基于中国股市高频数据的流动性风险研究
发布时间：2018-09-04 17:08
【摘要】：在金融市场全球化与衍生品交易不断繁荣的背景下,多起金融危机事件的发生促使了VaR的诞生,它已成为市场风险的标准计量方法,在金融界与学术界广受关注。虽然VaR具有综合性、量化性、通俗性等优点,但仍然存在诸多缺陷,其中之一即为假设投资组合不论头寸大小都能以当前市场价格瞬间出清,从而忽略了流动性风险。本文将流动性风险纳入VaR基本框架,并且分别基于高频数据与超高频数据构建流动性调整的VaR模型(La-VaR),最后作实证研究与对比分析。 本文的核心工作之一是针对高频数据建立La-VaR模型。对高频数据的La-VaR建模是源于BDSS模型基本框架,但是针对其正态分布假设、同方差假设、相对价差与中间价格不相关假设、市场风险与流动性风险同步最大化假设、流动性非动态假设等缺陷作出了改进,不仅可计算单个资产,还可计算投资组合的La-VaR。 具体地,首先构建GJR-GARCH-EVT-kernel模型刻画收益率序列的尖峰厚尾性、异方差性、波动非对称性以及上下尾极值分布特点,并采用多元Copula模型捕捉不同资产序列之间的相关结构,接着采用蒙特卡罗模拟对收益率序列进行一步预测。第二,类似地采用GJR-GARCH-EVT-kernel模型拟合相对价差的边缘分布,然后分别在每个单一资产中,采用二元阿基米德Copula模型刻画相对价差与收益率之间的相关结构,再基于预测的收益率序列生成相对价差的伪随机数作为其一步预测序列。第三,通过多条路径的预测最后得出交易价格的分位数,进一步便求出La-VaR值。 实证分析表明VaR存在一定程度上的风险低估,La-VaR计量的风险中流动性风险比例一般为10%-20%,并且经无条件覆盖性、独立性、条件覆盖性等检验,La-VaR基本上不存在风险低估现象,要明显优于VaR模型。 本文的第二个核心工作是基于超高频数据建立La-VaR模型。前述基于高频数据的La-VaR模型实际上是对相等时间间隔序列进行建模,但是超高频数据与传统时间序列相比存在着本质区别,即非等时间间隔性,因此无法直接应用传统时间序列模型,需要引入持续期序列,并对非等时间间隔的收益率与相对价差进行转换后再建模。 具体地,首先建立WACD模型拟合持续期序列,并迭代预测多步持续期直至累计持续期之和达到高频时间间隔(如1分钟)。第二,采用持续期对超高频数据序列进行转换后得到单位收益率与单位相对价差以满足相等时间间隔性,再将其纳入前述的GJR-GARCH-EVT-kernel-Copula框架中进行参数估计与多步预测,其步数与持续期的预测步数相同。第三,将多步预测的单位收益率与单位相对价差经持续期转换得到多步预测的分笔收益率与相对价差,再聚合得到预测的高频时间间隔(如1分钟)的收益率与相对价差。最后,通过多条路径的预测得出交易价格的分位数进而求出La-VaR值。 实证研究不仅能得出类似基于高频数据计算La-VaR的结论,而且对于流动性较差的资产,VaR度量的风险存在显著的低估现象,而基于超高频数据的La-VaR能较准确地反映流动性风险,不会产生低估。对比基于高频与超高频数据的计算结果后发现：后者的VaR、La-VaR失败时间节点数与理论值更接近、波动范围更小,说明超高频数据包含更精确的市场信息,风险度量具有更高的准确性与鲁棒性。经过检验也表明,后者在无条件覆盖性、条件覆盖性上要占优,在独立性检验上两者持平,总体上依然占优。
[Abstract]:Under the background of the globalization of financial market and the prosperity of derivatives trading, a number of financial crisis incidents prompted the birth of VaR, which has become a standard measurement method of market risk and has attracted wide attention in the financial and academic circles. In this paper, liquidity risk is incorporated into the basic framework of VaR, and a Liquidity-Adjusted VaR model (La-VaR) is constructed based on high-frequency data and ultra-high-frequency data respectively. Finally, empirical research and comparative analysis are conducted.
