Prediction of the Disappearance of Companies From the Market in Bogotá, Colombia Using Machine Learning

Introduction

The recent advances in machine learning and data availability have increased the capacity to create computational models for decision-making in the corporate realm. These models will have the ability to reduce uncertainty and make informed decisions for the organizations' benefits and the people in general. Economic stability is of paramount importance for society, and the use of these methods can be critical from both private and public perspectives.

The previous global financial crisis has shown how the lack of control from governmental institutions can hardly impact all societal network levels, especially the lower layers. Many regulatory frameworks arise aftermath, targeting mainly financial institutions and insurance operators (Acemoglu, 2015). Thoughtful modeling of companies' future capabilities to remain competitive in the market has become a necessary element, not only from the management of the individual companies but also from governmental and financial institutions. Companies can use these models to establish future business strategies or strengthen the current ones. Banks, public institutions, and private investors can predict future capabilities or assess future business risks, capital investing, loans, subsidies, among others. The goal is to keep the market's stability, generate profit, and in general, produce a steady and continuous wealth for the whole society.

The rapid development of centralized information databases and the processing power and the accessible specialized software have shifted the classical analysis of companies from statistical modeling to more data-centric approaches since Altman's seminal work (Altman, 1968). A decade later, Ohlson (1980) pioneered machine learning strategies for corporate bankruptcy modeling. Nowadays, most predictions are made by adopting optimal classification algorithms, ranging from logistic regression, support vector machines, and decision trees to more complex deep neural network modeling (Qu, 2019). Moreover, other methods such as genetic algorithms (Jiang, 2009), particle swarm optimization (Chen, 2013), or even reinforcement learning (Espinosa-Leal, 2020) algorithms have been proposed in combination with machine learning to improve the models' prediction capacity. Furthermore, an additional challenge is finding the right models where explainability and interpretability balance with the prediction accuracy and the possibility of the ethical assessment to avoid biases in the algorithms' outcomes.

In this work, we study and present the results concerning the prediction of the disappearance of companies registered in the commerce of Bogota, Colombia, during the period 2016 until 2019. This data is updated every year. Therefore, if a company fails or does not add its name in the chambers' registry, it is automatically declared closed or out of business starting that year. This fact can be considered an arbitrary generalization. The original information contains other valuable information such as size, legal type, economic sector, number of employees, as well as other features. This data can be used for analytics purposes using cutting-edge statistical and artificial intelligence methods. Hence, we model companies' disappearance following the registry, as mentioned earlier, if they have or have not been renewed in a specific year. Upon a full statistical description of the dataset, where each variable is presented and described, a feature selection process of the data is performed. In the final stage, we split the dataset into two subsets, one for training and the other for testing, then fit four different machine learning methods: Logistic Regression (LR), Extreme Learning Machine (ELM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) with a complete parameter's optimization to ensure the final model is optimal. The obtained scores are presented with the testing subset.