Can startups be designed with algorithms?

Working Paper (coming soon)

(with Rembrand Koning)

Prediction tasks are fundamental to startup team design. To create a high-performance team, founders, incubators and venture capitalists must use available information about potential team members to make predictions about how a new team might perform. Presently, team design in such contexts relies on limited information and simple heuristics. We introduce and evaluate a machine learning based framework for data-driven team design. Our empirical evaluation leverages data from the `Innovate Delhi’ bootcamp and field experiment. The structure of the experiment, which collected detailed data on 112 individuals and the performance of 117 teams in which they worked over three weeks, provides an opportunity to evaluate the proposed framework. We find that out-of-sample predictions can explain between 10% to 15% of the variation in the actual performance of teams. The best performing models use `social’ information about team members, specifically, their social networks and evaluations by peers on previous teams.