数学杂志  2016, Vol. 36 Issue (1): 183-190   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
李永明
张文婷
蔡际盼
正相协下风险度量VaR样本分位数估计的渐近性质
李永明1, 张文婷2, 蔡际盼2     
1. 上饶师范学院数学与计算机科学学院, 江西 上饶 334001;
2. 广西师范学院数学科学学院, 广西 南宁 530023
摘要:本文研究了正相协严平稳样本下, 风险度量VaR样本分位数估计的问题.利用其指数不等式和协方差不等式, 获得了风险度量VaR的样本分位数估计的相合性和渐近正态性, 并给出Bahadur表示.
关键词正相协样本    VaR风险度量    样本分位数    Bahadur表示    
THE ASYMPOTIC PROPERTIES OF THE SAMPLE QUANTILE ESTIMATOR OF VAR UNDER POSITIVE ASSOCIATED SAMPLES
LI Yong-ming1, ZHANG Wen-ting2, CAI Ji-pan2     
1. School of Math. & Computer Science, Shangrao Normal University, Shangrao, 334001, China;
2. College of Math. Science, Guangxi Teachers Education University, Nanning 530023, China
Abstract: In this paper, we consider the sample quantile estimator of VaR based on a stationary and positively associated sequence. For this setting, applying the exponential inequality of positively associated random variables, we prove the consistency and asympotic normality of the sample quantile estimator of VaR, and also give its Bahadur representation.
Key words: positive association     VaR     quantile estimates     Bahadur representation    
1 引言

风险度量VaR (Value at risk) 是指在正常的市场环境下, 金融资产或证劵组合在一定的持有期内和一定的置信水平下, 预期的最大损失称为风险价值或在险价值, 其定义一般表述为设$\{Y_t\}_{t=0}^n$$n$个时间段资产价格序列, $X_t=-\log({Y_t}/{Y_{t-1}})$是第$t$个时间段对数收益.假设$\{X_t\}_{t=1}^n$是相依严平稳序列, 其边缘分布函数为$F(x)=P(X_t\leq x),\ x\in R.$对任意给定$p\in (0,1),$ $F(x)$$p$-分位数定义为$\xi_p=F^{-1}(p)=\inf\{u: F(u)\geq p\},$其中函数$F^{-1}(t)$, $0﹤t﹤1$$F(\cdot)$的广义逆函数.易知$\xi_p$满足$F(\xi_p-)\leq p\leq F(\xi_p)$.置信水平为$1-p$的风险度量VaR值定义为$ v_p=-\xi_p=-\inf\{u: F(u)\geq p\}.$

在许多金融模型中, 风险度量VaR方法能够较准确地度量由不同风险来源及其相互作用而产生的潜在损失, 而且与其它风险度量方法相比, VaR方法的计算也相对简便.自1990年以来, 国内外学者对风险度量VaR方法进行了一定程度的研究.如: Dowd [1]给出的风险度量VaR样本分位数估计为$Q_{n,p}=-X_{([np]+1)},$其中$X_{(r)}$为样本$X_1, \cdots, X_n,$的第$r$个次序统计量. Koji et al [2]提出了用样本分位数$-X_{([np])}$估计VaR.谢佳利和杨善朝等[3]采用加权样本分位数估计量估计VaR, 该方法的估计精度相对文献[1, 2]有较好的改进.

$F_n(x)=\frac{1}{n}\sum\limits_{i=1}^{n}I(X_i\leq x)$为总体$X$的经验分布函数, 其样本分位数记为$Z_{n,p}$.显然$Z_{n,p}=-Q_{n,p}$.由于风险度量VaR与分位数只相差一个负号, 所以分位数的各种非参数估计自然也是VaR的非参数估计.而对于样本分位数估计的研究, 自Bahadur[4]给出了独立样本下分位数估计的Bahadur表示之后, 许多学者对混合和负相协 (NA) 以及负象限相依 (NOD) 样本下研究了分位数估计的Bahadur表示及渐近性质对此进行了研究, 也取得了不少成果, 这里就不一一列出.

定义1.1 称随机变量$\{X_1, \cdots, X_n, n\geq 2\}$是正相协 (PA) 的, 设$A_1$$A_2$$\{1,\cdots ,n\}$任何两个不相交的非空子集, 如果$ {\rm Cov}(f_1(X_i,i\in A_1),f_2(X_j,j\in A_2))\geq 0$成立, 其中$f_1$$f_2$是任何两个使得协方差存在且对每个变元均非降 (或对每个变元均非升) 的函数.称随机变量序列$\{X_i,i\geq1\}$是PA序列, 如果对任何$n\geq 2, X_1,\cdots ,X_n$是PA的.

