--- title: "Soft Information versus Bias in New Venture Finance: Machine Intelligence versus Human Judgment" authors: ["Christian Catalini", "Chris Foster", "Ramana Nanda"] venue: "Work in Progress" year: 2017 area: entrepreneurship featured: false abstract: "Across thousands of applications to a top global accelerator, expert scores barely correlate with startup outcomes — and aggregating scores cancels what useful variance experts have. A machine-learning model freed from imitating expert judgment selects on broader dimensions and picks markedly stronger portfolios." --- In the paper, we explore the types of errors domain experts make in screening startups. When outcomes are extremely skewed even skilled evaluators may struggle. Score aggregation - often the default within angel groups, accelerators and VC firms - only exacerbates the issue by cancelling useful variance. Over time, pattern matching may also lead experts to favor idiosyncratic types of ventures. Our analysis of a large set of thousands of applications to a top global accelerator shows little correlation between expert scores and startup outcomes. When we then train a machine learning model to replicate human decision making, the algorithm achieves similar performance to the humans by heavily weighting keywords related to the founding team. Once unconstrained from imitating human judgment, the algorithm selects a broader set of dimensions, and delivers dramatically higher returns. Our results highlight that while the team might be important for venture success, early-stage investors may be overemphasizing it. Moreover, the higher performance achieved by the machine learning algorithm hints at the existence of additional traits of high potential startups that are currently understudied.