SME Bankruptcy Prediction Using Convolutional Neural Networks

Abstract

Failure to repay obligations to creditors, whether credit institutions or business partners, causes serious economic problems not only for the debtor but also for its stakeholders. Preventing this problem requires identifying the potential threat. This paper explores the potential use of Convolutional Neural Networks (CNN) in identifying businesses at risk of bankruptcy. It is based on a graphical representation of differences in company performance and selected macroeconomic indicators. In our research, we used the GoogLeNet neural network architecture. The approach used allowed to display the financial situation of a company so that the generated CNN could identify active companies and companies at risk of bankruptcy with high accuracy. The procedure was applied to data of companies operating in the construction industry in the Czech Republic. The accuracy of the model was evaluated using receiver operating characteristic (ROC) curve and area under the curve (AUC). The use of CNN has yielded high forecast accuracy, demonstrating the ability to efficiently process graphical displays of financial data and capture differences between healthy and risky companies. The indicators identified in the constructed model can be used as input variables in an early warning system for financial distress.Failure to repay obligations to creditors, whether credit institutions or business partners, causes serious economic problems not only for the debtor but also for its stakeholders. Preventing this problem requires identifying the potential threat. This paper explores the potential use of Convolutional Neural Networks (CNN) in identifying businesses at risk of bankruptcy. It is based on a graphical representation of differences in company performance and selected macroeconomic indicators. In our research, we used the GoogLeNet neural network architecture. The approach used allowed to display the financial situation of a company so that the generated CNN could identify active companies and companies at risk of bankruptcy with high accuracy. The procedure was applied to data of companies operating in the construction industry in the Czech Republic. The accuracy of the model was evaluated using receiver operating characteristic (ROC) curve and area under the curve (AUC). The use of CNN has yielded high forecast accuracy, demonstrating the ability to efficiently process graphical displays of financial data and capture differences between healthy and risky companies. The indicators identified in the constructed model can be used as input variables in an early warning system for financial distress.