Machine Learning the Cross-Section of Corporate Bond Excess Returns

Abstract

We use machine learning, to predict the cross-section of corporate bond excess returns, both directly and via rank-transformations. We find some evidence that non-linearities matter for corporate bond excess return predictions and that using untransformed fea-tures produces the best out-of-sample results for the non-linear models. However, the fragility of the models in terms of the han-dling of outliers and transformation of features is something that needs careful consideration when trying to predict corporate bond excess returns. We also find that a forecast’s OOS-R2 is not nec-essarily a good indication of portfolio performance.