A new behavioral credit-risk model that integrates credit and debit transaction data significantly outperforms current machine learning methods in predicting credit card delinquency while providing clearer insight into the behavioral drivers behind repayment problems. Researchers from BI Norwegian Business School and NHH Norwegian School of Economics developed the hierarchical Bayesian model, which consistently outperforms leading algorithms including XGBoost, GBM, neural networks, and stacked ensembles according to a study published in The Journal of Finance and Data Science.
The research demonstrates that combining credit card data with customers' debit transactions substantially improves delinquency prediction capabilities. "Credit data alone gives only a partial picture of a customer's financial situation," explains first author Håvard Huse. "By integrating debit transactions, we gain insight into payday spending, repayment behavior, and income patterns—factors that strongly influence whether someone is at risk of missing payments." The study draws on detailed credit and debit transaction data from a large Norwegian bank, moving beyond traditional models that rely on monthly aggregates like balance and credit limit.
Traditional credit-risk models fail to reveal how customers actually manage their finances day-to-day, while the new model captures behavioral dynamics such as how repayment patterns evolve over time and how spending spikes after payday. "By capturing behavioral dynamics, the new model explains both why delinquency occurs and who is likely to default," shares Huse. The model improves prediction accuracy at the individual level and identifies distinct behavioral segments with different "memory lengths"—the extent to which past financial states affect current repayment behavior.
"Customers in financial distress tend to be more influenced by earlier months' behavior, and our model captures this dynamic far better than standard machine-learning tools," notes co-author Auke Hunneman. The team's approach not only performs better than state-of-the-art algorithms but is also more interpretable. "Banks not only need accurate predictions—they also need to understand which behavioral patterns drive risk," adds Hunneman. This interpretability represents a significant advancement over black-box machine learning models that offer predictions without clear explanations.
The practical implications are substantial. Using a three-month prediction horizon, early detection of at-risk cardholders could generate significant cost savings by enabling timely intervention and reducing losses. "For banks, this is more than an accuracy improvement—it is a way to proactively help customers avoid serious financial problems," says co-author Sven A. Haugland. The findings highlight an emerging shift in credit scoring from traditional static models toward richer behavioral analytics based on a complete picture of customer transactions. The original research is available at https://doi.org/10.1016/j.jfds.2025.100166.
