MIT Researchers Offer Algorithm for Picking 'Winning' Startups
Their method beats top IRL venture capitalists.
In parsing histories of the Third Industrial Revolution—that's this one—future academics will have to contend with the phenomenon of venture capital. Somehow, in this bygone era, much of the economic fuel for new information technologies was distributed according to a system that looks a lot like gambling. A small group of people with lots of money dictated a future by providing often large sums of money to companies likely to fail. The trade-off was that these companies offered at the same time the potential for not just success, but rapid, exponential growth.
In fairness, the gambles made by venture capitalists today are rather more educated and even more predictable than those made at horse-tracks and slot machines. But, still, there's an awareness that most investments will simply drown for reasons that are often unpredictable.
But maybe there's a way to better maximize startup investments. To this end, a pair of researchers from MIT, David Scott Hunter and Tauhid Zaman, have developed a framework for venture capital investment that is based on the randomness of Brownian motion coupled to large volumes of data collected on startup founders, investors, and performance. The resulting model, according to a paper posted earlier this month to the arXiv preprint server, is able to achieve 60 percent exit rates on optimized investment portfolios, a figure about twice that of top venture capital firms. (An "exit" is where a startup either goes public or is acquired by a bigger company, scenarios in which investors make a lot of money all at once.)
"In betting on startup companies, all I care about is one of my companies hitting it big," Zaman explained in an interview. "One IPO and the portfolio is profitable. When that's the goal, the way you pick the companies is different. I'm basically betting on moonshots. Either this company is going to be insanely awesome or it's going to be total garbage. I'd rather have a bunch of companies like that, that are going to be amazing or total junk, rather than a bunch of solid predictable companies that are going to do well but aren't going to be sensational."
Modeling the sort of company you want in that sort of portfolio isn't obvious. The desired trait is unpredictability, but the whole point of making statistical models is prediction.
The property the researchers wanted to optimize for companies in a VC investment portfolio is volatility. This is measured in terms of growth, generally. In the startup world, companies progress up a staircase of funding rounds with names you've probably heard like series A and series B and so forth. At each step, the startup gets a new infusion of cash as a new batch of investors pour in new money.
Volatility is measured here in terms of how unpredictable this progression is for a given company. If some startup sits around at one funding level for a while but then all of a sudden makes a huge leap through several funding rounds, that's better than if it had progressed through the same funding rounds in the same period but at a steady rate.
So, in other words, for a startup growth is good, but erratic growth is better.
Hunter and Zaman were able to collect data on over 83,000 companies via startup databases Crunchbase and Pitchbook. They then correlated this data with information on LinkedIn, which gave information on the educational and job histories of startup employees. With all of this data, they came up with a set of startup properties (or features) that could be used to develop a predictive model. So: machine learning.
What Hunter and Zaman found is that the things that positively impact volatility are pretty normal features of a solid company. More than anything else this is an experienced founder. Which is something that investors already use to judge potential investments, right?
That's the kind of funny thing about the results of Hunter and Zaman's work. While their framework handily beats real-world VC investors it's doing so by maximizing some pretty obvious features, like founder experience.
"I want to say like I found some secret sauce that VC investors weren't aware of, but our analysis showed that the things VCs look for are the things we found that are the keys to success," Zaman said. "No big revelation."
As to why the framework is able to make better predictions about startup success even though it's honing in on the same desired features as human investors, Zaman thinks it's a matter of how the algorithm is measuring things like founder experience compared to how investors measure it in the real-world. In the latter case, it's likely that experience is measured more quantitatively and intuitively while the algorithm is using only quantitative data (like, say, years of education). The message there would be to make investment decisions based on measurable quantities and not gut feelings.