Glass box machine learning and corporate bond returns

Machine learning methods in asset pricing are often criticized for their black box nature. We study this issue by predicting corporate bond returns using interpretable machine learning on a high-dimensional bond characteristics data set. We achieve state-of-the-art performance while maintaining an interpretable model structure, overcoming the accuracy–interpretability trade-off. The estimation uncovers nonlinear relationships and economically meaningful interactions in bond pricing, notably related to term structure and macroeconomic uncertainty. Subsample analysis reveals stronger sensitivities to these effects for small firms and long-maturity bonds. Finally, we demonstrate how interpretable models enhance transparency in portfolio construction by providing ex ante insights into portfolio composition.