CHAPTER 13.

Time Series Models:

Further topics

13.1. Seasonal ARIMA Models

Economic time series often exhibit strong

seasonal variation.

Fig 13.1 below: Logarithms of quar-

terly urban housing starts in Canada.
(a) Time series plot. (b) ACF. (c)
Boxplots by quarter.

1
 (a) (b) (c)
10.0

10.0
●
●●●●●●●● ●●
●●●●●
●●● ●● ●●●●●●
0.8

● ●● ●● ●●● ●●●●●
9.5

9.5
●●● ● ● ●
●●●● ● ●● ●● ●●● ●●●●● ●●●●
● ●●● ●●●● ● ● ●●●●●●● ●●●●●
log(starts)

log(starts)
●●● ● ● ●●●●● ●●●●●●
●
●●●● ●
ACF

● ● ●● ● ●●
9.0

9.0
0.4

●● ● ● ● ● ● ● ●●
● ● ●●
● ● ●● ●● ●●
●● ●
●
●
8.5

8.5

●
●● ●
●●
0.0
8.0

8.0

● ●

1960 1980 2000 0 5 10 20 1 2 3 4

year lag quarter

2
The series ∆sYt = (1 − B s)Yt = Yt −
Yt−s is called the seasonally differenced
series.

Fig 13.2 below: Time series (left col-

umn) and ACF plots (right column) of
the logarithms of urban housing starts
with nonseasonal difference (top low),
seasonal difference (middle row), and
both seasonal and nonseasonal differ-
encing (bottom row).

3
 (a) nonseasonal differencing (b) nonseasonal differencing

●
● ●

0.8
● ●● ● ● ● ●
● ● ● ● ● ●● ●●● ● ●
●● ●● ●
0.5

● ●● ● ● ● ● ● ● ●●
diff(x)

● ●
● ● ●

ACF
● ● ●
●● ● ● ●
● ● ●●● ●●

0.2
●●●● ●● ● ●●
●● ● ●●● ●●●● ● ●●
●● ● ●●● ● ●
●● ●● ●●● ●●●
● ● ●●
●●● ●
●
● ● ●
● ●●● ●
●
● ● ●●●●●● ● ● ● ● ● ●
● ● ●
−0.5

● ●
●● ● ●
● ● ● ● ● ● ●●● ● ● ●
●●●● ● ● ● ●● ● ● ●

−0.4
●
● ● ●

1960 1970 1980 1990 2000 0 5 10 15 20

year lag

(c) seasonal differencing (d) seasonal differencing

● ●
●
0.5

0.8

●●● ●● ● ●
● ●
● ●● ● ● ●● ● ●
● ● ●
diff(x, 4)

●
●● ● ●● ● ●●● ● ●●●●●
● ●●
●
● ●
● ● ●● ●● ● ●● ● ●●● ● ●
● ●●
●●●●●
ACF

● ●
● ●●● ● ● ●
●
●● ●● ● ● ● ●●●●
● ●● ● ● ● ● ● ● ●●
●
●
●
● ●
●●●
●● ●
●● ●●●
●●● ●●●
0.2

●● ● ●● ● ●
●● ●
● ●●●●●● ●● ●
●●
●
●● ● ● ●●
−0.5

● ● ●●
●
● ● ●
●
●
●
−0.4

1960 1970 1980 1990 2000 0 5 10 15 20

year lag

(e) seasonal & nonseasonal differencing (f) seasonal & nonseasonal differencing

● ●
diff(diff(x, 1), 4)

●
0.5

● ● ●
●●
●● ● ●
● ●
0.5

● ●
● ●● ● ●● ●
● ●●● ● ●
ACF

● ●●
● ●
●●● ●●● ●
●
●●● ● ● ●● ●
●● ●●
●● ● ●●
● ●●
●
●●●● ● ●●●
●●●
●
●● ●
●●●●● ●●● ●
●●
●● ● ● ●●● ●●●●●
● ●●●
● ● ● ●●●
● ●
● ● ● ●●●● ● ● ● ●● ●● ● ●●
● ●● ●● ● ●
● ● ● ●
● ● ● ● ●
●●
● ● ●●●● ● ●
−0.5

● ● ●● ●
−0.5

●
●

1960 1970 1980 1990 2000 0 5 10 15 20

year lag

4
Nonseasonal ARIMA(1,1,0) model:

(1 − φB)(∆Yt − µ) = t

Purely seasonal ARIM A(1, 1, 0)s model:

(1 − φ∗B s)(∆sYt − µ) = t

Put these two together and get the

ARIMA(1, 1, 0) × (1, 1, 0)s model:

(1 − φB)(1 − φ∗B s){∆s(∆Yt) − µ} = t

Note that ∆s(∆Yt) = ∆(∆sYt).

