Enhancing credit scoring accuracy with a comprehensive evaluation of alternative data.

Area of Science:

• Financial Risk Management
• Data Science
• Machine Learning

Background:

• Traditional credit scoring models rely heavily on credit bureau data.
• Alternative data sources are often overlooked in credit risk assessment.
• Enhancing predictive accuracy in credit scoring remains a key challenge.

Purpose of the Study:

• To investigate the impact of alternative data on credit scoring model accuracy.
• To compare the performance of models using traditional versus combined data sources.
• To identify novel predictors for improved credit risk evaluation.

Main Methods:

• Analysis of a comprehensive home loan portfolio dataset from Home Credit Group.
• Application of the model-X knockoffs framework for systematic variable selection.
• Inclusion of alternative predictors such as social network default status, regional economic ratings, and local population characteristics.

Main Results:

• Credit scoring models incorporating alternative data demonstrated improved predictive performance.
• The enhanced models achieved an area under the curve (AUC) of 0.79360 on the Kaggle Home Credit default risk dataset.
• Performance surpassed models relying solely on traditional credit bureau data.

Conclusions:

• Leveraging diverse, non-traditional data sources significantly augments credit risk assessment.
• Alternative data enhances overall credit scoring model accuracy and predictive power.
• The study underscores the value of incorporating overlooked data for robust financial modeling.