First order moving average process is defiend as: \[ X_t \hspace3mm = \hspace3mm \epsilon_t - \theta_1 \epsilon_{t-1}, \hspace15mm \epsilon_t \sim WN(0, \sigma^2) \] and \(\theta_1\) is real valued constant.
Compare this to the first order Autoregressive process \[ X_t - \phi_1 X_{t-1} \hspace3mm = \hspace3mm \epsilon_t \]
Using the backward operator, MA(1) can be written as \[ X_t \hspace3mm = \hspace3mm \underbrace{ (1 - \theta_1 B) }_{\Theta(B) } \, \epsilon_t. \]
Watch the sign in front of \(\theta_1\)! This is Cryer’s convension. See
ARIMA functions page of the website.
\[
\begin{align}
E(X_t)
&= \hspace3mm E(\epsilon_t - \theta_1 \epsilon_{t-1}) \\\\
&= \hspace3mm E(\epsilon_t) - \theta_1 E(\epsilon_{t-1})
&= \hspace3mm 0
\end{align}
\]
\[
\begin{align}
\mbox{Var}(X_t)
&= \hspace3mm \mbox{Var}(\epsilon_t - \theta_1 \epsilon_{t-1})\\
\\
&= \hspace3mm \mbox{Var}(\epsilon_t) + \theta_1^2
\mbox{Var}(\epsilon_{t-1})\\ \\
&= \hspace3mm \sigma^2 + \theta_1^2 \sigma^2 \\ \\
&= \hspace3mm (1+\theta_1^2) \sigma^2
\end{align}
\]
\[
\begin{align}
\gamma(1)
&= \hspace3mm \mbox{ Cov }(X_t, X_{t+1} ) \\ \\
&= \hspace3mm
\mbox{ Cov }(\epsilon_t- \theta_1 \epsilon_{t-1}, \hspace3mm
e_{t+1}-\theta_1 e_{t} )\\ \\
&= \hspace3mm \mbox{ Cov }(\epsilon_t , \, e_{t+1} )
- \mbox{ Cov }(\epsilon_t , \, \theta_1 e_{t} )
- \mbox{ Cov }(\theta_1 \epsilon_{t-1}, \, e_{t+1} )
+ \mbox{ Cov }(\theta_1 \epsilon_{t-1}, \, \theta_1 e_{t} ) \\ \\
&= \hspace3mm -\theta_1 \mbox{ Cov }(\epsilon_t , \, e_{t} ) \\
\\
&= \hspace3mm -\theta_1 \, \sigma^2
\end{align}
\]
\[
\begin{align}
\gamma(2)
&= \hspace3mm \mbox{ Cov }(Y_t, Y_{t+2} ) \\ \\
&= \hspace3mm \mbox{ Cov }(\epsilon_t- \theta_1 \epsilon_{t-1},
\hspace3mm e_{t+2}-\theta_1 e_{t+1} )\\ \\
&= \hspace3mm 0
\end{align}
\]
ACVF \[ \gamma(h) = \left\{ \begin{array}{ll} \sigma^2(1+\theta_1^2) & \mbox{ if } h=0 \\ -\sigma^2 \theta_1 & \mbox{ if } h= \pm 1 \\ 0 & \mbox{ if } |h|>1 \\ \end{array} \right. \]
ACF \[
\rho(h) =
\left\{
\begin{array}{ll}
1 & \mbox{ if } h=0 \\
-\theta_1/ (1+\theta^2_1) & \mbox{ if } h= \pm 1 \\
0 & \mbox{ if } |h|>1 \\
\end{array} \right.
\]
Moving Average process of order \(q\) is \[ X_t = \epsilon_t - \theta_1 \epsilon_{t-1} - \theta_2 \epsilon_{t-2} - \cdots - \theta_q e_{t-q}, \] where \(\epsilon_t \sim WN(0,\sigma^2)\) and \(\theta_1, \ldots, \theta_q\) is real valued constant.
Using the backward operator, we write this as \[ X_t = (1 - \theta_1 B - \theta_2 B^2 - \cdots - \theta_q B^q) \, \epsilon_t. \]
We shorten the notation further, and write \[
X_t = \Theta(B) Z_t.