PA序列的一个基本性质: PA序列经过单调函数变换得到的序列仍为PA序列.

对于NA序列的Rosenthal型矩不等式和Bernstein型概率不等式的研究已经有很好的成果, 这为研究负相协下统计大样本性质提供了很好的研究工具.然而, PA序列受到其协方差系数影响, 比较理想的Rosenthal型矩不等式和Bernstein型概率不等式目前比较少见, 这给研究PA下估计量的统计大样本性质带来一定程度的不便.

由于PA序列包括正相关的正态随机变量族,而且PA序列在与实际应用有关的模型中(如可靠性理论, 渗透理论及多元统计分析等) 有广泛的应用, 所以受到了学者们的广泛关注, 具体文献在此就不一一列举.但是PA样本下风险度量VaR分位数估计的大样本性质及其Bahadur表示的研究尚未见文献涉及.因此本文将采用文献[5]中PA序列指数不等式以及PA序列协方差性质, 讨论PA序列下风险度量VaR分位数估计的强相合性和渐近正态性及其Bahadur表示.所得结果推广了相关文献的结果.

2 辅助引理

下面我们给出PA序列指数不等式以及一些辅助结论.文中$C, C_1, C_2, \cdots$表示不依赖于$n$的正常数, 且在不同的地方取值可能相同也可能不同.

引理2.1[5] (i) 设$X_1,X_2,\cdots $是零均值正相协序列, $\max\limits_{1\leq i\leq n}|X_i|\leq c_n <\infty {\rm a.s.}\ (n=1,2,\cdots)$;

(ii) 设$u(n)=\sup\limits_{i\geq 1}\sum\limits_{j:j-i\geq n}{\rm cov}(X_i,X_j)$, 满足$\sum\limits_{i=1}^\infty u^{1/2}(2^i)<\infty$;

(iii) 设$\{p_n:n\geq 1\}$是一串正整数序列, 满足$p_n\leq \frac{n}{2}$成立.又设$\theta >0$, 满足$0<\theta p_nc_n\leq 1$.则存在一个不依赖于$n$的正常数$C_1$, 使得对任意的$\varepsilon >0$, 有

$P(|\sum\limits_{i = 1}^n {{X_i}} | > n\varepsilon ) \le 4\{ {\theta ^2}nu({p_n}){e^{\theta n{c_n}}} + {e^{{C_1}{\theta ^2}nc_n^2}}\} {e^{ - n\theta \varepsilon /2}}.$

引理2.2[6] 设$F(x)$是右连续分布函数, 则广义逆函数$F^{-1}(t)$, 在$0<t<1$非降且左连续, 并且满足

(i) $F^{-1}(F(x))\leq x,-\infty <x<+\infty;$

(ii) $F(F^{-1}(t))\geq t,0<t<1;$

(iii) $F(x)\geq t\Longleftrightarrow x\geq F^{-1}(t).$

引理2.3[7] 令$p\in (0,1),\ Z_{p,n}=F_n^{-1}(p)=\inf \{x:F_n(x)\geq p\}$.假设$P(X_i=X_j)=0,i\neq j$, 那么$p<F_n(Z_{p,n})<p+1/n\ \ {\rm a.s.}$.

引理2.4 (i) 如果$\{X_n,n\geq 1\}$具有相同分布函数$F(x)$和有界密度函数$f(x)$的严平稳PA序列, $F(x)$$\xi_p$的邻域$N_p$内连续可导 ($p\in(0,1)$), 且$0<d=\sup\{f(x):x\in N_p\}<\infty$;

(ii) 设$v(n)=\sup\limits_{i\geq 1}\sum\limits_{j:j-i\geq n}{\rm cov}^{1/3}(X_i,X_j)$满足$v(n)=O(e^{-2\sqrt{n}})$;

(iii) 对任意满足$d_n\rightarrow 0, n^\frac{1}{4}d_n/\log n \rightarrow \infty,n\rightarrow \infty$的正实数序列$\{d_n\}_{n\geq 1}$, 记$D_n=[\xi_p-d_n, \xi_p+d_n]$.则

