Validation of credit rating systems using multi-rater information

We suggest a new framework for the use of multi-rater information in the validation of
credit rating systems, applicable in any validation process where rating information from
different sources is available. As our validation framework does not rely on historical default
information it appears to be particularly useful in situations where such information is
inaccessible. We focus on the degree of similarity or - more generally - proximity of rating
outcomes stemming from different sources and show that it is important to analyze three
major aspects of proximity: agreement, association, and rating bias. We suggest using a
weighted version of Cohen's k to measure the agreement between two rating systems and we
introduce a new measure for rating bias. In contrast to the existing literature we suggest x as a measure of association which is based on the Kemeny-Snell metric and, opposed
to other measures, is consistent with a set of basic axioms and should therefore be used in
the context of multi-rater information. We provide an illustrative empirical example using
rating information stemming from the Austrian Credit Register on partially overlapping sets
of customers of 27 banks. Using a multi-dimensional scaling technique in connection with
a minimal spanning tree we show that it is possible to consistently detect "outliers", i.e.,
banks with a low degree of similarity to other banks. The results indicate that banks which
are less diversified across the size of their loans are more likely to be outliers than others.