The Future of Venture Capital: Humans vs. Machines
As venture capital firms look to gain a competitive advantage in an increasingly crowded space, GV has turned to a computer algorithm, named "The Machine", to guide investment decision making.
If you ever wanted to see a collection of Unicorns, you might want to take a look at the portfolio of GV (launched in 2009 as Google Ventures), the venture capital arm of Alphabet, Inc. . GV’s current and past investments impressively include Uber, Nest, Lime and Jet; all private companies with over one billion-dollar valuations (Unicorns) .
Given my background in investing in private companies, I found myself fascinated with how GV’s investment process has yielded such successful results. Fortunately for my curiosity, a few months ago, Axios shined a light on GV’s investment approval process aid called “The Machine”, a machine learning algorithm that opines on potential investments . This begs the question: Is GV’s incredible portfolio a result of human investment judgement or The Machine?
According to Axios, The Machine is fed numerous inputs on an investment opportunity (i.e. industry, fundraise size, investors, etc.), then the algorithm scores it on a 10-point scale and gives each a green (scores above an 8), yellow or red rank . The Machine recommends only proceeding with green ranked investment opportunities. Given venture investments are in young companies without much history to analyze, investors tend to rely on human judgment (“gut”). Decisions left up to gut are prone to human biases and error that can derail a firm like GV from making successful investments . In order to try and gain a competitive advantage against other venture capital firms, GV is attempting to create a data-driven investment process that outperforms simply using gut. Unlike the issue with company specific data of young firms, industry, founder, round and investor data is almost limitless! Through The Machine, GV is using machine learning to help analyze the infinite amount of data surrounding an investment opportunity to find the factors that make certain investments successful. If cultivated correctly, The Machine would create a massive advantage for GV in its investment approval process.
In 2013, GV’s former CEO, Bill Maris, stated “We have access to the world’s largest data sets you can imagine… It would be foolish to just go out and make gut investments” . In order to maximize The Machine’s potential, GV is presumably focused on harnessing the information access they uniquely have as a subsidiary of Alphabet today (it is important to note GV declined to comment on Axios’ story). There are other venture capital firms trying to also use machine learning to improve their investment process (most notably Hone Capital and EQT Ventures [3,6]), but they all lack the incredible data sets privy to Alphabet. With vast amounts of data, GV needs to now focus on deciphering between signals for good investments and noise to build The Machine to be a flexible model that is able to look at a myriad of different investment opportunities . As GV continues to fine tune its machine learning process, it certainly seems that the longer-term goal is to have The Machine be the “de facto investment committee” , removing the gut aspect of investing. Even with superior results, a machine-based process being the final arbiter of investments will take time to get full buy-in.
The benefits of a data-driven approach to the venture investing is undeniable, but relying entirely on algorithms, like The Machine, can also bare risks. In the more immediate term, GV needs to focus on the casual relationship that data inputs create and how they define successful investments. For example, including previous/current investors as inputs can create a bias to invest in firms with backing from prestigious capital providers (especially if success is measured via a company’s record of raising money). The reality is that these types of companies might just be able to raise capital due to their enhanced reputation, not because they are actually a good investment opportunity. Moreover, some of venture capital’s most lucrative opportunities are the ones that might seem most risky to a machine-learned process looking for common fact patterns . As to making final investment decisions, there should always be an integration of machine-learning and gut. Through studies conducted by the aforementioned Hone Capital on venture investments from 2015, combined human and machine teams sustainably outperformed teams with only humans or machines (which each interestingly performed at the same level on a standalone basis) . Machine learning provides relatively consistent result, while gut provides potential for home-run investment. In my opinion, the right mix of both is essential for GV to succeed in the future.
As GV moves forward with The Machine, there are a few critical questions. When a human investment professional and The Machine disagree on an investment opportunity in the future, what should GV do? Furthermore, given the shift in venture investing toward more machine-learning processes, what are the biggest differences between today’s human venture capitalist and the future’s?
1) GV. “Our Story”. https://www.gv.com/, accessed November 2018
2) GV. “Portfolio”. https://www.gv.com/portfolio/, accessed November 2018
3) Primack, Dan. “Scoop: Inside Google’s Venture Capital ‘Machine’”. https://www.axios.com/, July 18, 2018, accessed November 2018.
4) Yeomans, Mike. “What Every Manager Should Know About Machine Learning”. Harvard Business Review Digital Articles, July 7, 2015. pp.5-6.
5) Miller, Claire Cain. “ Google Ventures Stress Science of Deal, Not Art of the Deal” New York Times, June 23, 2013. https://www.nytimes.com/2013/06/24/technology/venture-capital-blends-more-data-crunching-into-choice-of-targets.html?utm_source=newsletter&utm_medium=email&utm_campaign=newsletter_axiosprorata&stream=top.
