It is known (see e.g., Brockwell and Davis, Introduction to Time Series and Forecasting, p. 95) that the $h$th partial autocorrelation $\phi_{hh}$ of a stationary process can be derived by first finding the prediction errors of $Y_h$ given $Y_{h-1},\ldots,Y_1$ and of $Y_0$ given $Y_{h-1},\ldots,Y_1$, and then finding the correlation of the prediction errors. (The second errors are thus "backcast" errors resulting from predicting $Y_0$ by future $Y_t$.)

(Note that I am aware that we may also compute $\phi_{hh}$ as the last element of $\Gamma^{-1}\widetilde\gamma_h$ where $\Gamma$ is the Toeplitz matrix of the autocovariances $\gamma_0,\ldots,\gamma_{h-1}$ and $\widetilde\gamma_h=(\gamma_1,\ldots,\gamma_{h})'$, but that is not the point of the question.)

I want to verify this for $h=2$ for the MA(1) case $Y_t=\epsilon_t+\theta\epsilon_{t-1}$, where we know that the second partial autocorrelation is $$ \phi_{22}=-\frac{\theta^2}{1+\theta^2+\theta^4} $$ So for $h=2$ I find the optimal prediction of $Y_2$ given $Y_1$ to be $Y_2^1=\rho Y_1$ (see e.g. slides 12-16 here), where $$ \rho=\frac{\theta}{1+\theta^2} $$ is the autocorrelation.

Likewise, $Y_0^1=\rho Y_1$. Plugging in for the $Y_t$ yields prediction errors \begin{align*} Y_2-Y_2^1&=\epsilon_2+\theta\epsilon_{1}-\frac{\theta}{1+\theta^2}\left(\epsilon_{1}+\theta\epsilon_{0}\right)\\ &=\epsilon_2+\frac{\theta^3}{1+\theta^2}\epsilon_1-\frac{\theta^2}{1+\theta^2}\epsilon_0 \end{align*} and likewise \begin{align*} Y_0-Y_0^1&=\epsilon_0+\theta\epsilon_{-1}-\frac{\theta}{1+\theta^2}\left(\epsilon_{1}+\theta\epsilon_{0}\right)\\ &=-\frac{\theta}{1+\theta^2}\epsilon_1+\frac{1}{1+\theta^2}\epsilon_{0}+\theta\epsilon_{-1} \end{align*} Some manipulation then gives me \begin{align*} Corr(Y_2-Y_2^1,Y_0-Y_0^1)&=Corr\left(\frac{\theta^3}{1+\theta^2}\epsilon_1-\frac{\theta^2}{1+\theta^2}\epsilon_0,-\frac{\theta}{1+\theta^2}\epsilon_1+\frac{1}{1+\theta^2}\epsilon_{0}\right)\\ &=-\frac{\theta^4}{(1+\theta^2)^2}-\frac{\theta^2}{(1+\theta^2)^2}\\ &=-\frac{\theta^2(1+\theta^2)}{(1+\theta^2)^2}\\ &=-\frac{\theta^2}{1+\theta^2}, \end{align*} which evidently is not the correct result.

Is it an algebraic mistake, a misunderstanding of this approach to compute $\phi_{22}$ or some intermediate mistake in implementing, e.g., backcasting?

  • $\begingroup$ just out of curiosity, should it not be that $\phi_{22}=\frac{\theta^2}{\theta^2-1}$ for MA(1)? And why do you set $\rho=\frac{\theta}{1+\theta^2}$, since this is not from back-casting but the OLS estimate of AR(1) fit to an MA(1) model... $\endgroup$ – Math-fun Jul 5 '16 at 14:32
  • $\begingroup$ Thanks for having a look. As for your first comment, see for example Example 3.2.7 in Brockwell and Davis (see above citation). As for the second, that was my understanding of slide 13 of the slides I linked in my question. $\endgroup$ – Christoph Hanck Jul 5 '16 at 14:54

I found my mistake in calculating $\phi_{22}$. Your expression is correct. I took a closer look at the slides you referred to and I find your expression for $\rho$ also fine. But I found your mistake in calculation too:

\begin{align*} Corr(Y_2-Y_2^1,Y_0-Y_0^1)&=\frac{Cov\left(\frac{\theta^3}{1+\theta^2}\epsilon_1-\frac{\theta^2}{1+\theta^2}\epsilon_0,-\frac{\theta}{1+\theta^2}\epsilon_1+\frac{1}{1+\theta^2}\epsilon_{0}\right)}{\sqrt{Var(Y_2-Y_2^1)}\sqrt{Var(Y_0-Y_0^1)}}\\ &=\frac{Cov\left(\frac{\theta^3}{1+\theta^2}\epsilon_1-\frac{\theta^2}{1+\theta^2}\epsilon_0,-\frac{\theta}{1+\theta^2}\epsilon_1+\frac{1}{1+\theta^2}\epsilon_{0}\right)}{Var(Y_2-Y_2^1)}\\ &=\frac{-\frac{\theta^2}{1+\theta^2}}{\frac{1+2\theta^2+2\theta^4+\theta^6}{(1+\theta^2)^2}}\\ &=\color{red}{-\frac{\theta^2}{1+\theta^2+\theta^4}}\\ \end{align*}

  • $\begingroup$ my pleasure! :-) $\endgroup$ – Math-fun Jul 6 '16 at 10:48

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.