This study develops an explainable machine-learning framework for survey-based consumer-credit late-payment risk identification. Using public microdata from Mexico's 2021 Encuesta Nacional de Inclusión Financiera (ENIF), the outcome is defined as self-reported late payment on at least one formal credit product since July 2020. The analytical cohort contains 4549 respondents with valid repayment responses, including 1216 late-payment reports (26.73%). Outcome components, contemporaneous repayment information, formal-credit-holding information, COVID-related late-payment items, routing variables, and survey-design identifiers are excluded from the predictors. Fourteen statistical-learning, bagging, boosting, neural-network, and ensemble candidates are compared under one leakage-controlled protocol: Each training fold selects 25 raw predictors, and the observed class distribution is retained without class resampling. Gradient boosting is selected using repeated cross-validation on training data only, with a mean out-of-fold area under the receiver operating characteristic curve (AUC) of 0.695. On the independent test set, its AUC, Kolmogorov–Smirnov (KS) statistic, and Brier score are 0.706, 0.312, and 0.174, respectively. Permutation importance and Shapley additive explanations (SHAP) analyses identify financial resilience, prior credit constraints, and informal borrowing as important correlates of reported late payment. The findings concern survey-based risk identification and should not be interpreted as an administrative-data probability-of-default score or an automated lending rule.
Citation: Yangting Huang, Yifan Liao, Ying Li, Zihuan Yue, Keyue Yan. Machine-learning-based prediction and interpretable analysis of survey-based consumer credit default risk: Evidence from self-reported late payment[J]. Data Science in Finance and Economics, 2026, 6(4): 622-643. doi: 10.3934/DSFE.2026022
This study develops an explainable machine-learning framework for survey-based consumer-credit late-payment risk identification. Using public microdata from Mexico's 2021 Encuesta Nacional de Inclusión Financiera (ENIF), the outcome is defined as self-reported late payment on at least one formal credit product since July 2020. The analytical cohort contains 4549 respondents with valid repayment responses, including 1216 late-payment reports (26.73%). Outcome components, contemporaneous repayment information, formal-credit-holding information, COVID-related late-payment items, routing variables, and survey-design identifiers are excluded from the predictors. Fourteen statistical-learning, bagging, boosting, neural-network, and ensemble candidates are compared under one leakage-controlled protocol: Each training fold selects 25 raw predictors, and the observed class distribution is retained without class resampling. Gradient boosting is selected using repeated cross-validation on training data only, with a mean out-of-fold area under the receiver operating characteristic curve (AUC) of 0.695. On the independent test set, its AUC, Kolmogorov–Smirnov (KS) statistic, and Brier score are 0.706, 0.312, and 0.174, respectively. Permutation importance and Shapley additive explanations (SHAP) analyses identify financial resilience, prior credit constraints, and informal borrowing as important correlates of reported late payment. The findings concern survey-based risk identification and should not be interpreted as an administrative-data probability-of-default score or an automated lending rule.
| [1] |
Arrieta AB, Díaz-Rodríguez N, Del Ser J, et al. (2020) Explainable artificial intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Inform Fusion 58: 82–115. Available from: https://www.sciencedirect.com/science/article/abs/pii/S1566253519308103. |
| [2] |
Bussmann N, Giudici P, Marinelli D, et al. (2020) Explainable AI in fintech risk management. Front Artif Intell 3: 26. https://doi.org/10.3389/frai.2020.00026 doi: 10.3389/frai.2020.00026
|
| [3] |
Bussmann N, Giudici P, Marinelli D, et al. (2021) Explainable machine learning in credit risk management. Comput Econ 57: 203–216. https://doi.org/10.1007/s10614-020-10042-0 doi: 10.1007/s10614-020-10042-0
|
| [4] |
Chawla NV, Bowyer KW, Hall LO, et al. (2002) SMOTE: Synthetic minority over-sampling technique. J Artif Intell Res 16: 321–357. https://doi.org/10.1613/jair.953 doi: 10.1613/jair.953
|
| [5] |
Doshi-Velez F, Kim B (2017) Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv: 1702.08608. https://arXiv.org/abs/1702.08608 |
| [6] |
Fernández-Delgado M, Cernadas E, Barro S, et al. (2014) Do we need hundreds of classifiers to solve real world classification problems? J Mach Learn Res 15: 3133–3181. Available from: https://jmlr.org/papers/v15/delgado14a.html |
| [7] |
Fisher A, Rudin C, Dominici F (2019) All models are wrong, but many are useful: Learning a variable's importance by studying an entire class of prediction models simultaneously. arXiv preprint arXiv: 1801.01489. https://arXiv.org/abs/1801.01489 |
| [8] |
Golec M, AlabdulJalil M (2025) Interpretable LLMs for credit risk: A systematic review and taxonomy. arXiv preprint arXiv: 2506.04290. https://arXiv.org/abs/2506.04290 |
