Modelling bank customers' behaviour using data warehouses and incorporating economic indicators

Abstract

A credit risk monitoring model using Markov Chains was first prescribed by Cyert, Davidson and Thomson (1962) (the CDT model). It is used to monitor transition of a credit account from one performance state to another, as an alternative to scorecard methodologies. The propensity of such transition is called transition probability. Successive variants ofthe CDT model assumed a few outdated assumptions although proper tests had been available. Moreover no solutions were offered despite many had long suspected the dependency oftransition probability on economic conditions. In this empirical research, using real, substantial retail bank data, and adopting the Mover-Stayer notion (Frydman et al, 1985): 1. the unquestioned assumptions are proved invalid; 2. the true functional dependency of a transition probability time series on selected economic indicators is established; 3. the parameter associated to each explanatory variable is estimated using a non linear optimisation technique on the maximum logarithmic likelihood of a transition probability; 4. segments based on different transitional behaviour are identified for the given portfolio; 5. a pilot scorecard scheme is carried out to investigate membership to the segments identified in (4), given existing application and behavioural information.