Comparative Analysis of Methods for Forecasting Bankruptcies of Russian Companies in the Non-Financial Sector

Currently, instability of companies in the construction sector more often becomes a subject for controversial discussions. This paper is devoted to comparison of abilities of various methods to predict bankruptcy of construction industry companies on the one-year horizon. The authors considered the following algorithms: logit and probit models, classification trees, random forests, artificial neural networks. Special attention is paid to peculiarities of the training machine learning models, impact of data imbalance to the predictive ability of models, and analysis of ways to deal with these imbalances. In their study, the authors used data on Russian construction companies for the period from 2011 to 2017. As a result, it was concluded that the models considered show acceptable quality for use in forecasting bankruptcy problems. It is revealed that neural networks outperform other methods in predictive power, while logistic regression models in combination with discretization follow them closely. At the same time, negative impact of the training set imbalance on predictive ability of the model was not found. In addition, significant impact of non-financial indicators on bankruptcy probability has not been confirmed.