Univariate Stationary Time Series Models
Applied Econometrics
by
Sunil Paul
Madras School of Economics
01-07-16
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Autoregressive Process
AR(1) process
I
A first order autoregressive process(without
constant):Yt = Yt1 + t , t WN(0, 2 )
Solution by recursive substitution:
Yt = Y0 +
t1
1 +
t2
t1
X
2 +...+t1 +t = Y0 +
i ti
t
i=0
I
Alternatively, solving forward j period from time t
Yt+j = j+1 Yt1 + j t + j1 t+1 + ... + t+j1 + t+j
= j+1 Yt1 +
j
X
i=0
Sunil Paul
Lecture Notes
i t+ji for j > 0
Univariate Stationary Time Series Models
Some Important Interpretations
I
Dynamic Multiplier (the effect of 1 on Yt ):
dYt+j
dt
dYt
dtj
= j , and
= j (The dynamic multiplier depends on j)
Explosive and stable system
Impulse response function(IRF): plot j against time j
Cumulative effect upto horizon j:
Pj
dYt+j
dYt+j
dYt+j
j
j1 + ... + + 1 =
j
i=o
dt + dt+1 + ... + dt+j = +
Long run
< 1):
i
h effect as j goes to (given||
dYt+j
dYt+j
dYt+j
2
lim
dt + dt+1 + ... + dt+j = 1 + + + .... =
1/(1 )
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Stability and Stationarity conditions
The AR(1) process is stable and stationary only if || < 1
If || < 1 then
I
lim j = 0
j
I
AR(1) can be expressed in wold form :Yt =
P
j=0 || = 1/(1 ) <
j=0
j tj
Remember
if || = 1 then AR(1) is a random walk and
P
j=0 || ( The process is not stable)
Sunil Paul
Lecture Notes
and
Univariate Stationary Time Series Models
AR(1) Process in Lag Operator Notation
I
(1 L)Yt = t
to get the wold form multiply both sides with (1 L)1
Note that if || < 1 then (1 L)1 = (1+ L + 2 L2 + ...)
(1 L)1 (1 L)Yt = (1 L)1 t
Yt = (1 + L + 2 L2 + ...)t = t + t1 + 2 t2 + ...
I
For stability root of (1 L) should be greater than one,
L = (1/) > 1. or < 1.
If = 1 its is unit root case, if > 1 then the process is
explosive and has unbounded mean and variance.
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
AR(1) Process with drift
Yt = c + Yt1 + t , t WN(0, 2 )
Solution to this equation is:
Yt = t Y0 + c
t1
X
i=0
i +
t1
X
i ti
i=0
Pt
i
Assuming
||
<
1
we
have:Y
=
c/(1
)
+
t
i=0 ti with
P
j=0 || < ( i.e. the process converges if and only if lies
strictly inside the unit interval)
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Moments of Stationary AR(1) Process with drift
= E [Yt ] = c/(1 )
0 = E (Yt )2 = E (t + t1 + ...)2 =
(1 + 2 + 4 + ...) 2 = 2 /(1 2 )
I
Pt
Pt
Note that both E [Yt ] = c i=0 i and var [Yt ] = 2 i=0 2i
will be finite and independent of time only if || < 1..
j = E (Yt )(Ytj ) = E (t + t1 + 2 t2 + ... +
j tj + j+1 tj1 + ...) (tj + tj1 + 2 tj2 + ...) =
(j + j+2 + j+4 + ...) 2 = j [1 + 2 + 4 + ...) 2 =
[j /(1 2 )] 2
j = j /0 = j
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Moments of Stationary AR(1) Process with drift
We can also derive moments of AR(1) process by using
stationarity properties:
E (Yt ) = c + E (Yt1 ) + E (t ) = c + E (Yt ) = E (Yt ) =
c/(1 ) = [Stationarity means E (Yt ) = E (Yt1 )]
Solving the expression = c/(1 ) for c we get
(1 ) = c
Substituting this into Yt = c + Yt1 + t we have
Yt = (1 ) + Yt1 + t = Yt = (Yt1 ) + t
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Moments of Stationary AR(1) Process with drift
To get variance square the above expression and take
expectation on both sides:E (Yt )2 =
2 E (Yt1 )2 + E (t )2 + 2E [(Yt1 )t ]
Assuming stationarity we get
E (Yt )2 = E (Yt1 )2 = 0 = 2 0 + 2 + 0.
