Comparing Machine Learning Models for Short-Term U.S. Treasury Yield Forecasting

This study examines historical trends in the U.S. 10-year Treasury yield and evaluates the effectiveness of four machine learning models, linear regression, decision tree, random forest, and multi-layer perceptron (MLP) neural networks, for short-term yield forecasting. As a key benchmark in global financial markets, the 10-year Treasury yield is influenced by multiple economic factors, including core inflation, the federal funds rate, GDP growth, and the U.S. Federal government's debt growth rate. Leveraging historical data from the Federal Reserve Economic Database (FRED), this study develops predictive models to assess the impact of these factors on yield fluctuations. Empirical results indicate that the random forest model outperforms the other approaches, achieving the lowest mean squared error (MSE) and mean absolute error (MAE), alongside an R² of 0.6073. This suggests its superior ability to capture nonlinear relationships in yield movements. The decision tree model also demonstrates competitive accuracy but is more susceptible to overfitting. Conversely, linear regression provides useful interpretability but struggles to capture complex economic interactions, leading to lower predictive accuracy. Despite its potential for handling nonlinear dependencies, the MLP model underperforms compared to the random forest, yielding an R² of 0.5058. The findings underscore the advantages of machine learning, particularly ensemble-based methods, in short-term Treasury yield forecasting.