Time Series: Assignment 1

Hang Nguyen

March 5, 2026

Question 1

Consider the difference equation

yt = a0 + a1 yt−1
with the initial condition y0 . Jill solved the difference equation by iterating backward:

yt = a0 + a1 yt−1
= a0 + a1 (a0 + a1 yt−2 )
= a0 + a0 a1 + a0 a21 + · · · + a0 at−1
1 + at1 y0 .

Bill added the homogeneous and particular solutions to obtain

 
a0 t a0
yt = + a1 y0 − .
1 − a1 1 − a1

1. Show that the two solutions are identical for |a1 | < 1.

2. Show that for a1 = 1, Jill’s solution is equivalent to

yt = a0 t + y0 .

How would you use Bill’s method to arrive at this same conclusion in the case that
a1 = 1?

Solution:

(a) Jill’s solution is

t−1
X
yt = a0 ak1 + at1 y0 .
k=0

Using the geometric series formula

t−1
X 1 − at1
ak1 =
1 − a1
k=0
we obtain
1 − at1
 
t a0 t a0
yt = a0 + a 1 y0 = + a 1 y0 −
1 − a1 1 − a1 1 − a1
which is exactly Bill’s solution.

(b) When a1 = 1,
yt = a0 + yt−1 .
Iterating gives
yt = y0 + a0 t.
Using Bill’s method:

Homogeneous equation
(h) (h) (h)
yt = yt−1 ⇒ yt = C.
(p)
Particular solution (try yt = kt):

kt = a0 + k(t − 1)

k = a0 .
Thus
yt = C + a0 t.
Using y0 :
C = y0 .
Therefore
yt = y0 + a0 t.

Question 2

The unit root problem in time-series econometrics is concerned with characteristic roots that
are equal to unity. In order to preview the issue:

1. Find the homogeneous solution to each of the following: (Hint: Each has at least one
unit root.)
(i) yt = 1.5yt−1 − 0.5yt−2 + εt
(ii) yt = yt−2 + εt
(iii) yt = 2yt−1 − yt−2 + εt
(iv) yt = yt−1 + 0.25yt−2 − 0.25yt−3 + εt
2. Show that each of the backward-looking solutions is not convergent.
3. Show that Equation (i) can be written entirely in first differences; that is,

∆yt = 0.5∆yt−1 + εt .

Find the particular solution for ∆yt .

2
 4. Similarly transform the other equations into their first-difference form. Find the par-
ticular solution, if it exists, for the transformed equations.

5. Write Equations (i) through (iv) using lag operators.

6. Given an initial condition y0 , find the solution for

yt = a0 − yt−1 + εt .

Solution:

The unit root problem in time-series econometrics is concerned with characteristic roots equal
to unity.

(a) Homogeneous solutions

To find the homogeneous solution, set εt = 0. (i)

yt = 1.5yt−1 − 0.5yt−2

Characteristic equation:
r2 − 1.5r + 0.5 = 0
Factorizing:
(r − 1)(r − 0.5) = 0
Thus the roots are r1 = 1 and r2 = 0.5. Therefore the homogeneous solution is
(h)
yt = C1 + C2 (0.5)t

(ii)
yt = yt−2
Characteristic equation:
r2 − 1 = 0
(r − 1)(r + 1) = 0
Roots:
r = 1, r = −1
Hence
(h)
yt = C1 + C2 (−1)t
(iii)
yt = 2yt−1 − yt−2
Characteristic equation:
r2 − 2r + 1 = 0
(r − 1)2 = 0
This is a repeated root r = 1. Thus
(h)
yt = C1 + C2 t

3
(iv)
yt = yt−1 + 0.25yt−2 − 0.25yt−3
Characteristic equation:
r3 − r2 − 0.25r + 0.25 = 0
Testing r = 1:
1 − 1 − 0.25 + 0.25 = 0
So (r − 1) is a factor:
(r − 1)(r2 − 0.25) = 0
r = 1, r = 0.5, r = −0.5
Thus
(h)
yt = C1 + C2 (0.5)t + C3 (−0.5)t

(b) Backward-looking solutions are not convergent. A backward-looking solution writes the
process as
X∞
yt = ψj εt−j
j=0

This representation requires the autoregressive polynomial to be invertible and the coefficients
ψj to converge. However, each of the above models has at least one unit root (r = 1). When
a root lies on the unit circle, the homogeneous component of the solution does not decay over
time. For example:
(h)
yt = C1
or
(h)
yt = C1 + C2 t

which do not converge as t → ∞. Therefore, the infinite backward representation does not
converge.

(c) First difference form of equation (i)

Start from
yt = 1.5yt−1 − 0.5yt−2 + εt
Subtract yt−1 from both sides:

yt − yt−1 = 0.5yt−1 − 0.5yt−2 + εt

∆yt = 0.5(yt−1 − yt−2 ) + εt

∆yt = 0.5∆yt−1 + εt

This is an AR(1) process for ∆yt . Since |0.5| < 1, it has the convergent particular solution
∞
X
∆yt = (0.5)j εt−j
j=0

4
(d) First difference forms of the other equations

(ii)
yt − yt−2 = εt
Using the lag operator:
(1 − L2 )yt = εt
(1 − L)(1 + L)yt = εt
Since (1 − L)yt = ∆yt :
(1 + L)∆yt = εt
∆yt + ∆yt−1 = εt
This process has coefficient −1 and therefore lies on the boundary of stability, so no stationary
particular solution exists.

