2025 年 12 月 15 日 总量研究 基于堆叠 LSTM 模型的十年期国债收益率预测 ——量化学习笔记之一 要点 1、 金融时序预测和神经网络模型 针对金融时间序列的预测，经历了从传统计量模型、到传统机器学习模型、再到 深度学习模型的三个主要发展阶段。深度学习模型能够较好适应金融时间序列的 非平稳、非线性、高噪声和长记忆性等复杂特征，是当前主流的金融时序预测方 法之一。 模型设计优化。在现有模型的基础上，针对时间窗口、数据处理、网络架构和训 练策略等相关设计进行调整优化。 输入多维度变量。将输入变量从单一的收益率序列扩展至宏观、市场、情绪等多 维度变量，使模型预测更加符合经济逻辑，捕捉信息更加全面。 构建混合模型。将 LSTM 模型与传统计量模型或其他机器学习模型相结合，构建 如 ARIMAX-LSTM、CNN-LSTM-ATT 等混合模型，能够发挥不同模型优势，弥补 单一 LSTM 模型缺陷，提升预测精度。 引入滚动回测机制。采用滚动时间窗口回测机制，固定样本时间窗口并随时间推 移实现模型的动态更新和持续预测，使模型能更好适应市场变化，提升其稳健性。 4、 风险提示 模型结构简单导致当前预测误差较大； ...

Summary of the Conference Call on the Bond Market Analysis Industry Overview - The analysis focuses on the Chinese bond market, specifically the ten-year government bond yield predictions for 2025 [2][4]. Key Points and Arguments - **Yield Prediction Model**: The model decomposes the ten-year government bond yield into trend and cycle components, achieving a fitting goodness of 0.98. The predicted yields for June and December 2025 are approximately 1.91% and 1.61%, respectively [2][6]. - **Market Behavior**: The market has shown hesitation around the 1.6% yield level, influenced by macroeconomic data improvements and tightening funds. The model aims to analyze these factors to better understand current yield levels and forecast future trends [4][11]. - **CPI and PMI Correlation**: The relationship between CPI and bond market cycles has changed over time. Before 2013, CPI growth was positively correlated with bond cycles. From 2013 to 2019, PMI data became the key indicator, while post-2020, CPI showed a negative correlation with bond cycles due to monetary policy effects [8][10]. - **Interest Rate Predictions**: The model forecasts other maturities based on the ten-year yield, predicting one-year, three-year, five-year, and seven-year yields to be approximately 0.99%, 1.26%, 1.42%, and 1.59% by December 2025 [10][12]. - **Market Sentiment**: The model serves as a neutral anchor, with actual market values expected to fluctuate around this anchor. Expectations of interest rate cuts may lower yields, while rate hikes could increase them [10][11]. - **Model Reliability**: Backtesting shows the model's fitting deviation is within 10%, indicating its reliability. This deviation can help identify overly optimistic market conditions, aiding investors in adjusting strategies [3][13][15]. Additional Important Insights - **Market Dynamics**: The early 2025 market behavior has led to a pessimistic outlook, with the neutral space being consumed early in the year. A breakthrough below 1.6% will require unexpected market stimuli [11][12]. - **Investment Decision Making**: The model provides a reliable benchmark for assessing market sentiment and potential overvaluation. When market deviations exceed 10%, it signals a need for strategy adjustments [14][16]. - **Future Adjustments**: The model is a tool for judgment and should be used alongside market assessments, especially in the face of unexpected events [7][9]. This comprehensive analysis highlights the dynamics of the Chinese bond market and the predictive capabilities of the model developed by the research team.