Bond Risk Premia with Machine Learning
96 Pages Posted: 15 Jun 2019 Last revised: 23 Jun 2019
Date Written: April 1, 2019
We compare a variety of machine learning methods for bond return predictability in the context of regression-based forecasting. Neural networks outperform both linear penalized regressions and non-linear shallow learners like random forests. These results hold both for annual and monthly holding-period bond excess returns. A novel approach based on ensembled deep neural networks shows that macroeconomic information substantially improves the out-of-sample predictability of bond excess returns. Non-linear features within, rather than across, economic categories are key to our results. Finally, using a simple trading strategy we show that the statistical performance of neural networks translates into large economic gains.
Keywords: Machine Learning, Ensembled Networks, Forecasting, Bond Return Predictability, Empirical Asset Pricing.
JEL Classification: C38, C45, C53, E43, G12, G17
Suggested Citation: Suggested Citation