Modelling retail bank customers' needs with Bayesian Belief Networks

Retail bank customer relationships, like many systems observed in real life, cannot be modelled deterministically - they have multiple variables, interdependent through various channels of cause and effect and are subject to exogenous effects. The methods used with such data extract a model by uncovering stochastic relationships between variables. Bayesian Belief Networks accomplish this effectively because they are capable of integrating experts' knowledge, discovering causal relationships in the data and introducing hidden variables to represent exogenous effects. In this paper we present a method of building a customer relationship model using Bayesian Belief Networks. In addition, we emphasise the need to use concepts employed by company experts. Financial executives think in terms of discretionary income, affinity for technology, card vs. cash preferences and lifestyle when they form a profile of a customer and their needs - factors which, though not directly observable, are easy to capture with tools designed for situations where concepts and basic data have a stochastic relationship. Models employing concepts that are meaningful in terms of customer profile, demand and transactional patterns are more accurate and easy to interpret. Concerning the methodology, the right approach depends on the question. We use offline data mining where problems are hard to standardize or where models are generated within a constantly changing environment. In cases where an integrated tool is needed for answering well-formulated, periodically occurring questions, we use analytical CRM.