\] where Characteristic Polinomial, \(\Theta(z) = (1 - \theta_1 z - \theta_2 z^2 -
\cdots - \theta_q z^q)\).
Mean \[ E(X_t) \hspace3mm = \hspace3mm E(\epsilon_t - \theta_1 \epsilon_{t-1} - \theta_2 \epsilon_{t-2} - \cdots - \theta_q e_{t-q}) \hspace3mm = \hspace3mm 0. \]
Variance \[
\begin{align}
\mbox{Var}(X_t)
&= \hspace3mm \mbox{Var}(\epsilon_t - \theta_1 \epsilon_{t-1} -
\theta_2 \epsilon_{t-2} - \cdots - \theta_q e_{t-q}) \\\\
&= \hspace3mm \sigma^2 + \theta_1 \, \sigma^2 + \theta_2 \,
\sigma^2 + \cdots + \theta_q \, \sigma^2 \\\\
&= \hspace3mm \Big(1 + \theta_1 + \theta_2 + \cdots + \theta_q
\Big) \, \sigma^2
\end{align}
\]
\[ \begin{align} \gamma(h) &= \hspace3mm \mbox{ Cov }(X_t, \hspace3mm X_{t+h} )\\ &= \hspace3mm \mbox{ Cov }\Big(\epsilon_t - \theta_1 \epsilon_{t-1} - \theta_2 \epsilon_{t-2} - \cdots - \theta_q e_{t-q}, \hspace5mm e_{t+h} - \theta_1 e_{t-1+h} - \theta_2 e_{t-2+h} - \cdots - \theta_q e_{t-q+h}\Big) \end{align} \]
Theoretical ACVF: \[ \gamma(h) = \left\{ \begin{array}{ll} \sigma^2 (1+\theta_1^2 \cdots + \theta_q^2) & \mbox{ if } h=0 \\ \sigma^2 (-\theta_1 + \theta_2 \theta_1 + \theta_3 \theta_2 + \cdots + \theta_{q-1} \theta_{q}) & \mbox{ if } h=1 \\ \sigma^2 (-\theta_2 + \theta_3 \theta_1 + \cdots + \theta_{q-2} \theta_{q}) & \mbox{ if } h=2 \\ \vdots & \\ \sigma^2 (-\theta_q ) & \mbox{ if } h=q \\ 0 & \mbox{ if } |h|>q\\ \end{array} \right. \]
Theoretical ACF: \[
\rho(h) =
\left\{ \begin{array}{ll}
1 & \mbox{ if } h=0 \\
\gamma(h) / \gamma(0) & \mbox{ if } -q \leq h \leq q \\
0 & \mbox{ if } |h|> q \\
\end{array} \right.
\]
Every stationary q-corr TS with mean 0 can be represented as MA(\(q\)) process.
\[ X_t = \epsilon_t - \theta_1 \epsilon_{t-1} - \theta_2 \epsilon_{t-2} - \theta_3 \epsilon_{t-3} \]
\[ \gamma(h) = \left\{ \begin{array}{ll} \sigma^2 (1+\theta_1^2 +\theta_2^2 + \theta_3^2 ) & \mbox{ if } h=0 \\ \sigma^2 (-\theta_1 + \theta_2 \theta_1 + \theta_3 \theta_2 ) & \mbox{ if } h=1 \\ \sigma^2 (-\theta_2 + \theta_3 \theta_1 ) & \mbox{ if } h=2 \\ \sigma^2 (-\theta_3 ) & \mbox{ if } h=3 \\ 0 & \mbox{ if } |h|> \\ \end{array} \right. \]
\[ \rho(h) = \gamma(h) / \gamma(0) \]
Watch the Sign!
