Big Data Approach to Startup Selection

Abstract

Ninety percent of startups fail. Of the other ten percent, returns follow a power law distribution. To quote Peter Thiel, ''A VC portfolio will only make money if [its] best company investment ends up being worth more than the whole fund.'' Easier said than done, of course. Startup equity is highly illiquid, and VCs can’t make money off their investments until they sell all or part of their ownership during a liquidity event, such as an acquisition or IPO. Since investment outcomes follow a power law, VCs also cannot expect to make money simply by cutting checks. The competition is cutthroat; they not only have to spot high potential portfolio companies to fund but also help them in a differentiated way by leveraging their network on the companies’ behalf or advising founders well. Granted, it’s exponentially easier to wait for a startup to become successful first before jumping onto the boat in a later funding round, but you probably won’t be able to recoup your investments that way. The only way to survive in the VC world is to spot the decacorn when it’s still a nobody and fund it all the way from seed to IPO, because then the returns are exponential. Doing this is already difficult as a VC, but can we make an algorithm do it? To today, there has been very little literature on this topic, so in this paper, we’ll frame venture capital as an optimization problem which attempts to minimize the error encountered when classifying a large dataset of startups into those that will potentially return a profit and those that most likely will not. Just how important is it to have that MBA in your startup team? Is it better to have a co-founder or should you be a lone wolf? Armed with data downloaded from Crunchbase and a set of features that we came up with, join us on this adventure as we attempt to build a logistic model to predict the future of startups.