How Do You Automate Venture Capital

By Lucinda Shen at Fortune · interview with Chris Farmer, MD and CEO of SignalFire

Kyle’s read-through of the Fortune interview on whether venture can be automated — SignalFire’s data-first, “more Google than venture fund” positioning, and where machine learning falls down in seed investing. Notes captured verbatim on the April 17th, 2020 Roam daily note.

Notes

• Read “How Do You Automate Venture Capital” - interview between Lucinda Shen at Fortune and Chris Farmer the CEO of SignalFire
  • In recent years, quantitative hedge funds have made a killing in public markets with their data-first approach (coronavirus aside). But skeptics say the strategy makes less sense for venture capital, where public financial reports are nonexistent.
  • Six-year-old venture firm SignalFire is among those seeking to challenge that assumption. The firm, which tends to invest in the seed stage, says it is using a plethora of data to source deals and help portfolio companies find early employees. The firm raised $500 million for a pair of funds in December.
  • Managing director and CEO Chris Farmer thinks of the firm as more of a Google than a venture fund—saying that they focus on finding opportunities using alternative data from their constantly shifting database of some 15 million companies and 10 million engineers.
  • Granted, the method comes at a premium. When asked if the fund uses the traditional 2-and-20 model—2% management fee and 20% carry—Farmer declined to specify, but added that everything is at a higher price.
  • Here is our conversation, lightly edited for clarity, on how to (or how not to) infuse data into venture capital’s experience.
  • How have you used the data to source potential deals?
    • Grammer.ly is one of them. They were way off the radar at the time, bootstrapped and avoiding venture capitalists, but we saw it in the credit card data. Similarly, OneSignal, a messaging platform—we picked it up based on the adoption rate of the product.
    • There is no single company score [to determine yes or no]. We use the data for benchmarking and for tracking companies, but it is a human decision at the end of the day. There are tons of times when we get founders based on referrals—and the data is then super helpful in terms of triaging.
  • SignalFire invests heavily in the seed stage—an area that traditionally doesn’t have a lot of data. How do you try to predict if someone has the potential to be a founder or early employee?
    • We tend not to talk about our data sources, so I’m trying to state the more obvious ones—but ==It’s everything from tracking people through multiple social networks, to academic papers, to patents, to regulatory filings, to H1B visas to company websites.==
    • We predict, for instance, the probability of someone changing their job. There are unknown billions of factors that go into someone changing their jobs—is the stock down, has a senior executive that you followed left the team, do you have a higher propensity for changing jobs than your peers, have you tuned up your resume—we predict that and put it in the hands of our founders so that they know not only if someone might be a good candidate [as an early employee], but how likely they are to move.
    • [In the case of finding founders], it also depends on the sector. What you look for in a social media company is completely different from in a storage infrastructure business. Social media tends to have young, first-time founders without a history—so the background is underweighted. It’s all about traction and virality. Twitter at the beginning didn’t have a super notable engineering team—they were known for the “fail whale”—but they had such a strong product-model fit that Twitter grew a stronger team over time.
    • If you’re building core database infrastructure, though, very rarely is it going to be a 22-year-old that has no history. It’s going to be someone with deep expertise: the guy who, say, built TensorFlow at LinkedIn or the woman who built the largest community of WordPress websites and is now building a WordPress infrastructure company—or, say, the team that built YouTube’s infrastructure that is now building PlanetScale, but they were totally anonymous engineers that were known inside [YouTube] to be total badasses, and we picked them up with the data that they had this massive open-source following.
  • What are the shortcomings of machine learning in venture capital?
    • We have invested in travel companies and the signals were positive. We didn’t predict the coronavirus pandemic. We also can’t predict when a competitor could pivot, or when a business model might not work. Early traction does not mean long-term success—but without early traction, it’s hard to have long-term success.
    • And at the end of the day, you can’t tell the grit and determination of a founder using data, nor can you see their vision from historical data. So you have to take a hybrid approach. Even if you’ve found opportunities [using data], you need people who can do traditional venture, who can assess people and understand where markets are going, and who can win the deal.
  • So much of this industry is about who you know and how you know them. How do you avoid the issue of baking in systemic discrimination in datasets?
    • We are gender- and race-blind. But there are also discriminatory elements we try to be cognizant of, such as rate of promotion within a company. We can’t change biases in history, so we try hard to weigh other things. For example, they may be big in the open-source community. You also can’t skew to tenure, because senior executives are always going to rank higher and you’re never going to catch the next Collison brothers [ of Stripe] because they have just never risen inside a company.
    • We specifically don’t use the data of who your investors are. Almost by definition, if another top investor is in there, you’ve missed [the deal]. Those firms make mistakes too—every one of them do. It’s not about where you went to school: There’s a lot of catch-up mechanisms. You may not have gone to Harvard, but you may have been a top person in the open-source community, or you may have great metrics around your business or great retention or customer dynamics.
  • How are your portfolio companies faring through coronavirus?
    • We have some that are doing fine—we have a large telemedicine portfolio or companies like ClassDojo that have significant tailwind. But some companies are going to have slow quarters, and then there are companies we don’t know how are going to get out of this and will need to go down to a skeleton crew and go into cockroach mode, particularly in the travel sector.
    • We’ve been generally bearish. That’s not the data speaking, but the general experience of having been through two downturns myself and my partner having been through five. We knew when we started that statistically we were likely to hit a correction, so we avoided capital-intensive businesses like autonomy even though we were super bullish on autonomous driving in the long term. We knew these businesses would depend on research and development dollars from the auto industry, which gets clobbered in every downturn.

Connections

• SignalFire · Chris Farmer — the subject.
• Data in Investing — the recurring thread this feeds.
• Read on April 17th, 2020.