|Abstract:||The presented dissertation considers the oldest form of financial risk, namely the credit risk. When modern banks are exposed to credit risk in the context of the highly interconnected financial system, it creates room for the emergence of systemic risk. That being said, the capital of the bank is set in the function of the buffer against unexpected losses and finally preventing transfer of losses among banks. With the implementation of Basel II capital accord, the amount of capital is directly linked to the risk of a bank’s single portfolio. From this definition the need for very precise quantification of risks arises. Good quality of risk evaluation and the exact estimation of portfolio’s risk parameters is a prerequisite for effective capital regulation at the microprudential as well as macroprudential level. An improvement of the credit classification model quality would therefore contribute to an improvement of the process of capital requirement’s assessment. In this thesis we propose an alternative methodology when modeling credit risk for a retail consumer credit portfolio of a commercial bank in order to raise the quality of the risk assessment.
The dissertation contributes to the knowledge regarding credit risk in a portfolio of consumer loans to households. For the purpose of credit application classification there were many quantitative methods presented in the literature, among which the most popular include statistical methods. In business practice most widely used method for modeling credit risk in retail portfolios is the logistic regression. The reasons for its popularity may lie on one hand in the fact that it is wide-known among bankers and on the other hand in its simple application and satisfactory results when considering the undemanding level of expert knowledge needed for the implementation. Due to its popularity, we applied the logistic regression for modeling benchmarking models. However, it is neither the best nor the only and final solution. Logistic regression has many shortcomings, which are also commented on in this dissertation. A review of the literature suggests that a better classification accuracy can be achieved by applying non-linear methods such as support vector machines, neural networks, fuzzy logic etc.
The thesis of this dissertation proposes that a classification model for a retail portfolio of consumer loans, which exclusively or partly uses self-organizing networks, may outperform a standard solution, namely a logistic regression model. First, we trained a network using the learning vector quantization, a method from the family of self-organizing networks. Next, we estimated the benchmarking models. In this dissertation we focus also on the issue of the definition of a model’s quality, which was crucial for the ability of the comparison between two very different solutions. We argue that the classification accuracy is the best measure of the model’s quality, as there is no subjectivity and prejudice to the outcome. Provided that the interpretation of the model results does not include any human interference, the user gets a unique and transparent response to the classification problem of credit applications. When comparing both alternatives the thesis of this dissertation turns out to be proper, since the LVQ model achieved higher classification accuracy compared to benchmarking model. We assign the success of the LVQ method to the presence of non-linearity in the data.
The importance of this dissertation’s contribution manifests itself also in the ability of immediate transfer of findings into business practice. Using the method that improves the classification accuracy of a credit risk model introduces for banks the possibility of reducing the future costs of non-performing loans and reduction in opportunity costs of a rejected potentially good credit. Furthermore, improved credit risk models contribute to greater stability of the banking system and an improved relationship between the actual risk in the portfolio and the required capital. The disserta|