Multiscale Nonlinear Forecasting of Government Bond Yields and Volatility via a Hybrid VMD–LSTM Framework

Government bond yields and volatility exhibit nonlinearity, complexity, and noise, making accurate forecasting challenging for conventional econometric or deep learning models alone. This study develops a multiscale nonlinear forecasting framework that combines variational mode decomposition (VMD) with a long short-term memory (LSTM) model to forecast China’s government bond yields and volatility. By decomposing the time series into trend, periodic, and disturbance components, the hybrid model effectively captures both linear and nonlinear patterns while mitigating overfitting. In the empirical analysis, five loss functions—MSE, RMSE, MAE, MAPE, SMAPE—and the DM test are used as evaluation criteria to compare the predictive performance of ARIMA, SVM, LSTM, VMD-SVM, and VMD-LSTM models. Using the yields and volatility of 3-year government bonds as the benchmark case and 1-year government bonds for robustness tests, the results indicate that the VMD-LSTM model achieves superior predictive accuracy, demonstrating its effectiveness and robustness. The proposed hybrid model offers a novel framework for government bond yield forecasting, providing valuable insights for monetary policy and financial risk monitoring.