being a little bit overvalued by the public. ... Especially in a bigger bracket pool, Duke is looking like a good value pick to win the tournament. If you want to get even more wild and crazy, a couple ones we picked out are Kansas State, Baylor, and even Maryland and Wisconsin. ... Their odds of winning the tournament are clearly very slim, but ... you’d do better to just pick, say, Wisconsin to win the tournament if you were competing against a million people. By the freak chance that they actually do pull it out, you’ll almost certainly finish out in the top one or two or three.
What makes these models so complex? They were developed by my partner, Mike Greenfield, who’s a Stanford mathematical computer scientist major, about 10 years ago, and we’ve been refining them every year since. ... We take into account betting odds for games, statistics for teams, how teams have played historically in similar situations. Much more than a human mind could ever process on its own, obviously. And then it’s just the sophistication of all the modeling ... It’s really done at a scientist level, versus ... people building their own Excel spreadsheets.
How do you account for a star coach or player? Obviously, there are a bunch of intangibles that are difficult to model with just hard data. Things like morale or a ... coach who’s historically performed really well in the tournament. ... It might not really be the isolated effect of the coach’s ability. There could be a million other factors that went in there. ... Typically, rather than looking at personal matchups of players or a coach’s history, the vast majority of our analysis is driven on hard data: How did the team perform this year; what were their margins of victory against various opponents; what was their schedule strength.