| Abstract [eng] |
Small and medium enterprises take important place in nowadays economics, yet due to dynamic market they are often considered to be on high risk to go bankrupt. Therefore, to guarantee functioning of an enterprise a strong position in financial safety is essential. To be fully prepared for all planned expanses enterprises are willing to take loans from banks. Banks wants to lower the risk of giving loans to companies that will not be able to pay their payments on time or will go bankrupt. This paper is going to investigate the issue of enterpresis going bankrupt by creating model that would predict probability of default based on behavior data. The main objectives of this paper are: define the bakruptcy of small and medium enterprises problem and review ralated works to gather knowledge about used methods to predic probability of default an lastly create a binary classification model for bankruptcy of enterprise. Models were built on one of Scandinavian bank‘s behavior and transactions history data. Three classification algorithms (random forest, neural networs, xgboost) were tested using 182 variable. Best model was obtained by using xgboost algorithm with 5 fold cross validation with the result AUC = 0,817, while random forest‘s AUC = 0,763, neural network‘s AUC = 0,69. After investigating most important variables extracted from best performed model there can be made a conclusion that out of 182 variables related with costumers behavior, dutifulness for paying payments for their loans are more important and informative than behavior with money compared with liabilities and income. |