$\sup \limits _{x\in D_n}|(F_n(x)-F(x))-(F_n(\xi_p)-p)|\leq (1+d)n^{-\frac{1}{4}}d_n \ \ {\rm a.s.}. $

注2.1 如果$v(n)=O(e^{-2\sqrt{n}})$, 则必有

$\sum\limits_{n = 1}^\infty {{v^{1/2}}} ({2^n})<\infty .$

事实上, 由于$\sum\limits_{n=1}^\infty v^{1/2}(2^n)\leq C\sum\limits_{i=1}^\infty e^{-2^{n/2}}$, 易知当$n>N$$2^{n/2}>n$.故$\sum\limits_{n=N}^\infty e^{-2^{n/2}} <\infty, $从而$\sum\limits_{n=1}^\infty v^{1/2}(2^n)<\infty$.

引理2.4的证明 设$t_n=d_nn^{-\frac{1}{4}}, S_{r,n}=\xi_p+rt_n,$

$\Delta_{r,n}=F_n(S_{r,n})-F(S_{r,n})-F_n(\xi_p)+p, \ r=0,\pm1,\pm2,\cdots ,\pm m_n,\ m_n=[n^\frac{1}{4}]+1,\\ D_n\subset [\xi_p-m_nt_n, \xi_p+m_nt_n]=\bigcup\limits_{r=-m_n}^{m_n-1}[\xi_p+rt_n, \xi_p+(r+1)t_n]. $

由于

$\sup\limits_{x\in D_n}|F_n(x)-F(x)-F_n(\xi_p)+p| \\ \leq \sup\limits_{\xi_p-m_nt_n\leq x\leq \xi_p+m_nt_n}|F_n(x)-F(x)-F_n(\xi_p)+p| \\=\max\limits_{-m_n\leq r\leq m_n-1}\sup\limits_{\xi_p+rt_n\leq x\leq \xi_p+(r+1)t_n}|F_n(x)-F(x)-F_n(\xi_p)+p|.$

$x\in [\xi_p+rt_n,\xi_p+(r+1)t_n]$时, 由$F_n(x)$是非降函数以及微分中值定理得

$F_n(x)-F(x)-F_n(\xi_p)+p \leq F_n(S_{r+1,n})-F(S_{r,n})-F_n(\xi_p)+p \\ = \Delta_{r+1,n}+F(S_{r+1,n})-F(S_{r,n})\leq \Delta_{r+1,n}+dt_n$ (2.1)

$F_n(x)-F(x)-F_n(\xi_p)+p \geq F_n(S_{r,n})-F(S_{r+1,n})-F_n(\xi_p)+p \\ = \Delta_{r,n}+F(S_{r,n})-F(S_{r+1,n}) \geq \Delta_{r,n}-dt_n.$ (2.2)

根据 (2.1) 和 (2.2) 式得

$\sup\limits_{x\in D_n}|F_n(x)-F(x)-F_n(\xi_p)+p|\leq \max\limits_{-m_n\leq r\leq m_n}|\Delta_{r,n}|+dt_n.$ (2.3)

$P(|\Delta_{r,n}|>t_n)=P(|F_n(S_{r,n})-F(S_{r,n})-F_n(\xi_p)+p|>t_n) \nonumber\\ \leq P(|F_n(S_{r,n})-F(S_{r,n})|>\frac{t_n}{2})+P(|F_n(\xi_p)-p|>\frac{t_n}{2}) \nonumber\\=:I_1+I_2.$ (2.4)

$\eta_{ni}=I(X_i\leq \xi_p+rt_n)-EI( X_i\leq \xi_p+rt_n),(i=1,2,\cdots ,n).$ (2.5)

由PA序列性质知$\eta_{n1},\cdots,\eta_{nn}$仍是平稳PA随机变量, 且$E\eta_{ni}=0,\ |\eta_{ni}|\leq 2\stackrel {\Delta}{=}c_n,\ i=1,\cdots,n$, 故引理2.1中的条件 (i) 满足.又令$w(n)=\sup\limits_{i\geq 1}\sum\limits_{j:j-i\geq n} {\rm Cov}(\eta_{ni},\eta_{nj})$, 根据引理2.4条件 (i) 及Roussas [8]中的引理2.6 (也可见文献[9]的 (4) 式) 知, 存在某正常数$M$

${\rm Cov}(\eta_{ni},\eta_{nj})\leq M {\rm Cov}^{1/3}(X_i,X_j). $ (2.6)

