AI for Digital Finance Workshop

Credit risk prediction in peer-to-peer (P2P) lending is traditionally performed using statistical or machine learning models that treat borrower observations as independent data points. However, borrowers in credit markets often exhibit latent similarities in financial characteristics and behavioral patterns, suggesting that relationships between loan applications may contain additional predictive information. This presentation explores how such relationships can be incorporated into credit risk modeling by transforming structured financial data into a graph representation. Using a large dataset of P2P loans, each loan application is represented as a node in a graph, while edges capture similarity between loans based on selected financial attributes. The graph structure is constructed by computing pairwise Gower distances and retaining edges through an α-threshold filtering strategy that controls graph sparsity. This transformation enables the application of Graph Neural Networks (GNNs), specifically the GraphSAGE architecture, to perform node-level default prediction. Empirical results show that the graph-based representation facilitates pattern recognition in structured data by allowing the model to incorporate information from similar loan instances. The proposed approach demonstrates competitive predictive performance compared with traditional machine learning baselines and provides insights into how relational structures within financial datasets can enhance risk assessment. The study highlights the broader potential of graph-based learning methods for structured financial data, particularly in contexts where latent relationships between observations may carry predictive value.