From Intern to Venture-Backed BIZ Experiences With a Little Help From Big Data Competition site Kaggle makes predictive modeling more effective and efficient--and a lot more fun.
By Gwen Moran Edited by Frances Dodds
Opinions expressed by BIZ Experiences contributors are their own.
What it is
Think of Kaggle as The Amazing Race for brainiacs. The site manages competitions in which statisticians, scientists, academics and other whizzes compete to uncover the best predictive modeling--also called data mining--to solve specific challenges. The winner receives a money prize in return for intellectual property rights to the model. Kaggle has helped an insurance company find out which drivers are more likely to crash a car, worked with a financial services company to gauge default risks for particular customers and figured out how to rate chess players.
How it started
Through an internship at The Economist magazine, founder Anthony Goldbloom was able to interview chief information officers about the role of "big data." The biggest complaint he heard was that predictive modeling, a top priority for large firms, was confusing, expensive and came with no guarantee of accuracy.
Goldbloom, who had worked in macroeconomic modeling at the Reserve Bank of Australia and the Australian Treasury, saw an opportunity to tap an international pool of academics and scientists itching to get their hands on the giant data sets held by enterprise businesses.
Why it took off
Goldbloom understood that the competitive nature of scientists--not to mention potential prize money--would spur participants to develop the best possible model. Then there was the appeal of the data itself. Early contests promised access to projects ranging from data on HIV patients to voting for the Eurovision Song Contest.
Goldbloom claims that winning teams have never failed to improve upon industry benchmarks and produce more accurate predictive models, boasting, "I can put a problem up on Kaggle, and in some cases, five or 10 years of academic research could be outdone in two or three weeks."
The business
Last year customers were charged a monthly fee based on the contest's complexity--normally $10,000. This year the site has plans to move to a fixed fee and will take a double-digit percentage cut of the prize money for private competitions.
Payouts have ranged from a few hundred dollars to $3 million, the latter put up by a healthcare company looking to predict policyholders' hospitalizations.
What's next
Goldbloom's goal is to host as many as 10,000 competitions a year. To get there, the company closed an $11 million round of Series A venture capital funding last November. Ultimately, he hopes the site will raise the value and reputations of data scientists. "The world's best data scientists should be earning a more substantial portion of the value they create," he says, "not totally dissimilar from the top wages that, say, a hedge fund manager might take home." --G.M.
