Combining [[Autoregressive Model]] and [[Moving Average Model]] models into $\text{ARMA}(p,q)$ where $p$ defines the number of autoregressive lags and $q$ defines the number of lagged [[White Noise Model|White Noise]] terms.
$
X_t=\phi_1X_{t-1}+\dots +\phi_pX_{t-p}+
W_t+ \theta_1W_{t-1}+\dots+\theta_qW_{t-q}
$
For the following derivation of moments, consider a simpler $\mathrm{ARMA}(1,1)$ process of the form:
$ r_t = \phi_0 + \phi_1 r_{t-1} + \epsilon_t- \theta \epsilon_{t-1} $
where:
- $\epsilon$ is the white noise that follows $\mathcal N(0, \sigma^2)$
- $\lvert \phi_1 \rvert < 1$ to ensure [[Stationarity]] in the process
- $\{\phi_0, \phi_1, \theta, \sigma^2\}$ are constants
## Expectation
In a stationary process $\mathbb E[r_t]= \mu$. Thus, we simply write:
$
\begin{align}
\mu &= \mathbb E[\phi_0 + \phi_1 r_{t-1} + \epsilon_t- \theta \epsilon_{t-1}]\tag{1}\\[4pt]
&=\phi_0+ \phi_1 \mathbb E[r_{t-1}]+\mathbb E[\epsilon]- \theta \mathbb E[\epsilon_{t-1}] \tag{2}\\[4pt]
&=\phi_0+ \phi_1 \mu \tag{3}\\[4pt]
\mu&=\frac{\phi_0}{1-\phi_1} \tag{4}
\end{align}
$
where:
- (2) The expectation of white noise terms is zero.
- (3) The expectation of $r_{t-1}$ is the same as for $r_t$ due to stationarity, and hence $\mu$.
## Variance
For simplicity we assume $\phi_0=0$, which means that the process is centered (i.e. has zero mean).
$ r_t = \phi_1 r_{t-1} + \epsilon_t- \theta \epsilon_{t-1} $
For such a process with zero mean, the [[Variance]] is equal to the second moment $\mathbb E[r_t^2]$. Hence we can write:
$
\begin{align}
\mathrm{Var}(r_t) &= \mathbb E\left[(\phi_1 r_{t-1} + \epsilon_t- \theta \epsilon_{t-1})^2\right]\tag{1}\\[6pt]
&= \mathbb E\left[\phi_1^2 r_{t-1}^2 + \epsilon_t^2 + \theta^2 \epsilon_{t-1}^2 + 2 \,\phi_1 \, r_{t-1}\, \epsilon_t -2\phi_1r_{t-1} \theta \epsilon_{t-1}-2\epsilon_t\theta\epsilon_{t-1}\right]\tag{2}\\[6pt]
&= \phi_1^2\mathbb E[r_{t-1}^2]+\mathbb E[\epsilon_t^2]+\theta^2 \mathbb E[\epsilon_{t-1}^2]+2\phi_1\mathbb E[r_{t-1}\epsilon_t]-2\phi_1\theta \mathbb E[r_{t-1} \epsilon_{t-1}] - 2\theta \mathbb E[\epsilon_t \epsilon_{t-1}]\tag{3}\\[6pt]
&= \phi_1^2 \mathrm{Var}(r_t)+\sigma^2+\theta^2 \sigma^2-2\phi_1\theta \sigma^2 \tag{4}
\end{align}
$
In (4) multiple simplifications can be made:
| Term | Explanation |
| -------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| $\mathbb E[\epsilon_t^2]$ | Has zero mean for all $t$. Therefore its second moment is the variance, which by definition is $\sigma^2$. |
| $\mathbb E[\epsilon_t \epsilon_{t-1}]$ | Represents the covariance between error terms, which by definition of the model are independent. Hence covariance is zero. |
| $\mathbb E[r_{t-1}^2]$ | Represents the $\mathrm{Var}(r_{t-1})$ since the $\mathbb E[r_{t-1}]$ is zero. Due to stationarity the variance for all $t$ is the same. |
| $\mathbb E[r_{t-1} \epsilon_t]$ | Represents the [[Covariance]] between the two [[Random Variable\|r.v's]] as both have zero mean. Since $e_t$ happens after $r_{t-1}$ the two terms are completely independent (i.e. covariance of zero). |
| $\mathbb E[r_{t-1} \epsilon_{t-1}]$ | Since $r_{t-1}$ contains $\epsilon_{t-1}$ in its definition, they are not uncorrelated. More explicitly their covariance is $\sigma^2$. |
Finally we can move over the variance from the right to the left side, and express it as follows:
$
\begin{align}
\mathrm{Var}(r_t) *(1-\phi_1^2) = \sigma^2+\theta^2 \sigma^2-2\phi_1\theta \sigma^2 \\[2pt]
\mathrm{Var}(r_t) =\frac{\sigma^2(1 + \theta^2 - 2\phi_1 \theta)}{1-\phi_1^2}
\end{align}
$
## Autocovariance
We want to derive the 1-lag autocovariance $\gamma_1$ of the defined $\mathrm{ARMA}(1,1)$ process. Since the mean of $r$ for all $t$ is zero, we can write the covariance as the product of the expectation.
$ \gamma_1 = \mathrm{Cov}(r_t, r_{t-1})=\mathbb E[r_tr_{t-1}]$
By filling in the recursion we can write:
$
\begin{align}
\gamma_1 &= \mathbb E[(\phi_1 r_{t-1} + \epsilon_t - \theta \epsilon_{t-1}) *r_{t-1}] \tag{1}\\[2pt]
&=\mathbb E[\phi_1r_{t-1}^2+ \epsilon_tr_{t-1}- \theta \epsilon_{t-1}r_{t-1}] \tag{2}\\[2pt]
&=\phi_1\mathbb E[r_{t-1}^2]+ \mathbb E[\epsilon_tr_{t-1}]- \theta \mathbb E[\epsilon_{t-1}r_{t-1}] \tag{3}\\[2pt]
&=\phi_1 \mathrm{Var}(r_t)- \theta \sigma^2 \tag{4}
\end{align}
$
where:
- (2) Multiply out the terms
- (3) Apply [[Linearity of Expectations]] to write expectations separately
- (4) Apply simplifications as above.
Finally we fill in the derived definition of the variance, to get:
$ \gamma_1 = \phi_1*\frac{\sigma^2(1 + \theta^2 - 2\phi_1 \theta)}{1-\phi_1^2} - \theta \sigma^2 $