(iii)
yt − 2yt−1 + yt−2 = εt
(1 − L)2 yt = εt
(1 − L)∆yt = εt
∆yt − ∆yt−1 = εt
This implies
∆yt = ∆yt−1 + εt
which is a random walk and therefore nonstationary.

(iv)
yt − yt−1 − 0.25yt−2 + 0.25yt−3 = εt
(1 − L − 0.25L2 + 0.25L3 )yt = εt
Factorizing:
(1 − L)(1 − 0.25L2 )yt = εt
Thus
(1 − 0.25L2 )∆yt = εt
∆yt = 0.25∆yt−2 + εt
Iterating backward gives the particular solution
∞
X
∆yt = (0.25)j εt−2j
j=0

(e) Lag operator representation

(i)
(1 − 1.5L + 0.5L2 )yt = εt

5
(ii)
(1 − L2 )yt = εt
(iii)
(1 − 2L + L2 )yt = εt
(iv)
(1 − L − 0.25L2 + 0.25L3 )yt = εt
(f) Solution for yt = a0 − yt−1 + εt

Rewrite the equation:

yt + yt−1 = a0 + εt
The homogeneous solution is
(h)
yt = C(−1)t
(p)
Assume a constant particular solution yt = µ:

µ = a0 − µ

2µ = a0
a0
µ=
2
Thus the general solution is
t−1
a0 X
yt = + C(−1)t + (−1)j εt−j
2
j=0

Using the initial condition y0 :

a0
C = y0 −
2
Final solution:
t−1
a0  a0  X
yt = + (−1)t y0 − + (−1)j εt−j
2 2
j=0

Question 3

Consider the stochastic process

yt = a0 + a2 yt−2 + εt .

1. Find the homogeneous solution and determine the stability condition.

2. Find the particular solution using the method of undetermined coefficients.

3. Find the particular solution using lag operators assuming |a2 | < 1.

6
Solution:

(a) Homogeneous solution and stability Set εt = 0 and ignore the constant term a0 to obtain
the homogeneous equation
yt = a2 yt−2 .
Seek a solution of the form yt = rt . Substitution gives

rt = a2 rt−2 =⇒ r2 = a2 .

Hence the characteristic roots are

√ √
r1 = a2 , r2 = − a2 ,

and the homogeneous solution is

(h) √ √
yt = C1 ( a2 )t + C2 (− a2 )t .

The process is stable (in the usual AR sense) if all characteristic roots lie inside the unit
circle:
√
|r1 | < 1, |r2 | < 1 ⇐⇒ | a2 | < 1 ⇐⇒ |a2 | < 1.

(b) Particular solution by undetermined coefficients Because the nonhomogeneous term is a

(p)
constant, propose a constant particular solution yt = µ. Substituting yt = µ and yt−2 = µ
into the model (and using E(εt ) = 0) yields

µ = a0 + a2 µ.

Solving,
a0
(1 − a2 )µ = a0 =⇒ µ= .
1 − a2
Thus
(p) a0
yt = .
1 − a2

(c) Particular (stationary) solution using lag operators, assuming |a2 | < 1. Using the lag
operator L defined by Lyt = yt−1 , we have yt−2 = L2 yt . The model becomes

yt = a0 + a2 L2 yt + εt ,

or equivalently
(1 − a2 L2 )yt = a0 + εt .
When |a2 | < 1, the inverse admits the geometric expansion

(1 − a2 L2 )−1 = 1 + a2 L2 + a22 L4 + a32 L6 + · · · .

Therefore
yt = (1 − a2 L2 )−1 (a0 + εt ) = (1 + a2 L2 + a22 L4 + · · · )(a0 + εt ).

7
Applying this to the constant term gives
a0
(1 + a2 + a22 + · · · )a0 = ,
1 − a2
and applying it to the shock term gives
∞
aj2 εt−2j .
X
εt + a2 εt−2 + a22 εt−4 + ··· =
j=0

Hence the stationary solution is

∞
a0 X j
yt = + a2 εt−2j , |a2 | < 1.
1 − a2
j=0

Question 4

Part 1: For each of the following, determine whether {yt } represents a stable process. De-
termine whether the characteristic roots are real or imaginary.

(a) yt − 1.2yt−1 + 0.2yt−2 = 0

(b) yt − 1.2yt−1 − 1.2yt−2 = 0
(c) yt − 0.7yt−1 − 0.25yt−2 + 0.175yt−3 = 0

Part 2: Write each of the above equations using lag operators. Determine the characteristic
roots of the inverse characteristic equation.

Solution:

Consider the stochastic process

yt = a0 + a2 yt−2 + εt ,

where {εt } is white noise with E(εt ) = 0 and Var(εt ) = σε2 .

(a) Homogeneous solution and stability.

Set εt = 0 and ignore the constant term a0 to obtain the homogeneous equation

yt = a2 yt−2 .

Seek a solution of the form yt = rt . Substitution gives

rt = a2 rt−2 =⇒ r2 = a2 .

Hence the characteristic roots are

√ √
r1 = a2 , r2 = − a2 ,

8
and the homogeneous solution is
(h) √ √
yt = C1 ( a2 )t + C2 (− a2 )t .
The process is stable if all characteristic roots lie inside the unit circle:
√
|r1 | < 1, |r2 | < 1 ⇐⇒ | a2 | < 1 ⇐⇒ |a2 | < 1.
(b) Particular solution by undetermined coefficients.

