Core Viewpoint - The report focuses on the development and evaluation of a hybrid neural network model for predicting the long-term trends of the ten-year government bond yield, utilizing various macroeconomic, monetary policy, and market sentiment indicators to enhance predictive performance [3][5]. Group 1: Hybrid Neural Network Model - A hybrid neural network model integrates multiple neural network architectures to improve learning capabilities and performance, commonly combining CNN, GRU, LSTM, and attention mechanisms [4]. - CNN captures short-term local features, while GRU and LSTM are adept at learning long-term trends, and the attention mechanism helps the model focus on significant time points [4]. Group 2: Research Design - The research compares single neural network models with hybrid models, training multiple networks to assess their predictive performance on the ten-year government bond yield [5]. - Innovations include multi-dimensional input and output for time series prediction, the introduction of hybrid models, and a focus on medium to long-term yield trends rather than short-term fluctuations [5]. Group 3: Research Conclusions - The single GRU model demonstrated the best overall predictive performance [6]. - The accuracy of the optimal model in predicting the direction of yield changes improves with longer prediction time spans [6]. - The optimal model predicts a decline of approximately 3 basis points (BP) in the ten-year government bond yield from the end of January 2026 to the end of February 2026, and a decline of about 6 BP from the end of 2025 to the end of 2026 [6].