Intertemporal defaulted bond recoveries prediction via machine learning

Abstract

The recovery rate on defaulted corporate bonds has a time-varying distribution, a topic that has received limited attention in the literature. We apply machine learning approaches for intertemporal analysis of U.S. corporate bonds’ recovery rates. We show that machine learning techniques significantly outperform traditional approaches not only out-of-sample as documented in the literature but also in various out-of-time prediction setups. The newly applied sparse power expectation propagation approach provides the most compelling out-of-time prediction results. Motivated by the association of systematic factors with the time-varying characteristic of recovery rates, we study the effect of text-based news measures to account for bond investors’ expectations about the future which translate into market-based recovery rates. Especially during recessions, government-related news are associated with higher recovery rates. Although machine learning is a data-driven approach rather than considering economic intuition for ranking a group of predictors, the most informative groups of predictors for recovery rate prediction are nevertheless economically meaningful.