Banks and credit unions continue to struggle with elevated new-to-bank (NTB) customer attrition, particularly within the first 12 months of account opening. Traditional analytics approaches rely heavily on tabular, account-level features and often fail to capture the relational, behavioral, and engagement patterns that determine whether a newly acquired customer will remain active or quietly disengage.
This white paper documents a production-oriented proof of concept (POC) developed to test the feasibility and performance of using graph-based machine learning, including Graph Neural Networks (GNNs) and related advanced analytics techniques, to predict NTB customer attrition. Using one year of real customer data from a financial institution, the team constructed a multi-entity graph representation of customer behavior, trained multiple models of increasing sophistication, and evaluated their ability to predict attrition for a future cohort of newly onboarded customers.
The results demonstrate that graph-aware models materially improve predictive performance over traditional approaches by explicitly modeling how customers interact with accounts, products, channels, and services over time.
