Forecasting Bankruptcy of Small and Medium-Sized Businesses

In this paper there are several classical (linear probability model, logit regression) and modern machine learning bankruptcy forecasting models (K-nearest neighbors, Support vector machines and Random Forest) built with up to one year prediction power on a sample consisting of Russian small and medium-sized businesses (SME) and proves that traditional insolvency prediction models have become outdated, and should be replaced by statistical learning ones for the purpose of predicting likelihood of a firm going bankrupt. The study resulted in the Random Forest model with features selected by the Least Absolute Shrinkage and Selection Operator technique being the dominant one. It performed with the lowest Type II error rate for the testing set. Moreover, the tuned Random Forest models constructed by the author of this paper outperformed the Random Forest model proposed by Barboza et al. in 2017 and any of the model examined by Fedorova et al. (2013) in the terms of accuracy, sensitivity, specificity and Type II error rate. Key words: SME, bankruptcy forecast.