Help with First-Order Autoregressive Autocorrelation

49 Views Asked by At

According to my lecturer:

$\text{Cov}(u_t,u_{t-1})=\rho(\sigma_u)^2$

where

$u_t=\rho u_{t-1}+\epsilon_t$

We have been given no more information than this. I would like to prove and understand this result, but to be honest, I do not even know where to start. Hoping someone can point me in the right direction.

1

There are 1 best solutions below

3
On BEST ANSWER

Without some extra assumptions the stated result is not necessarily true. In such models it is typically assumed that $\ \epsilon_t\ $ are independent and identically distributed (commonly with a normal distribution) with zero mean, constant variance, and are independent of $\ u_0\ $. These assumptions will imply that $\ \epsilon_t\ $ and $\ u_{t-1}\ $ are uncorrelated, which is all you really need to show that $\ \text{Cov}\big(u_t,u_{t-1}\big)=\rho\sigma_{u_{t-1}}^2\ $. However, the notation $\ \sigma_u\ $ on the right side of the stated result suggests that your lecturer is taking the variance of $\ u_t\ $ to be constant. This cannot be the case unless $\ |\rho|<1\ $ and $\ \epsilon_t\ $ has constant variance. If these conditions do hold, then $\ \sigma_{u_t}^2\rightarrow\frac{\sigma_\epsilon^2}{1-\rho^2}\ $ as $\ t\rightarrow\infty\ $, and if $\ \sigma_{u_0}^2=\frac{\sigma_\epsilon^2}{1-\rho^2}\ $ then $\ u_t\ $ will have that same variance for all $\ t\ $.

Let \begin{align} e_t&=\mathbb{E}\big(u_t\big)\ \ \text{ and}\\ m_t&=\mathbb{E}\big(\epsilon_t\big)\ . \end{align} Then taking expectations of both sides of the equation $$ u_t=\rho u_{t-1}+\epsilon_t $$ gives $$ e_t=\rho e_{t-1}+m_t\ , $$ and subtracting this latter equation from the former gives $$ (1)\hspace{5em}u_t-e_t=\rho\big(u_{t-1}-e_{t-1}\big)+\epsilon_t-m_t\ . $$ If we now multiply this equation by $\ u_{t-1}-e_{t-1}\ $ and take expectations of the result, we get $$ \text{Cov}\big(u_t,u_{t-1}\big)=\rho\sigma_{u_{t-1}}^2+\text{Cov}\big(u_{t-1},\epsilon_t\big)\ , $$ which reduces to the equation given by your lecturer when $\ u_t\ $ has constant variance and $\ u_{t-1},\epsilon_t\ $ are uncorrelated.

For the results on the variance of $\ u_t\ $, square both sides of equation $(1)$ and take expectations to obtain $$ \sigma_{u_t}^2=\rho^2\sigma_{u_{t-1}}^2+2\rho\,\text{Cov}\big(u_{t-1},\epsilon_t\big)+\sigma_{\epsilon_t}^2\ , $$ which reduces to $$ \sigma_{u_t}^2=\rho^2\sigma_{u_{t-1}}^2+\sigma_\epsilon^2 $$ when $\ u_{t-1},\epsilon_t\ $ are uncorrelated and $\ \epsilon_t\ $ has constant variance.

If $\ \rho=\pm1\ $, it follows by induction from this recurrence that $\ \sigma_{u_t}^2=\sigma_{u_0}^2+\sigma^2_\epsilon t\ $. Otherwise, the recurrence is known to have a solution of the form $\ \sigma_{u_t}^2=A+B\rho^{2t}\ $. Substituting this expression for $\ \sigma_{u_t}^2\ $ into the recurrence gives $\ A=\frac{\sigma_\epsilon^2}{1-\rho^2}\ $, and the initial condition at $\ t=0\ $ gives $\ B=\sigma_{u_0}^2-A\ $. In general, therefore, $$ \sigma_{u_t}^2=\cases{\sigma_0^2+\sigma^2_\epsilon t&if $\ |\rho|=1$\\\frac{\sigma_\epsilon^2}{1-\rho^2}+\left(\sigma_{u_0}^2-\frac{\sigma_\epsilon^2}{1-\rho^2}\right)\rho^{2t}&otherwise.} $$