Résumé

Credit risk prediction presents significant challenges due to data imbalance, complex nonlinear relationships, and stringent regulatory requirements for model interpretability. This research introduces a novel interpretable framework that integrates quantum-inspired feature engineering with classical machine learning to enhance default prediction accuracy while maintaining transparency. By encoding classical financial features into quantum state representations via a ZZFeatureMap circuit and extracting principal components, we create enhanced feature spaces that capture complex interactions beyond conventional methods. We conduct comprehensive benchmarking of eight classifiers (HistGradientBoosting, XGBoost, LightGBM, CatBoost, Random Forest, SVM, MLP, and Logistic Regression) on a substantial dataset of 32,581 credit observations. Our quantum-enhanced models demonstrate significant performance improvements, with HistGradientBoosting achieving the highest performance (ROC AUC: 0.942, specificity: 0.993, precision: 0.968) and LightGBM offering optimal efficiency-accuracy trade-offs (ROC AUC: 0.934 with only 1.1s training time). Gradient boosting models consistently outperformed other approaches, with all quantum-enhanced variants exceeding ROC AUC scores of 0.930 compared to less than 0.900 for non-tree-based models. A key innovation is our reverse interpretation analysis, which maps quantum features back to interpretable classical risk dimensions, addressing a critical gap in quantum-enhanced model explainability for regulatory compliance. Statistical validation through ANOVA and Friedman tests confirms the significance of performance differences between model architectures, with tree-based models showing particular synergy with quantum-augmented feature spaces. This work establishes quantum-inspired feature engineering as a practical enhancement for credit risk assessment, providing financial institutions with immediately applicable methods to improve predictive accuracy while meeting transparency requirements. By demonstrating measurable performance gains without sacrificing interpretability, our framework bridges the gap between advanced machine learning techniques and the practical constraints of regulated financial environments, paving the way for responsible adoption of quantum-enhanced analytics in the financial sector.