再根据引理2.4的条件 (ii) 以及注2.1得$w(n)\leq v(n)<\infty$, 且$\sum\limits _{i=1}^\infty w^{1/2}(2^i)<\infty,$从而引理2.1条件 (ii) 满足.又令$\theta =\frac{1}{\sqrt{n}}>0,\ p_n\leq \frac{\sqrt{n}}{2}\leq n/2$, 则$0<\theta p_nc_n\leq 1 $, 故引理2.1条件 (iii) 满足.由此, 根据引理2.1可得

$\begin{array}{l} {I_1} = P(|\sum\limits_{i = 1}^n {{\eta _{ni}}} | > {t_n})\\ \le 8\{ {\theta ^2}nu({p_n}){e^{2n\theta }} + {e^{4{C_1}n{\theta ^2}}}\} {e^{ - n{t_n}\theta /2}}\\ = 8\{ u({p_n}){e^{2\sqrt n }} + {e^{4{C_2}}}\} {e^{ - \sqrt n {t_n}/2}}\\ \le {C_3}\exp \{ - {n^{\frac{1}{4}}}{d_n}/2\} = {C_4}\exp \{ - 2\log n\} \le \frac{{{C_5}}}{{{n^2}}}(当n足够大时) \end{array}$ (2.7)

类似于$I_1$的计算方法, 得

$I_2\leq \frac{C_6}{n^2}.$ (2.8)

故由 (2.4), (2.7) 和 (2.8) 式得

$P(\max\limits_{-m_n\leq r\leq m_n}|\Delta_{r,n}|>t_n)\leq\sum\limits_{r=-m_n}^{m_n}P(|\Delta_{r,n}|>t_n) \leq \frac{C_7}{n^{7/4}}.$

从而$\sum\limits_{n=1}^\infty P(\max\limits_{-m_n\leq r\leq m_n}|\Delta_{r,n}|>t_n)<\infty.$故由Borel-Cantelli引理及 (2.3) 式得

$\sup\limits_{x\in D_n}|(F_n(x)-F(x))-(F_n(\xi_p)-p)|\leq (1+d)n^{-\frac{1}{4}}d_n {\rm a.s..}$

引理2.4得证.

3 主要结论

下面给出本文的主要结果.

定理3.1 (1) 设$\{X_n,n\geq 1\}$是具有相同分布函数$F(x)$和有界密度函数$f(x)$的严平稳PA序列, $F(x)$$\xi_p$处可导且$F'(\xi_p)=f(\xi_p)>0$;

(2) 设$v(n)=\sup\limits_{i\geq 1}\sum\limits_{j:j-i\geq n}{\rm Cov}^{1/3}(X_i,X_j)$满足$v(n)=O(e^{-2\sqrt{n}})$;

(3) 设$\{d_n\}_{n\geq 1}$为满足$d_n\rightarrow 0,n^\frac{1}{4}d_n/\log n\rightarrow \infty,\ n\rightarrow \infty$的任意正实数序列.则

$v_p-Q_{n,p}=o(n^{-\frac {1}{4}}d_n)\ {\rm a.s.}. $

定理3.2 设$\{X_n,n\geq 1\}$是具有相同分布函数$F(x)$和有界密度函数$f(x)$的严平稳PA序列, $F(x)$$\xi_p$的邻域$N_p$内连续可导 ($p\in(0,1)$), 且$0<d=\sup\{f(x):x\in N_p\}<\infty$.如果定理3.1的条件 (2)-(3) 满足, 则

$v_p-Q_{n,p}=\frac{p-F_n(\xi_p)}{f(\xi_p)}+O(d_nn^{-\frac{1}{4}})\ \ {\rm a.s.}. $

定理3.3 在定理3.2的条件下, 如果$\forall x,y\in R$, $j \geq 1$, 有

$\sum\limits_{j=n}^{\infty}\sup\limits_{(x,y)\in R}|F_{1,j+1}(x,y)-F(x)F(y)|<\infty,$

其中$F_{1,j+1}(x,y)$$(X_1, X_j)$的联合分布函数.令$\sigma_0^2=F(x)[1-F(x)]+\sum\limits_{j=1}^{\infty}E(Z_{n1}Z_{nj+1})$, 记$\stackrel {d}{\to} $为依分布收敛.则

