Design of a Scoring Model for Assessing the Credit Quality of Corporate Borrowers for Development Banks

The current research focused on practical issue of building scoring model for credit risk assessment. Credit risk assessment is an essential part for financial institutions, because it makes an impact its capital and RWA, as well as, it affects management decision-making process. In this paper were build several models to assess credit risk for Development Banks, using only open sources of information. The study systemize of the scientific literature devoted to the analysis of methods and models for assessing credit quality. On the basis of this literature methods and models for IRB were identified, besides, necessary data normalization to gain needed quality. This research contain information on advantages and drawbacks of each method of assessing credit quality, inclusion practical application of presented methods. To select the best model for assessing credit quality, a comparison of predictive ability is carried out. The predictive ability criterion is the probability of determining a credit rating “exactly” and the probability of determining a credit rating withing one notch deviation. In order to calculate the predictive power, the data was divided into the training and test samples in proportion of 70 to 30. Models were constructed on the data of the training sample, and the rating was predicted on the data of the test sample using the function of the constructed model. According to the results of the research, the best predictive ability was shown by the support vector method and neural network, which made it possible to accurately predict the credit rating for an average of 32% of cases, and showed predictive ability within one notch rating of more than 73%. However, due to the difficult interpretability and the need for “training” on a large data set, the neural network is inferior to the support vector method in terms of applied use. Econometric models have shown less high predictive power. The best specification of the linear regression model allowed us to determine exactly 26% of the ratings, and 59% of the ratings with an error in one rating. The best specification of ordered logistic regression allowed us to determine exactly 27% of the ratings, and 62% of the ratings with an error in one rating. The quadratic discriminatory analysis was excluded from results due to qualitative limitation.