Because the nonhomogeneous term is a constant, propose a constant particular solution

(p)
yt = µ. Substituting yt = µ and yt−2 = µ into the model (and using E(εt ) = 0) yields
µ = a0 + a2 µ.
Solving,
a0
(1 − a2 )µ = a0 =⇒ µ= .
1 − a2
Thus
a0
(p)
yt .=
1 − a2
(c) Particular (stationary) solution using lag operators, assuming |a2 | < 1.

Using the lag operator L defined by Lyt = yt−1 , we have yt−2 = L2 yt . The model becomes
yt = a0 + a2 L2 yt + εt ,
or equivalently
(1 − a2 L2 )yt = a0 + εt .

When |a2 | < 1, the inverse admits the geometric expansion

(1 − a2 L2 )−1 = 1 + a2 L2 + a22 L4 + a32 L6 + · · · .
Therefore
yt = (1 − a2 L2 )−1 (a0 + εt ) = (1 + a2 L2 + a22 L4 + · · · )(a0 + εt ).

Applying this to the constant term gives

a0
(1 + a2 + a22 + · · · )a0 = ,
1 − a2
and applying it to the shock term gives
∞
aj2 εt−2j .
X
εt + a2 εt−2 + a22 εt−4 + · · · =
j=0

Hence the stationary solution is

∞
a0 X j
yt = + a2 εt−2j , |a2 | < 1.
1 − a2
j=0

9
Question 5

Consider the stochastic difference equation:

yt = 0.8yt−1 + εt − 0.5εt−1 .

1. Suppose that the initial conditions are such that: y0 = 0 and ε0 = ε−1 = 0. Now
suppose that ε1 = 1. Determine the values y1 through y5 by forward iteration assuming
that all future values of {εt } = 0.

2. Find the homogeneous and particular solutions.

3. Trace out the time path of an εt shock on the entire time path of the {yt } sequence.

Solution:

(a) Forward iteration given the specified initial conditions Assume the initial conditions are

y0 = 0, ε0 = ε−1 = 0,

and suppose that

ε1 = 1, εt = 0 for all t ≥ 2.
Compute yt sequentially. For t = 1:

y1 = 0.8y0 + ε1 − 0.5ε0 = 0.8(0) + 1 − 0 = 1.

For t = 2:
y2 = 0.8y1 + ε2 − 0.5ε1 = 0.8(1) + 0 − 0.5(1) = 0.3.
For t = 3:
y3 = 0.8y2 + ε3 − 0.5ε2 = 0.8(0.3) + 0 − 0 = 0.24.
For t = 4:
y4 = 0.8y3 + ε4 − 0.5ε3 = 0.8(0.24) = 0.192.
For t = 5:
y5 = 0.8y4 + ε5 − 0.5ε4 = 0.8(0.192) = 0.1536.
Hence
y1 = 1, y2 = 0.3, y3 = 0.24, y4 = 0.192, y5 = 0.1536.

(b) Write the model in lag-operator form. Using Lyt = yt−1 ,

yt − 0.8yt−1 = εt − 0.5εt−1 ⇐⇒ (1 − 0.8L)yt = (1 − 0.5L)εt .

The homogeneous equation sets the forcing term to zero:

(h) (h) (h)
(1 − 0.8L)yt =0 ⇐⇒ yt = 0.8yt−1 .

Therefore
(h)
yt = C(0.8)t .

10
Particular (shock-driven) solution.

Assuming |0.8| < 1, the inverse operator exists:

(1 − 0.8L)−1 = 1 + 0.8L + 0.82 L2 + · · · .
Hence
yt = (1 − 0.8L)−1 (1 − 0.5L)εt .
Expanding,
yt = 1 + 0.8L + 0.82 L2 + · · · (1 − 0.5L) εt .

Compute the first few MA coefficients {ψj } in

∞
X
yt = ψj εt−j .
j=0

We have
ψ0 = 1,
and the coefficient on L is 0.8 − 0.5 = 0.3, so
ψ1 = 0.3.
For j ≥ 2, multiplying by (1 − 0.8L)−1 implies the recursion
ψj = 0.8ψj−1 , j ≥ 2,
so
ψj = 0.3(0.8)j−1 , j ≥ 1.
Thus the particular (causal) solution is
∞
X
yt = εt + 0.3εt−1 + 0.24εt−2 + 0.192εt−3 + · · · = εt + 0.3(0.8)j−1 εt−j .
j=1

(c) Consider a one-unit shock at time t:

εt = 1, εt+k = 0 for k ̸= 0,
and assume no other shocks.

From the MA(∞) representation

∞
X
yt+h = ψj εt+h−j ,
j=0

only the term with t + h − j = t survives, i.e. j = h. Hence the response at horizon h is
yt+h = ψh .
Therefore the impulse response function is
ψ0 = 1, ψ1 = 0.3, ψh = 0.3(0.8)h−1 for h ≥ 1.
Equivalently, the shock produces the path
yt = 1, yt+1 = 0.3, yt+2 = 0.24, yt+3 = 0.192, . . .
which decays geometrically because |0.8| < 1.

11
Question 6

Consider the second-order autoregressive process

yt = a0 + a2 yt−2 + εt , where |a2 | < 1.

Find

1. Et−2 yt , Et−1 yt , and Et yt+2 .

2. cov(yt , yt−1 ) and cov(yt , yt−2 ).

Solution:

(a) Conditional expectations Et−2 (yt ), Et−1 (yt ), and Et (yt+2 ).

Let Ft denote the information set through time t, and write

Et (·) ≡ E(· | Ft ).

