Laboratoire d'Informatique de Grenoble

Abstract:

=====================

Startups play an increasingly important role in the modern economy. In this thesis, we study startup valuation and fundraising problems with machine learning and causal discovery methods. After reviewing the existing machine learning approaches to startup success prediction and the literature on startup valuation factors, we present a domain adaptation-based approach to predict startup valuations in funding rounds with known funding amounts. We show that funding rounds in which startup valuation is announced to the public are statistically different from those in which the valuation is kept secret. We mine a novel data source, Companies House, to learn the startup valuation in the later funding rounds and show that domain adaptation methods yield the best results for our task. Further, we collect a rich dataset of United Kingdom startups and their valuations and discover which variables make the best valuation predictors. Also, we apply causal discovery methods to learn which variables, directly and indirectly, affect startup valuation. We draw the connection to the previous startup valuation factors research and provide evidence for further theoretical studies. Finally, we propose a method for predicting whether a startup will secure a funding round based on publicly freely available information on the web. We propose methods to collect information about the startups and their funding rounds from different sources. Since it is impossible to collect the information about all the funding rounds, we propose to tackle the funding round prediction problem in the positive-unlabeled setting and show that this setting is beneficial for the neural network model.