股票市场收益率的分位数回归分析外文翻译资料

 2023-01-13 09:01

股票市场收益率的分位数回归分析

原文作者 Michelle L. Barnes and Anthony W. Hughes

单位 Federal Reserve Bank of Boston

摘要:传统的回报和资本资产定价模型 (CAPM) 测试条件分布的平均数。相反,这个模型通过利用技术对分位数回归 (Koenker 和Bassett 1978 年)的分析返回和测试条件下CAPM分布规律的其它点有效。而这里我们主要讲分位数回归及其一般模型,以及在Errors-in-Variables 前的分位数回归。

关键词:分位数回归;Errors-in-Variables;横截面分析.

  1. 分位数回归

分位数回归,是由Koenker和巴塞特(1978)开发的,是条件均值的经典最小二乘估计到的模式不同的条件分位数函数集合的延伸。中位数(分位数)回归估计能最小化绝对误差的均权加权总和(其中权重等于0.5)从而估计条件中值(分位数)函数,其他条件位数函数由最小化的非对称加权的绝对误差之和推定,这里所述的权重是利益分位数的函数。因此,分位数回归是加强异常值的存在。这种技术已经广泛在过去十年中在许多领域的应用计量经济学 ;应用包括调查工资结构 (Buchinsky 和Leslie 1997 年),收入流动性 (Eide和Showalter 1999 年 ;Buchinsky 和Hunt 1996 年),和受教育程度 (Eide和Showalter 1998年)。财务应用包括Engle和Manganelli (1999 年) 和Morillo (2000 年) 问题的风险价值及其期权定价,而到目前为止,Koenker 和Bassett的 (1978) 的方法股票市场收益对截面的的应用已不被考虑。

一般分位数回归模型,是Buchinsky(1998年)所描述的模型:

或者,

其中, 是一个未知的k times; 1与 th 的百分位数相关的回归参数向量,xi是一个k times; 1的自变量,yi是因变量,uI 是一个未知的误差项。th 的条件分位数由y 确定x:Quant(yi|xi) = xi.其估计值由xi 给出。随着 连续增大,由y确定x的条件分布也随之绘出。虽然很多的实验上的分位数回归的论文假设的误差是独立的恒等分布(i.i.d.),关于uI 惟一必要的假设是

即,假定 th 的误差项分位数等于0. 因此,分位数回归法包含允许边际效应为企业在条件分布的不同点通过不同的( (0, 1)).值估计的值从而作出改变。以这种方式, 在不同类型的资产分位数回归允许参数类型的不均匀性。

因此,分位数回归估计量可以找到如下最小化问题的解决方案:

i.e.,当 = 0.5时在权重平均的中位数回归和权重不均匀时,通过将绝对误差的加权和减到最小。此最小化可以化为线性规划或GMM问题进行论证。前面的方法计算简单而后者意味着

这样测试可以使用正态分布渐近线理论的临界值来构建。这样的关于的曲线称为分位数图。如果图显示有显著的变化,那么意味着自变量xi的影响随着股票业绩条件的提高而改变。请注意,所有的数据观测用于构建每个分位数回归估计;没有对因变量进行区分,因为这会导致样本选择偏差。对的部分估计是有效的;事实上最常用的也最被广为接受的是Buchinsky(1995)的方法,因为它是基于设计矩阵自举,在小的样本中更有效,且在回归因子和回归误差之间有更稳定的依赖性。这各方法包含下列计算

此处,是基于jth (j =1, · · · , B)的子样本的分位数回归估计量。这个子样本说明(yiBS, xirsquo;BS )是通过从原始样本(yi , xirsquo; )中有放回抽样得到的。这个过程可以用Stata软件实现。为了构造包括限制条件下m的分位数的联合测试,我们可以假定:

其中, = (1,1, · · · , k,1, · · · , 1,m, · · · , k,m )rsquo;。这里,R是一个q times; km 阶矩阵,r 是一个具有限制利益特征的q times; 1向量,其中,q是限制条件下强加的小于0的一个数字。我们可以用Buchinsky(1998年)描述的F—统计,

这里,F (q, n k 1)小于且渐近于H0。这种情况下,是一个关于估计方差的协方差矩阵,其中,可以通过应用设计矩阵引导得到。我们用这个来引导整个造纸计算标准误差和方差 - 协方差矩阵。在所有情况下,我们用的进都是来自大样本的样本容量为100的子样本。

  1. 在Errors-in-Variables 前的分位数回归

衰减问题已经充分证明,在金融文献(见Fama和MacBeth 1973和Jagannathan和Wang 1996)。简单地说,因为BETA变量包含OLS时间序列回归估计,但它有一个EIV问题导致的BETA的平均收益的影响估计要做到零偏差。处理这个问题的标准方法是聚集单个企业进入投资组合。人们普遍认为,投资组合可以更精确地估计比单个公司做到更精确地估计,从而减少,但不能消除,所得到的偏差。

这种方法有几个相关的问题。首先,集合体可以减少固有的个体信息内容的返回数据。此外,建设大型投资组合可能导致一个综合偏倚。仅仅因为个别公司自己投资组合的个人资产,由于将估计的参数更精确地使用一个更大的数据集,最优解之间的衰减问题涉及贸易减少偏差(通过组成集体)和提高效率(通过解散集体)。在标准条件下,使用OLS回归,主要关心的是偏差,所以通常大约只有20 – 25个组合(1995年Kothari Shanken,斯隆和Ferson哈维1999)。

在我们的应用中,尽量少使用代表性(投资组合)的观测值可避免在一个广泛的分位数(见Buchinsky(1998))中辨别分位数的回归估计。这些估计可以计算,特别是在条件分布的末尾,将被测量地很不严密并使之后的测试的意义缺乏说服力。进一步的,利用分位数回归,我们试图描绘的行为坏消息比好消息公司(即表现不佳和偏离剧本表演公司)。如果我们公司聚合到大型投资组合,这将有多样化的真菌性病害的真菌性病害的冲击或省略变量为特定公司的真菌性病害,将稀释分位数回归结果的影响。因为我们更感兴趣的相对价值参数在分位数的绝对数量的偏见是次要问题;在这一节中,我们研究当变化时分位数回归估计量的偏离。

分位数回归估计量的性质在errors-in-variables面前没有被认为是在统计或计量经济学文献。如果分位数的方法可以证明有助于处理EIV造成的偏见,这些方法可能获得更广泛的认可与那些分析聚合数据自组合形成减少,但并不能消除,EIV问题。在这里,一个采用蒙特卡洛模拟研究小样本的属性。渐近性质不明,超出了本文的范围,但是像OLS,分位数回归估计量不一致,当应用于省略变量的存在。

实验设计如下。设计矩阵是取自一个横截面考虑经验的例子。然后,生成一系列的平均回报率根据方程

RET U RNi = 0 1BET Ai Zi ui, ui sim; IN (0, u 2)

在这种情况下,确定一个现实的价值((difcult因此允许不同从0(没有errors-in-variables),通过0.2到0.4。样本量是100年和1093年的值(本文示例用于各个横截面)。我们使用5000年蒙特卡洛复制。这里的主要目的是评估分位数回归估计量的偏差(比较OLS估计量。经典线性回归模型的假设下(即错误和解释变量的独立性),真正的分位数边坡参数是相同的所有分位数和分位数回归估计量是一致的(Buchinsky(1998))。我们必须使用这种方法来生成数据与已知的分位数属性,因为与生成数据相关的困难,山坡上危险点的分布,这个性质是Buchinsky (1995)论述的。此中包括2 = 0允许测定小样本偏差渐近的偏差为零,因此代表了一个对照实验的结果的情况下EIV可以相比。蒙特卡洛的输出包含在表1。

结果表明考虑每个实验中,1分位数回归估计较大,平均来看,低值的比 的高值多。其中, 是零和古典假设举行,那里是一个小样本偏差明显分位数回归估计的这就是积极 lt; 0.5 和负面在哪里 gt; 0.5。当 = 0.5时为中位数回归的特殊情况,那么理想情况下分位数回归估计是无偏 (最小的Monte Carlo误差)。 通常情况下在所有分位数中n = 1,093与n = 100相比偏差更小,这证明了渐近收敛性。这种小样本偏差尽管超出了本文的范围,但值得更多的关注。

当 EIV 推导时,功能有关偏见的形状是大致相当于观察到的 = 0,即平均系数是高 = 0.1,比 = 0.9。然而,在 EIV 中, 位数回归估计是OLS比实际上少偏见,并因此可能是首选的估计,在处理 EIV 问题时。当 = 0.5,改进的中位数回归估计手段偏倚,它遭受不少平方的误差损失比苏丹生命线行动,尽管效率较低。这一结果提供了强有力的证据至少对于现时的问题中, 位数回归是一种更好的估计比OLS。

当 EIV 问题呈现更极端的考虑= 0.4分位数回归估计量的偏差小于OLS甚至当 gt; 0.5。对于 n = 100,1= 0.1 时,OLS 估计量的偏差超过了所有正在审议该分位数估计。相同的试验运行情况下在1 = -0.05 和-0.1,结果被发现是相同的定性但方向的衰减偏见仍然趋向于零。分位数情节 (关于的图像) 通常会比它是在现实中,即模式偏置的方法效果似乎不那么极端样本中,比他们在人口中。

还有证据表明分位数回归估计的偏差关于的图像随着n增加到 1,093将lsquo;变平rsquo;。 这一发现了一个小实验涉及混合横截面,表明连接池是青睐相对于基于单个横截面相结合的方法。因为我们在本文使用池,分位数变化完全是因为跨可以排除偏见之别。

外文文献出处:Michelle L. Barnes and Anthony W. Hughes. A Quantile Regression Analysis of the Cross Section of Stock Market Returns,2002.

1. Quantile Regression

Quantile regression, developed by Koenker and Bassett (1978), is an extension of the classical least squares estimation of the conditional mean to a collection of models for diferent conditional quantile functions. As the median (quantile) regression estimator minimizes the symmetrically weighted sum of absolute errors (where the weight is equal to 0.5) to estimate the conditional median (quantile) function, other conditional quantile functions are estimated by minimizing an asymmetrically weighted sum of absolute errors, where the weights are functions of the quantile of interest.Thus, quantile regression is robust to the presence of outliers. This technique has been used widely in the past decade in many areas of applied econometrics; applications include investigations of wage structure (Buchinsky and Leslie 1997), earnings mobility (Eide and Showalter 1999; Buchinsky and Hunt 1996), and educational attainment (Eide and Showalter 1998). Financial applications include Engle and Manganelli (1999) and Morillo (2000) to the problems of Value at Risk and option pricing respectively, but to date the application of Koenker and Bassetts (1978) method to the cross section of stock market returns has not been considered.

The general quantile regression model, as described by Buchinsky (1998), is

or, alternatively,

where is an unknown k times; 1 vector of regression parameters associated with the th percentile,

剩余内容已隐藏,支付完成后下载完整资料


A Quantile Regression Analysis of the Cross Section

of Stock Market Returns

Michelle L. Barnes and Anthony W. Hughes

Federal

Reserve Bank of Boston,

T-8, Research

600 Atlantic Avenue,

Boston, MA 02106

USA

bSchool of Economics,

University of Adelaide

Adelaide, SA 5005

Australia

Current Draft: November 2002.

Key Words: Capital Asset Pricing Model (CAPM); semi-parametric regression; errors-in-variables; Monte Carlo simulation; cross-section analysis; underper- forming stocks and overperforming stocks.

JEL Classifications: G12; C14; C21.

