Gradient Boosting Model for Corporate Default

Abstract

Estimates of corporate default risk have improved from early agency rating scales to regression-based models, and more recently to Merton/structural and hybrid models. Despite their increasing accuracy and timeliness, access to default models is limited by high costs and computational complexity. In this study, we use extreme gradient boosting (XGBoost) to mimic the 1-year default probabilities generated by existing hybrid structural/statistical models. The dataset consists of over 1 million monthly, model-based, 1-year probability-of-default (PD) estimates from 2010 to 2019. A decision tree model with 50 input variables, including agency rating, spread-duration, industry sector, profitability, and other financial indicators is trained on PDs from 2010 to 2013, and tested on PDs from 2014 to 2019. PDs from the XGBoost model exhibit correlations of 0.8 with both DRISK and StarMine PDs, demonstrating its potential to provide consistent, timely, and accurate estimates of changes in credit risk. When PDs from the XGBoost model are substituted for hybrid-model PDs as input to relative value trading strategies, returns are similar in magnitude and monotonic, with returns increasing with relative value deciles. This is indicative of effectiveness of the XGBoost model in estimating the risk and relative value of corporate bonds.