α-threshold networks in credit risk models

Abstract

Peer-to-peer (P2P) lending markets offer risky investment opportunities, for which accurate credit risk models are in high demand. Loan books offer a broad spectrum of loan and borrower characteristics, making it challenging to construct high-dimensional systems that make the use of traditional credit scoring models. In this study, we propose two network-based feature extraction methods that extract complex relationships between risky assets, namely, loans, which are represented as vertices, and weighted edges, which correspond to the feature-based similarity between loans. Our two methods differ with respect to how similar loans are identified. The traditional approach uses partitioning based on the medoid algorithm to identify similar loans (the k-PAM model). A much faster alternative is to eliminate (Formula presented.) of the largest distances (the α-threshold model). The resulting network structure is used to extract features that augment profit scoring models. Utilizing P2P loan data, we find that forecasting models that use network-based features consistently outperform the benchmarks in a statistical sense and lead to higher returns and risk-adjusted returns.