Explore the scholarly record.

Abstract

Abstract Background Peer-to-peer (P2P) lending has expanded rapidly, exposing individual investors to credit default risk typically absorbed by traditional bank intermediation. Methods This study proposes a hybrid deep learning framework for credit default prediction using 1,335,455 leakage-audited records from the LendingClub accepted-loans dataset. A systematic feature-timing audit and multi-criteria ranking procedure selected 64 discriminative features, transformed into structure-aware grayscale images via a data-driven row-ordering strategy (IGTD) that places correlated features in adjacent rows, benchmarked against hierarchical-clustering seriation, the original composite-rank ordering, and a 1D-CNN baseline. Fourteen classifiers, spanning tabular models (including XGBoost and LightGBM), a standalone CNN, and five CNN-based hybrids, were evaluated under repeated 10×5 stratified cross-validation across three balanced dataset conditions, with performance differences assessed via the Friedman test and the corrected resampled paired t-test. Results The hybrid CNN-Random Forest architecture achieved the strongest and most stable performance across all conditions, reaching an accuracy of 0.971 and AUC-ROC of 0.990 on the original imbalanced dataset, confirmed by the Friedman test and feature-row-aggregated Grad-CAM analysis of the convolutional backbone, and closely matched by CNN-LightGBM and CNN-XGBoost. An ablation removing the platform-assigned risk grade reduced AUC-ROC to 0.930, confirming a meaningful but non-dominant reliance on this signal, while nested hyperparameter sensitivity analysis showed only marginal variation around the selected configuration. A censoring-corrected chronological out-of-time validation, enforcing a fixed loan-maturity horizon and testing on the most recent 20% of matured loans, retained an accuracy of 0.951 and AUC-ROC of 0.967, indicating generalization to future loan cohorts rather than temporal leakage. Conclusion Structure-aware feature-to-image encoding combined with convolutional feature extraction and ensemble learning provides a measurable, statistically supported, and temporally stable improvement over conventional tabular credit-risk models, although a substantial share of this advantage derives from the platform’s own risk grade rather than from borrower-behavioral features alone. Beyond its methodological contribution, the study offers lenders and investors an empirically grounded basis for a two-tier credit-scoring strategy, pairing transparent, grade-based decisioning for regulatory reason-coding with CNN-based risk stratification as a secondary tool for portfolio-level monitoring.