Transforming Borrower Profiles into Risk Signals: A Hybrid Convolutional Framework for Peer-to-Peer Credit Default Prediction

Purpose: Accurate credit default prediction is essential for sustainable risk management in peer-to-peer (P2P) lending platforms. However, the presence of class imbalance, high-dimensional borrower information, and complex nonlinear relationships among financial variables continues to limit the effectiveness of conventional prediction models. This study proposes a hybrid deep learning framework that transforms borrower attributes into structured grayscale images, enabling convolutional feature extraction for enhanced credit risk assessment. Material and Methods: A LendingClub dataset containing 115,487 completed loan records was analyzed. Data preprocessing involved feature cleaning, transformation, and statistical feature selection, reducing 151 original variables to 64 informative predictors through a composite scoring framework integrating five complementary statistical tests. The selected features were encoded into 64×64 grayscale images and processed using a Convolutional Neural Network (CNN). Seven predictive models were evaluated under five-fold stratified cross-validation across three dataset configurations: the original imbalanced dataset, SMOTE-balanced training data, and randomly undersampled training data. The evaluated models included conventional machine learning classifiers (Support Vector Machine, Random Forest, and Decision Tree), a Deep Neural Network (DNN), a standalone CNN, and three hybrid architectures combining CNN-based feature extraction with downstream classifiers (CNN+SVM, CNN+RF, and CNN+DT). Statistical significance was assessed using the Friedman test and Wilcoxon signed-rank tests at α = 0.05. Results: The CNN+RF model consistently achieved the best performance across all experimental settings. On the original dataset, it attained an AUC-ROC of 1.000, Accuracy of 0.987, F1-Score of 0.991, and MCC of 0.977, significantly outperforming all competing approaches (p < 0.05). The standalone CNN also demonstrated strong predictive capability, supporting the effectiveness of the proposed feature-to-image representation strategy. Furthermore, SMOTE-based balancing yielded superior performance compared with random undersampling across all models. Among conventional classifiers, SVM exhibited the greatest sensitivity to class imbalance, whereas DNN produced comparatively balanced class-level predictions. Conclusion: The proposed CNN–Random Forest hybrid framework provides a highly effective and statistically validated solution for P2P credit default prediction. By converting borrower information into image-based representations and leveraging convolutional feature learning, the model achieves superior predictive accuracy, robustness, and stability compared with both traditional machine learning and standalone deep learning methods. The framework offers practical value for loan approval decisions, borrower screening, and portfolio risk management in digital lending environments.