Intelligent credit card default prediction algorithm based on ensemble learning

Credit default prediction constitutes a fundamental task in financial risk management. Although machine learning methods have demonstrated effectiveness in this domain, the heterogeneity and complexity of feature variables introduce substantial challenges. In particular, individual models often exhibit limitations in fully capturing features from large-scale, multi-type, and high-dimensional datasets. Additionally, data imbalance results in reduced detection rates for minority class instances. To mitigate these issues, a Stacking ensemble learning framework is proposed, incorporating LightGBM, XGBoost, Random Forest, and a BP neural network. The methodology begins from exploratory data analysis and feature engineering, followed by an application of the Borderline-SMOTE technique to correct class imbalance. Recognizing that ensemble models inherently increase algorithmic complexity, five-fold cross-validation and early stopping mechanisms are implemented to reduce the risk of overfitting. Subsequent model evaluation is conducted using the UCI_Credit_Card datasets. The results indicate that the proposed framework attains the accuracy of 0.9146, precision of 0.9117, recall of 0.9187, an AUC of 0.9370, and an F1-score of 0.9152, surpassing the performance of benchmark algorithms. Furthermore, model performance is examined across varying sample sizes, thereby confirming its robustness and applicability to large-scale datasets.