1. Compute Et−2 (yt ). Start from the model:

yt = a0 + a2 yt−2 + εt .

Taking conditional expectation given Ft−2 yields

Et−2 (yt ) = a0 + a2 Et−2 (yt−2 ) + Et−2 (εt ).

Since yt−2 is Ft−2 -measurable, Et−2 (yt−2 ) = yt−2 .

Also, εt is a one-step-ahead innovation (indeed two steps ahead relative to t − 2), so

Et−2 (εt ) = 0.

Therefore,
Et−2 (yt ) = a0 + a2 yt−2 .
2. Compute Et−1 (yt ).

Again,
yt = a0 + a2 yt−2 + εt .
Taking conditional expectation given Ft−1 ,

Et−1 (yt ) = a0 + a2 Et−1 (yt−2 ) + Et−1 (εt ).

Because yt−2 is already known at time t − 1, Et−1 (yt−2 ) = yt−2 , and since εt is the innovation
at time t,
Et−1 (εt ) = 0.

12
Hence,
Et−1 (yt ) = a0 + a2 yt−2 .

3. Compute Et (yt+2 ). Use the model at time t + 2:

yt+2 = a0 + a2 yt + εt+2 .

Taking Et (·) gives

Et (yt+2 ) = a0 + a2 Et (yt ) + Et (εt+2 ).
Because yt is known at time t, Et (yt ) = yt , and since εt+2 is a future innovation,

Et (εt+2 ) = 0.

Therefore,
Et (yt+2 ) = a0 + a2 yt .

(b) Because |a2 | < 1, the process is stable and admits a causal MA(∞) representation.

Iterating the model backward in steps of 2 gives

yt = a0 + a2 yt−2 + εt = a0 + a2 (a0 + a2 yt−4 + εt−2 ) + εt = a0 (1 + a2 ) + a22 yt−4 + εt + a2 εt−2 ,

and continuing indefinitely (using |a2 | < 1 so the powers aj2 decay) yields
∞
a0
aj2 εt−2j ,
X
yt = µ + µ= .
1 − a2
j=0

Similarly,
∞ ∞
aj2 εt−1−2j , aj2 εt−2−2j .
X X
yt−1 = µ + yt−2 = µ +
j=0 j=0

Define the mean-adjusted variables ỹt = yt − µ. Then

∞
aj2 εt−2j .
X
ỹt =
j=0

1. Compute cov(yt , yt−1 ). Because µ is constant,

cov(yt , yt−1 ) = cov(ỹt , ỹt−1 ).

Now note the index sets of shocks entering each series:

ỹt depends on {εt , εt−2 , εt−4 , . . .},

ỹt−1 depends on {εt−1 , εt−3 , εt−5 , . . .}.

These two sets are disjoint (even vs. odd time indices), and white noise shocks are uncorrelated
across time. Therefore every cross-product has zero expectation, implying

cov(yt , yt−1 ) = 0.

13
2. Compute cov(yt , yt−2 ).

Again,
cov(yt , yt−2 ) = cov(ỹt , ỹt−2 ).
Write the MA expansions:
∞ ∞
aj2 εt−2j ,
X X
ỹt = ỹt−2 = ak2 εt−2−2k .
j=0 k=0

Then
∞ X
∞
aj+k
X
cov(ỹt , ỹt−2 ) = 2 cov(εt−2j , εt−2−2k ) .
j=0 k=0

Because {εt } is white noise,

(
σε2 , m = n,
cov(εm , εn ) =
0, m ̸= n.

Thus the only nonzero terms occur when the indices match:

t − 2j = t − 2 − 2k ⇐⇒ 2j = 2 + 2k ⇐⇒ j = k + 1.

So the double sum collapses to

∞ ∞
(k+1)+k 2
X X
cov(ỹt , ỹt−2 ) = a2 σε = σε2 a2k+1
2 .
k=0 k=0

This is a geometric series in a22 :

∞ ∞
X X a2
a2k+1
2 = a2 (a22 )k = , |a2 | < 1.
1 − a22
k=0 k=0

Therefore,
a2
cov(yt , yt−2 ) = σ2.
1 − a22 ε
Final answers

Et−2 (yt ) = a0 + a2 yt−2 , Et−1 (yt ) = a0 + a2 yt−2 , Et (yt+2 ) = a0 + a2 yt ,

a2
cov(yt , yt−1 ) = 0, cov(yt , yt−2 ) = σ2.
1 − a22 ε

Question 7

Consider an MA(3) process

xt = εt + β1 εt−1 + β2 εt−2 + β3 εt−3 .

14
Prove that this process is stationary.

Solution:

A process {xt } is covariance stationary (weakly stationary) if:

1. E(xt ) is constant in t;
2. Var(xt ) is finite and constant in t;
3. the autocovariance function
γ(k) ≡ Cov(xt , xt−k )
depends only on the lag k, not on time t.

We verify each property directly.

1. Constant mean

Using linearity of expectation and E(εt ) = 0,

E(xt ) = E(εt ) + β1 E(εt−1 ) + β2 E(εt−2 ) + β3 E(εt−3 ) = 0.

Hence the mean is constant:

E(xt ) = 0 for all t.
2. Finite, time-invariant variance

Because εt is white noise, different-time shocks are uncorrelated. Therefore the variance
of a finite linear combination is the sum of the variances of each component:

Var(xt ) = Var(εt + β1 εt−1 + β2 εt−2 + β3 εt−3 ) .