$\sqrt{n}(v_p-Q_{n,p})\stackrel {d}{\to} N(0, \sigma_0^2f^{-2}(\xi_p)).$
4 定理证明

定理3.1的证明 对于任意的$\varepsilon >0$

$P(|v_p-Q_{n,p}|>\varepsilon n^{-\frac{1}{4}}d_n) \\ =P(v_p-Q_{n,p}>\varepsilon n^{-\frac{1}{4}}d_n)+P(v_p-Q_{n,p}﹤-\varepsilon n^{-\frac{1}{4}}d_n)\stackrel {\Delta}{=}I_1+I_2.$ (4.1)

由引理2.2可得

$I_1= P(Z_{p,n}>\xi_p+\varepsilon n^{-\frac{1}{4}}d_n) = P\{p>F_n(\xi_p +\varepsilon n^{-\frac{1}{4}}d_n)\} \\ = P\{1-F_n(\xi_p+\varepsilon n^{-\frac{1}{4}}d_n)-(1-F(\xi_p+\varepsilon n^{-\frac{1}{4}}d_n))>F(\xi_p+\varepsilon n^{-\frac{1}{4}}d_n)-p\} \\ = P\{\frac{1}{n}\sum\limits_{i=1}^n(w_i-Ew_i)>\delta_1\},$ (4.2)

其中$w_i=I(X_i>\xi_p+\varepsilon n^{-\frac{1}{4}}d_n),\ \delta_1=F(\xi_p+\varepsilon n^{-\frac{1}{4}}d_n)-p$.同理可得

$I_2= P(Z_{p,n}<\xi_p-\varepsilon n^{-\frac{1}{4}}d_n)=P\{p﹤F_n(\xi_p-\varepsilon n^{-\frac{1}{4}}d_n)\} \nonumber\\ = P\{F_n(\xi_p-\varepsilon n^{-\frac{1}{4}}d_n)-F(\xi_p-\varepsilon n^{-\frac{1}{4}}d_n)>p-F(\xi_p-\varepsilon n^{-\frac{1}{4}}d_n)\} \nonumber\\ = P\{\frac{1}{n}\sum\limits_{i=1}^n(v_i-Ev_i)>\delta_2\},$ (4.3)

其中$v_i=I(X_i\leq \xi_p-\varepsilon n^{-\frac{1}{4}}d_n), \ \delta_2=p-F(\xi_p-\varepsilon n^{-\frac{1}{4}}d_n).$

类似于 (2.5) 式关于$\{\eta_{ni}, i\geq 1\}$的性质讨论, 知$\{w_i-Ew_i\}_{1\leq i\leq n}$$\{v_i-Ev_i\}_{1\leq i\leq n}$都为0均值平稳PA序列, 且满足$|w_i-Ew_i|\leq 2\stackrel {\Delta}{=}c_n$, $|v_i-Ev_i|\leq 2\stackrel {\Delta}{=}c_n.$$\theta =1/\sqrt{n}>0$, $p_n\leq \sqrt{n}/2$, 有$\theta p_nc_n\leq 1.$从而由引理2.1及 (4.1)-(4.3) 式可得

$P(|v_p-Q_{n,p}|>\varepsilon n^{-\frac {1}{4}}d_n) = P\{\frac{1}{n}\sum\limits_{i=1}^n(w_i-Ew_i)>\delta_1\}+P\{\frac{1}{n}\sum\limits_{i=1}^n(v_i-Ev_i)>\delta_2\} \nonumber\\ \leq 8\{\theta^2nu(p_n)e^{\theta n c_n}+e^{C_1\theta^2nc_n^2} \}e^{-n\theta \min(\delta_1,\delta_2)/2} \nonumber\\ = 8\{u(p_n)e^{2\sqrt{n}}+e^{4C_2}\}e^{-\sqrt{n}\min(\delta_1,\delta_2)/2} \leq C_3e^{-\sqrt{n}\min(\delta_1,\delta_2)/2},$ (4.4)

其中$c,c_1$为常数.又因为$F(x)$$\xi_p$点处连续, $F'(\xi_p)>0$, 此处$\xi_p$是不等式$F(x-)\leq p\leq F(x)$的唯一解且$F(\xi_p)=p$, 由Taylor展开得

