Deep Learning for Corporate Bonds

61 Pages Posted: 1 Aug 2023 Last revised: 12 Sep 2023

See all articles by Jetlir Duraj

Jetlir Duraj

Department of Economics, University of Pittsburgh

Oliver Giesecke

Stanford University - Hoover Institution

Date Written: July 31, 2023

Abstract

We estimate an asset pricing model for the U.S. corporate bonds market using bond portfolios, as well as a large longitudinal dataset of individual bonds that we augment with fundamental characteristics of the issuer. We further enrich the information set with a large set of macroeconomic time series. We estimate diverse model architectures with two approaches: (1) minimizing the mispricing loss, and (2) maximizing the Sharpe ratio. We find that, contrary to the equivalence of these two approaches in the sense of financial theory, maximizing the Sharpe ratio performs better for individual bonds, whereas the difference is smaller for bond portfolios. The out-of-sample annual SDF portfolio Sharpe ratios are in the range of .59 to 1.00, and show statistically significant excess returns (alphas) relative to conventional risk factors. Our results are robust to the exclusion of financials and REITs.

Keywords: conditional asset pricing, corporate bonds, deep learning, general method of moments, big data, stochastic discount factor, Sharpe ratio

JEL Classification: G12, G11, G13

Suggested Citation

Duraj, Jetlir and Giesecke, Oliver, Deep Learning for Corporate Bonds (July 31, 2023). Available at SSRN: https://ssrn.com/abstract=4527372 or http://dx.doi.org/10.2139/ssrn.4527372

Jetlir Duraj (Contact Author)

Department of Economics, University of Pittsburgh ( email )

Pittsburgh, PA 15260
United States

Oliver Giesecke

Stanford University - Hoover Institution ( email )

Stanford, CA 94305
United States

Do you have negative results from your research you’d like to share?

Paper statistics

Downloads
853
Abstract Views
2,060
Rank
51,527
PlumX Metrics