One of the key tasks of this paper is to build a La-VaR model for high-frequency data. La-VaR modeling for high-frequency data is derived from the basic framework of BDSS model, but for its normal distribution assumption, the assumption of the same variance, the assumption that the relative price difference is not related to the intermediate price, the assumption of maximizing the synchronization of market risk and liquidity risk, the assumption of non-dynamic liquidity. Improvements such as defects can not only calculate individual assets, but also calculate La-VaR. of portfolios.
Specifically, the GJR-GARCH-EVT-kernel model is first constructed to characterize the spike-tail, heteroscedasticity, volatility asymmetry and the distribution characteristics of the upper and lower tail extremes of the return series, and the multivariate Copula model is used to capture the correlation structure between different asset sequences. Then the Monte Carlo simulation is used to predict the return series in one step. Similarly, GJR-GARCH-EVT-kernel model is used to fit the marginal distribution of the relative price difference, and then the binary Archimedes Copula model is used to describe the correlation structure between the relative price difference and the yield in each single asset. Then the pseudo-random number of the relative price difference is generated based on the predicted yield sequence as its one-step prediction sequence. Three, through the prediction of multiple paths, we finally get the quantile of transaction price, and further calculate the La-VaR value.
Empirical analysis shows that VaR has a certain degree of underestimation of risk. The proportion of liquidity risk measured by La-VaR is generally 10%-20%, and the unconditional coverage, independence, conditional coverage tests show that La-VaR basically does not exist the phenomenon of underestimation of risk, which is obviously better than VaR model.
The second core work of this paper is to build a La-VaR model based on ultra-high frequency data. The La-VaR model based on high frequency data is actually to model the same time interval sequence, but ultra-high frequency data is essentially different from the traditional time series, that is, non-equal time interval, so it can not be directly applied to the traditional time series. In column model, we need to introduce duration sequence, and transform the unequal interval yield and relative price difference to model again.
Specifically, a WACD model is established to fit the duration sequence and iteratively predict the high frequency interval (such as 1 minute) from the sum of the multi-step duration to the cumulative duration. Secondly, the UHF data sequence is converted by the duration to obtain the unit yield and the unit relative price difference to satisfy the equal time interval, and then it is incorporated into the model. In the GJR-GARCH-EVT-kernel-Copula framework, the steps of parameter estimation and multi-step prediction are the same as those of duration prediction. Thirdly, the multi-step forecast of unit yield and unit relative price difference is converted into multi-step forecast of fractional yield and relative price difference by duration conversion, and the high frequency interval of prediction is obtained by aggregation. Yield and Relative Price Spread. Finally, the quantiles of the transaction price are predicted through multiple paths and the La-VaR value is calculated.
Empirical research can not only draw a conclusion similar to the calculation of La-VaR based on high-frequency data, but also significantly underestimate the risk of VaR measurement for assets with poor liquidity. La-VaR based on ultra-high-frequency data can accurately reflect the liquidity risk without underestimation. It is found that the number of VaR and La-VaR failure time nodes of the latter is closer to the theoretical value and the fluctuation range is smaller, which indicates that the UHF data contains more accurate market information and the risk measurement has higher accuracy and robustness. Ping, overall, is still dominant.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：F832.51;F224
本文编号：2222810
[Abstract]:Under the background of the globalization of financial market and the prosperity of derivatives trading, a number of financial crisis incidents prompted the birth of VaR, which has become a standard measurement method of market risk and has attracted wide attention in the financial and academic circles. In this paper, liquidity risk is incorporated into the basic framework of VaR, and a Liquidity-Adjusted VaR model (La-VaR) is constructed based on high-frequency data and ultra-high-frequency data respectively. Finally, empirical research and comparative analysis are conducted.
