MA(q)的参数估计——公式推导

$MA(q)$的参数估计——公式推导

---

设$MA(q)$模型为$y_t={\theta }_0+{\theta }_1{ϵ}_{t-1}+\dots +{\theta }_q{ϵ}_{t-q}+{ϵ}_t$，其中${ϵ}_t\stackrel{i.i.d.}{\sim }N(0,{\sigma }_{ϵ}^2)$

极大似然估计

---

假设观测数据集为 { y 1 , … , y T } \{y_1,\ldots,y_T\} {y1​,…,yT​}，令$\theta =({\theta }_0,\dots ,{\theta }_q,{\sigma }_{ϵ}^2{)}^{\prime }$，$y=(y_1,\dots ,y_T{)}^{\prime }$，
似然函数为：
$$L(\theta )=f_{\theta }(y_1,\dots ,y_T)$$
其中，$(y_1,\dots ,y_T)\sim N(\mu ,\Sigma )$

在精确似然估计中，我们假设$(y_1,\dots ,y_T)$服从多元正态分布
$$\mu =(E{y}_1,\dots ,E{y}_T{)}^{\prime }=({\theta }_0,\dots ,{\theta }_0{)}^{\prime }$$
$$\begin{array}{rl}\Sigma & =\left(\begin{array}{cccccccc}Cov(y_1,y_1)& Cov(y_1,y_2)& \dots & Cov(y_1,y_{q+1})& Cov(y_1,y_{q+2})& \dots & Cov(y_1,y_{T-1})& Cov(y_1,y_T)\\ Cov(y_2,y_1)& Cov(y_2,y_2)& \dots & Cov(y_2,y_{q+1})& Cov(y_2,y_{q+2})& \dots & Cov(y_2,y_{T-1})& Cov(y_2,y_T)\\ ⋮& ⋮& & ⋮& ⋮& & ⋮& ⋮\\ Cov(y_{q+1},y_1)& Cov(y_{q+1},y_2)& \dots & Cov(y_{q+1},y_{q+1})& Cov(y_{q+1},y_{q+2})& \dots & Cov(y_{q+1},y_{T-1})& Cov(y_{q+1},y_T)\\ Cov(y_{q+2},y_1)& Cov(y_{q+2},y_2)& \dots & Cov(y_{q+2},y_{q+1})& Cov(y_{q+2},y_{q+2})& \dots & Cov(y_{q+2},y_{T-1})& Cov(y_{q+2},y_T)\\ ⋮& ⋮& & ⋮& ⋮& & ⋮& ⋮\\ Cov(y_{T-1},y_1)& Cov(y_{T-1},y_2)& \dots & Cov(y_{T-1},y_{q+1})& Cov(y_{T-1},y_{q+2})& \dots & Cov(y_{T-1},y_{T-1})& Cov(y_{T-1},y_T)\\ Cov(y_T,y_1)& Cov(y_T,y_2)& \dots & Cov(y_T,y_{q+1})& Cov(y_T,y_{q+2})& \dots & Cov(y_T,y_{T-1})& Cov(y_T,y_T)\end{array}\right)\\ & \\ & =\left(\begin{array}{cccccccc}(1+\sum _{t=1}^q{\theta }_t^2){\sigma }_{ϵ}^2& {\theta }_1{\sigma }_{ϵ}^2+\sum _{t=1}^{q-1}{\theta }_t{\theta }_{t+1}{\sigma }_{ϵ}^2& \dots & {\theta }_p{\sigma }_{ϵ}^2& 0& \dots & 0& 0\\ {\theta }_1{\sigma }_{ϵ}^2+\sum _{t=1}^{q-1}{\theta }_t{\theta }_{t+1}{\sigma }_{ϵ}^2& (1+\sum _{t=1}^q{\theta }_t^2){\sigma }_{ϵ}^2& \dots & {\theta }_{q-1}{\sigma }_{ϵ}^2+{\theta }_1{\theta }_2{\sigma }_{ϵ}^2& {\theta }_p{\sigma }_{ϵ}^2& \dots & 0& 0\\ ⋮& ⋮& & ⋮& ⋮& & ⋮& ⋮\\ {\theta }_p{\sigma }_{ϵ}^2& {\theta }_{q-1}{\sigma }_{ϵ}^2+{\theta }_1{\theta }_2{\sigma }_{ϵ}^2& \dots & (1+\sum _{t=1}^q{\theta }_t^2){\sigma }_{ϵ}^2& {\theta }_1{\sigma }_{ϵ}^2+\sum _{t=1}^{q-1}{\theta }_t{\theta }_{t+1}{\sigma }_{ϵ}^2& \dots & Cov(y_{q+1},y_{T-1})& Cov(y_{q+1},y_T)\\ 0& {\theta }_p{\sigma }_{ϵ}^2& \dots & {\theta }_1{\sigma }_{ϵ}^2+\sum _{t=1}^{q-1}{\theta }_t{\theta }_{t+1}{\sigma }_{ϵ}^2& (1+\sum _{t=1}^q{\theta }_t^2){\sigma }_{ϵ}^2& \dots & Cov(y_{q+2},y_{T-1})& Cov(y_{q+2},y_T)\\ ⋮& ⋮& & ⋮& ⋮& & ⋮& ⋮\\ 0& 0& \dots & Cov(y_{T-1},y_{q+1})& Cov(y_{T-1},y_{q+2})& \dots & (1+\sum _{t=1}^q{\theta }_t^2){\sigma }_{ϵ}^2& {\theta }_1{\sigma }_{ϵ}^2+\sum _{t=1}^{q-1}{\theta }_t{\theta }_{t+1}{\sigma }_{ϵ}^2\\ 0& 0& \dots & Cov(y_T,y_{q+1})& Cov(y_T,y_{q+2})& \dots & {\theta }_1{\sigma }_{ϵ}^2+\sum _{t=1}^{q-1}{\theta }_t{\theta }_{t+1}{\sigma }_{ϵ}^2& (1+\sum _{t=1}^q{\theta }_t^2){\sigma }_{ϵ}^2\end{array}\right)\end{array}$$

