A1112
Title: Default risk prediction using a Cox model with moderately clipped lasso penalty
Authors: Keongmuk Lim - Chungnam National University (Korea, South) [presenting]
Sangin Lee - Chungnam National University (Korea, South)
Abstract: Cox proportional hazard model is a fundamental tool in survival analysis. For high dimensional survival data with right censoring, various survival models have been proposed to analyze large-scale survival data, including lasso, ridge, or scad penalized Cox models. The penalized Cox models enable both predicting relative risk for each individual and selecting relevant predictive variables. A novel penalized method is introduced for the Cox model, called the moderately clipped lasso (MCL), which integrates the strengths of the minimax concave penalty (MCP) and lasso penalty. Hence, it preserves the advantages of both the lasso and MCP methods. An efficient optimization algorithm is developed that combines the concave-convex procedure, modified local quadratic approximation, and coordinate descent. Simulation studies demonstrate that the proposed method achieves superior variable selection and prediction performances compared to the other existing methods. Its practical utility is further validated by applying the MCL model to a time-to-default prediction task using Bondoras peer-to-peer loan dataset, where it consistently outperforms the existing methods in credit risk modeling.
