Non stationary processes

September 6, 2025

1 Stochastic trend versus deterministic trend

1.1 Deterministic trend
Consider the following model
yt = + t + "t (1)
where "t is a white noise with mean zero and standard-deviation 2 and t is a
time. The variable is non-stationary because of the presence of deterministic
trend. The mean is given by:
E(yt ) = E ( + t + "t )
= + t (2)
the variance is
2
V ar(yt ) = V ar( + t + "t ) = V ar("t ) = (3)
and the correlation of lag j 6= 0 is:

j = Corr(yt ; yt j ) = Corr("t ; "t j ) = 0:

The process is non-stationary because the mean depends on time and we
have
lim E(yt ) = lim E( + t) = 1:
t!1 t!1
To render the process stationary, we just remove the trend
y•t = yt t = "t :
The new variable y•t is a white noise and thus becomes stationary.

1
1.2 Stochastic trend

Consider the following AR(1) model

yt = + yt 1 + "t (4)
2
where "t is a white noise with mean zero and standard-deviation .

We can rewrite the process as

X
t 1
t 2 t 1 i
yt = y0 + 1+ + + ::: + + "t i + "t (5)
i=1

dyt i
=
d"t i
i
dyt = d"t i

It j j < 1; lim i = 0 and the recent shocks have higher e¤ect than old
i!1
shocks.
If = 1, then we have
X
t 1
yt = y0 + t + "t i : (6)
i=0

dyt
= 1
d"t i
dyt = d"t i

If y0 = 0,
X
t 1
yt = t + "t i : (7)
i=0
The mean is given by:
X
t 1
E(yt ) = t+ E ("t i ) (8)
i=0
= t (9)

2
the variance is
Xt 1 X
t 1
2
V ar(yt ) = V ar( "t i ) = V ar("t i ) = t (10)
i=0 i=0

and the correlation of lag j 6= 0 is:

t j
j = Corr(yt ; yt j ) = :
t

We notice that not only the mean depends on time but also the variance
and all autocorrelations.
Also we have
lim E(yt ) = lim t = 1;
t!1 t!1

lim V ar(yt ) = lim (t 2 ) = 1

t!1 t!1

and
t j j
lim j = lim = lim 1 = 1:
t!1 t t!1 t!1 t
If = 1, the AR(1) model becomes

yt = + yt 1 + "t : (11)

Thus we have a random walk with drift .

The forecast

E (yt nIt 1 ) = E ( + yt 1 + "t nIt 1 )

= + yt 1 : (12)

To render the process stationary, we take the …rst di¤erence

y•t = yt = yt yt 1 = + "t :

The process becomes a white noise plus a constant. Thus, it is stationary.

Bonus question: Derive the E (• yt ) and Corr (•
yt ), V ar (• yt ; y•t 1 ).

3
2 Autoregressive integrated moving average
model: ARIMA(p,q)
Consider an ARMA(p,q) process
)
(L)yt = (L)"t (13)
)
where the polynomial (L) is given by
)
(L) = (L)(1 L): (14)
Pp j
where the roots of (L) = 1 i=1 j L are all outside of the unity
circle.
One of the roots is equal to one. It follows that the process is not a
stationary ARMA model. To see this, we can rewrite the model as

(L)(1 L)yt = (L) (yt yt 1 ) = (L) yt (15)

p
!
X
j
= 1 jL yt = (L)"t (16)
i=1
p
X
yt = i yt i + (L)"t (17)
i=1
p q
X X
yt = i yt i+ j "t j (18)
i=1 j=1
p q
X X
y•t = iy
•t + j "t j (19)
i=1 j=1

It follows that, y•t = yt has an ARMA(p,q) model. This means that the
…rst di¤erence yt of follows an ARMA(p,q) model.
Finally, we have
p
X
yt = yt 1 + i yt i + (L)"t : (20)
i=1

The coe¢ cient related to yt 1 is equal to one.

Recall that

4
 (L) yt = (L)"t :
so that yt is stationary process. In this case, the non-stationarity is elim-
inated by the …rst di¤erence. We say that yt is integrated of order one and
we call is an I(1) process (has one unit root).
Remarks:

If yt is stationary without di¤erencing then it is called an I(0) process.

We say also that it is stationary in level.

If yt is stationary after …rst di¤erencing is called an I(1) process (inte-

grated of order one).

If yt is stationary after d times di¤erencing is called an I(d) process

(integrated of order d).

The series that are I(1) are called di¤erence stationary, while the series
with a deterministic trend are called to be trend stationary.

If for example yt has two umit roots, then we di¤erentiate the variable
twice to make it stationary. That is,
2
( yt ) = yt = (yt yt 1 ) = yt yt 1
= (yt yt 1 ) (yt 1 yt 2 )
= yt 2yt 1 + yt 2
3
yt = (yt 2yt 1 + yt 2 )

3 Testing for the presence of unit roots

Consider the following AR(1) model

yt = + yt 1 + "t (21)
2
where "t is a white noise with mean zero and standard-deviation .
The null hypothesis to test is

= 1:

5
 The model can be rewritten by subtracting yt 1 from both sides as,

yt = + yt 1 + "t (22)

where = 1:The null hypothesis to test becomes

= 0:

We can estimate by OLS and test for its signi…cance.

Nevertheless, as shown by Dickey and Fuller (1979), under the null, the
standard t-statistic doesn’t follow a t distribution anymore. However, we can
test for the validity the null = 1 or = 0 using the t-stat

t = (b 1)=sb or t = b=sb

but the statistic follows the Dickey-Fuller distribution and the critical values
must be taken from the Dickey-Fuller tables. This test is called Dickey-Fuller
test.

