How Good Is „Good“ ? - Making Better Use of Subjective Information in Bank Internal Credit Scoring Systems

Lenders experience positive net revenue impacts from lending if they increase the classification power of their credit scoring systems. If loan officers’ subjective assessments of otherwise intangible borrower characteristics contain additional information about a borrower, a lender may improve the default forecast quality of his internal credit scoring systems by utilizing this subjective information. The Basel II regulatory framework requires lenders to use all available information about a borrower, both subjective and nonsubjective, but at the same time produce consistent and objectified borrower ratings. However, soft information is often laden with inconsistencies due to the lack of comparability of different raters’ assessments and the existence of incentives to manipulate the soft rating. These inconsistencies leave soft information expensive to acquire and with only limited power to improve the forecast quality of lenders’ credit scoring systems. It is the objective of this thesis to introduce empirical methods that allow lenders to analyze the available soft information in a more sophisticated way, treat the inconsistencies in the data and improve the classification power of soft facts. Instead of using total scores from credit scorecards as an indicator of a customer’s probability of default, we analyze different rating patterns by applying latent trait models borrowed from psychometrics. We use a data set of 20,000 SME (Small and Medium Enterprises) credit scoring observations, including hard scores (financials, account behavior) and soft scores (scorecard responses). Applying a Mixed Rasch Model, six latent response pattern classes are identified in our data set such that, within each pattern class, the item responses are independent and there are no item redundancies. The interpretation and analysis of the pattern classes provide credit managers with information about the loan officers’ usage of the scorecard, allow them to develop monitoring tools, and to mitigate adverse rater behavior. A new soft score is constructed by utilizing the information about the pattern classes’ individual default rates and classification power. To compare alternative scoring models we use ROC (Receiver Operating Curve) inspection and related measures. We find that, by making better use of already existing subjective information, the forecast quality of a lender’s credit scoring system can be significantly increased without affecting front end lending processes.