Improving corporate bond recovery rate prediction using multi-factor support vector regressions

In the multi-factor framework described in this paper, we use instrument-specific characteristics, several macroeconomic variables, and industry-specific characteristics as our explanatory variables for predicting recovery rates for corporate bonds. By including the principal components derived from a large number of macroeconomic variables, all three least-squares support vector regression methods, as well as the ordinary linear regression, exhibit higher out-of-sample predictive accuracy than the models that included only the few macroeconomic variables suggested in the literature. We compare the prediction accuracies of all techniques by incorporating sparse principal components, nonlinear principal components from an auto-associative neural network, and kernel principal components. Our results show that sparse principal components generate more interpretable and accurate estimations compared to the other principal component techniques. Moreover, we apply gradient boosting to generate a ranking of the 104 macroeconomic variables, from best to worst, based on their prediction power in recovery rate estimation. The three categories with the most informative macroeconomic predictors are micro-level factors, business cycle variables, and stock market indicators.