A recent article in The Journal of Finance and Data Science reports that combining customers’ credit card information with their debit account transactions markedly improves the ability to forecast credit card delinquency. The study, carried out by Håvard Huse of BI Norwegian Business School, Sven A. Haugland of NHH, and Auke Hunneman of BI, introduces a hierarchical Bayesian behavioural model that consistently surpasses prominent machine-learning systems, including XGBoost, GBM, neural networks, and stacked ensemble methods.
Huse notes that relying solely on credit data offers only a limited view of an individual’s financial circumstances. By incorporating debit-side activity, the researchers gain visibility into payday-driven spending, repayment routines, and patterns in income flows—elements that play a decisive role in whether a customer may struggle to meet payment obligations.
Their analysis draws on granular transaction-level data from a central Norwegian bank. Where standard credit-risk models depend predominantly on monthly summary indicators such as balances or credit limits, these traditional markers reveal little about the day-to-day financial habits that underpin repayment outcomes. By modelling behavioural trajectories—how repayment behaviour shifts over time, or how expenditure rises immediately after payday—the new framework provides a richer explanation of both the mechanisms behind delinquency and the individuals most likely to default.
The model’s advantages extend to its ability to generate more precise predictions for individual customers. It also uncovers distinct behavioural groups characterised by differing “memory lengths”, referring to how strongly past financial states influence present repayment patterns. According to Hunneman, customers under financial strain tend to be more affected by their earlier behaviour, and this dynamic is captured far more effectively by the Bayesian specification than by conventional machine-learning tools.
A further strength of the approach lies in its interpretability. While outperforming cutting-edge algorithms, the model remains transparent enough for practitioners to understand the behavioural drivers of risk. As Hunneman observes, accuracy alone is insufficient for financial institutions; they must also be able to trace the patterns that shape customer vulnerability.
The authors illustrate the model’s practical significance by showing that, over a three-month prediction window, financial institutions could realise considerable savings by identifying at-risk cardholders earlier and taking timely action. Haugland emphasises that this improvement is not just a technical gain in predictive power but a means of offering more proactive support to customers who might otherwise slide into serious financial difficulty.
Together, these findings signal an essential evolution in credit-scoring practice: a movement away from static, aggregate measures toward deeper behavioural analytics grounded in the full spectrum of customer transactions.
More information: Håvard Huse et al, Integrating credit and debit data for enhanced insights into borrowing behavior and predictive modeling of credit card delinquency, The Journal of Finance and Data Science. DOI: 10.1016/j.jfds.2025.100166
Journal information: The Journal of Finance and Data Science Provided by KeAi Communications Co., Ltd.