MA(q)的参数估计——公式推导

原创

已于 2023-06-14 23:40:38 修改 · 939 阅读

于 2023-06-14 23:39:12 首次发布

文章介绍了MA(q)时间序列模型的参数估计方法，包括基于极大似然估计的公式推导，以及在已知一部分滞后误差项条件下的条件极大似然估计。通过对观测数据集进行分析，利用多元正态分布的似然函数来估计模型参数。

$M A (q)$ 的参数估计——公式推导

设 $M A (q)$ 模型为 $y_t=\theta_0+\theta_1\epsilon_{t-1}+\ldots+\theta_q\epsilon_{t-q}+\epsilon_t$ ，其中 $\epsilon_t\mathop{\sim}\limits^{i.i.d.}N(0,\sigma_\epsilon^2)$

极大似然估计

假设观测数据集为 $\{y_1,\ldots,y_T\}$ ，令 $\theta=(\theta_0,\ldots,\theta_q,\sigma_\epsilon^2)^\prime$ ， $y=(y_1,\ldots,y_T)^\prime$ ，
似然函数为：
$L(\theta)=f_\theta(y_1,\ldots,y_T)$
其中， $(y_1,\ldots,y_T)\sim N(\mu,\Sigma)$

在精确似然估计中，我们假设 $(y_1,\ldots,y_T)$ 服从多元正态分布
$\mu=(Ey_1,\ldots,Ey_T)^\prime=(\theta_0,\ldots,\theta_0)^\prime$
$\begin{align*} \Sigma&=\left(\begin{matrix} Cov(y_1,y_1)&Cov(y_1,y_2)&\ldots&Cov(y_1,y_{q+1})&Cov(y_1,y_{q+2})&\ldots&Cov(y_1,y_{T-1})&Cov(y_1,y_T) \\ Cov(y_2,y_1)&Cov(y_2,y_2)&\ldots&Cov(y_2,y_{q+1})&Cov(y_2,y_{q+2})&\ldots&Cov(y_2,y_{T-1})&Cov(y_2,y_T) \\ \vdots&\vdots&&\vdots&\vdots&&\vdots&\vdots \\ Cov(y_{q+1},y_1)&Cov(y_{q+1},y_2)&\ldots&Cov(y_{q+1},y_{q+1})&Cov(y_{q+1},y_{q+2})&\ldots&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)&\ldots&Cov(y_{q+2},y_{q+1})&Cov(y_{q+2},y_{q+2})&\ldots&Cov(y_{q+2},y_{T-1})&Cov(y_{q+2},y_T) \\ \vdots&\vdots&&\vdots&\vdots&&\vdots&\vdots \\ Cov(y_{T-1},y_1)&Cov(y_{T-1},y_2)&\ldots&Cov(y_{T-1},y_{q+1})&Cov(y_{T-1},y_{q+2})&\ldots&Cov(y_{T-1},y_{T-1})&Cov(y_{T-1},y_T) \\Cov(y_T,y_1)&Cov(y_T,y_2)&\ldots&Cov(y_T,y_{q+1})&Cov(y_T,y_{q+2})&\ldots&Cov(y_T,y_{T-1})&Cov(y_T,y_T) \end{matrix}\right) \\ \\&=\left(\begin{matrix} (1+\sum\limits_{t=1}^{q}\theta_t^2)\sigma_\epsilon^2&\theta_1\sigma_\epsilon^2+\sum\limits_{t=1}^{q-1}\theta_t\theta_{t+1}\sigma_\epsilon^2&\ldots&\theta_p\sigma_\epsilon^2&0&\ldots&0&0\\ \theta_1\sigma_\epsilon^2+\sum\limits_{t=1}^{q-1}\theta_t\theta_{t+1}\sigma_\epsilon^2&(1+\sum\limits_{t=1}^{q}\theta_t^2)\sigma_\epsilon^2&\ldots&\theta_{q-1}\sigma_\epsilon^2+\theta_1\theta_2\sigma_\epsilon^2&\theta_p\sigma_\epsilon^2&\ldots&0&0 \\ \vdots&\vdots&&\vdots&\vdots&&\vdots&\vdots \\ \theta_p\sigma_\epsilon^2&\theta_{q-1}\sigma_\epsilon^2+\theta_1\theta_2\sigma_\epsilon^2&\ldots&(1+\sum\limits_{t=1}^{q}\theta_t^2)\sigma_\epsilon^2&\theta_1\sigma_\epsilon^2+\sum\limits_{t=1}^{q-1}\theta_t\theta_{t+1}\sigma_\epsilon^2&\ldots&Cov(y_{q+1},y_{T-1})&Cov(y_{q+1},y_T) \\ 0&\theta_p\sigma_\epsilon^2&\ldots&\theta_1\sigma_\epsilon^2+\sum\limits_{t=1}^{q-1}\theta_t\theta_{t+1}\sigma_\epsilon^2&(1+\sum\limits_{t=1}^{q}\theta_t^2)\sigma_\epsilon^2&\ldots&Cov(y_{q+2},y_{T-1})&Cov(y_{q+2},y_T) \\ \vdots&\vdots&&\vdots&\vdots&&\vdots&\vdots \\ 0&0&\ldots&Cov(y_{T-1},y_{q+1})&Cov(y_{T-1},y_{q+2})&\ldots&(1+\sum\limits_{t=1}^{q}\theta_t^2)\sigma_\epsilon^2&\theta_1\sigma_\epsilon^2+\sum\limits_{t=1}^{q-1}\theta_t\theta_{t+1}\sigma_\epsilon^2\\0&0&\ldots&Cov(y_T,y_{q+1})&Cov(y_T,y_{q+2})&\ldots&\theta_1\sigma_\epsilon^2+\sum\limits_{t=1}^{q-1}\theta_t\theta_{t+1}\sigma_\epsilon^2&(1+\sum\limits_{t=1}^{q}\theta_t^2)\sigma_\epsilon^2 \end{matrix}\right) \end{align*}$