Tackling change and uncertainty in credit scoring.
Credit scoring methods summanse information on credit applicants. An
assessment of creditworthiness is derived from this summary. This thesis is
concerned with statistical methods of credit scoring.
Much of the existing literature on credit scoring is concerned with comparing the
predictive power of a wide variety of classification techniques. However, much
of the published work concludes that classifier performance on credit data is
relatively insensitive to the choice of statistical technique. Consequently, the
techniques used in commercial credit scoring have remained broadly similar
during recent years. This thesis investigates credit scoring from a more
fundamental level, by considering the formulation of the credit problem.
A review of the credit literature is given, focusing on areas that have been
subjected to much recent research activity. Details of the data sets used
throughout this thesis are provided and analysed using techniques common to the
Methods that capitalise on the uncertainty and flexibility in the definitions of the
classes used to represent 'good' and 'bad' credit risks are proposed. Firstly, a
class of models is described that permits the choice of class definition to be
deferred until the time at which the classification is required. Secondly, a strategy
for choosing a suitable definition which optimises some external criterion is
introduced. In addition, an approach is presented that combats classifier
deterioration resulting from the evolution of the underlying populations.
This thesis is essentially concerned with the uncertainties and change inherent in
credit scoring. We present novel ways in which these properties may be
incorporated in the formulation of the credit problem.