1
$\begingroup$

Consider this regression model $$Y_t=X_t\beta+\epsilon_t, ~~~~~~~~~~\epsilon_t \sim WN(0, \sigma^2_{\epsilon})$$

with 3 different specifications of the error term:

  1. $\epsilon_t=\alpha_1\epsilon_{t-1}+\alpha_2\epsilon_{t-2}+\eta_t,~~~~~~~~~~~~~~~~~~~~~~~~~~\eta_t \sim WN(0, \sigma^2_{\eta}) $
  2. $\epsilon_t=\eta_t+\theta_1\eta_{t-1}+\theta_2\eta_{t-2},~~~~~~~~~~~~~~~~~~~~~~~~~~\eta_t \sim WN(0, \sigma^2_{\eta})$
  3. $\epsilon_t=\alpha_1\epsilon_{t-1}+\eta_t~~~~~~~~~~$ with $|\alpha|=1, ~~~~~~~~~~\eta_t \sim WN(0, \sigma^2_{\eta})$


a) Compute the expected value, the variance and the autocorrelation coefficients of $\epsilon_t$ for each specification

b) Transform the model $Y_t=X_t\beta+\epsilon_t$ in first differences and let $\varepsilon_t$ be the transformed error. Show that $\varepsilon_t$ is not white noise.



My attempt was the following:

POINT A) - (1) FIRST SPECIFICATION

$E(\epsilon_t) =E(\alpha_1\epsilon_{t-1}+\alpha_2\epsilon_{t-2}+\eta_t)$

$~~~~~~~~~~~~~~~~=\alpha_1E(\epsilon_{t-1})+\alpha_2E(\epsilon_{t-2})+E(\eta_t)$

$~~~~~~~~~~~~~~~~=0$


$\operatorname{Var}(\epsilon_t)=\operatorname{Var}(\alpha_1\epsilon_{t-1}+\alpha_2\epsilon_{t-2}+\eta_t)$

$~~~~~~~~~~~~~=\alpha_1^2 \operatorname{Var}(\epsilon_{t-1})+\alpha_2^2\operatorname{Var}(\epsilon_{t-2})+\operatorname{Var}(\eta_t)$

$~~~~~~~~~~~~~=\alpha_1^2\cdot\sigma^2_{\epsilon}+\alpha_2^2\cdot\sigma^2_{\epsilon}+\sigma^2_{\eta}$


$\rho(h)=\frac{\operatorname{Cov}(\epsilon_t,\epsilon_{t-h})}{\operatorname{Var}(\epsilon_t)}$ From this point, how can I compute $\operatorname{Cov}(\epsilon_t,\epsilon_{t-h})$?


For the other ones

Basically, I have to follow the same reasoning (hoping that the previous calculus are correct)


POINT B) - SHOW THAT $\varepsilon_t$ IS NOT WHITE NOISE

$$Y_t-Y_{t-1}=(X_t-X_{t-1})\beta+\varepsilon_t,~~~~~~~~~~\varepsilon_t=\epsilon_t-\epsilon_{t-1}$$

How can I prove that the transformed error term is not white noise?

Any help would be appreciated

$\endgroup$
  • $\begingroup$ I don't understand your model. How can $\epsilon_t$ be both white noise and follow the specifications? $\endgroup$ – Riccardo Sven Risuleo Feb 27 '19 at 15:28
1
$\begingroup$

Your variance is wrong: $\epsilon_{t-1}$ and $\epsilon_{t-2}$ are not independent!

Hint: $$ \mathbf{E}\big[\eta_t \epsilon_t] = \mathbf{E}\big[\eta_t(\alpha_1\epsilon_{t-1} + \alpha_2\epsilon_{t-2} + \eta_t)] = \sigma_\eta^2; $$ because $\epsilon_k$ is independent of $\eta_j$ if $j>k$ (think about why this is the case!); similarly, $$\begin{aligned} \mathbf{E}\big[\eta_{t-1} \epsilon_t] &= \mathbf{E}\big[\eta_{t-1}(\alpha_1\epsilon_{t-1} + \alpha_2\epsilon_{t-2} + \eta_t)] = \alpha_1\sigma_\eta^2;\\ \mathbf{E}\big[\eta_{t-2} \epsilon_t] &= \mathbf{E}\big[\eta_{t-2}(\alpha_1\epsilon_{t-1} + \alpha_2\epsilon_{t-2} + \eta_t)] = \alpha_2\sigma_\eta^2;\\ \end{aligned}$$ Hence, $$\begin{aligned} \mathrm{Var}(\epsilon_t) &= \mathbf{E}\big[\epsilon_t \epsilon_t\big]\\ &= \mathbf{E}\big[(\alpha_1 \epsilon_{t-1} + \alpha_2\epsilon_{t-2} + \eta_t) \epsilon_t\big]\\ &= \mathbf{E}\big[(\alpha_1 (\alpha_1 \epsilon_{t-2} + \alpha_2 \epsilon_{t-3} + \eta_{t-1}) + \alpha_2(\alpha_1\epsilon_{t-3} + \alpha_2\epsilon_{t-4} + \eta_{t-2}) +\eta_t\big) \epsilon_t\big] \\\ &=\cdots \end{aligned} $$

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.