Data-Driven Investment: Formalizing The Early-Stage Venture Capital Process Using Machine Learning

Abstract

We frame venture capital as an optimization problem and use its constraints to motivate using machine learning techniques to make investment decisions. Through this theoretical framework for sourcing and evaluating startups, we perform an empirical analysis to show how we can use data to predict the future success of companies. We create a dataset of independent variables such as past investors and founder alma maters and use it to regress on the amount of funding the startup will raise in the future. Using precision and recall as our evaluation metrics, we compare the performance of a variety of models in the task of predicting startup success. We find that both a regular logistic regression and a kernel logistic regression provide the best generalization error. When analyzing the coefficients of our regression, we find that some positive significant correlations between certain investors and schools with future funding raised. In our analysis, we also consider notions of fairness and bias in our data and methodology through the use of protected attributes.