I've read that a time series should be a weakly stationary for the Autocorrelation function (ACF) to make sense. The definition for weakly stationary series that I have is that all the observations should have the same mean and variance and that $\mathrm{Cov}(x_t,x_{t-1})$ should be same for all $t$. But ACF for a natural number $h$ is defined to be the correlation of $x_t$ and $x_{t-h}$. So, for it to be defined as a function, we should have $\mathrm{Cov}(x_t,x_{t-h})$ independent of $t$. How is this implied by $\mathrm{Cov}(x_t,x_{t-1})$ is independent of $t$??

$\mathrm{Cov}(x_t,x_{t-1})\text{ is independent of }t\implies\mathrm{Cov}(x_t,x_{t-h}) \text{ is independent of } t\,\forall h\in\Bbb N$??

Please help.

  • $\begingroup$ Can you give a reference for that claim? All sources I am aware of state the condition for all $h$, as you do on the r.h.s. of the presumed implication. $\endgroup$ Jun 17, 2020 at 14:07

1 Answer 1


Consider an AR(2) process $$ Y_t=\phi_{2}Y_{t-2}+\epsilon_t, $$ where $\phi_{2}=\phi_{2,1}$ for $t=-\infty,\ldots,[sT]$ and $\phi_{2}=\phi_{2,2}$ for $t=[sT]+1,\ldots,\infty$. Hence, the process is not stationary, because the properties of the process change after $[sT]$. However, before and after the break, $\phi_1$, the coefficient on the first lag, is zero.

By properties of AR(2) processes, the first autocorelation of this process is $$ \rho_1=\frac{\phi_1}{1-\phi_2}=0, $$ which is evidently independent of $t$. The second autocorrelation, however, is now given by $$ \rho_2=\phi_2, $$ which is not stable in view of the structural change in the process.

  • $\begingroup$ Thank you very much! $\endgroup$
    – Martund
    Jun 19, 2020 at 5:09

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.