Predicting company bankruptcy using artificial neural networks. : Visualization and ranking of key features

Abstract: This paper presents a deep learning model that challenges what is known in the financial field of company bankruptcy. Specifically, a Multilayer Perceptron (MLP) model for predicting corporate bankruptcies is constructed and analyzed to visualize which input parameters that are most important for the accuracy of the model. The model uses approximately 55,000 rows of data, data cleaning and hyperparameter optimization to achieve an average accuracy of 82.8% and a standard deviation of 0.0678% after 120 epochs and 30 trials, which is an outstanding result. The model outperformed two support vector machine (SVM) models that were compared and showed good generalization ability. However, the non-linear SVM model generated 20.48% false positives and had an accuracy of 71.96% while the MLP model generated 25.1%. So if it is more important to reduce the number of false positives, the SVM model may be preferable despite having a lower accuracy. After analysing the input parameters, it was found that the number of employees, the turnover group and the equity ratio were among the inputs that had the greatest impact on the bankruptcy prediction. This thus resulted in the conclusion that these parameters could potentially be the most important factors to look at when analysing whether a firm will go bankrupt or not.