Digital credit governance requires lending systems that are not only accurate but also auditable and operationally interpretable. This study examines how prediction, risk evaluation, and portfolio allocation can be organized as a reviewable decision chain in digital credit governance. Using publicly available U.S. listed-firm data from 2010 to 2016 as a validation setting rather than as a direct proxy for Chinese bank portfolios, we combine XGBoost-SHAP, EWM-TOPSIS, and a hybrid simulated annealing-neighborhood algorithm (SA-NA) to connect default-risk estimation with portfolio-level choice. The model attains an AUC of 0.851; relative to a rules-based baseline, the selected portfolio improves RAROC from 12.5% to 17.5% and reduces 95% CVaR from 3.10 M to 2.53 M. The substantive contribution is not the algorithm alone, but the way the framework makes credit-allocation choices more auditable, reviewable, and open to supervisory scrutiny. The external validity claim is therefore analytic rather than statistical: the study offers a transferable design logic for linking prediction, quantification, and allocation, while direct application to Chinese commercial banks would require local recalibration of ratings, RI weights, supervisory constraints, and review procedures.
Pan et al. (Mon,) studied this question.