Predicting start-up success with machine learning

Dissertation presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Knowledge Management and Business IntelligenceStart-ups are becoming the motor that moves our economy. Google, Apple, or more recently Airbnb and Uber are companies with tremendous impact in worldwide economy, social interactions and government. Over the past decade, both in the US and Europe, there has been an exponential growth in start-up formation. Thus, it seems a relevant challenge understanding what makes this type of high-risk ventures successful and as such, attractive to investors and entrepreneurs. Success for a start-up is defined here as the event that gives a large sum of money to the company’s founders, investors and early employees, specifically through a process of M&A (Merger and Acquisition) or an IPO (Initial Public Offering). The ability to predict success is an invaluable competitive advantage for venture capitals on the hunt for investments since first-rate targets are those who have the potential for growing rapidly soon, which ultimately, allows investors to be one step ahead of competition. We explored the world’s largest structured database for start-ups – provided by the website CrunchBase.com, with the objective of building a predictive model, through supervised learning, to accurately classify which start-ups are successful and which aren’t. Most of the studies regarding the prediction of processes of M&A or an alternative definition of a company’s success tend to focus on traditional management metrics provided by financial reports and thus using a low number of observations compared with the present study. As technologies of information evolve it became possible to achieve highly reliable results in data analysis by manipulating it with complex machine learning algorithms or data mining techniques to define features and characterize robust models. Further developments on previous studies such as the development of new features and a new definition for the target variable were applied. Using Random Forests on our dataset, a general model (as including all categorical features) achieved a True Positive Rate (TPR) of 94%, which is the highest recorded with this data source, and a False Positive Rate (FPR) of 8%. The author also generated models per each category of a company to provide results comparable with previous studies the values achieved ranged between 61% and 96% compared with 44% and 80%. As a novelty, models for each of the five geographical regions selected (all from USA) are provided, with TPRs ranging between 90% and 96%. The new features, focused on the impact of venture capital in a company, proved pivotal to the overall performance of the models by being some of the most important to the final models showing the critical importance this type of investment has on these ventures