Predictive Model Design. The excess returns of assets, specifically bonds, are characterized by a model with an additive prediction error. A panel approach is used to predict bond returns, assuming a homogeneous conditional expectation functional form that applies to all bond-return observations. The conditional expectation is expressed as the sum of bond characteristics and aggregate predictors. The model is recursively estimated with a rolling window of the past 20 years at the end of each year, and then fixed for the next 12 months for prediction. Different predictive models, including combination forecast, penalized linear regression, dimension reduction, and ensemble tree models, are surveyed for performance evaluation.
ASSET CLASS: ETFs, funds | REGION: United States | FREQUENCY:
Monthly | MARKET: bond | KEYWORD: Corporate Bonds, Machine learning
I. STRATEGY IN A NUTSHELL
This strategy predicts corporate bond returns using machine learning models, including random forest, lasso, and partial least squares. A panel approach combines bond-specific characteristics and aggregate economic predictors. The models are evaluated with out-of-sample R² metrics and Fama-MacBeth regressions, identifying the most influential predictors for public and private bonds across different rating and maturity groups. Market-timing is implemented by longing bonds with positive return predictions and shorting those with negative predictions, creating long-short portfolios optimized for risk-adjusted performance.
II. ECONOMIC RATIONALE
The strategy’s effectiveness stems from improved data availability, long-span historical datasets, and the adoption of advanced machine learning techniques. By capturing nonlinear relationships and differentiating public versus private bond behaviors, it delivers robust return forecasts, enabling better-informed investment and risk management decisions across the corporate bond market.
III. SOURCE PAPER
Predicting Individual Corporate Bond Returns [Click to Open PDF]
Xin He, Hunan University; Guanhao Feng, City University of Hong Kong (CityU); Junbo Wang, City University of Hong Kong (CityU) – Dept. of Economics and Finance; Chunchi Wu, SUNY at Buffalo – School of Management
<Abstract>
This paper documents substantial evidence of return predictability and investment gains for individual corporate bonds via machine learning. The forecast-implied long-short and market-timing strategies deliver significant risk-adjusted returns over transaction costs. Random Forest has the best performance as the ensemble of regression trees helps reduce overfitting. Using a long-span sample from 1976 to 2020, we can evaluate return predictability over business cycles and find aggregate predictors (e.g., corporate bond market return and TERM factor) that show higher forecasting power than bond characteristics. Finally, we find return predictability differs between bonds issued by private and public firms, with higher investment gains in private bonds.
IV. BACKTEST PERFORMANCE
| Annualised Return | 19.7% |
| Volatility | 11.45% |
| Beta | N/A |
| Sharpe Ratio | 1.72 |
| Sortino Ratio | N/A |
| Maximum Drawdown | N/A |
| Win Rate | N/A |
