3.

4: ARMA(p, q ) Model

Homework 3a

AR(p) + MA(q ) = ARMA(p, q )

ARMA(p, q ) Model
Denition (ARMA(p, q ) Model) A time series is ARMA(p, q ) if it is stationary and satises Zt = + 1 Zt 1 + + p Zt p + 1 at 1 + + q at q +at
AR part MA part

( )

Using the AR and MA operators, we can rewrite ( ) as t = (B )at (B )Z The textbook denes the ARMA process with a slightly different parametrization given by Zt = + 1 Zt 1 + + p Zt p 1 at 1 q at q + at ( )

Arthur Berg

AR(p) + MA(q ) = ARMA(p, q )

3/ 12

3.4: ARMA(p, q ) Model

Homework 3a

3.4: ARMA(p, q ) Model

Homework 3a

Causality and Invertibility of ARMA(p, q )

Parameter Redundancy
The process

Facts: The ARMA(p, q ) model given by (B )Zt = (B )at is causal if and only if |(z )| = 0 when |z | 1 i.e. all roots including complex roots of (z ) lie outside the unit disk. The ARMA(p, q ) model given by (B )Zt = (B )at is invertible if and only if |(z )| = 0 when || 1 i.e. all roots including complex roots of (z ) lie outside the unit disk. is equivalent to

Zt = at .5Zt 1 = .5at 1 which is also equivalent to ( ) ( ), i.e. Zt .5Zt 1 = at .5at 1 Zt in the last representation is still white noise, but has an ARMA(1,1) representation. We remove redundancies in an ARMA model (B )Zt = (B )at by canceling the common factors in (B ) and (B ). (

( ) )

Arthur Berg

AR(p) + MA(q ) = ARMA(p, q )

4/ 12

Arthur Berg

AR(p) + MA(q ) = ARMA(p, q )

5/ 12

3.4: ARMA(p, q ) Model

Homework 3a

3.4: ARMA(p, q ) Model

Homework 3a

Example Removing Redundancy

Example Consider the process

Example Causality and Invertibility

Example (cont.)
Zt = .4Zt 1 + .45Zt 2 + at + at 1 + .25at 2 which is equivalent to (1 .4B .45B 2 ) Zt = (1 + B + .25B 2 ) at
(B ) (B )

The reduced model is (B )Zt = (B )at where (z ) = 1 .9z (z ) = 1 + .5z There is only one root of (z ) which is z = 10/9, and |10/9| lies outside the unit circle so this ARMA model is causal. There is only one root of (z ) which is z = 2, and | 2| lies outside the unit circle so this ARMA model is invertible.

Is this process really ARMA(2,2)? No! Factoring (z ) and (z ) gives (z ) = 1 .4z .45z 2 = (1 + .5z )(1 .9z ) (z ) = (1 + z + .25z 2 ) = (1 + .5z )2 Removing the common term (1 + .5z ) gives the reduced model Zt = .9Zt 1 + .5at 1 + at
Arthur Berg AR(p) + MA(q ) = ARMA(p, q ) 6/ 12

Arthur Berg

AR(p) + MA(q ) = ARMA(p, q )

7/ 12

3.4: ARMA(p, q ) Model

Homework 3a

3.4: ARMA(p, q ) Model

Homework 3a

ARMA Mathematics
Start with the ARMA equation (B )Zt = (B )at To solve for Zt , we simply divide! That is Zt = (B ) at (B )
(B )

Example MA() Representation

Example (cont.) Note that (z ) = MA() representation (z ) 1 + .5z = (z ) 1 .9z for |z | 1 for |z | 1 for |z | 1
2 3

= (1 + .5z )(1 + .9z + .92 z 2 + .93 z 3 + )

2

= 1 + (.5 + .9)z + (.5(.9) + .92 )z 2 + (.5(.92 ) + .93 )z 3 + = 1 + (.5 + .9)z + (.5 + .9)(.9)z + (.5 + .9)(.9 )z +

Similarly, solving for at gives at = (B ) Zt (B )

(B )

AR() representation

= 1 + (.5 + .9)
1.4 j =1

.9j 1 z j

for |z | 1

The key is to guring out the j and j in

(B ) = 1 +
j =1
Arthur Berg

j B j

(B ) = 1 +
j =1

j B j
8/ 12

This gives the following MA() representation: (B ) Zt = at = 1 + 1.4 .9j 1 B j at = at + 1.4 (B )

j =1
Arthur Berg AR(p) + MA(q ) = ARMA(p, q )

.9j 1 at j
j =1
9/ 12

AR(p) + MA(q ) = ARMA(p, q )

3.4: ARMA(p, q ) Model

Homework 3a

3.4: ARMA(p, q ) Model

Homework 3a

Example AR() Representation

Example (cont.) Similarly, one can show (z ) = (z ) (z ) 1 .9z = 1 + .5z . . .

Homework 3a

Read 4.1 and 4.2. Derive the equation for (z ) on slide 10. Do exercise #3.14(a) only for models (i), (ii), and (iv) Do exercise #3.14(b,c) only for model (ii) (.5)j 1 z j
j =1

= 1 1.4

for |z | 1

This gives the following AR() representation:

Zt = 1.4
Arthur Berg

(.5)j 1 at j + at
j =1
10/ 12 Arthur Berg AR(p) + MA(q ) = ARMA(p, q ) 12/ 12

AR(p) + MA(q ) = ARMA(p, q )