基于贝叶斯分位回归理论的截面相依面板协整研究

发布时间：2018-08-21 11:35

【摘要】：非平稳面板数据研究是目前计量经济学领域中的前沿问题,其中,面板单位根和协整研究,作为时间序列单位根与传统协整理论在面板数据中的发展和延伸,更具有重要意义。由于全球国际趋势和国际经济周期等共同驱动的影响,宏观经济、管理或金融面板数据尤其是国家(地区或个体单元)的面板数据之间通常存在截面相依特征,因此,考虑截面相依假设条件的面板协整更加符合实际应用背景,也是面板数据研究中亟待解决的一个热点问题。与传统的面板协整不同,本文针对具有截面相依条件的面板协整进行研究,在贝叶斯理论框架中,假设各个截面个体具有截面相依特征,结合贝叶斯分位回归估计方法,提出了面板数据的贝叶斯分位协整模型。贝叶斯分位协整模型可以充分发挥贝叶斯方法考虑了参数不确定性风险的优势,并且体现了分位回归方法不仅可以刻画响应变量的中心趋势,还可以刻画变量尾部行为的优点,从而为更全面地刻画响应变量与协变量的长期均衡关系提供了方法和工具支撑,在理论上扩展面板协整的研究方法和研究视角,在实践上为经济管理问题的定量分析和决策提供技术支持和有力依据。针对面板数据之间通常存在截面相依性,首先应用动态公共因子结构刻画面板数据的截面相依特征,结合贝叶斯决策理论,提出一类考虑了截面相依假设条件的协整模型,利用贝叶斯分位回归方法,通过把非对称Laplace分布表示成指数分布和正态分布的线性组合,获得了条件分位函数后验估计量的解析表达形式,并设计Kalman滤波与Gibbs抽样算法对模型参数进行估计和协整检验。同时,Monte Carlo仿真实验结果表明,贝叶斯分位协整可以更加全面地对变量间的协整关系进行判断。经济金融变量因为战争,政府政策以及自然灾害等因素的影响,往往表现出结构突变性,这种结构性变化的发生会影响传统线性协整检验的判断。放松线性假设条件,本文提出一类考虑了结构变化特征的面板协整模型—平滑变结构面板协整模型,利用傅立叶级数展开形式来刻画变结构特征,并采用去除截面均值的方法消除面板数据的截面相依性,以避免参数过多的问题,进而结合贝叶斯分位回归方法得到相应条件分位函数后验估计量的解析形式,并设计MCMC抽样算法对模型进行参数后验估计和协整检验。仿真实验结果表明,贝叶斯分位变结构协整能够有效全面地刻画各个分位水平下的变结构长期关系。与变结构协整不同,门限协整主要研究协整回归模型是线性,而其相应的误差修正项是非对称时的情形。针对传统门限协整模型由于似然函数具有多峰、不连续特征,导致冗余参数识别存在困难,最优化计算相对复杂的问题,本文从贝叶斯的角度出发,提出面板数据的贝叶斯分位门限协整模型,通过去除截面均值以消除面板数据间潜在的相依性,并对参数的先验分布进行灵敏度分析以选择合适的参数先验,结合贝叶斯分位回归方法对面板门限协整模型进行参数估计,得到条件分位函数后验估计量的解析表达式,同时,利用MCMC算法对协整模型的参数进行估计,计算出协整检验的后验概率以进行更加全面的门限协整检验。将上述考虑了面板数据截面相依特征的贝叶斯分位协整方法应用到原油与股票市场的关系研究中,并与传统面板协整方法进行比较,发现贝叶斯分位协整方法对原油与股票市场之间联动性关系的刻画更加全面,验证了贝叶斯分位协整方法的可行性和有效性,说明贝叶斯分位方法能够提供全方面的便捷的模型参数估计和协整检验信息。
[Abstract]:Nonstationary panel data is a frontier issue in econometrics. Among them, panel unit root and cointegration, as the development and extension of time series unit root and traditional cointegration theory in panel data, are of great significance. The cross-sectional dependence between economic, management or financial panel data, especially the panel data of a country (region or individual unit), is a common feature. Therefore, panel co-integration considering the assumption of cross-sectional dependence is more suitable for practical application and is also a hot issue in panel data research. In this paper, we study the panel co-integration with cross-section dependence. In the Bayesian framework, we assume that each section has cross-section dependence characteristics. Combined with Bayesian quantile regression estimation method, we propose a Bayesian quantile co-integration model for panel data. Bayesian quantile co-integration model can give full play to Bayesian method. The advantages of parametric uncertainties and the advantages of quantile regression not only can depict the central trend of the response variables, but also can depict the tail behavior of the variables are illustrated. The method and tool support are provided for describing the long-term equilibrium relationship between the response variables and the covariates more comprehensively, and the research on Panel Cointegration is expanded theoretically. Research methods and research perspectives, in practice, provide technical support and strong basis for quantitative analysis and decision-making of economic management issues.
Aiming at the cross-section dependence between panel data, a class of co-integration model considering the assumption of cross-section dependence is proposed by using the cross-section dependence characteristics of dynamic common factor structural panel data and Bayesian decision theory. The asymmetric Laplace distribution is expressed as exponential by Bayesian quantile regression method. The linear combination of distribution and normal distribution obtains the analytical expression of conditional quantile function posterior estimator, and designs Kalman filter and Gibbs sampling algorithm to estimate and test the model parameters. At the same time, Monte Carlo simulation results show that Bayesian quantile co-integration can be more comprehensive to the co-integration relationship between variables. Make a judgement.
Economic and financial variables often exhibit structural catastrophe because of war, government policies and natural disasters. The occurrence of such structural changes will affect the judgment of traditional linear cointegration test. In the co-integration model, the Fourier series expansion is used to characterize the variable structure features, and the cross-section dependence of panel data is eliminated by removing the cross-section mean, so as to avoid the problem of too many parameters. The simulation results show that Bayesian fractional variable structure co-integration can effectively and comprehensively describe the long-term relationship of the variable structure at each fractional level.
Unlike variable structure co-integration, threshold co-integration mainly studies the case when the co-integration regression model is linear and the error correction term is asymmetric. In this paper, a Bayesian thresholding co-integration model for panel data is proposed. The potential dependence between panel data is eliminated by removing the cross-sectional mean, and the prior distribution of parameters is analyzed to select the appropriate prior parameters. The parameters of the model are estimated by Bayesian quantile regression method. At the same time, the MCMC algorithm is used to estimate the parameters of the co-integration model, and the posterior probability of the co-integration test is calculated to conduct a more comprehensive threshold co-integration test.
The Bayesian fractional cointegration method considering the cross-sectional dependence of panel data is applied to the study of the relationship between crude oil and stock market. Compared with the traditional panel cointegration method, it is found that the Bayesian fractional cointegration method is more comprehensive in describing the linkage relationship between crude oil and stock market, which verifies the Bayesian fractional cointegration. The feasibility and validity of the whole method show that Bayesian grading method can provide all-round and convenient information of model parameter estimation and co-integration test.
【学位授予单位】：湖南大学
【学位级别】：博士
【学位授予年份】：2012
【分类号】：F831.51;F416.22;F224

【参考文献】