A Rigorous Statistical Comparison of Deep Learning Models for US Treasury Yield Prediction

The intrinsic nonlinearity and dynamic relationships in interest rate fluctuations present a substantial challenge when forecasting financial time series, particularly US Treasury yields. These intricate relationships are sometimes not adequately captured by traditional econometric models. In recent years, deep learning (DL) methodologies have gained prominence in the financial market, offering advanced predictive capabilities by modeling high-dimensional dependencies and nonlinear interactions inside yield curves. To enhance the predictive accuracy of short-term (13-week) and long-term (5-year) US Treasury yields, this study leverages advanced deep learning models, including convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and gated recurrent units (GRUs). A comprehensive statistical evaluation is performed to assess model performance through key error metrics such as root mean squared error (RMSE), mean squared error (MSE), mean absolute error (MAE), the coefficient of determination ( $$R^2$$ R 2 ), maximum error, and minimum error, as well as SAFE metrics (Sustainability, Accuracy, Fairness, Explainability) for a holistic assessment. To ensure a robust comparison, we employed the paired t-test to determine if the differences in model predictions are statistically significant. Additionally, we analyzed correlation metrics using Pearson and Spearman coefficients, which evaluate the models’ ability to capture both linear dependencies and ranking trends in yield fluctuations. This rigorous framework not only benchmarks the predictive power of each model but also provides deeper insights into their effectiveness in forecasting treasury yields across different time horizons.