基于CNN-LSTM-Attention的国债收益率预测研究
Research on the Prediction of Treasury Bond Yield Based on CNN-LSTM-Attention Model

国债是利率债的一种重要形式，其到期收益率在经济金融体系中扮演着关键的参考角色。当前，市场对国债收益率的预测精度提出了更高的要求，这不仅体现在对宏观经济趋势的判断上，也体现在微观利率债交易中。然而，宏观经济分析涉及的指标众多且错综复杂，指标间的领先滞后关系也在不断变化，这使得传统的定性分析和定量分析方法在预测的精确度和准确性方面存在一定的局限性。鉴于此，本文采用了一种新型的深度学习预测模型来预测长期国债的收益率，并将其与基础的LSTM模型进行比较。该模型融合了卷积神经网络(CNN)、长短期记忆网络(LSTM)和注意力机制(AM)。首先，通过CNN对输入数据进行特征提取；然后，LSTM利用这些特征数据来预测国债次日的收益率；最后，借助调幅法来捕捉不同时间特征状态对国债收益率的影响，以提升预测的精度。根据机器学习时序预测技术的研究成果，集成CNN-LSTM-Attention模型相比基础LSTM模型能得到更好的预测效果，本文对此进行了验证，得到的结果与此相符。
Treasury bond is an important form of interest rate debt, and its yield to maturity plays a key reference role in the economic and financial system. At present, the market has put forward higher requirements for the prediction accuracy of treasury bond yield, which is not only reflected in the judgment of macroeconomic trends, but also reflected in the trading of micro interest rate bonds. However, macroeconomic analysis involves numerous and complex indicators, and the leading and lagging relationships between indicators are constantly changing, which limits the accuracy and precision of traditional qualitative and quantitative analysis methods in forecasting. In view of this, this paper adopts a new deep learning prediction model to predict the yield of long-term treasury bond, and compares it with the basic LSTM model. This model integrates Convolutional Neural Network (CNN), Long Short Term Memory Network (LSTM), and Attention Mechanism (AM). Firstly, feature extraction is performed on the input data using CNN; Then, LSTM uses these characteristic data to predict the yield of treasury bond the next day; Finally, the amplitude modulation method is used to capture the impact of different time characteristic states on treasury bond yield, so as to improve the accuracy of prediction. According to the research results of machine learning time series prediction technology, the integrated CNN-LSTM-Attention model can achieve better prediction performance compared to the basic LSTM model. This paper verifies this and the results obtained are consistent with this.