Predicting customer attrition using binge trading patterns: Implications for the financial services industry

Abstract

Human activities tend to burst at specific times, followed by long dormant periods. This study analyzes detailed trading records, as well as the demographic profiles of 486,049 customers from a major securities company and shows that an entropy measure of non-Poisson trading patterns has advantages over the canonical recency, frequency, and monetary value framework in the financial services sector. The LASSO logistic regression, the information gain metric in gradient boosting decision trees, and the relative importance method in neural networks all lend support to the conclusion that the clumpiness measure of trade clustering plays a significant role in explaining customers’ future churning. Furthermore, it appears that recently developed statistical learning techniques reduce churn prediction errors to a greater extent. A metric-based parsimonious RFMC approach coupled with machine learning techniques can be effectively used to better gauge customer lifetime value.