Peer-to-peer loan acceptance and default prediction with artificial intelligence.

Area of Science:

• Machine Learning
• Financial Modeling
• Data Science

Background:

• Traditional credit risk models face challenges in accurately predicting loan defaults.
• Lender acceptance and default prediction are critical aspects of loan portfolio management.

Purpose of the Study:

• To develop and evaluate a two-phase artificial intelligence model for predicting loan rejection and default risk.
• To compare the performance of logistic regression (LR) and deep neural networks (DNNs) in credit risk assessment.
• To investigate the impact of dataset subsets (all loans vs. small business loans) on model performance.

Main Methods:

• Application of logistic regression (LR) and support vector machine algorithms.
• Implementation of linear and nonlinear deep neural networks (DNNs).
• Development of a two-phase model: Phase 1 for loan rejection prediction, Phase 2 for approved loan default risk prediction.

Main Results:

• LR demonstrated strong performance in predicting loan rejection (Phase 1).
• DNNs achieved superior performance in predicting loan defaults for approved loans (Phase 2).
• The models showed potential to reduce loan default risk by a significant margin.
• Phase 1 performance was better on the entire dataset, while Phase 2 excelled on the small business loan subset.

Conclusions:

• Artificial intelligence, particularly DNNs, offers significant improvements over traditional methods for credit risk modeling.
• The two-phase approach effectively separates loan rejection and default prediction.
• Discrepancies in performance between overall and small business loan datasets suggest nuanced analysis is required for different loan types.