In the field of entrepreneurial finance, both academic researchers and venture capital firms are exploring the use of data-driven approaches to the analysis of entrepreneurial projects. For example, using the data provided by Crunchbase, some researchers have developed machine learning models aimed at predicting the exit event of startup companies. However, these previous contributions have always looked at ventures as static entities over time, only considering the values assumed by the key variables at the time of data extraction. This paper aims to propose a new modelling approach, based on the analysis of the evolution of companies over time. The work considers a sample of 10,211 US-based companies, appropriately selected through a sequence of data processing activities. The rationale applied to reorganize the information and design a database ready to be used for a temporal analysis is described. In particular, each firm is modelled considering three different groups of features whose values change as the company evolve and therefore describe the key milestones achieved. In this regard, the number and amount of funding rounds over time, the number of investors involved and the number of patents obtained over the years are considered. To highlight the importance of the evolution of these variables over time, their statistical trends are reported within a 10-year time window from the companies’ foundation. Considering a binary classification problem aimed at predicting whether or not a startup exit event will occur, statistics are presented for the two groups of companies, those that have made an exit or not. Figures show how this approach makes it possible to achieve a greater level of detail on the characteristics of the companies, not otherwise obtainable without considering the time factor. The obtained dataset is then used to train a binary deep learning classifier designed to perform time series analysis. The results obtained confirm the effectiveness of the applied modelling strategy. The obtained model is in fact able to predict whether a company will make an exit within 10 years of its foundation with a recall equal to 93%.
A Deep Learning Model for Startups Evaluation Using Time Series Analysis
Francesco Ferrati
;Moreno Muffatto
2021
Abstract
In the field of entrepreneurial finance, both academic researchers and venture capital firms are exploring the use of data-driven approaches to the analysis of entrepreneurial projects. For example, using the data provided by Crunchbase, some researchers have developed machine learning models aimed at predicting the exit event of startup companies. However, these previous contributions have always looked at ventures as static entities over time, only considering the values assumed by the key variables at the time of data extraction. This paper aims to propose a new modelling approach, based on the analysis of the evolution of companies over time. The work considers a sample of 10,211 US-based companies, appropriately selected through a sequence of data processing activities. The rationale applied to reorganize the information and design a database ready to be used for a temporal analysis is described. In particular, each firm is modelled considering three different groups of features whose values change as the company evolve and therefore describe the key milestones achieved. In this regard, the number and amount of funding rounds over time, the number of investors involved and the number of patents obtained over the years are considered. To highlight the importance of the evolution of these variables over time, their statistical trends are reported within a 10-year time window from the companies’ foundation. Considering a binary classification problem aimed at predicting whether or not a startup exit event will occur, statistics are presented for the two groups of companies, those that have made an exit or not. Figures show how this approach makes it possible to achieve a greater level of detail on the characteristics of the companies, not otherwise obtainable without considering the time factor. The obtained dataset is then used to train a binary deep learning classifier designed to perform time series analysis. The results obtained confirm the effectiveness of the applied modelling strategy. The obtained model is in fact able to predict whether a company will make an exit within 10 years of its foundation with a recall equal to 93%.
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
