Likelihood Estimation on ARCH process

169 Views Asked by At

on the seminal paper of Robert Engle about Arch Process (Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica: Journal of the Econometric Society, 987-1007.) the first order conditions for maximum likelihood estimation for an ARCH(1) process is derived. Consider the next statements:

\begin{equation} y_t = \epsilon_t\cdot h_t^{\frac{1}{2}} \\ h_t = \alpha_0 + \alpha_1 \cdot y_{t-1}^2 \end{equation}

Where $E{[}\epsilon_t{]}=0$ and $Var{[}\epsilon_t{]}=1$. Now, the log-likelihood estimator of $\alpha$ named $l_t$ is:

\begin{equation} l_t = -\frac{1}{2} log\, h_t - \frac{1}{2} \frac{y_t^2}{h_t} \end{equation}

where $log\,x$ is the natural logarithm of $x$. Well, the paper state that the first order conditions for the likelihood estimation are:

\begin{equation} \frac{\partial l_t}{\partial \alpha} = \frac{1}{2h_t} \frac{\partial h_t}{\partial \alpha} \left(\frac{y_t^2}{h_t} - 1\right) \end{equation}

along with the Hessian. My question is how get the expression for $ \frac{\partial l_t}{\partial \alpha}$ from $l_t$? I tried to deduce the expression but I get a different result. In the calculation of the partial derivative of $y_t^2$ with respect to $\alpha$, is $y_t$ treated as independent of $\alpha$? and if the answer is yes why is that? (considering that $y_t = \epsilon_t \cdot h_t^{\frac{1}{2}}$ where $h_t$ itself depends on $\alpha$).

In advance, thanks for your time.

1

There are 1 best solutions below

0
On BEST ANSWER

Well, checking my calculations after a long time, there's the answer. First, the model used is:

\begin{align} y_t | \psi_{t-1} &\sim \mathcal{N}(0, h_t)\\ h_t &= \alpha_0 + \alpha_1 y_{t-1}^2 \\ h_t &= h(y_{t-1}, y_{t-2}, \ldots,y_{t-p}, \alpha) \end{align}

where $p$ is the order of the process. As you can see $y_t$ is not equal to $\epsilon_t \cdot h^{1/2}$. So:

\begin{align} \frac{\partial{h_t}}{\partial\alpha} &= - \frac{1}{2} \frac{\partial{log(h_t)}}{\partial\alpha} - \frac{1}{2} \frac{\partial{\frac{y^2_t}{h_t}}}{\partial\alpha}\\ \frac{\partial{h_t}}{\partial\alpha} &= - \frac{1}{2 h_t} \frac{\partial{h_t}}{\partial\alpha} - \frac{1}{2} \frac{-y^2_t}{h^2_t} \frac{\partial h_t}{\partial \alpha}\\ \frac{\partial{h_t}}{\partial\alpha} &= - \frac{1}{2 h_t} \frac{\partial{h_t}}{\partial\alpha} \left(1 - \frac{y^2_t}{h^2_t}\right) \\ \frac{\partial{h_t}}{\partial\alpha} &= \frac{1}{2 h_t} \frac{\partial{h_t}}{\partial\alpha} \left(\frac{y^2_t}{h^2_t} - 1\right) \end{align}

which is the expression needed.