Transaction-based machine learning under restricted feature availability for short-term credit distress under a structural shock

This paper studies whether a transaction-based machine learning model for short-term credit distress, developed under a restricted transaction-only feature set, remains predictive under a structural shock. Using granular daily account data from a large European bank, we construct a model based on an event-level representation of overdraft dynamics and a feature set that excludes customer-level characteristics and relies exclusively on transactional data. We evaluate model performance out-of-time during the COVID-19 pandemic and the associated mortgage moratorium, which provide a demanding real-world stress environment. The complete predictive framework achieves an out-of-time ROC AUC of 0.83 during the COVID-19 mortgage moratorium, indicating that the transaction-based approach retains predictive performance under this specific exogenous policy shock. These results show that a transaction-based model developed using a restricted feature set can remain predictive and operationally effective in this specific validation setting, outperforming a human-based benchmark in the Bank's validation exercise. The evidence pertains to one exogenous, policy-driven shock, and does not establish comparable performance under endogenous credit downturns or monetary tightening. The analysis does not test whether the feature restrictions themselves reduce proxy effects, improve fairness, or causally enhance predictive robustness.