Pelle Bergsten och Thomas Gunnarsson
	
Validation of Credit Risk Models

A Sensitivity Analysis of the Validation Methods and Estimation of a Logistic Regression Model

Abstract

To create and promote a stable and a more risk-sensitive financial
system, Basel II introduces new approaches for the calculation of
capital requirements for credit risks. In two of these the banks have
to estimate their own internal credit risk models, but to use the
models the banks have to prove to the supervising authority that they
are reliable. Consequently there is an enormous need for well
functioning validation methods for credit risk models and for the
knowledge about how these are interpreted and how they behave. In the
thesis different quantitative validation methods and measures for
statistical credit risk models are described and examined. The focus
is on the sensitivity of the measures, i.e. their relative change due
to changes in the characteristics of the data sample used for
validation. This includes, among other things, to study the measures
when applied to portfolios of different risk, closely related to how
the apprehension of the goodness of a model can change over time due
to the phenomenon of credit migration. Furthermore, the thesis deals
with the estimation theory of credit risk models. The model estimated
is also re-estimated over time with the purpose to observe how its
parameters change and how this affects the validation measures
previously described and analyzed. The main conclusion drawn is that
the validation methods and measures are highly sensitive to the
special characteristics of the sample used for validation, and always
associated with a certain degree of uncertainty. By choosing the
distribution of borrowers over the risk scale of the model in a
certain way, the measures can be manipulated in both a positive and a
negative direction, thus implying that samples that maximize the
measures and pass the model in different statistical tests always can
be found.