Semi-supervised adapted HMMs for P2P credit scoring systems with reject inference

Area of Science:

• Credit Risk Assessment
• Machine Learning Applications
• Financial Data Science

Background:

• Current credit scoring models often suffer from selection bias due to excluding rejected applicants.
• This bias impacts the statistical and economic validity of credit scoring, particularly in high-rejection-rate environments like peer-to-peer lending.
• Reject inference is crucial for developing more representative and accurate credit scoring models.

Purpose of the Study:

• To propose and evaluate a novel semi-supervised learning framework for reject inference.
• To address the selection bias inherent in traditional credit scoring models.
• To enhance the performance and reliability of credit scoring models, especially for peer-to-peer lending platforms.

Main Methods:

• Development of a semi-supervised hidden Markov model (SSHMM) framework.
• Application of the SSHMM to infer information from rejected credit applicants.
• Utilizing real-world data from the Lending Club peer-to-peer lending platform for empirical validation.

Main Results:

• The proposed SSHMM method demonstrates superior performance compared to existing reject inference approaches.
• The model exhibits significant stability and adaptability across different datasets and conditions.
• Empirical results confirm the effectiveness of SSHMM in mitigating selection bias.

Conclusions:

• The SSHMM framework offers a robust and effective solution for reject inference in credit scoring.
• This novel approach improves the accuracy and fairness of loan approval processes.
• The study highlights the potential of semi-supervised learning for addressing data limitations in financial modeling.