Aggregation of Diverse Approaches to Improve Predictive Power of Bankruptcy Models for Russian Firms

The chief aim of this research is to compare performance of different statistical models of bankruptcy prediction for Russian private small and medium-sized entreprises (SMEs) prior and subsequent to the global economic crisis of 2008 - 2009. We use the following methods: Linear Discriminant Analysis, Quadratic Discriminant Analysis, Mixture Discriminant Analysis, logistic regression, probit regression, Support Vector Machine, Classification Tree and Random Forest. We extend the definition of default to financial difficulties: we analyse voluntary liquidated firms together with firms that were liquidated as a result of legal bankruptcy. Our dataset comprises about 1,000,000 companies from the Ruslana database and covers the period from 2004 to 2012. We study four industries: construction, manufacturing, real estate activities and retail and wholesale trade. In addition to financial ratios derived from financial statements we include non-financial variables such as regional distribution, age, size and legal form into statistical models. Evaluation of the prediction performance is done with the help of out-of-sample forecasts. For out-of-sample forecasting we achieved specificity and sensitivity of about 0.7. Random Forest outperforms all other methods. Furthermore, adding non-financial variables improves predictive power, while mixing predictions does not improve the results. Finally, imputation of missing data slightly improves predictive power, but enables to significantly increase the size of both estimation and forecasting samples. This research will be of vital importance especially to banks and other credit organisations providing loans to SMEs.