National Culture and Venture Capital Monitoring
Neural network implementations of venture capital/corporate innovation research
Tools used: Keras, Hyperas, Cox-NNet (Theano and Numpy), SciKit imbalanced-learn for SMOTE, and Scikit manifold for the T-SNE visualization of before-and-after SMOTE scatterplots. R's survival package for the original inferential Cox-hazard analysis, the dplyr package for data munging, and Seaborn and GGPlot2 for plots.
This repository contains a series of predictive objects made using the data from my finance research paper "National Culture and Venture Capital Monitoring." This research studies how national-level cultural differences between venture capital firms and foreign portfolio companies affect the success and innovation of those companies (with national culture measures by Geert Hofstede).
Although inferential statistics are the standard in academic business research, predictive analytics are increasingly being used in practice and in the area of venture capital. For example, Google Ventures actively uses machine learning to predict whether or not a venture capital investment should be undertaken. Other venture capital firms are beginning to follow suit as well as venture capital shifts overseas. To that end, I am making predictive objects available for practitioners to use to make my research more useful to VC firms that are undertaking foreign investments. A voluminous management literature surrounding Hofstede's cross-cultural dimensions provides prescriptions for taking action in the event that the neural network objects in this repository consistently predict positive or negative chances of innovation and/or success.
Two separate directories in this repository contain code and neural network weights for the two separate analytic methods from this research:
- Difference-in-difference analysis
- Cox Proportional Hazards models
The difference-in-difference analyses are similar to linear regressions and are suitable for use with Keras/TensorFlow. Hyperas hyperparameter tuning of network depth, network width, and activation functions was conducted for these particular neural networks as well. The Cox-PH analyses are more exotic, and required the use of a different neural network library based on Theano and numpy; Cox-NNet.
The choice of a neural network predictive versions of the research analyses fits the data as well as my background. My dissertation chapters both make use of left-censored tobit models, which are essentially the same as the rectified linear unit (ReLU) functions that are used to activate hidden layers of neural networks. In 2018, I co-authored merchandising research that makes use of structural equation models, which are essentially Bayesian graph neural networks that are inferential in nature--networks that do not predict on data and instead focus on the coefficients between nodes. The neural networks in this repository simply represent deeper, predictive versions of analytic methods that I have been applying for years.