Developing the Bankruptcy Prediction Models for SMEs

This paper considers the problem of bankruptcy prediction of small and medium enterprises in the agricultural industry based on financial statements analysis. The definition of default is expanded to the fact of termination of business due to serious financial difficulties, including liquidation and exclusion from the register. The generated sample contains the financial statements of 1,407 bankrupt companies and 15,900 healthy companies from 2015 to 2017. On this basis, two classes of models are compared: logistic regression and random forest algorithm. Explanatory variables are chosen from the following categories: profitability, coverage, liquidity, activity, and leverage. The heterogeneity of firms by age, size, and region of activity is also taken into account. The quality metric of models is the expected losses from incorrect classification, which is a weighted sum of the costs of type I and type II errors. Random forest outperformed logistic regression, demonstrating high quality on cross-validation and out-of-sample validation. Meanwhile, the models have similar classification principles, mainly based on indicators of profitability and business activity. These results display the presence of non-linear relationships between explanatory variables and the fact of bankruptcy, as well as the applicability of machine learning models to the problem of bankruptcy prediction.