于是，似然函数便可借由多元正态密度函数推导如下：
$$L(\theta )=\frac{1}{\sqrt{2\pi |\Sigma |}}exp\left(-\frac{1}{2}(y-\mu {)}^{\prime }{\Sigma }^{-1}(y-\mu )\right)$$

条件极大似然估计

---

在给定 { ϵ 1 − q , … , ϵ 0 } \{\epsilon_{1-q},\ldots,\epsilon_0\} {ϵ1−q​,…,ϵ0​}的条件下，我们可以迭代计算得到 { ϵ 1 , … , ϵ T − 1 } \{\epsilon_1,\ldots,\epsilon_{T-1}\} {ϵ1​,…,ϵT−1​}，因此：
$$\begin{array}{rl}L(\theta )& =f_{\theta }(y_1,\dots ,y_T|{ϵ}_{1-q},\dots ,{ϵ}_0)\\ & =f_{\theta }(y_1,\dots ,y_T|{ϵ}_{1-q},\dots ,{ϵ}_{T-1})\\ & =\prod _{t=2}^T{f}_{\theta }(y_t|y_{t-1},\dots ,y_1,{ϵ}_{1-q},\dots ,{ϵ}_{T-1})f_{\theta }(y_1|{ϵ}_{1-q},\dots ,{ϵ}_{T-1})\\ & =\prod _{t=1}^T{f}_{\theta }(y_t|{ϵ}_{t-1},\dots ,{ϵ}_{t-q})\end{array}$$其中，$y_t|{ϵ}_{t-1},\dots ,{ϵ}_{t-q}\sim N({\theta }_0+\sum _{i=1}^q{\theta }_i{ϵ}_{t-i},{\sigma }_{ϵ}^2)$
故似然函数为：
$$L(\theta )=\prod _{t=1}^T\frac{1}{\sqrt{2\pi {\sigma }_{ϵ}^2}}exp\left(-\frac{(y_t-{\theta }_0-\sum _{i=1}^q{\theta }_i{ϵ}_{t-i}{)}^2}{2{\sigma }_{ϵ}^2}\right)$$