Applied Machine Learning for Forecasting Bond Excess Returns and Building Portfolios

Maria Francisca Faria de Carvalho Azevedo Roque

Key information

Authors:

Maria Francisca Faria de Carvalho Azevedo Roque (Maria Francisca Faria de Carvalho Azevedo Roque)

Supervisors:

Nuno Cavaco Gomes Horta (Nuno Cavaco Gomes Horta); Duarte Caetano Rodrigues

Published in

November 18, 2025

Abstract

The bond market offers returns that compensate investors for bearing credit and liquidity risk. Accurately forecasting these returns could enhance investment decisions and portfolio allocation, yet these premia remain challenging to model. This work tackles two questions: (i) can Machine Learning models predict bond returns one month ahead; and (ii) can trading strategies based on model derived signals deliver robust and economically meaningful investment performance. This work combines bond and issuer characteristics with macro financial predictors and compare regularized regressions, Random Forest, Neural Networks and a glass-box method, Explainable Boosting Machine. Forecasting results show that EBM delivers performance on par with the flexible alternatives achieving out-of-sample R^2 of 9.81% against a naive zero benchmark and 6.73% against a rating-average benchmark. Its additive structure reveals intuitive drivers, with financial uncertainty emerging as the dominant factor, followed by credit spreads. Interactions between macro conditions and bond characteristics are crucial, especially the link between uncertainty with spreads and yields, so the payoff to wide spreads or high yields strengthens when uncertainty falls and weakens when it rises. Signals are then mapped into portfolios under explicit turnover costs. The strongest results arise from Sorted Quantiles built from an interpretable EBM Top K factor, which achieve a net Sharpe of 2.11 in a buy-and-hold long-only setup and 2.03 in a long-short spread, consistently outperforming characteristics sorted quantiles. This analysis uses a monthly panel of euro-denominated bonds from 2010-2025, combining Investment Grade and High Yield constituents from the Bloomberg LECPTREU and LP02TREU indices, respectively.