Vývoj unsecured cross-portfolio modelu pro predikci ztráty

The main aim of this thesis is to develop a joint model predicting the loss of recovered contracts for three unsecured credit products, compare this model with individual models for the products and decide whether it can be used in live operation for the bank. In a bank, a myriad of situations are predicted, such as the probability of default or loss given default. A variety of models are used for prediction. Therefore, the aim of the work is to create one model for multiple products at once and consequently reduce the number of models. The work describes the entire process of creating the joint model, comparison with separate models, and discussion of whether the joint model is suitable. The first step in the work is data preparation, such as classing or WOE calculation. Subsequently, the preparation of variables such as manual classing is discussed in the paper. General principles of manual classing are provided, followed by examples of specific situations that differ from the general principles. Furthermore, the stepwise regression approach is discussed, which is used for selecting significant variables on which the model is trained. Once the variables are selected, the first logistic regression model loss prediction is developed. The model is then subjected to what is known as fine-tuning to further improve it. Finally, the model is compared with newly created models for each product for objective comparison from two perspectives: predictive ability measured by Gini and Somers' D metrics and the time taken to build. As a result of the work, the joint model developed is compared with individual models. The joint model has a lower predictive ability than the stand-alone models, but its creation is significantly simpler and less time-consuming. Therefore, it can be concluded that the joint model is suitable for use in the banking environment in order to save time.