One of the key tasks of this paper is to build a La-VaR model for high-frequency data. La-VaR modeling for high-frequency data is derived from the basic framework of BDSS model, but for its normal distribution assumption, the assumption of the same variance, the assumption that the relative price difference is not related to the intermediate price, the assumption of maximizing the synchronization of market risk and liquidity risk, the assumption of non-dynamic liquidity. Improvements such as defects can not only calculate individual assets, but also calculate La-VaR. of portfolios.
Specifically, the GJR-GARCH-EVT-kernel model is first constructed to characterize the spike-tail, heteroscedasticity, volatility asymmetry and the distribution characteristics of the upper and lower tail extremes of the return series, and the multivariate Copula model is used to capture the correlation structure between different asset sequences. Then the Monte Carlo simulation is used to predict the return series in one step. Similarly, GJR-GARCH-EVT-kernel model is used to fit the marginal distribution of the relative price difference, and then the binary Archimedes Copula model is used to describe the correlation structure between the relative price difference and the yield in each single asset. Then the pseudo-random number of the relative price difference is generated based on the predicted yield sequence as its one-step prediction sequence. Three, through the prediction of multiple paths, we finally get the quantile of transaction price, and further calculate the La-VaR value.
Empirical analysis shows that VaR has a certain degree of underestimation of risk. The proportion of liquidity risk measured by La-VaR is generally 10%-20%, and the unconditional coverage, independence, conditional coverage tests show that La-VaR basically does not exist the phenomenon of underestimation of risk, which is obviously better than VaR model.
The second core work of this paper is to build a La-VaR model based on ultra-high frequency data. The La-VaR model based on high frequency data is actually to model the same time interval sequence, but ultra-high frequency data is essentially different from the traditional time series, that is, non-equal time interval, so it can not be directly applied to the traditional time series. In column model, we need to introduce duration sequence, and transform the unequal interval yield and relative price difference to model again.
Specifically, a WACD model is established to fit the duration sequence and iteratively predict the high frequency interval (such as 1 minute) from the sum of the multi-step duration to the cumulative duration. Secondly, the UHF data sequence is converted by the duration to obtain the unit yield and the unit relative price difference to satisfy the equal time interval, and then it is incorporated into the model. In the GJR-GARCH-EVT-kernel-Copula framework, the steps of parameter estimation and multi-step prediction are the same as those of duration prediction. Thirdly, the multi-step forecast of unit yield and unit relative price difference is converted into multi-step forecast of fractional yield and relative price difference by duration conversion, and the high frequency interval of prediction is obtained by aggregation. Yield and Relative Price Spread. Finally, the quantiles of the transaction price are predicted through multiple paths and the La-VaR value is calculated.
Empirical research can not only draw a conclusion similar to the calculation of La-VaR based on high-frequency data, but also significantly underestimate the risk of VaR measurement for assets with poor liquidity. La-VaR based on ultra-high-frequency data can accurately reflect the liquidity risk without underestimation. It is found that the number of VaR and La-VaR failure time nodes of the latter is closer to the theoretical value and the fluctuation range is smaller, which indicates that the UHF data contains more accurate market information and the risk measurement has higher accuracy and robustness. Ping, overall, is still dominant.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：F832.51;F224
【参考文献】
相关期刊论文 前6条
1 王春峰;卢涛;房振明;;最小报价单位对我国股票市场流动性影响——基于高频数据的实证研究[J];系统工程;2005年12期
2 林辉;;流动性调整的风险价值模型及其实证研究[J];科技与经济;2010年02期
3 宋逢明,谭慧;VaR模型中流动性风险的度量[J];数量经济技术经济研究;2004年06期
4 胡经生,王荣,丁成;VaR方法及其拓展模型在投资组合风险管理中的应用研究[J];数量经济技术经济研究;2005年05期
5 刘晓星;;流动性调整地风险价值度量:基于金融高频数据的实证分析[J];系统工程理论与实践;2009年07期
6 朱小斌;;股票投资组合流动性风险度量模型:构建与检验[J];中国管理科学;2007年01期
，本文编号：2222810
本文链接：https://www.wllwen.com/guanlilunwen/zhqtouz/2222810.html
最近更新
教材专著
