| dc.description.abstract | U.S. Treasury yields have a significant impact on the economy, and the term structure of interest rates and changes in the yield curve are often used to predict economic cycles. Generally, U.S. Treasury yields are closely related to the federal funds rate, with the market predicting yields based on information released by the Federal Reserve. However, in recent years, the severe inflation problems caused by the pandemic, the Russia-Ukraine war, and the energy crisis have led to frequent adjustments in monetary policy by the Federal Reserve to stabilize the economy, increasing the difficulty of making accurate predictions.
This study aims to construct a model to predict U.S. Treasury yields using a more diverse set of macroeconomic, monetary policy, and financial variables. Traditional linear prediction methods often suffer from overfitting, affecting accuracy. Therefore, machine learning methods are introduced to address nonlinear issues and handle high-dimensional data. By comparing different machine learning prediction models with traditional regression models, the study seeks to identify the best prediction method for short, medium, and long-term Treasury yields.
The research results indicate that traditional benchmark models are suitable for extremely short-term predictions, but their predictive ability declines as the forecast period extends. Machine learning models excel in handling high-dimensional nonlinear data, particularly showing greater accuracy in predictions beyond two periods. Among them, the random forest model performs best and remains stable in predicting short, medium, and long-term U.S. Treasury yields. The deep neural network model performs well in predictions beyond eight periods but shows larger errors when faced with emergency policy adjustments. Overall, the random forest model demonstrates high stability and accuracy in predicting short, medium, and long-term U.S. Treasury yields. | en_US |