An integrated explainable machine learning framework for prediction explanation and audit in high stakes binary classification

Machine learning classifiers are increasingly deployed in high-stakes decision systems, yet their opacity creates governance risks that undermine accountability and trust. This study proposes an integrated framework unifying predictive modelling, explainability, and model auditing, evaluated on a primary public tabular dataset (Lending Club) and an independent benchmark dataset (Give Me Some Credit) drawn from the consumer-credit domain. The framework enforces a strict decision-time feature-availability policy under which all post-origination repayment variables are excluded, defines default over a fixed 24-month outcome horizon, and combines class-imbalance handling, seven benchmark classifiers (including a traditional scorecard), probability-calibration assessment, global and local explainability analysis with quantified SHAP–LIME agreement, and a six-dimension audit covering performance stability, feature reasonableness, error structure, subgroup fairness screening, threshold policy, and robustness to missing inputs. On the temporally held-out Lending Club test set, LightGBM achieved the strongest discrimination, with an AUC-ROC of 0.8451 (95% CI 0.8421–0.8481), AUC-PR of 0.6118, F1-score of 0.6097, and KS statistic of 0.5127; its advantage over CatBoost was not statistically significant, whereas both exceeded the non-boosting baselines. A similar ordering was observed on the independent benchmark, where severe class imbalance made AUC-PR more informative than accuracy. SHAP explanations identified interest rate, loan grade, revolving utilization, debt-to-income ratio, credit-history length, and income as the dominant predictors, and SHAP–LIME attributions agreed strongly across a systematic evaluation sample (mean Spearman correlation 0.84). The audit revealed stable cross-validation performance, mild temporal drift, adequate post-calibration reliability, and elevated false-positive rates for low-income and short-credit-history subgroups; because the datasets contain no directly observed protected attributes, these results constitute socioeconomic subgroup screening rather than a demonstration of demographic fairness. The study shows that predictive performance, interpretability, calibration, and auditability can be jointly evaluated in a single reproducible workflow, providing an operational technical-audit template that complements, but does not replace, legal and institutional review.