# [Cryer]'s sign convention is (1-phi B) Y_t = (1-theta) \epsilon_t.
# [Brockwell]'s is (1-phi B) Y_t = (1+theta) \epsilon_t.
#
# For AR(1) (1-.5 B) Y_t = \epsilon_t, ACF should be all positive.
# (1+.5 B) Y_t = \epsilon_t, ACF should be alternating in sign.
#
# For MA(1) Y_t = (1-.5 B) \epsilon_t ACF at lag 1 should be negative
# Y_t = (1+.5 B) \epsilon_t ACF at lag 1 should be positive
#--- Theoretical ACF and PACF of MA ---
# ARMAacf() uses sign of MA parameters like [Brockwell]: Y_t = (1+theta B) \epsilon_t
= c(.2, .3, .2, .2)
Theta
= ARMAacf(ma = Theta, lag.max=20, pacf=FALSE)
MArho1 <- ARMAacf(ma = Theta, lag.max=20, pacf=TRUE)
MApacf1
layout(matrix(1:2, 1,2))
plot(0:20, MArho1, type="h", col="red"); abline(h=0)
plot(1:20, MApacf1, type="h", col="red"); abline(h=0)
#--- Theoretical ACVF of MA
#
# tacvfARMA() uses sign of MA parameters like [Cryer]: Y_t = (1 - theta B) \epsilon_t
#
library(ltsa)
= tacvfARMA(theta= c(.2, .3, .2, .2), maxLag=20, sigma2=1) #- Theoretical ACVF
MAgam2 = MAgam2/MAgam2[1] #- Theoretical ACF
MArho2
plot(0:20, MArho2, type="h", col="red"); abline(h=0)
= c(.5)
T1 = c(.2, -.3, .2, -.2)
T2 = c(.2, .3, .2, .2)
T3 = c(-.2, -.3, -.2, -.2)
T4
= T1
Theta
# arima.sim() uses sign of MA parameters like [Brockwell]: Y_t = (1+theta B) \epsilon_t
= arima.sim( list(ma = c(0.5) ), n=250 )
x = ARMAacf(ma = Theta, lag.max=20, pacf=FALSE)
MArho1 <- ARMAacf(ma = Theta, lag.max=20, pacf=TRUE)
MApacf1
plot(x, type="o"); abline(h=0)
layout(matrix(1:2, 1, 2) )
acf(x); lines(0:20, MArho1, type="p", col="red")
pacf(x); lines(1:20, MApacf1, type="p", col="red")
layout(1)
ACF cuts off at \(q\)
PACF tails off
MA(\(q\)) process is defiend as \[ X_t \hspace3mm = \hspace3mm (1 - \theta_1 B - \theta_2 B^2 - \cdots - \theta_q B^q) \epsilon_t. \\ \hspace3mm = \hspace3mm \Theta(B) \epsilon_t \] where \(\Theta(z)\) is the characteristic polynomial of MA(\(q\)), and \(\epsilon_t ~ WN(0,\sigma^2)\).
MA(\(q\)) has ACF that cuts-off after lag \(q\), and PACF that decays down.
MA(\(q\)) has mean of 0. Theoretical ACVF: \[ \gamma(h) = \left\{ \begin{array}{ll} \sigma^2 (1+\theta_1^2 \cdots + \theta_q^2) & \mbox{ if } h=0 \\ \sigma^2 (-\theta_1 + \theta_2 \theta_1 + \theta_3 \theta_2 + \cdots + \theta_{q-1} \theta_{q}) & \mbox{ if } h=1 \\ \sigma^2 (-\theta_2 + \theta_3 \theta_1 + \cdots + \theta_{q-2} \theta_{q}) & \mbox{ if } h=2 \\ \vdots & \\ \sigma^2 (-\theta_q ) & \mbox{ if } h=q \\ 0 & \mbox{ if } |h|>q\\ \end{array} \right. \] \[ \rho(h) = \left\{ \begin{array}{ll} 1 & \mbox{ if } h=0 \\ \gamma(h) / \gamma(0) & \mbox{ if } -q \leq h \leq q \\ 0 & \mbox{ if } |h|> q \\ \end{array} \right. \]
You can find code for Theoretical ACF and ACVF of MA(q) on page .
You can find code for Simulating MA(\(q\)) process and looking at Saompe ACF and PACF on page .