A Fast Survival Support Vector Regression Approach to Large Scale Credit Scoring via Safe Screening.

Area of Science:

• Computational statistics
• Machine learning for finance

Background:

• Survival models are increasingly used in credit scoring to analyze time-dependent risk.
• Existing support vector regression survival models face challenges with censored data and efficiency in large datasets.

Purpose of the Study:

• To propose a novel algorithm, Buckley-James safe sample screening support vector regression (BJS4VR), for modeling large-scale survival data in credit scoring.
• To improve the accuracy and computational efficiency of survival analysis for credit risk assessment.

Main Methods:

• The BJS4VR algorithm combines the Buckley-James transformation with support vector regression.
• Censored samples are imputed using a Buckley-James estimator, ensuring unbiased estimation.
• Safe sample screening is employed to remove non-active samples, enhancing computational efficiency.

Main Results:

• The BJS4VR model demonstrated superior prediction accuracy compared to RSFM, CoxRidge, and CoxBoost on large-scale lending data.
• The proposed method significantly improved time efficiency in modeling large survival datasets.
• Key variables influencing credit risk were effectively identified.

Conclusions:

• The BJS4VR algorithm offers a robust and efficient approach for survival analysis in credit scoring.
• This method provides accurate risk prediction and aids in identifying critical credit risk factors.