Applying Complementary Credit Scores to Calculate Aggregate Ranking
Researchers have been improving credit scoring models for decades. The reason for this is following: an increase in the predictive ability of scoring even by small values can save a financial institution from a significant losses. As a result, many researchers have conclude that ensembles of classifiers or aggregated scorings have greater performance. However, ensembles outperform the base classifiers by thousandths of a percent on unbalanced samples.
This article suggests building an aggregated scoring. Unlike previously proposed aggregate scores, its baseline classifiers are focused on identifying different types of borrowers. The purpose of this study is to illustrate the effectiveness of such scoring aggregation on real unbalanced data.
We use one performance measure as effectiveness indicator - the area under the ROC curve. The DeLong, DeLong and Clarke-Pearson test is used to measure the statistical difference between the two or more areas. In addition, this study uses a logistic model of defaults (logistic regression), which is applied to the data of companies financial statements. This model is usually focused on identifying default borrowers. To obtain a scoring aimed at non-default borrowers, a modified Kemeny median is used, which was conceived by the authors to rank companies with credit ratings. Both scores are aggregated by logistic regression.
Our data contains most of the observations of Russian banks that existed and defaulted from 01.07.2010 to 01.07.2015. The sample of banks is highly unbalanced, a concentration of defaults is about 5%. However, the aggregation is carried out for the banks that have several ratings. As a result, it was found that aggregated classifiers based on different types of information improves significantly the discriminatory power of scoring even on an unbalanced sample.
This aggregated scoring and the approach to its construction could be applied in financial institutions as part of credit risk assessment, as well as an auxiliary tool for decision-making process because of relatively high interpretability of these scores.