| Abstract [eng] |
The main objective of this master’s thesis is to develop a credit risk assessment model for individuals by applying machine learning methods and to quantitatively evaluate the suitability of these methods for predicting consumer loan credit risk in the Lithuanian market. The tasks of the thesis included the analysis of the concept, challenges, and economic significance of credit risk, the systematization of credit risk assessment model development principles, the justification of the model development methodology, as well as the presentation and evaluation of model results. The thesis consists of three main parts: a scientific literature review, an empirical study and its results, and formulated conclusions and recommendations. The literature review draws on both international and national sources to explore the concept of credit risk, its challenges and economic importance, the conceptual framework of quantitative risk management tools, and trends in the application of modern classification methods. The emphasis is placed on different classification methods, their advantages and disadvantages, and the importance of data feature exploration, data transformation and sample balancing. The empirical part of the thesis is based on consumer loan data issued on the IPPT platform during the period from 2021-01-01 to 2024-01-28. The data sources included national institutions and registries, loan application information, and client behavior-related data. In total, 36 models were developed. The best-performing credit risk assessment model for individuals was selected, which was based on the XGB classification method, combined with WOE data transformation and SMOTE sample balancing techniques. The selected model achieved a predictive performance of 0.83 in terms of AUC. The empirical results revealed that the most significant predictors of credit risk level were client debt indicators and the frequency of loan inquiries. A client rating scale system was created to segment clients according to risk levels from A1 to E3, and a statistically validated threshold of 48% was identified, from which a client is assigned to the “bad” borrower category. The empirical findings confirmed that advanced machine learning methods can ensure high accuracy in credit risk modelling. The results also highlighted that managing selection of independent variables, sample imbalance, and data transformation has a substantial impact on the model’s effectiveness. The results of this thesis may be published in scientific journals focused on individual credit risk management, machine learning applications and consumer lending markets. The practical significance of the study lies in the potential for financial institutions and IPPT platforms to apply the developed model in decision-making processes, while its theoretical contribution provides a basis for further research integrating the mortgage loan market, macroeconomic indicators, and major commercial bank data. |