6) Wu, Veronica. Interviewed by Chandra Gnanasambandam. McKinsey Quarterly. https://www.mckinsey.com/. June 2017
Student comments on The Future of Venture Capital: Humans vs. Machines
In the IBM Watson case, we saw that the “promise of AI lies in the combination of man and machine”(https://www.ibm.com/blogs/think/2017/05/deep-blue/). IBM’s Deep Blue highlighted that teams with both machine and human performed better in chess games as compared to exclusively human or machine teams. I would be interested to understand how, if at all, Google will approach integration of The Machine with human investment committee members. Will investment committee members only discuss opportunities given a green rank? Will The Machine’s recommendation be given the same weight as other committee members? How will Google structure the investment committee going forward (ie. Does The Machine sit on the investment committee and actively vote, or does simply recommend opportunities for them to discuss)?
To answer your question, I think when machine and human reach to different conclusions, it’s important to understand where the difference comes from, similar to understand why Watson asks the question “what is Toronto” in Jeopardy -for example does it come from the machine putting too much weight on former investors, etc. But I assume the challenge would be it’s difficult to answer that question in most cases. Another potential concern of machine learning-based investment decision is how can machine quantify certain unquantifiable but important aspects, such as the determination and execution capability of the founding team. Maybe this is another area where the human can provide input based on their interview with the founders, etc.
Thanks for sharing, AEG91.
Would be curious to know how Google defines a successful investment. Oftentimes an exit valuation will only be announced to the public if it’s a strong valuation or an IPO (i.e., databases like Pitchbook and Crunchbase are extremely susceptible to survivorship bias). Further, valuations which funds invest at are often not publicly disclosed. Without these goal posts (i.e., the ability to calculate an investment return), how can GV know which investments should be given a green score?
It’s particularly interesting to think about how “The Machine” might help combat biases investors have. Women, unfortunately, have a tougher time raising venture capital and I’d imagine having an objective check on human investors can only help in this regard (so long as the algorithm isn’t taught biases, of course).
However, as I read, I begin to think of highly backed startups which have failed. For instance, Yik Yak, Beepi, Quixey. As I think about the stages of investing, from VC to growth equity to more traditional private equity, I often associate the risk at each “stage” with the amount of information available. If The Machine makes investment decisions based on data such as industry, founder, round and investor data, is that not the same as making investment decisions based on “past performance?”
The competitive advantage of The Machine is its ability to process massive amounts of information. But, if the information doesn’t exist, I am skeptical that The Machine can accurately predict future success. Therefore, I would use The Machine in later rounds (Series B, C, etc.) and keep the “gut” in angel investing.
Right now, supervised learning machines are only as good as the information fed to them. While that could change in the near future, the human investor’s ability to gauge the idea, motivation, intellect, and capability of the startup’s CEO is the key to finding a unicorn.
I’d definitely be concerned about the problem you raise about whether a “machine learning investor” would be able to identify a successful investment that bears huge risk, without regressing to the mean. I also suspect that a large part of the venture investing decision is based on the quality of the team, which would be represented by qualitative factors that aren’t necessarily perceptible in an algorithm. Machine learning could identify patterns (e.g., Stanford computer science major = good), but it would be difficult to act on inputs such as someone’s trustworthiness, value system, etc. Regardless, this is a great case for the combined power of human and machine, and there definitely seems to be a place for the machine to outperform humans in making purely data-driven decisions without bias in the context of an investment decision.
I did not know that VC firms were using machine learning to guide their investments and found that fascinating. I was impressed by the scoring system. It makes sense that Google would use the vast data it has available. The question of how much weight to give to the machine’s recommendation versus a human’s is a tricky one.
I agree that the best approach is to combine human investment professional and The Machine to analyze investments because they complement each other very well. The Machine relies on historically available data and may not be appropriate for coming up with new innovation, while human investment professional can be more creative and hence innovative. I do not think there is a right answer whether to always choose one side between human investment professional and The Machine when they disagree. However, I think it is extremely important to try to understand why they disagree and start analyzing from there. This approach should yield good results as it allows human to try to understand why The Machine disagrees and initiates further analysis in order to make better decisions.
Very interesting article.
Given these algorithms largely rely on past data to predict future performance, I wonder whether they will end up rejecting many more companies than a human would have otherwise simply because there have not been similar companies raise capital in the past. I wonder to what extent these algorithms will reduce the upper bound of investment success (i.e., fail to approve the biggest unicorns) in turn for taking on less risk. If the industry moves towards this, I could see the venture industry suffer from a lack of capital for the most risky or unproven concepts…the ones that may be backed today based on someone’s “gut”.
The balance between deal sourcing and diligence are extremely important in the space. Bill Maris, the founder of GV, wrote in an article in 2015 about why the company decided to pass on Theranos, a unicorn healthcare company that turned out to be a scam . From The Machine model, I’m sure Theranos was flagged as a “go” due to its industry and funding profile. The important thing is to not solely rely on the data of this model for due diligence. By sending in experts in the field, GV was able to pass on the investment and save millions of dollars in losses. The combination of human and machine-powered tasks in the venture world lead to the best outcomes rather than simply one or the other.