统计与数据科学学院知识库

一直以来，金融风险的管控都是金融界的焦点，而金融资产波动率作为量化风险的重要指标，它的精准预测就成为了研究热点。在理论层面上，金融资产波动率研究是金融市场学科的一个重要分支领域；实践层面上，预测波动率并把握其高波动率区间，有利于国家宏观调控以及使投资者规避风险，也是防范贸易、投资等领域风险的重要方式。在实际研究过程中波动率的预测具有不小的挑战，一方面是其影响因素众多，不同影响因素之间具有复杂的非线性关系；另一方面原因在于大数据时代的来临及互联网技术的应用发展，使金融数据级别达到了海量单位，数据之间的复杂性及非结构化形式进一步加强，导致使用传统的经济计量预测模型已很难得到较为精确的结果。因此本文尝试将传统计量经济学模型和深度学习模型相结合，使用融合模型的框架进行证券价格波动率的预测研究，以期达到融合模型中的计量模型部分提供合理的参数解释，而深度学习模型的非线性处理能力带来预测精度提升的目的。

论文首先对金融市场波动率的预测历史和现状做出综述，随后介绍了相关理论基础以及波动率预测模型，最后使用单独模型和融合模型分别进行实证研究，该过程使用的是沪深300指数和中证500指数的5分钟高频收盘价格数据。具体的研究工作和结果如下：（1）对两指数的波动率序列进行相关统计性检验，发现已实现波动率可以很好刻画真实波动率的分布特征，即可作为代理变量使用；（2）比较Realized GARCH模型在不同残差分布假设下的拟合预测能力，基于多种损失函数的结果证明，当假设残差服从于t分布或偏斜t分布时，能提高模型预测精确性；（3）对LSTM模型的最优参数进行选择，发现LSTM模型和Realized GARCH模型相比预测能力更优；（4）使用Realized GARCH模型得到模型参数后，将样本内拟合值和真实值的残差作为LSTM模型输入，将两部分预测结果相加后得到的融合模型表现优于单一模型。文章末尾进行了结论总结和未来研究展望，发现由于不同数据集数据分布特点的原因，不存在无差别的最优模型，模型选择和参数设定等都能影响模型的表现，这对未来的波动率建模研究具有一定的借鉴意义。

The management and control of financial risks has always been the focus of the financial community. The volatility of financial assets is an important indicator to quantify risks and its precise prediction has become a research hotspot. At the theoretical level, the study of financial asset volatility is an important branch of the financial market discipline; at the practical level, predicting volatility and grasping its high volatility range is conducive to the country’s macro-control and allows investors to avoid risks, as well as preventing trade , Investment and other fields of risk. In the actual research process, the prediction of volatility is not a small challenge. On the one hand, there are many influencing factors and the complex nonlinear relationship between different influencing factors; on the other hand, the reason is the advent of the era of big data and the application of Internet technology. Development has enabled the level of financial data to reach massive units, and the complexity and unstructured forms of data have been further strengthened. As a result, it is difficult to obtain more accurate results using traditional econometric forecasting models. Therefore, this article attempts to combine the traditional econometrics model with the deep learning model, and use the framework of the fusion model to predict the volatility of securities prices, in order to achieve reasonable parameter explanations for the econometric model part of the fusion model, and the deep learning model Non-linear processing capabilities bring the purpose of improving prediction accuracy.

The thesis first summarizes the forecast history and current situation of financial market volatility, then introduces the relevant theoretical basis and volatility forecast model, and finally uses a separate model and a fusion model to conduct empirical research respectively. The process uses the Shanghai and Shenzhen 300 index and The 5-minute high-frequency closing price data of the CSI 500 Index. The specific research work and results are as follows: (1) Carry out correlation statistical test on the volatility series of the two indices, and find that the realized volatility can well describe the distribution characteristics of the true volatility, which can be used as a proxy variable; (2) Comparing the fitting prediction ability of the Realized GARCH model under different residual distribution assumptions, the results based on various loss functions prove that when the residuals are assumed to obey the t distribution or the skewed t distribution, the prediction accuracy of the model can be improved; (3) Select the optimal parameters of the LSTM model, and find that the LSTM model has better predictive ability than the Realized GARCH model; (4) After the Realized GARCH model is used to obtain the model parameters, the residuals between the fitted values ​​and the true values ​​in the sample are taken as The LSTM model is input, and the fusion model obtained by adding the two parts of the prediction results is better than the single model. At the end of the article, a conclusion summary and future research prospects are made. It is found that due to the data distribution characteristics of different data sets, there is no undifferentiated optimal model. Model selection and parameter settings can affect the performance of the model, which will affect future fluctuations. Rate modeling research has certain reference significance.