1. Report Industry Investment Rating No relevant content provided. 2. Core View of the Report The report systematically reviews the evolution of financial time - series forecasting models and constructs a prediction model for China's 10 - year treasury bond yield using a long - short - term memory (LSTM) neural network with historical time series as the single input variable, initially exploring the application of this deep - learning model in the fixed - income quantitative field [10]. 3. Summary by Relevant Catalog 3.1 Financial Time - Series Forecasting and Neural Network Models 3.1.1 Evolution of Financial Time - Series Forecasting Models Financial time - series forecasting has gone through three main development stages: traditional econometric models, traditional machine - learning models, and deep - learning models. Traditional econometric models have clear forms and strong interpretability but struggle to depict nonlinear and complex dynamic relationships. Traditional machine - learning models can perform nonlinear fitting and automatic feature screening but need manual feature extraction. Deep - learning models can automatically extract features from raw data and capture complex long - term time - series patterns, adapting well to the complex characteristics of financial time series [11][12]. 3.1.2 Neural Network Models and LSTM Models Neural network models are machine - learning models imitating the connection structure of human brain neurons. Recurrent neural networks (RNN) and their variants, such as LSTM, are designed for processing sequence data. LSTM solves the long - term dependence problem of traditional RNN through a "gating mechanism" and memory units, enhancing robustness to irregular data and being suitable for bond yield prediction [13][18]. 3.2 Treasury Bond Yield Prediction Based on Stacked LSTM Model 3.2.1 Stacked LSTM Model Stacked LSTM connects multiple LSTM layers in sequence, having advantages in long - sequence processing and multi - dimensional feature extraction, more suitable for complex time - series forecasting in financial scenarios [23]. 3.2.2 Construction of Treasury Bond Yield Prediction Model The report uses a classic and robust architecture of three - layer stacked LSTM + Dropout regularization to build a neural network model for predicting the 10 - year treasury bond yield. It only uses the historical time series of the 10 - year treasury bond yield as a single variable for prediction. The data is from the beginning of 2021 to December 12, 2025. After data processing and sample construction, a medium - complexity LSTM neural network model with about 130,000 adjustable parameters is built. The optimal model is obtained at the 27th training iteration, with an average absolute error of 1.43BP for the test set. The predicted yield on December 19, 2025, is 1.8330%, slightly lower than 1.8396% on December 12, 2025 [2][24][30]. 3.3 Follow - up Optimization Directions - Optimize model design: Adjust and optimize the design related to time windows, data processing, network architecture, and training strategies [3][36]. - Input multi - dimensional variables: Expand input variables from a single yield sequence to multi - dimensional variables such as macro, market, and sentiment to make the model more in line with economic logic and capture more comprehensive information [3][36]. - Build hybrid models: Combine the LSTM model with traditional econometric models or other machine - learning models to build hybrid models like ARIMAX - LSTM and CNN - LSTM - ATT, enhancing prediction accuracy [3][36]. - Introduce a rolling back - testing mechanism: Use a rolling time - window back - testing mechanism to update the model dynamically and make continuous predictions, improving the model's adaptability to market changes [3][36].