Stacking-based heterogeneous genetic programming for interpretable credit risk evaluation

The development of advanced ensemble models to handle complex and large-scale datasets has become a central focus in credit risk prediction. Although ensemble methods offer strong predictive performance, stacking lacks a standardized construction pipeline and its complex structure often reduces transparency and robustness. To address these challenges, this study proposes a stacking based heterogeneous genetic programming classifier (SH-GPC), an end to end pipeline in which genetic programming serves as the meta-classifier to enhance interpretability and generalization. Through experiments on a large-scale credit risk dataset comprising millions of records, SH-GPC is shown to significantly outperform conventional homogeneous ensemble methods and several emerging models, including XGBoost, LightGBM, NGBoost, and TabNet, in terms of AUC. Compared to stacking frameworks with logistic regression or XGBoost as the meta-classifier, SH-GPC achieves better predictive accuracy while relying on a smaller number of base classifiers, thereby improving simplicity and interpretability. Transparency is further enhanced by representing the GP meta-classifier as evolved symbolic expressions and syntax trees. Additionally, the incorporation of the Shapley additive explanations (SHAP) technique enables visualization and attribution of each base classifier’s contribution, offering insights into the model’s internal decision logic. This study demonstrates the applicability of evolutionary algorithms in ensemble learning and introduces a new framework for credit risk modeling that achieves a balance between accuracy, stability, and interpretability. • An end-to-end pipeline for generating a heterogeneous ensemble model is proposed. • Genetic programming is used as the meta-classifier in a stacking credit risk model. • NGBoost and TabNet are used as base-classifiers. • Our SH-GPC model outperforms other meta-classifiers and provides better interpretability. • Visualization methods are used to enhance the trustworthiness of the stacking model.