An Evaluation of Machine Learning Models for Forecasting Short-Term U.S. Treasury Yields

This study explores the historical evolution and short-term predictive modeling of the U.S. 10-year Treasury bond yield, a critical indicator in global financial markets. Recognizing its sensitivity to macroeconomic conditions, the research integrates economic variables, including the federal funds rate, core Consumer Price Index (CPI), real Gross Domestic Product (GDP) growth rate, and the U.S. federal debt growth rate, to assess their influence on yield movements. Four forecasting models are employed for comparative analysis: linear regression (LR), decision tree (DT), random forest (RF), and multilayer perceptron (MLP) neural networks. Using historical data from the Federal Reserve Economic Data (FRED), this study finds that the RF model offers the most accurate short-term predictions, achieving the lowest mean squared error (MSE) and mean absolute error (MAE), with an R2 value of 0.5760. The results highlight the superiority of ensemble-based nonlinear models in capturing complex interactions between economic indicators and yield dynamics. This research not only provides empirical support for using machine learning in economic forecasting but also offers practical implications for bond traders, system developers, and financial institutions aiming to enhance predictive accuracy and risk management.