Predicting financial well-being using behavioral transactional data

Official URL: http://risc01.sabanciuniv.edu/record=b1857580 (Table of Contents)

The recent introduction of using customers' spatio-temporal mobility patterns to predict their financial well-being proved to show significant results when examined on an OECD country’s bank data. In this research, we attempt to validate the same concept using another large bank’s transactional data set and see if it can be generalized. We examine a 1-year dataset spanning 2014 and 2015, calculate the relevant features from the literature and run prediction models using the bagging algorithm. The results show that the models built on spatio-temporal mobility features are still significant when predicting a customer's overspending and the status of financial trouble. In the case of late credit card payments as signs of financial trouble, demographics prove to be more significant than the spatio-temporal mobility features. We conduct further analysis to introduce new input variables related to shopping and channel categories, in an effort to improve the prediction accuracies of these models. The results show that among all the new features we experiment with, shopping categories used as an entropy variable and used as a binary indicator variable were the most significant ones in predicting overspending. The results of this study further validate that spatio-temporal mobility and other behavioral features can successfully predict financial well-being across different datasets, and hence can be used by decision makers in the financial industry.