PolyTrunc-ANN: Polynomial Features and TruncatedSVD for Optimized Neural Network Performance for Predicting Loan Defaults in P2P Lending Platforms

Abstract

Peer-to-peer (P2P) lending networks have revolutionized the financial landscape by providing borrowers with alternative credit options and offering lenders new investment opportunities. Nonetheless, predicting loan defaults remains a critical challenge that demands advanced predictive models for effective risk management. This research evaluates the performance of a sophisticated predictive algorithm for loan defaults in P2P lending platforms, utilizing detailed borrower, loan, and historical default data. We identify and analyze key borrower characteristics influencing default likelihood, demonstrating that our model significantly improves prediction accuracy compared to traditional methods. Additionally, the study explores the practical implications of integrating this model into P2P platforms, including its impact on stakeholders and associated ethical considerations. The goal is to enhance the stability and reliability of P2P loan ecosystems, benefiting both investors and borrowers.

The distinctiveness of this research lies in its comprehensive approach, integrating advanced pre-processing techniques—such as parallel variable pre-processing pipelines, TruncatedSVD for dimensionality reduction, and SMOTE for class imbalance correction—with a custom-designed ANN architecture optimized through RandomizedSearchCV. Furthermore, this study employs the RandomForestClassifier to provide valuable insights into the significance of features and individual prediction explanations.