Expand using Var(aZ) = a2 Var(Z) and Cov(εt−i , εt−j ) = 0 for i ̸= j:

Var(xt ) = Var(εt ) + β12 Var(εt−1 ) + β22 Var(εt−2 ) + β32 Var(εt−3 )

+ 2β1 Cov(εt , εt−1 ) + 2β2 Cov(εt , εt−2 ) + 2β3 Cov(εt , εt−3 )
+ 2β1 β2 Cov(εt−1 , εt−2 ) + 2β1 β3 Cov(εt−1 , εt−3 ) + 2β2 β3 Cov(εt−2 , εt−3 ).

All covariance terms are zero (different time indices), and each variance term equals σε2 , so

Var(xt ) = σε2 1 + β12 + β22 + β32 ,

which is finite and does not depend on t. Thus

Var(xt ) = σε2 1 + β12 + β22 + β32 .

3. Autocovariances depend only on lag

Because E(xt ) = 0, for any integer k,

γ(k) = Cov(xt , xt−k ) = E(xt xt−k ).

15
Write
3
X 3
X
xt = βi εt−i , xt−k = βj εt−k−j ,
i=0 j=0

where β0 ≡ 1.

Then ! 3 
3
X X 3 X
X 3
xt xt−k = βi εt−i  βj εt−k−j =
 βi βj εt−i εt−k−j .
i=0 j=0 i=0 j=0

Taking expectations gives

3 X
X 3
γ(k) = βi βj E(εt−i εt−k−j ) .
i=0 j=0

By the white-noise property, (

σε2 , m = n,
E(εm εn ) =
0, m ̸= n.
Hence E(εt−i εt−k−j ) is nonzero only when

t−i=t−k−j ⇐⇒ k = i − j.

Therefore X
γ(k) = σε2 βi βj ,
i,j∈{0,1,2,3}
i−j=k

which depends only on k (not on t).

Moreover, because i and j can only be 0, 1, 2, 3, the difference i−j must lie in {−3, −2, −1, 0, 1, 2, 3}.
Thus, for |k| > 3 there are no pairs (i, j) satisfying i − j = k, so

γ(k) = 0 for |k| > 3.

It is often useful to write the autocovariances explicitly for k ≥ 0 (and use symmetry
γ(−k) = γ(k)).

Matching common shocks gives:

γ(0) = σε2 (1 + β12 + β22 + β32 ),

γ(1) = σε2 (β1 + β1 β2 + β2 β3 ),
γ(2) = σε2 (β2 + β1 β3 ),
γ(3) = σε2 β3 ,
γ(k) = 0, k ≥ 4.

Each γ(k) depends only on the lag k, confirming stationarity.

16
Conclusion:

We have shown that E(xt ) is constant, Var(xt ) is finite and constant, and γ(k) = Cov(xt , xt−k )
depends only on k (indeed γ(k) = 0 for |k| > 3). Therefore the MA(3) process is covariance
stationary.

The MA(3) process is stationary (weakly/covariance stationary).

Question 8

Given an initial condition for y0 , find the solution for yt .

(a) yt = 1.1yt−1 + εt

(b) yt = yt−1 + 1 + εt

(c) yt = yt−1 + εt + 0.5εt−1

Solution:

(a) yt = 1.1yt−1 + εt

Rewrite the equation as a first-order linear difference equation. Iterating forward:

y1 = 1.1y0 + ε1

y2 = 1.1y1 + ε2 = 1.1(1.1y0 + ε1 ) + ε2 = 1.12 y0 + 1.1ε1 + ε2

y3 = 1.1y2 + ε3 = 1.13 y0 + 1.12 ε1 + 1.1ε2 + ε3
Continuing by induction, for any integer t ≥ 1:

t−1
X
yt = 1.1t y0 + 1.1j εt−j .
j=0

Equivalently (re-indexing with k = t − j):

t
X
t
yt = 1.1 y0 + 1.1t−k εk .
k=1

(b) yt = yt−1 + 1 + εt

Iterate forward:
y1 = y0 + 1 + ε1
y2 = y1 + 1 + ε2 = y0 + 2 + (ε1 + ε2 )
y3 = y2 + 1 + ε3 = y0 + 3 + (ε1 + ε2 + ε3 )

17
Thus, in general,
t
X
yt = y0 + t + εs .
s=1

(c) yt = yt−1 + εt + 0.5εt−1

This is a difference equation with a moving-average term in the innovation. Iterate:

yt − yt−1 = εt + 0.5εt−1 .

Summing both sides from s = 1 to t,

t
X t
X t
X
(ys − ys−1 ) = εs + 0.5 εs−1 .
s=1 s=1 s=1

The left-hand side telescopes:

t
X
(ys − ys−1 ) = yt − y0 .
s=1

Therefore,
t
X t
X
yt − y 0 = εs + 0.5 εs−1 .
s=1 s=1
Re-index the second sum by letting u = s − 1:
t
X t−1
X
εs−1 = εu .
s=1 u=0

Hence the solution is:

t
X t−1
X
yt = y0 + εs + 0.5 εu .
s=1 u=0

It is also common to combine the overlapping sums (noting that ε1 , . . . , εt−1 appear twice):
t−1
X
yt = y0 + 0.5ε0 + 1.5 εs + ε t , (t ≥ 1).
s=1

Final answers
t−1
X
(a) yt = 1.1t y0 + 1.1j εt−j
j=0

t
X
(b) yt = y0 + t + εs
s=1

t
X t−1
X
(c) yt = y0 + εs + 0.5 εu
s=1 u=0

18
Question 9

Given the initial condition y0 , find the solution and the forecast function for the following
variant of the trend plus irregular model:

yt = µt + νt ,

where
µt = µt−1 + εt
and
νt = (1 + β1 L)ηt .
Solution:

1. Expand the model explicitly

First write νt in time-domain form. Since (1 + β1 L)ηt = ηt + β1 ηt−1 ,

νt = ηt + β1 ηt−1 .

