Deep learning models as decision support in venture capital investments : Temporal representations in employee growth forecasting of startup companies

Abstract: Venture capital investors are constantly exposed to high levels of risk in investment scenarios. To lower that risks, various decision support tools can be exploited, such as machine learning models aimed at predicting startup success. In the related work, we observe a lack of deep learning models and solutions that cater to the time-dependent features that are naturally present in the data. This thesis compares two state-of-the-art deep learning models, and inherently their different temporal representations, in their ability to capture long-term dependencies in both time-dependent and static data. We consider the sequential Long Short-Term Memory (LSTM) and the attention- based Transformer in the comparison, motivated by their popularity and contrasting approaches to temporal representation. The method solves a binary classification problem: given time-dependent and static features describing a company, predict whether this company will have intense employee growth in the coming period. The thesis raises questions regarding the suitability of the Transformer under certain data conditions, and establishes that the LSTM is successful in representing long-term dependencies in the data, both with and without the influence of static features.