German startup Anydesk was already cash-flow positive when the founders opened an email from Stockholm’s EQT Ventures. Anydesk’s founders weren’t looking for outside investment, but were nevertheless intrigued. The partners at EQT Ventures’ two-year-old fund, with more than $650 million to invest, are not the erstwhile bankers who typically run European VC funds. Instead, they’re all company founders or high-level executives from European tech firms such as Spotify, Huddle, and Rebtel.
At a meeting in Berlin with venture lead Ashley Lundström, Anydesk cofounder Philipp Weiser learned about Motherbrain, a machine learning system EQT Ventures built to find under-the-radar startups. “She told us we were among the first companies that were discovered by this software,” says Weiser. In May, AnyDesk, which sells remote desktop software powered by a proprietary compression system, closed a funding round of $7.6 million with EQT Ventures.
Whether the money EQT is putting into Anydesk will turn into a success story remains to be seen. But the firm has applied the Motherbrain algorithm to historical data and shown that it would have identified some of today’s highest-flying tech companies as promising investment candidates before they became phenoms. For instance, the system would have flagged Airbnb, Snapchat, and Stripe when they had received only angel and seed funding.
Henrik Landgren was Spotify’s VP of analytics before he joined EQT Ventures. “You have millions of companies out there,” he says. “How would you ever know who to talk to? The old way is to talk to the ones that seek you out through your networks, but the more modern approach is to use the latest technology and data and algorithms to proactively reach out to the ones that have the highest probability of becoming good investments.”
Motherbrain monitors several million companies using financial data such as funding, web ranking and app ranking data, social network activity, and much more. EQT Ventures continuously adds data about its own assessments of companies in order to train Motherbrain to focus on the right opportunities.
The software is used during every stage of the investment process, but its most important function is prioritization: suggesting which companies the fund should look at now. “Anydesk came out clearly because of the traction that they show in their metrics,” says Landgren. “They were not experienced fundraisers. They were a really smart, clever team that built this product that had amazing traction.”
Motherbrain also speeds up the assessment of a company once it’s on the fund’s radar. Its ranking is used even if the source of the recommendation is not Motherbrain itself, and the system also contains useful information such as competitors and market size. It can even help companies once EQT Ventures has invested in them, because it contains so much data about investors, competitors, emerging technologies, and trends in the market. “For example, with B2B companies, we can use Motherbrain to help them to find leads for new customers,” says Landgren.
EQT Ventures is not the only VC fund to use data analysis. Most funds at least gather basic company data and create filters, rules, or trend alerts based on that data. InReach Ventures, which invests in early-stage European technology companies, also harnesses machine learning for the discovery process. San Francisco-based SignalFire has been using a data-driven investment model since 2013. But Motherbrain is notable both for the sophistication of the system and the size of the fund that it helps to allocate.
How Motherbrain works
Motherbrain uses a mix of unsupervised and supervised deep learning algorithms. Unsupervised learning algorithms find significant patterns in data without any external guidance. Supervised learning algorithms require labeled training data. If the training data contains examples of animals labeled as “cat” or “not a cat,” for example, an algorithm tries to learn the characteristics of a cat in order to predict whether new animals are cats or not.
In the case of venture capital, data can be used to allocate companies to market sectors, a basic task conducted by any VC firm. The text a startup uses to describe itself and the sectors that other people have defined can help in this process, but the data is often noisy and contradictory, especially for emerging technology and sectors. Motherbrain uses unsupervised learning to find clusters or categories of companies. EQT staff then label the companies in each cluster–for instance, identifying a cluster as containing blockchain companies–and the labeled data is used to train a supervised algorithm. Motherbrain can then automatically categorize new companies.
“That’s actually a way to yield some meaningful information from very noisy data,” says Landgren. “You can build fewer categories that have better predictive power if you train the models right based on the noisy data.”
The evolution of a company over time–and how its relevant metrics change accordingly–is also an important factor in the assessment process. A time series is a series of data points indexed in time which captures this evolution. Motherbrain can learn from time series data to predict how a company will evolve in the future, based on the performance of companies with similar metrics. “For example, did they raise money fast or slow?” says Landgren. “The web trends, the app rankings, all those different things, are actually even more interesting when you can take time series into consideration.”
Landgren is also proud of the state-of-the-art data infrastructure that EQT Ventures has built for Motherbrain. A dedicated development team of five people works closely with Google to leverage technology like Kafka, Kubernetes, Google BigQuery, and BigTable.
“When the developers work, they’re constantly finding that they have are questions that no one else has asked yet,” says Landgren.”Then you feel like you’re really on the edge. We’re building an AI decision-making platform for decisions that involve humans and very high-volume data sets together. I think that combination is very new. We’re in the frontier.”
Big data, big problems
Motherbrain is far from perfect. Like any machine learning system, its predictions are only as good as the data it is trained on, and that data will always be incomplete and sometimes downright inaccurate. “We do see companies that have no data–they are in stealth mode–which even Motherbrain can’t see, so that’s going to be a problem,” says Landgren.
By stitching together multiple data sources EQT Ventures tries to increase the coverage of the system and the overlap between different data sources. Since the firm’s investors use Motherbrain every day, they can correct discrepancies between data sources themselves when they spot them.
Landgren is also keen to emphasize that Motherbrain is very far from making final investment decisions. “To build models where you can find great companies is great, but it’s not as if you just press a button and then you have the investment,” says Landgren. “It requires much, much more work.” Much of that work is in the traditional investor realm of building relationships. “It’s about knowing which relationships to build and when,” says Landgren. “That’s what we use Motherbrain for.”