Enhancing U.S. Treasury Yield Curve Forecasting Through Machine Learning and Forecast Combination Techniques

This study presents an innovative framework for improving US Treasury yield curve forecasting, with a dual emphasis on enhancing prediction accuracy and ensuring interpretability in machine learning models, particularly during periods of financial stress. The framework integrates advanced machine learning algorithms with adaptive forecast combination techniques, emulating the decision-making process of professional forecasters to improve economic transparency.

We introduce the Enhanced Factor-Augmented Dynamic Nelson-Siegel (Enhanced FADNS) model, which leverages 10 principal components derived from 151 macroeconomic variables (adjusted for stationarity) and the market value of Treasury debt. To capture non-linear dependencies and dynamic interactions among market drivers, we incorporate Random Forest and Long-Short-Term Memory (LSTM) models. Forecasts are combined using eight techniques, including Equal Weights, Rank-Weighted Combinations, three Thick Modeling approaches, and three variants of the Aggregated Forecast Through Exponential Reweighting (AFTER) algorithm. The AFTER algorithm, which dynamically adjusts weights based on historical performance, consistently outperforms other methods and demonstrates resilience to structural breaks.

Our empirical evaluation extends to sensitivity assessments, cointegration tests, and the development of error correction models to assess the robustness of the AFTER algorithm. Rigorous robustness assessment is conducted to ensure the validity of results. First, all analyses are replicated using end-of-month yield observations to verify consistency across data sampling methodologies. Second, model performance is evaluated without the LSTM framework to ensure that improvements in predictive accuracy are not reliant on deep learning models.

Although Treasury yields are influenced by short-term market fluctuations, our findings reveal that they ultimately converge to underlying macroeconomic and supply-demand fundamentals.