I want to find Yule-Walker equations for a causal ARMA(1,1)-process

$Y_t=\alpha Y_{t-1}+\epsilon_t+\theta\epsilon_{t-1}$.

My idea initially was to first multiply both sides by $Y_{t-1}$, so I would get

$\gamma_0=\alpha\gamma_1+\epsilon_t Y_t+\theta\epsilon_{t-1}Y_t$, where $\gamma_0$ is the variance and $\gamma_1$ is the first autocovariance.

But, I don't know how to continue from there.

Could someone help me out? Thanks in advance!

  • 3
    $\begingroup$ Yule walker is for AR models. This has an MA component $\endgroup$
    – Taylor
    May 6 '17 at 16:05
  • $\begingroup$ Go down the road of representing the process as an $AR(\infty)$? Does that get you anywhere? $\endgroup$ May 6 '17 at 17:32

Yule Walker (for parameter estimation) is usually only used for AR models, but this method you're using is still a valid technique for finding the autocovariance function. I'm assuming that's what you're after.

Multiply both sides of an ARMA(p,q) model by $Y_{t-k} = \sum_{j=0}^{\infty}\psi_j \epsilon_{t-j-k}$ (this is where causality assumption comes in)

$$ Y_t Y_{t-k} - \phi_1 Y_{t-1}Y_{t-k} - \cdots - \phi_p Y_{t-p}Y_{t-k} = \left[\sum_{l=0}^q \theta_l \epsilon_{t-l}\right]\left[\sum_{j=0}^{\infty}\psi_j \epsilon_{t-j-k} \right]. \tag{1} $$

Then take expectations. Note that $\psi_j = 0$ for $j < 0$ and $\theta_j = 0$ for $j \not \in \{1,\ldots,q\}$. If $0 \le k < \text{max}(p,q+1)$ you get these $p$ (1 in your case) equations here $$ \gamma(k) - \phi_1 \gamma(k-1) - \cdots - \phi_p\gamma(k-p) = \sigma^2 \sum_{j=0}^{\infty}\theta_{k+j}\psi_j. \tag{2} $$ And if $k \ge \text{max}(p,q+1)$ you get $$ \gamma(k) - \phi_1 \gamma(k-1) - \cdots - \phi_p\gamma(k-p) = 0 \tag{3} $$ ($0$ on RHS because $l > q$).

You can solve your two equations by using results of homogeneous difference equations, or by just solving the system directly.

For your specific ARMA(1,1) model (I'll turn the $\phi$s back into $\alpha$s) you have

  1. $\gamma(0) - \alpha \gamma(1) = \sigma^2[1 + \theta_1 (\theta + \alpha)]$
  2. $\gamma(1) - \alpha \gamma(0) = \sigma^2[\theta]$
  3. $\gamma(k) - \alpha \gamma(k-1) = 0$ for $k \ge 2$

Plugging (2) into (1) gives us $$ \gamma(0) = \sigma^2\left[ 1 + \frac{(\theta+\alpha)^2 }{1 - \alpha^2 } \right] $$ (which is the same as before) and plugging this into (2) gives us $$ \gamma(1) = \sigma^2\left[ \theta + \alpha + \frac{(\theta+\alpha)^2\alpha }{1 - \alpha^2 } \right]. $$ If you want to get it at higher lags, just use (3) repeatedly.

When the model is a pure AR model, the RHS of (1) is simplified tremendously, and the resulting set of equations to solve is linear in the parameters $\phi_1,\ldots,\phi_p$.

  • $\begingroup$ What are the $\psi$ terms that you introduce in the second paragraph? How are they related to the ARMA parameters? Is there no way to express the solution in a closed form, say, using matrix algebra? Thank you $\endgroup$
    – Confounded
    Nov 15 '20 at 14:27

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.