Forecasting the U.S. Treasury Yield Curve Using the Hybrid Dynamic Nelson-Siegel and Long Short-Term Memory (LSTM) Method

Abstract

U.S. Treasury (UST) securities are widely regarded as safe-haven assets and serve as global financial benchmarks, making the U.S. Treasury yield curve a key indicator of market expectations and economic risks, including recession probabilities. For Indonesia, where foreign exchange reserves are partly allocated to UST securities, accurate yield curve forecasts are essential for effective reserve management and monetary policy formulation. This study proposes a hybrid forecasting framework that integrates the dynamic Nelson–Siegel (DNS) model with long short-term memory (LSTM) networks to improve the accuracy and stability of U.S. Treasury yield curve forecasts. The decay parameter in the DNS model is estimated using the Newton–Raphson method, while the remaining parameters are estimated using ordinary least squares (OLS). The resulting DNS latent factors are subsequently used as input features for the LSTM model under various hyperparameter configurations. Forecasting performance is evaluated using the root mean squared error (RMSE) and benchmarked against a DNS–ARIMA model. The empirical results demonstrate that the proposed DNS-LSTM approach consistently outperforms DNS-ARIMA across all maturities, yielding lower forecasting errors and greater flexibility in capturing yield curve dynamics, particularly during the post-pandemic period. Overall, the DNS-LSTM model offers a more robust and data-driven alternative to traditional yield curve forecasting methods. These findings have practical implications for foreign reserve management, exchange rate stabilization, and investment decision-making. Future research may extend this framework by incorporating macroeconomic variables and exploring longer forecast horizons.

Keywords: Bonds, yield, dynamic Nelson-Siegel, long short-term memory, U.S. Treasury