Therefore, the observation equation becomes

yt = µt + ηt + β1 ηt−1 .

2. Solve for the trend component µt

The trend evolves as a random walk:

µt = µt−1 + εt .

Iterating forward,
µ1 = µ0 + ε1 ,
µ2 = µ1 + ε2 = µ0 + ε1 + ε2 ,
and by induction,
t
X
µt = µ0 + εs .
s=1

3. Obtain the solution for yt

Substitute the expression for µt into yt = µt + νt :

t
!
X
yt = µ0 + εs + ηt + β1 ηt−1 .
s=1

Hence the solution is

t
X
yt = µ0 + εs + ηt + β1 ηt−1 .
s=1

19
If the initial condition is given as y0 rather than µ0 , note that

y0 = µ0 + ν0 = µ0 + η0 + β1 η−1 ,

so without additional information about η0 and η−1 , y0 does not uniquely pin down µ0 . In
many treatments, the initial state is taken as µ0 = y0 (or ν0 = 0 in expectation), but the
exact mapping depends on what is assumed known at time 0.

4. Forecast function

Let Ft denote the information set through time t and write

Et (·) = E(· | Ft ).

A standard assumption in state-space forecasting is that future innovations are mean-zero

and unpredictable:
Et (εt+h ) = 0, Et (ηt+h ) = 0 (h ≥ 1).

Ph
4.1. Forecast the trend µt+h . From µt+h = µt + j=1 εt+j ,

h
X
Et (µt+h ) = µt + Et (εt+j ) = µt .
j=1

Thus the h-step-ahead forecast of the random-walk trend is

µ̂t+h|t = Et (µt+h ) = µt , h ≥ 1.

4.2. Forecast the irregular component νt+h . Recall

νt+h = ηt+h + β1 ηt+h−1 .

• For h ≥ 2, both ηt+h and ηt+h−1 are future shocks, so

Et (νt+h ) = 0, h ≥ 2.

• For h = 1,
νt+1 = ηt+1 + β1 ηt .
Here ηt+1 is future (mean zero), but ηt is time-t noise. In a typical one-step-ahead
forecasting setup, ηt is not directly observed; however, after observing yt , one can form
an estimate of ηt from the current forecast error (innovation). Denote the time-t estimate
by η̂t|t . Then
Et (νt+1 ) = Et (ηt+1 ) + β1 Et (ηt ) = 0 + β1 η̂t|t .
If instead one uses the simplifying assumption that ηt is not forecastable even at time
t (i.e. Et (ηt ) = 0), then Et (νt+1 ) = 0 as well.

20
4.3. Forecast yt+h . Since yt+h = µt+h + νt+h ,

Et (yt+h ) = Et (µt+h ) + Et (νt+h ).

For h ≥ 2:
Et (yt+h ) = µt + 0 = µt .
Thus,
ŷt+h|t = Et (yt+h ) = µt , h ≥ 2.

For h = 1:
Et (yt+1 ) = Et (µt+1 ) + Et (νt+1 ) = µt + β1 η̂t|t .
So the one-step-ahead forecast function can be written as

ŷt+1|t = µt + β1 η̂t|t .

Under the common simplification Et (ηt ) = 0 (so η̂t|t = 0), this reduces to

ŷt+1|t = µt .

Summary

t
X t
X
µt = µ0 + εs , νt = ηt + β1 ηt−1 , yt = µ0 + εs + ηt + β1 ηt−1 .
s=1 s=1

Forecasts:
µ̂t+h|t = µt (h ≥ 1),

ŷt+h|t = µt (h ≥ 2),

ŷt+1|t = µt + β1 η̂t|t (general), ŷt+1|t = µt if Et (ηt ) = 0.

Question 10

Consider the process

yt = yt−1 + εt .
Given an initial condition for y0 (y0 is deterministic):

(a) Derive Var(yt ) and Cov(yt , yt−s ).

(b) Show that the autocorrelation ρs → 1 as t → ∞.

(c) Show that yt is nonstationary.

21
Solution:

Iterate the recursion forward:

y1 = y0 + ε1 , y2 = y1 + ε2 = y0 + ε1 + ε2 , ...

By induction,
t
X
yt = y0 + εi .
i=1

Because y0 is deterministic, all randomness in yt comes from the partial sum of shocks.

(a) Derive Var(yt ) and Cov(yt , yt−s )

Variance. Using the solution above and the fact that adding a constant does not change
variance,
t
!
X
Var(yt ) = Var εi .
i=1
Since the εi are uncorrelated across time,
t t t
!
X X X
Var εi = Var(εi ) = σε2 = tσε2 .
i=1 i=1 i=1

Thus,
Var(yt ) = tσε2 .

Covariance. For an integer s ≥ 0 with s ≤ t, write

t−s
X
yt−s = y0 + εj .
j=1

Again constants do not affect covariance, so

 
Xt t−s
X
Cov(yt , yt−s ) = Cov εi , εj  .
i=1 j=1

Use bilinearity:  
t
X t−s
X t X
X t−s
Cov εi , εj  = Cov(εi , εj ).
i=1 j=1 i=1 j=1

Because the shocks are white noise,

(
σε2 , i = j,
Cov(εi , εj ) =
0, i ̸= j.

