Econometric Modelling of Retail Credit Risks Considering Time Effects

As part of this work was to analyze the direction of the banking business for the private Russian bank, namely, information about defaults on retail credit cards. Data was analyzed in terms of the structure of levels of defaults depending on different temporal parameters. We formulated the main research hypotheses about credit risk modeling, which were tested in the following chapters on the basis of a visual analysis. First solved the problem of classification of borrowers on "good" and "bad" based on the individual characteristics of known application forms for credit and credit history. different classification techniques were tested: the methods of logistic regression, linear and quadratic discriminant analysis, the method of k-nearest neighbors, decision trees (implementation of the algorithm of a random forest) and neural networks. For comparison, the forecast accuracy analysis method was used ROC-curve based on the values ​​of the AUC. It should be addressed risk function simulation problem of default, depending on the time and individual characteristics, which are used to model survival (survival models). We were introduced proportional hazards model and accelerate the time for which it was possible to use parametric and non-parametric approach (the assumption of a baseline hazard function).