# Why is an ARMA model a parsimonous approximation of an AR model?

I am reading a book on time series and I came across the following: "In addition to being a parsimonous approximation to a high-order AR(p) model, ARMA models...". Why is an ARMA model a (parsimonous) approximation to an AR model?

Also, somewhat later on I read that the coefficients of an MA model can be estimated by writing the model in an 'AR form'. Could anyone also please explain that statement to me?

1. It's not that all high-order AR models are well approximated by low-order ARMA models, but in practice it's often the case that the correlation structure is such that to achieve a good approximation by a pure-AR(p) requires a high-order model (p large), but a lower-order ARMA(p',q) may fit at least as well ((p'+q) < p). For example, an ARMA(1,1) might be a reasonable model for a series, but as a pure AR you might need an AR(10) to do about as well.

2. MA models may be written as infinite AR's (and vice-versa) by simple manipulation. For simplicity, I'll assume there's no constant term. Let $B$ be the backshift operator:

$y_t= e_t - \theta_1 e_{t-1} - \theta_2 e_{t-2} - ... - \theta_q e_{t-q}$

$y_t= e_t - \theta_1 B e_{t} - \theta_2 B^2 e_{t} - ... - \theta_q B^q e_{t}$

$y_t= (1 - \theta_1 B - \theta_2 - ... - \theta_q B^q )e_t$

$y_t= (1 - \mathbf{\theta}(B))e_t$ , where $\mathbf{\theta}(B)$ is a polynomial in $B$.

Hence

$y_t (1-\mathbf{\theta}(B))^{-1}= e_t$

where the series expansion for $(1-\mathbf{\theta}(B))^{-1}$ is an infinite series in powers of $B$ - an infinite AR.

For example, consider an MA(1):

$y_t= e_t - \theta e_{t-1}$

$y_t= (1 - \theta B )e_t$

$y_t (1-\theta B)^{-1}= e_t$

$y_t (1+\theta B +\theta^2B^2+\theta^3B^3+...)= e_t$

$y_t +\theta y_{t-1} +\theta^2y_{t-2}+\theta^3y_{t-2}+...)= e_t$

which is an infinite AR with $\phi_1=-\theta$, $\phi_2=-\theta^2$, $\phi_3=-\theta^3$ and so on. Note that if such an inversion is valid, these coefficients will decrease geometrically (for MA(q), eventually the same effect will be seen - a fairly rapid decrease in the magnitude of the high-order AR coefficients).

Such inversions are only possible under certain conditions on the parameters. (e.g. see the bottom of the page here)

• Thanks. But you took here an infinite order of an MA model as well (at least of the same order $q$). So why is this then more parsimonous than an AR model?
– rbm
Commented Jun 14, 2014 at 9:53
• Your edit clarified thinks a lot. Thanks for your help :)!
– rbm
Commented Jun 14, 2014 at 10:23