Credit Risk Modeling Using Transfer Learning and Domain Adaptation.

Area of Science:

• Financial Risk Management
• Machine Learning in Finance
• Credit Scoring Models

Background:

• Lenders often lack historical data for Micro, Small, and Medium Enterprises (MSMEs), hindering accurate credit risk assessment and Probability of Default (PD) prediction.
• This data scarcity restricts credit access and increases costs for MSMEs.
• Transfer learning offers a potential solution by leveraging data from related credit risk domains.

Purpose of the Study:

• To apply transfer learning and a novel Progressive Shift Contribution (PSC) concept to predict PD in the MSME lending domain.
• To explain transfer learning model behavior using Shapley values.
• To develop and test domain adaptation techniques for differing source and target datasets.

Main Methods:

• Utilized transfer learning from credit card and debt consolidation (CD) domains to predict PD for MSMEs.
• Applied Shapley values to interpret transfer learning model improvements.
• Developed and implemented a domain adaptation strategy, including feature selection and value adjustment algorithms.
• Proposed a combined strategy of feature selection and adaptation for improved domain alignment.

Main Results:

• Transfer learning successfully improved PD prediction accuracy for MSMEs.
• Shapley values provided insights into how transfer learning enhances model performance.
• Domain adaptation further boosted model accuracy beyond transfer learning alone.
• The combined feature selection and adaptation strategy effectively aligned source and target domain features.

Conclusions:

• Transfer learning, particularly with domain adaptation, is a viable method to improve credit risk assessment for MSMEs.
• The Progressive Shift Contribution (PSC) concept offers a novel way to understand and implement transfer learning.
• While percentage improvements may seem small, they hold significant economic value for real-world lending practices.
• Interpretable AI methods like Shapley values are crucial for understanding and trusting complex models in finance.