| [9] |
Gramegna A, Giudici P (2021) SHAP and LIME: An evaluation of discriminative power in credit risk. Front Artif Intell 4: 752558. https://doi.org/10.3389/frai.2021.752558 doi: 10.3389/frai.2021.752558
|
| [10] |
Guidotti R, Monreale A, Ruggieri S, et al. (2018) A survey of methods for explaining black box models. arXiv preprint arXiv: 1802.01933. https://arXiv.org/abs/1802.01933 |
| [11] |
Han S, Jung H (2024) NATE: Non-pArameTric approach for Explainable credit scoring on imbalanced class. PLOS ONE 19: e0316454. https://doi.org/10.1371/journal.pone.0316454 doi: 10.1371/journal.pone.0316454
|
| [12] |
Kipf TN, Welling M (2017) Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv: 1609.02907. https://arXiv.org/abs/1609.02907 |
| [13] |
Lipton ZC (2016) The mythos of model interpretability. arXiv preprint arXiv: 1606.03490. https://arXiv.org/abs/1606.03490 |
| [14] |
Lundberg SM, Lee SI (2017) A unified approach to interpreting model predictions. Adv Neural Inform Process Syst 30: 4765–4774. |
| [15] |
Mapfumo I, Shongwe T (2026) Performance evaluation of machine learning and deep learning models for credit risk prediction. J Risk Financ Manage 19: 210. https://doi.org/10.3390/jrfm19030210 doi: 10.3390/jrfm19030210
|
| [16] |
Misheva BH, Osterrieder J, Hirsa A, et al. (2021) Explainable AI in credit risk management. arXiv preprint arXiv: 2103.00949. https://arXiv.org/abs/2103.00949 |
| [17] |
Monje L, Carrasco RA, Sánchez-Montañés M (2026) Machine learning XAI for early loan default prediction. Comput Econ 67: 4033–4062. https://doi.org/10.1007/s10614-025-10962-9 doi: 10.1007/s10614-025-10962-9
|
| [18] |
Nwafor CN, Nwafor O, Brahma S (2024) Enhancing transparency and fairness in automated credit decisions: An explainable novel hybrid machine learning approach. Sci Rep 14: 25174. https://doi.org/10.1038/s41598-024-75026-8 doi: 10.1038/s41598-024-75026-8
|
| [19] |
Pedregosa F, Varoquaux G, Gramfort A, et al. (2011) Scikit-learn: Machine learning in Python. J Mach Learn Res 12: 2825–2830. |
| [20] |
Ribeiro MT, Singh S, Guestrin C (2016) Why should I trust you?: Explaining the predictions of any classifier. arXiv preprint arXiv: 1602.04938. https://arXiv.org/abs/1602.04938 |
| [21] |
Rudin C (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. arXiv preprint arXiv: 1811.10154. https://arXiv.org/abs/1811.10154 |
| [22] |
Saarela M, Jauhiainen S (2021) Comparison of feature importance measures as explanations for classification models. Sn Appl Sci 3: 272. https://doi.org/10.1007/s42452-021-04148-9 doi: 10.1007/s42452-021-04148-9
|
| [23] |
Schmitt M (2024) Explainable automated machine learning for credit decisions: Enhancing human-artificial intelligence collaboration in financial engineering. arXiv preprint arXiv: 2402.03806. https://doi.org/10.48550/arXiv.2402.03806 |
| [24] |
Shreya, Pathak H (2025) Explainable artificial intelligence credit risk assessment using machine learning. arXiv preprint arXiv: 2506.19383. https://doi.org/10.48550/arXiv.2506.19383 |
| [25] |
Szepannek G, Lübke K (2021) Facing the challenges of developing fair risk scoring models. Front Artif Intell 4: 681915. https://doi.org/10.3389/frai.2021.681915 doi: 10.3389/frai.2021.681915
|
| [26] |
Wang H, Huang J, Sua LS, et al. (2025) Enhancing credit risk prediction: A multi-stage ensemble pipeline. arXiv preprint arXiv: 2509.22381. https://doi.org/10.48550/arXiv.2509.22381 |
Figures(16) / Tables(2)
Yangting Huang, Yifan Liao, Ying Li, Zihuan Yue, Keyue Yan. Machine-learning-based prediction and interpretable analysis of survey-based consumer credit default risk: Evidence from self-reported late payment[J]. Data Science in Finance and Economics, 2026, 6(4): 622-643. doi: 10.3934/DSFE.2026022
Questionnaire domains retained for predictor governance
Analytical-cohort construction and leakage-controlled evaluation workflow
Fold-level stability of the common 25-predictor selection budget
Cross-validated comparison of candidate models under the common protocol
Independent-test AUC comparison of all candidate models under the common protocol
Repeated cross-validation AUC distributions for QDA and gradient boosting
Independent-test metrics for gradient boosting
Independent-test confusion matrix at the training-selected threshold
Global importance diagnostics for gradient boosting: (a) independent-test permutation importance and (b) model-based feature importance
Global SHAP importance ranking for gradient boosting
Cross-checks for global explanation: (a) Questionnaire-domain contribution summary and (b) Agreement of global importance measures
SHAP value distributions for the leading questionnaire variables
Frequency with which leading predictors are selected across folds
Independent-test probability diagnostics: (a) Calibration curve and (b) ROC curve
Local SHAP explanations for two representative lower-risk observations: (a) Predicted probability = 0.064 and (b) Predicted probability = 0.164
Local SHAP explanations near the operating threshold: (a) Predicted probability = 0.244 and (b) Predicted probability = 0.354
DownLoad: