<p>While machine learning models are acclaimed for their superior predictive performance in the era of big data, their practical application in credit scoring has been limited by a lack of interpretability comparable to that of logistic regression-based scorecards. This paper bridges this gap by generalizing the method of generating scorecards from logistic models using SHAP (SHapley Additive exPlanations) to the LightGBM model. This approach ensures monotonicity for each default predictor, a critical requirement for regulatory compliance and practical implementation. To demonstrate the robustness and generalizability of the proposed method, we conducted extensive experiments on three real-world datasets, encompassing both traditional credit bureau data and alternative credit scoring data. The proposed method showed improved predictive accuracy (measured by the Brier Score), discriminatory power (measured by the Area Under the ROC Curve), and classification quality (measured by the Kolmogorov-Smirnov statistic) compared to traditional logistic regression. These results suggest that our approach not only maximizes predictive performance but also offers a practical pathway for financial institutions to achieve more reliable credit assessment by significantly reducing misclassification costs while maintaining model transparency.</p>