$F(\xi_p+\varepsilon n^{-\frac{1}{4}}d_n)-p=f(\xi_p)\cdot \varepsilon n^{-\frac{1}{4}}d_n+o(\varepsilon n^{-\frac{1}{4}}d_n),$
$p-F(\xi_p-\varepsilon n^{-\frac{1}{4}}d_n)=f(\xi_p)\cdot \varepsilon n^{-\frac{1}{4}}d_n+o(\varepsilon n^{-\frac{1}{4}}d_n).$

从而$\min(\delta_1,\delta_2)=f(\xi_p)\cdot \varepsilon n^{-\frac{1}{4}}d_n$, ($n\rightarrow \infty$), 代入 (4.4) 式得

$ P(|v_p-Q_{n,p}|>\varepsilon n^{-\frac{1}{4}}d_n) \leq C_1\exp\{-\frac{f(\xi_p)\varepsilon n^\frac{1}{4}d_n}{2\log n}\cdot \log n\}\leq\frac{C_2}{n^2}. $

于是

$\sum\limits_{n=1}^\infty P(|v_p-Q_{n,p}|>\varepsilon n^{-\frac{1}{4}d_n})< \infty.$

故由Borel-cantelli引理得$v_p-Q_{n,p}=o(n^{-\frac{1}{4}}d_n) \ \ {\rm a.s.}. $定理得证.

定理3.2的证明 由定理3.1知

$ v_p-Q_{n,p}=o(n^{-\frac {1}{4}}d_n)\ \ {\rm a.s.},$ (4.5)

由引理2.4得

$ F_n(\xi_p)-p=F_n(Z_{p,n})-F(Z_{p,n})+O(n^{-\frac{1}{4}}d_n)\ \ {\rm a.s.}, $ (4.6)

又由引理2.3得

$ F_n(Z_{p,n})-p=O(n^{-1})\ \ {\rm a.s.}.$ (4.7)

$\theta_n$是介于$Q_{p,n}$$v_p$之间的随机变量, 结合 (4.5)-(4.7) 式, 利用Taylor展开得

$F_n(\xi_p)-p \\ =F(\xi_p)-F(Z_{p,n})+F_n(Z_{p,n})-F(\xi_p)+O(n^{-\frac{1}{4}}d_n) \\ =F(\xi_p)-F(Z_{p,n})+O(n^{-1})+O(n^{-\frac{1}{4}}d_n) \\ = F(\xi_p)-[F(\xi_p)+f(\xi_p)(Z_{p,n}-\xi_p)+\frac{1}{2}f'(\theta_n)(Z_{p,n}-\xi_p)^2]+O(n^{-1})+O(n^{-\frac{1}{4}}d_n) \\ =F(\xi_p)-[F(\xi_p)+f(\xi_p)(v_p-Q_{n,p})+\frac{1}{2}f'(\theta_n)(v_p-Q_{n,p})^2]+O(n^{-1})+O(n^{-\frac{1}{4}}d_n) \\ = -f(\xi_p)(v_p-Q_{n,p})-\frac{1}{2}f'(\theta_n)(v_p-Q_{n,p})^2+O(n^{-1})+O(n^{-\frac{1}{4}}d_n) \\ =-f(\xi_p)(v_p-Q_{n,p})+O(n^{-\frac{1}{4}}d_n),$

由此可得

$v_p-Q_{n,p}=\frac{p-F_n(\xi_p)}{f(\xi_p)}+O(d_nn^{-\frac{1}{4}})\ \ {\rm a.s.}. $

定理3.2证明完毕.

定理3.3的证明 由定理3.2得$v_p-Q_{n,p}=(p-F_n(\xi_p))/f(\xi_p)+O(d_nn^{-\frac{1}{4}})\ \ {\rm a.s.}. $故要证明定理3.3成立, 只需证明$F_n(\xi_p)-p$具有渐近正态性.记$Z_{ni}=I_{(X_i\leq \xi_p)}-EI_{(X_i\leq \xi_p)}$, 则

$\sqrt n ({F_n}({\xi _p}) - p) = {n^{ - 1/2}}\sum\limits_{i = 1}^n {({I_{({X_i} \le {\xi _p})}} - E{I_{({X_i} \le {\xi _p})}})} = {n^{ - 1/2}}\sum\limits_{i = 1}^n {{Z_{ni}}} \mathop = \limits^\Delta {n^{ - 1/2}}{S_n}. $ (4.8)

$r_{n2}<r_{n1}$为正整数, $r_{n2}/r_{n1}\to 0$, 且当$n\to \infty$时, $r_{n1}/n\to 0$成立.又取$k=[n/(r_{n1}+r_{n2}]$, 满足$k(r_{n1}+r_{n2})/n\to 1$, $k r_{n2}/n\to 0$, 且$r_{n1}^2/n\to 0$.利用Bernstein分块方法, $S_n$可分解为

$S_n=S_n^{\prime}+S_n^{\prime\prime}+S_n^{\prime\prime\prime}, $ (4.9)

其中

$S_n^\prime = \sum\limits_{m = 1}^k {{y_{nm}}} ,S_n^{\prime \prime } = \sum\limits_{m = 1}^k {y_{nm}^\prime } ,S_n^{\prime \prime \prime } = y_{nk + 1}^\prime ,\\ {y_{nm}} = \sum\limits_{i = {k_m}}^{{k_m} + {r_{n1}} - 1} {{Z_{ni}}} , y_{nm}^\prime = \sum\limits_{i = {l_m}}^{{l_m} + {r_{n2}} - 1} {{Z_{ni}}} , y_{nk + 1}^\prime = \sum\limits_{i = k({r_{n1}} + {r_{n2}}) + 1}^n {{Z_{ni}}} ,$

$k_m=(m-1)(r_{n1}+r_{n2})+1,\ \ l_m=(m-1)(r_{n1}+r_{n2})+r_{n1}+1,\ \ m=1,\cdots, k. $则在定理的条件下, 由文献Li and Yang[10]的引理3.4, 得到

$\frac{1}{n}S_n^{\prime\prime}\to 0,\ \ \frac{1}{n}S_n^{\prime\prime\prime}\to 0, \ \ \frac{k}{n}Ey_{m1}\to \sigma_0^2. $ (4.10)

再由文献Li et al.[11]的引理A.3及文献Li and Yang[10]的引理4.2和4.3得

$n^{-1/2}S_n^{\prime} \stackrel {d}{\to} N (0, \sigma_0^2). $ (4.11)

结合 (4.9)-(4.11) 式可得

$n^{-1/2}S_n \stackrel {d}{\to} N (0, \sigma_0^2).$

由此知$F_n(\xi_p)-p \stackrel {d}{\to} N (0,\sigma_0^2)$.这样定理3.3证明完毕.

参考文献
[1] Dowd K. Estimating VaR with order statistics[J]. J. Derivatives, 2001, 8(3): 23–30. DOI:10.3905/jod.2001.319154
[2] Koji Inui, Masaaki Kijima, Atsushi Kitano. VaR is subject to a signiflcant positive bias[J]. Statist.Probab. Letters, 2005(72): 299–311.
[3] 谢佳利, 杨善朝, 梁鑫. VaR样本分位数估计的偏差改进[J]. 数量经济技术经济研究, 2008, 25(12): 139–148.
[4] Bahadur R R. A note on quantiles in large samples[J]. Ann. Math. Stat., 1966, 37: 577–580. DOI:10.1214/aoms/1177699450
[5] 杨善朝, 陈敏. 相协随机变量的指数不等式与强大数律[J]. 中国科学 (A辑), 2007, 37(2): 200–208.
[6] Serfling R J. Approximation theorems of mathematical statistics[M]. New York: John Wiley and Sons, 1980.
[7] Wang Xuejun, Hu Shuhe, Yang Wenzhi. The Bahadur representation for sample quantiles under strongly mixing sequence[J]. J. Statist. Plan. Infer., 2011, 141: 655–662. DOI:10.1016/j.jspi.2010.07.008
[8] Roussas G G. Asymptotic normality of a smooth estimate of a random fleld distribution function under association[J]. Statist. Probab. Letters, 1995, 24: 77–90. DOI:10.1016/0167-7152(94)00151-W
[9] Azevedo C M, Oliveira P E. On the kernel estimation of a multivariate distribution function under positive dependence[J]. 2011, URI: http://hdl.handle.net/1822/12218.
[10] Li Yongming, Yang shanchao. Asymptotic normality of the empirical distribution under negatively associated sequencces and its applications[J]. J. Math. Res. Exp., 2006, 26(3): 457–464.
[11] Li Yongming, Yang Shanchao, Zhou Yong. Consistency and uniformly asymptotic normality of wavelet estimator in regression model with associated samples[J]. Statist. Probab. Letters, 2008, 78(17): 2947–2956. DOI:10.1016/j.spl.2008.05.004