The Comparative Analisis of Applicant´s Default Probability Evaluation Methods

As you can see, different methods to minimize borrower's credit risk were analyzed in this study. The importance of this research arises because banking system's stability is largely dependent on the reliance of the risk management, on which the entire banking business is based. Hereupon mathematical models of credit risk assessment were examined in this thesis work. On the example of real data a construction of scoring model of the clients’ assessment quality was carried out and its ability to forecast was checked as well. In addition, the contingency tables were used In order to construct logit model, which resulted in high quality scoring model. The ROC-curve was used to analyze the quality of the model. To estimate the probability of default by the method of logit regression the probability of correct results is 77.2% (in the training set) and 82,61% (on the test sample). This means that in the first case, 77.2% (82.61%) of the applicants for the loan will be classified as good borrowers and their loan application can be approved. It should be noted that when checked on the test sample with threshold limit of default set to 70% the model could recognize 396 of bad borrowers out 397 that were present in that test sample. Also, the method of self-organizing Kohonen's maps was studied in this work. The two methods give approximately the same result. Method of neural networks is easier since no data is needed to be cleaned, and secondly the results obtained using a standard regression are not always applicable to another sample, by geography, type, etc. hence, banks are forced to build a log regression for different geographical areas, types of customers, and so on, which is not always possible with a given sample size, so in this context, the use of neural networks to predict the possibility of customer's default seems to be more feasible than the others.