Dynamic Classification of Bank Clients by the Predictability of Their Transactional Behavior

Abstract

We propose a method for dynamic classification of bank clients by the predictability of their transactional behavior (with respect to the chosen prediction model, quality metric, and predictability measure). The method adopts incremental learning to perform client segmentation based on their predictability profiles and can be used by banks not only for determining predictable (and thus profitable, in a sense) clients currently but also for analyzing their dynamics during economical periods of different types. Our experiments show that (1) bank clients can be effectively divided into predictability classes dynamically, (2) the quality of prediction and classification models is significantly higher with the proposed incremental approach than without it, (3) clients have different transactional behavior in terms of predictability before and during the COVID-19 pandemics.