22
So only the diagonal terms with i = j survive. The common indices between the two sums
are
i = j = 1, 2, . . . , t − s,
which gives exactly (t − s) nonzero terms. Hence
t−s
X
Cov(yt , yt−s ) = σε2 = (t − s)σε2 .
i=1

Therefore,
Cov(yt , yt−s ) = (t − s)σε2 , 0 ≤ s ≤ t.

(If s > t, then yt−s is not defined for nonnegative time indexing; under the usual convention
we focus on 0 ≤ s ≤ t.)

(b) Show that the autocorrelation ρs → 1 as t → ∞

The autocorrelation at lag s (for fixed s and large t) is

Cov(yt , yt−s )
ρs (t) = p .
Var(yt ) Var(yt−s )

Using part (a),

Cov(yt , yt−s ) = (t − s)σε2 , Var(yt ) = tσε2 , Var(yt−s ) = (t − s)σε2 .

Substitute: r
(t − s)σε2 (t − s)σε2
r
t−s s
ρs (t) = p = p = = 1− .
tσε2 (t − s)σε2 σε2 t(t − s) t t
As t → ∞ (holding s fixed),
r
s s
→0 =⇒ ρs (t) = 1 − → 1.
t t
Thus,
ρs (t) → 1 as t → ∞ for each fixed lag s.

(c) A covariance-stationary process must have a time-invariant finite variance and an auto-
covariance function that depends only on the lag, not on t. Here,

Var(yt ) = tσε2 ,

which depends on t and grows without bound as t → ∞. Therefore the variance is not con-
stant over time, violating a basic requirement for covariance stationarity.

Equivalently, the autocovariance depends on calendar time:

Cov(yt , yt−s ) = (t − s)σε2 ,

23
which depends on t as well as on s.

Hence {yt } is not covariance stationary. In words, random walks have shocks that accu-
mulate permanently, so uncertainty increases with time.

{yt } is nonstationary.

Summary of results

t
X
yt = y0 + εi , Var(yt ) = tσε2 , Cov(yt , yt−s ) = (t − s)σε2 ,
i=1

r
s
ρs (t) = 1− → 1 as t → ∞, {yt } is nonstationary.
t

Question 11

A series follows
yt = 1 + yt−1 + 0.5εt−1 + εt .

(a) Express the model in first differences.

(b) Derive E(∆yt ) and Var(∆yt ).

(c) Is ∆yt stationary?

Solution:

(a) Recall the first-difference operator:

∆yt ≡ yt − yt−1 .

Subtract yt−1 from both sides of the original equation:

yt − yt−1 = 1 + 0.5εt−1 + εt .

Therefore the first-difference form is

∆yt = 1 + εt + 0.5εt−1 .

(b) Derive E(∆yt ) and Var(∆yt )

Using part (a),

∆yt = 1 + εt + 0.5εt−1 .

24
Mean. By linearity of expectation and E(εt ) = 0,
E(∆yt ) = E(1) + E(εt ) + 0.5E(εt−1 ) = 1 + 0 + 0 = 1.
Hence
E(∆yt ) = 1.

Variance. Since adding a constant does not change variance,

Var(∆yt ) = Var(εt + 0.5εt−1 ).
Expand the variance:
Var(εt + 0.5εt−1 ) = Var(εt ) + 0.52 Var(εt−1 ) + 2(0.5) Cov(εt , εt−1 ).
Because {εt } is white noise, Cov(εt , εt−1 ) = 0 and Var(εt ) = Var(εt−1 ) = σε2 . Thus,
Var(∆yt ) = σε2 + 0.25σε2 = 1.25σε2 .
Therefore
Var(∆yt ) = 1.25 σε2 .

(c) From (a),

∆yt = 1 + εt + 0.5εt−1 .
This is a constant plus a finite moving-average process of order 1 (MA(1)) in white-noise
shocks.

To verify covariance stationarity:

• The mean is constant: E(∆yt ) = 1 for all t.

• The variance is finite and constant: Var(∆yt ) = 1.25σε2 .
• The autocovariance depends only on the lag. Indeed, because an MA(1) uses only εt
and εt−1 , the autocovariance is zero for |k| > 1.

Therefore {∆yt } is covariance stationary.

∆yt is stationary (weakly/covariance stationary).

(Optional) Autocovariances of ∆yt

Let ut ≡ ∆yt − 1 = εt + 0.5εt−1 . Then ut is MA(1). Since E(ut ) = 0,

γu (k) = Cov(ut , ut−k ).
Compute:
γu (0) = Var(ut ) = σε2 (1 + 0.52 ) = 1.25σε2 ,
γu (1) = Cov(εt + 0.5εt−1 , εt−1 + 0.5εt−2 ) = 0.5 Var(εt−1 ) = 0.5σε2 ,

and γu (k) = 0 for |k| > 1. This further confirms stationarity.

25
Question 12

Consider the process

yt = y0 + 2t + 0.5t2 + εt .

(a) Is this process trend-stationary or difference-stationary?

(b) Suggest a method to remove the trend.
(c) What is the mean of the detrended series?