Hence 0 = 2 /(1 2 )
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Moments of Stationary AR(1) Process with Drift
To get autocovariance multiply Yt = (Yt1 ) + t
with Ytj on both sides and take expectations
i.e. E [(Yt )(Ytj )] =
E [(Yt1 )(Ytj )] + E [t (Ytj )]
Thus j = j1 and this is a first order diffidence equation
By solving we get j = j 0 = j [ 2 /(1 2 )](For instance;
2 = 1 , since1 = 0 , 2 can be writen as2 0 )
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Moments of Stationary AR(1) Process with drift
j
0
= j
Autocorrelation:j =
Half life (speed of mean reversion): lag at which IRF
decreases by one half, j = ln(0.5)
ln() [ take the log and rearrange
j
the equation = 0.5]
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Autocovarice generating function
The ACVGF of an AR(1)can be constructed as follows
gY (z) = (z)(z 1) =
1
1 z
1
1 z 1
= 2 (1 + z + 2 z 2 + ...) + (1 + z 1 + 2 z 2 + ...)
X
1
1 +
= 2
j z j + z j
2
1
j=1
Thus
j =
Sunil Paul
2 j
1 2
Lecture Notes
Univariate Stationary Time Series Models
AR(2) process
Yt = c + 1 Yt1 + 2 Yt2 + t or
(1 1 L 2 L2 )Yt = c + t .
Stationarity and stability conditions can be specified in terms
of the roots of lag polynomial (1 1 L 2 L2 ) = 0 (require
the roots to lie outside the unit circle)
The solution to the characteristic equation can be obtained by
factoring the quadratic equation into (1 1 L)(1 2 L) = 0,
where 1 + 2 = 1 and 1 2 = 2
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
AR(2) Process
In general we have a characteristic equation by replacing L
with z : (1 1 z 2 z 2 ) = (1 1 z)(1 2 z) = 0
The roots of
the quadratic characteristic
equation
1 21 +42
1 + 21 +42
, z2 =
are:z1 =
22
22
q
The roots can be real if 21 + 42 0 or complex if
q
21 + 42 < 0
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Moments of Stationary AR(2)
I
E (Yt ) = c + 1 E (Yt1 ) + 2 E (Yt2 ) + E (t )
Assuming stationarity we get
= c + 1 + 2 + 0 = = c/(1 1 2 )
Substituting the value of c from the above equation into
Yt = c + 1 Yt1 + 2 Yt2 + t we get
(Yt ) = 1 (Yt1 ) + 2 (Yt2 ) + t
For variance multiply both sides with (Yt ) and take
expectations on both sides:
E (Yt )2 =
1 E (Yt1 )(Yt )+2 E (Yt2 )(Yt )+E (t )(Yt )
I
i.e. 0 = 1 1 + 2 2 + 2
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Moments of Stationary AR(2)
I
For auto covariance multiply both sides with (Ytj ) and
take expectations on both sides:
E (Yt )(Ytj ) = 1 E (Yt1 )(Ytj ) +
2 E (Yt2 )(Ytj ) + E (t )(Ytj )
I
j = 1 j1 + 2 j2 for j = 1, 2, ....[Yule Walker equations]
To get autocorrelation j divide both sides of j by
0 : j = 1 j 1 + 2 j 2 for j 6= 0
I
I
0
1
1
2
=1
= 1 + 2 1 ( using the fact 1 = 1 ),thus
= 1 /(1 2 )
= 1 1 + 2
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Wold Representation of Stationary AR(2) process
(11 L2 L2 )Yt = c +t . = (11 L)(12 L)Yt = c +t .
Operating both sides of this equation
by(1 1 L)1 (1 2 L)1 we get:
I
I
Yt = (1 1 L)1 (1 2 L)1 c + (1 1 L)1 (1 2 L)1 t or
Yt = (L)c + (L)
t where (L)
P j j P j j
=
L
is the inverse of (1 1 L 2 L2 )
j=0 1
j=0 2 L
Sunil Paul
Lecture Notes
Univariate Stationary Time Series Models
Reference
Hamilton, Time series analysis Chapters 1,2 and 3
Sunil Paul
Lecture Notes
