To shine or not to shine: Startup success prediction by exploiting technological and venture-capital-related features

Abstract

Startups play a crucial role in driving economic growth, job creation, regional development, and technological innovation. However, they often encounter risks stemming from uncertainties in technology, unfamiliar markets, and limited resources. Given these challenges, effectively predicting startup success, defined as achieving a successful exit within a specific observation window, is vital for shaping investment decisions and facilitating the strategy formulation of stakeholders such as venture capitalists and startups themselves. In this study, we are interested in startups in high-tech industries. Existing startup success prediction research primarily focuses on exploiting features related to company profile, funding, founder, and top management team, and pays less attention to technological and venture-capital-related (VC-related) features that are prominent to high-tech startups. Furthermore, prior studies do not assess the effectiveness of startup success prediction over different prediction time points. To address these gaps, we design a startup success prediction method that incorporates three categories of features: basic, technological, and VC-related. For empirical evaluation purposes, we collected a dataset comprising 4415 startup cases and their corresponding feature values from the Securities Data Company's VentureXpert database and the USPTO database. Our evaluation results indicate the superiority of our method over the literature model, suggesting the predictive value of our proposed technological and VC-related features. Our results also show that the VC-related features are more salient in predicting high-tech startup success than the technological and basic features. Finally, our exploratory study of the deep learning approach reveals that using deep learning (e.g., graph convolutional network) to extract VC features automatically may not enhance prediction effectiveness at the very early stage of startups but shows a potential advantage over statistical and machine learning methods at a later prediction time point due to the increased number of VCs investing in the startups.