3.1 DF test for an AR(1) process

The Dickey-Fuller test can be run by estimating the equation

yt = + yt 1 + "t ; (23)
and test for the validity of the null hypothesis

H0 : = 0;
where "t is a white noise
Model with linear trend:

yt = + t + yt 1 + "t (24)
We can test for the presence of a unit root

H0 : = 0:

or jointly test for the validity of the following null hypothesis

H0 : = 0 and = 0:

If the last hypothesis is rejected, then we have a trend stationary series.

6
3.2 Augmented Dickey-Fuller test (ADF test)
If is not a white noise in AR(1) model ( yt = + yt 1 + "t ), the n we can
add the lags of until we get a white noise. That is, we consider the model

yt = + 1 yt 1 + 2 yt 2 + ::: + p yt p + "t

In this case, the unit-root hypothesis implies the joint hypotheses:

1 + 2 + ::: + p = 1 and j 1 + 2 + ::: + pj <1

We can also add t term to the model if we expect the presence of deter-
ministic linear trend.
Equivalently, we can subtract yt 1 from both sides as of the regression
equation to get,

yt = + yt 1 + 1 yt 1 + 2 yt 2 + ::: + p yt p + "t (25)

where = 1 + ::: + p 1: To test for the presence of unit root, we test

for the validity of the following hypothesis

H0 : = 0:

Again, we have to use DF tables.

If a series must be di¤erenced d times to get stationarity time series (I(d)
series) and if after di¤erencing the process is an ARMA(p,q) process, then
we call it ARIMA(p,d,q) and we denote it
d
(L) yt = (L)"t :
d
yt has an ARMA(p,q) then yt has has an ARIMA(p,d,q)
Remarks:

Usually the economic indicators are stationary in level or after …rst

di¤erence and rarely need the second di¤erence to become stationary.

Before we take a decision about the null hypothesis (unit root), we

need to identify the la lenght in the AR part of equation (25). The
lag lenght can selected using information criteria. We can also use
backward signi…cant t-statistic (we keep the …rst backward signi…cant
lag to be p).

7
Note:

V ar(yt ) = V ar( + t + "t )

= E((yt E(yt ))2 )
= E(( + t + "t t)2 )
= E("2t ) = V ar("t ) = 2

yt = + t + "t

E(yt ) = + t
yt j = + (t j) + "t j

E(yt j ) = + (t j)

j = Cov(yt ; yt j )
= E((yt E(yt ))(yt j E(yt j )))
= E(( + t + "t t)( + (t j) + "t j (t j)))
= E("t "t j ) = E(("t 0) ("t j 0))
= E(("t E ("t )) ("t j E ("t j ))) = Cov("t ; "t j )
= 0

Cov(yt ; yt j )
j = Corr(yt ; yt j ) = p
V ar(yt )V ar(yt j )
Cov("t ; "t j ) Cov("t ; "t j )
= p = p
V ar("t )V ar("t j ) 2 2

0
= 2
= 0:

yt = + yt 1 + "t
yt 1 = + yt 2 + "t 1

8
 yt = + ( + y t 2 + "t 1 ) + "t
= (1 + ) + 2 yt 2 + "t 1 + "t

yt 2 = + yt 3 + "t 2

2
yt = (1 + ) + y t 2 + "t 1 + " t
2
= (1 + ) + ( + yt 3 + "t 2 ) + "t 1 + "t
2
= 1+ + + 3 yt 3 + 2 "t 2 + "t 1 + "t

yt 3 = + yt 4 + "t 3

2 3 2
yt = 1+ + + yt 3 + "t 2 + "t 1 + "t
2 3 4 3 2
= 1+ + + + yt 4 + "t 3 + "t 2 + "t 1 + "t

2 3 j 1 j
yt = 1+ + + + ::: + + yt j
j 1 3 2
+ "t j+1 + ::: + "t 3 + "t 2 + "t 1 + "t

For j = t, we have
2 3 t 1 t
yt = 1+ + + + ::: + + y0
t 1 3 2
+ "1 + ::: + "t 3 + "t 2 + "t + "t 1
X
t 1
2 3 t 1 t i
= 1+ + + + ::: + + y0 + "t i
i=0

t-i=1,i=t-1

X
t 1
yt = t + "t i = t + "t + "t 1 + ::: + "1
i=0

X
t 1
yt 1 = (t 1) + "t i = (t 1) + "t 1 + ::: + "1
i=1

9
1 = Cov(yt ; yt 1 )
= E((yt E(yt ))(yt 1 E(yt 1 )))
= E(("t + "t 1 + "t 2 + ::: + "1 )("t 1 + "t 2 + ::: + "1 ))
= E "2t 1 + E "2t 2 + ::: + E "21
2
= + 2 + ::: + 2
= (t 1) 2

X
t 1
yt = t + "t i = t + "t + "t 1 + ::: + "t j + ::: + "1
i=0

X
t 1
yt j = (t j) + "t i = (t j) + "t j + ::: + "1
i=j

j = Cov(yt ; yt j )
= E((yt E(yt ))(yt j E(yt j )))
= E(("t + "t 1 + ::: + "t j + ::: + "1 )("t j + ::: + "1 ))
= E "2t j + E "2t 2 + ::: + E "21
= 2 + 2 + ::: + 2

= (t j) 2

j = Corr(yt ; yt j )
Cov(yt ; yt j ) (t j) 2
= p =p
V ar (yt ) V ar (yt j ) t 2 (t j) 2
p 2
(t j) 2 t j t j
= p =p = pp
2 t (t j) t (t j) t t j
p r
t j t j
= p = :
t t

yt = + yt 1 + "t (26)
yt yt 1 = + yt 1 yt 1 + "t (27)
yt = +( 1) yt 1 + "t (28)

10
 yt = + yt 1 + "t (29)
where = 1:

11