Solution:

(a) A process is trend-stationary if it can be decomposed as

yt = m(t) + ut ,
where m(t) is a deterministic function of time and ut is a stationary process. In contrast,
a process is difference-stationary if differencing is required to eliminate a stochastic trend
(typically a unit root), i.e. the nonstationarity comes from a random-walk-type component.

Here we can write

y0 + 2t + 0.5t2 +


yt = εt .
| {z } |{z}
deterministic trend m(t) stationary component

The nonstationarity in yt comes entirely from the deterministic polynomial trend m(t), while
the deviation from the trend is εt , which is stationary by assumption.

Therefore,
The process is trend-stationary (not difference-stationary).

(Equivalently, if one subtracts the deterministic trend, the remainder is stationary.)

(b) Because the trend is a quadratic polynomial in t, a natural way to remove it is to fit
and subtract a quadratic time trend by regression. Specifically, run the OLS regression
yt = α + β1 t + β2 t2 + ut .
If the data are generated exactly as given, the population values are
α = y0 , β1 = 2, β2 = 0.5, ut = εt .
Define the detrended series as the residual
 
ût ≡ yt − α̂ + β̂1 t + β̂2 t2 .

Then ût is an estimate of εt and, under correct specification, behaves as a stationary series.

Thus a valid detrending method is:

Regress yt on 1, t, t2 and use the residuals as the detrended series.

26
(If the coefficients y0 , 2, 0.5 are known, detrending can be done directly by subtraction:
ỹt = yt − (y0 + 2t + 0.5t2 ) = εt .)

(c) Let the detrended series be

ỹt ≡ yt − y0 + 2t + 0.5t2 .

Using the model definition,

ỹt = εt .
Therefore,
E(ỹt ) = E(εt ) = 0.
Hence,
The mean of the detrended series is 0.
Note: If detrending is performed by regression (unknown coefficients), then the residuals
satisfy the OLS normal equations and have sample mean zero (when an intercept is included),
and their population mean is also zero when the trend is correctly specified. In either case,
the intended answer is:
E(detrended series) = 0.

Question 13

Consider the third-order autoregressive process

yt = a0 + a1 yt−1 + a2 yt−2 + a3 yt−3 + bt + εt .

Rewrite the following AR(3) model in the Augmented Dickey–Fuller (ADF) form, expressing
the equation in terms of the first difference ∆yt , the lagged level yt−1 , and the lagged differ-
ences ∆yt−1 and ∆yt−2 .

Solution:

We want to rewrite it in the Augmented Dickey–Fuller (ADF) form, i.e. in terms of

∆yt , yt−1 , ∆yt−1 , ∆yt−2 ,

and deterministic terms.

1. Start by converting to first differences Subtract yt−1 from both sides:

yt − yt−1 = a0 + a1 yt−1 + a2 yt−2 + a3 yt−3 + bt + εt − yt−1 .

Since ∆yt ≡ yt − yt−1 , this becomes

∆yt = a0 + (a1 − 1)yt−1 + a2 yt−2 + a3 yt−3 + bt + εt .

At this point the right-hand side still contains yt−2 and yt−3 , but the ADF form replaces these
with yt−1 and lagged differences.

27
2. Express yt−2 and yt−3 using yt−1 and differences Use the identities

∆yt−1 = yt−1 − yt−2 =⇒ yt−2 = yt−1 − ∆yt−1 ,

and
∆yt−2 = yt−2 − yt−3 =⇒ yt−3 = yt−2 − ∆yt−2 .
Substitute the first identity into the second to express yt−3 entirely in terms of yt−1 and
differences:

yt−3 = yt−1 − ∆yt−1 − ∆yt−2 = yt−1 − ∆yt−1 − ∆yt−2 .
Therefore,
yt−2 = yt−1 − ∆yt−1 , yt−3 = yt−1 − ∆yt−1 − ∆yt−2 .

3. Substitute back into the differenced equation and collect terms

Start from
∆yt = a0 + (a1 − 1)yt−1 + a2 yt−2 + a3 yt−3 + bt + εt
and substitute the expressions above:



∆yt = a0 + (a1 − 1)yt−1 + a2 yt−1 − ∆yt−1 + a3 yt−1 − ∆yt−1 − ∆yt−2 + bt + εt .

Now collect the coefficients on each component.

Coefficient on yt−1
(a1 − 1) + a2 + a3 = (a1 + a2 + a3 − 1).

Coefficient on ∆yt−1

From −a2 ∆yt−1 and −a3 ∆yt−1 :

−(a2 + a3 ).

Coefficient on ∆yt−2

From −a3 ∆yt−2 :

−a3 .
Thus, the ADF form is

∆yt = a0 + bt + (a1 + a2 + a3 − 1)yt−1 − (a2 + a3 )∆yt−1 − a3 ∆yt−2 + εt .

4. Define
γ ≡ a1 + a2 + a3 − 1, δ1 ≡ −(a2 + a3 ), δ2 ≡ −a3 .

Then the ADF regression form is

∆yt = a0 + bt + γ yt−1 + δ1 ∆yt−1 + δ2 ∆yt−2 + εt .

28
Interpretation (why this is ADF form) The ADF regression isolates the lagged level yt−1 (the
term that tests for a unit root via γ) and includes lagged differences ∆yt−1 , ∆yt−2 to absorb
higher-order AR dynamics. In this model, the mapping from the original AR(3) parameters
(a1 , a2 , a3 ) to the ADF parameters (γ, δ1 , δ2 ) is exactly:

γ = a1 + a2 + a3 − 1, δ1 = −(a2 + a3 ), δ2 = −a3 .

29