Modeling the Probability of Credit Default of Clients of Microfinance Organizations: The Case of One MFI

Microfinance organizations have become widespread in the crisis years, issuing microloans (up to 100000 rubles) at high interest rates almost without documents. Today, the Central Bank of Russia actively regulates this market, more and more tightening requirements, limiting rates and pennies on loans. This necessitates the formation of a new strategy for assessing the risk of non-repayment of a loan or loan, based on the prevention of delinquency on the part of customers. To do this, first, it is necessary to obtain more informative data about customers, without complicating the relationship with them. Secondly, it is necessary to have a good understanding of the possibilities of certain methods of classification in solving various problems of evaluating potential customers. The authors of this study analyze the importance for the clients classification quality of those indicators that are traditionally collected by MFIs, as well as the importance of some new indicators based on data from social networks. In this case, the importance of indicators is interpreted in the context of specific classification algorithms (methods).To model credit default (delay of more than 30 days), the authors use several algorithms for constructing classification trees – CART and C 4.5 algorithms, logistic regression and Random Forest algorithm. Modeling is carried out based on a sample of customer profiles of real MFI. Ambiguous results were obtained. Depending on the formulation of the problem of classification of customers have advantage different algorithms descriptive analysis (CART, C4.5, Logit). At the same time, as you might expect, the non-interpreted predictive algorithm “Random Forest” provides the best quality of forecasts. According to the results of the analysis, it was revealed that the credit history of the borrower, as well as his age, is of great importance for the classification of MFI clients. Gender had no impact on the classification results. In addition, data from social networks turned out to be unimportant.