Real-Time Forecasting of U.S. Treasury Bond Excess Returns
An evaluation of statistical and machine-learning forecasts of Treasury bond excess returns using macroeconomic data available at each forecast date.
- Fixed Income
- Macro
- Machine Learning
Treasury return forecasts may appear stronger when they rely on macroeconomic data that was revised after the forecast date. Investors making decisions in real time only had the data available at the time, so revised series may overstate how well a model would have performed.
One-year excess holding-period returns on two- through five-year Treasury bonds will be constructed and validated before comparing historical-mean, Cochrane–Piazzesi, regularized linear, and gradient-boosting models. Forecasts using real-time ALFRED vintages will then be compared with forecasts using revised data, with 2020 through 2025 reserved for final out-of-sample evaluation.
The paper will be accompanied by the data loader, benchmark implementations, results, and reproduction instructions.