Abstract

Traditional methods of modelling returns and testing the Capital Asset Pricing Model (CAPM) do so at the mean of the conditional distribution. Instead, we model returns and test whether the conditional CAPM holds at other points of the distribution by utilizing the technique of quantile regression (Koenker and Bassett 1978). This method allows us to model the performance of firms or portfolios that underperform or overperform in the sense that the conditional mean under- or overpredicts the firms return. In the context of a conditional CAPM, the market price of beta risk is significant in both tails of the conditional distribution of returns - negative for firms that underperform and positive for firms that overperform - but is insignificant around the median, and the opposite pattern obtains for large firms. Underperforming firms exhibit a positive relationship between size and returns in support of Mertons (1987) prediction, and there is some evidence of a positive relationship between returns and financial paper for overperforming firms. Quantile regression alleviates some of the statistical problems which plague CAPM studies: errors-in- variables; omitted variables bias; sensitivity to outliers; and non-normal error distributions.

1. Quantile Regression

Quantile regression, developed by Koenker and Bassett (1978), is an extension of the classical least squares estimation of the conditional mean to a collection of models for diferent conditional quantile functions. As the median (quantile) regression estimator minimizes the symmetrically weighted sum of absolute errors (where the weight is equal to 0.5) to estimate the conditional median (quantile) function, other conditional quantile functions are estimated by minimizing an asymmetrically weighted sum of absolute errors, where the weights are functions of the quantile of interest.Thus, quantile regression is robust to the presence of outliers. This technique has been used widely in the past decade in many areas of applied econometrics; applications include investigations of wage structure (Buchinsky and Leslie 1997), earnings mobility (Eide and Showalter 1999; Buchinsky and Hunt 1996), and educational attainment (Eide and Showalter 1998). Financial applications include Engle and Manganelli (1999) and Morillo (2000) to the problems of Value at Risk and option pricing respectively, but to date the application of Koenker and Bassetts (1978) method to the cross section of stock market returns has not been considered.

The general quantile regression model, as described by Buchinsky (1998), is

or, alternatively,

where is an unknown k times; 1 vector of regression parameters associated with the th percentile, xi is a k times; 1 vector of independent variables, yi is the dependent variable of interest and ui is an unknown error term. The th conditional quantile of y given x is Quant(yi|xi) = xi. Its estimate is given by xi. As increases continuously, the conditional distribution of y given x is traced out. Although many of the empirical quantile regression papers assume that the errors are independently and identically distributed (i.i.d.), the only necessary assumption concerning ui is

that is, the conditional th quantile of the error term is equal to zero. Thus, the quantile regression method involves allowing the marginal efects to change for firms at different points in the conditional distribution by estimating using several diferent values of , (0, 1). It is in this way that quantile regression allows for parameter heterogeneity across different types of assets.

Thus, the quantile regression estimator can be found as the solution to the following minimization problem:

i.e., by minimizing a weighted sum of the absolute errors, where the weights are symmetric for the median regression case ( = 0.5) and asymmetric otherwise. This minimization can be formulated either as a linear programming or as a GMM problem. The former implies that the method is computationally straightforward while the latter implies that

so tests can be constructed using critical values from the normal distribution with asymptotic justification. A plot of against is called the quantile plot. If this indicates significant variation, it implies that the efect of the xi variable changes as the conditional performance of the stock improves. Note that all data observations are used to construct each quantile regression estimate; there is no partitioning of data performed on the dependent variable as this would incur sample selection bias. Several estimators are available for ; the most commonly used in practice, and the one most favored by Buchinsky (1995) because it is more effcient in small samples and is robust to dependence between the regressors and the regression errors, is based on the design matrix bootstrap. This method involves computing

where j is the quantile regression estimator based on the jth bootstrap sample, j =1, · · · , B. The bootstrap samples, say (yiBS, xiBS ), are obtained by sampling with replacement from the original sample, (yi, xi). This procedure can be implemented using the software program Stata. In order to construct joint tests, inclu

剩余内容已隐藏,支付完成后下载完整资料


资料编号:[286801],资料为PDF文档或Word文档,PDF文档可免费转换为Word

您需要先支付 30元 才能查看全部内容!立即支付

课题毕业论文、开题报告、任务书、外文翻译、程序设计、图纸设计等资料可联系客服协助查找。