5
ARIMA{(p, d, q) × (ps, ds, qs)s} model:

(1 − φ1B − · · · − φpB p){1 − φ∗1B s − · · ·

−φps (B s)ps }{∆d(∆ds s Yt) − µ}
= (1 + θ1B + · · · + θq B q ){1+
θ1∗ B s + · · · + θq∗s (B s)qs }t

Fig 13.3 below: Forecasting logarithms

of quarterly urban housing starts using
the ARIM A{(1, 1, 1)×(0, 1, 1)4} model.
The dashed line connects the data, the
dotted line connects the forecasts, and
the solid lines are the forecast limits.

6
 o data
* predictions
10.0

lower CL
upper CL
●
● * * *
9.5

● *
log(starts)

● ● *
● ● ●● ●● * * *
●
●
●
● ● ● ● ● * * * *
●
● ● ●
● ● ●
● ●
9.0

● *
● ● * * *
●
● ● ● ●
●

●
8.5

1992 1994 1996 1998 2000 2002 2004 2006

year

7
13.2 Box-Cox transformation for time
series

It is often desirable to transform a time

series to stabilize the size of the vari-
ability, both seasonal and random.

Box-Cox transformation:
(α)
(∆dYt − µ)
(α) (α)
= φ1(∆dYt−1 − µ) + · · · + φp(∆dYt−p − µ)
+t + θ1t−1 + · · · + θq t−q ,
where 1, · · · , is Gaussian white noise,
and
Y α−1
(α)
Yt = t .
α
8
Fig 13.5. Profile likelihood for α (called
λ in the legend) in the housing start
example. Values of λ with R(λ) (the
profile likelihood) above the horizontal
line are in the 95% confidence limit.

Relative Likelihood Analysis

95% Confidence Interval
1.0
0.8

^
λ = 0.337
0.6
R(λ)

0.4
0.2
0.0

0.0 0.2 0.4 0.6

9
13.3 Multivariate time series

Yt = (Y1,t, · · · , Yd,t)

A multivariate time series said to be

stationary if for every n and m, Y1, · · · , Yn
and Y1+m, · · · , Yn+m have the same
distribution.

10
Suppose that Yj and Yj 0 are the two
component series of a stationary multi-
variate time series. The cross-correlation
function (CCF) between Yj and Yj 0 is
defined as

ρYj ,Y 0 (k) = Corr{Yj (t), Yj 0 (t − k)}

j

and is the correlation between Yj at a

time t and Yj 0 at k time units earlier.
As with autocorrelation, k is called the
lag.

11
A multivariate time series Y1, · · · is said
to be weakly stationary if the mean and
covariance matrix of Yt do not depend
on t and if the right-hand side of the
above equation is independent of t for
all j, j 0, k.

12
A d-dimensional multivariate time se-
ries 1,· · · is a weak WN(µ, Σ) process
if

• E(t)=µ for all t;

• Cov(t)=Σ for all t;

• for all t 6= t0 all components of t are

uncorrelated with all components of
t0

13
A d-dimensional multivariate time se-
ries Y1, · · · is a multivariate ARMA(p,q)
process with mean µ if for p × p matri-
ces Φ1, · · · ,Φp and Θ1, · · · ,Θq ,

Yt − µ
= Φ1(Yt−1 − µ) + · · · + Φp(Y t−p − µ)
+Θ1t−1 + · · · + Θq t−q + t,

where 1, 2, · · · is a multivariate WN(0,

Σ) process.

A multivariate AR process is also called

a vector AR or VAR process.

14
An AR(1) process is stationary iff all of
the eigenvalues of Φ are less than 1 in
absolute value. The eigen function in
R can be used to find the eigenvalues.

Forecasting with multivariate AR pro-

cesses:
Ŷ n+k = µ̂ + Φ̂1(Ŷ n+k−1 − µ̂) + · · ·
+Φ̂p(Ŷ n+k−p − µ̂)

15
13.4 Long-Memory processes

Short memory: ρ(k) < Drk for all k and

some D > 0 and 0 < r < 1

Long memory: ρ(k) ∼ Dk−α for some

D > 0 and α > 0 (hyperbolic rate)

16
For d > −1 define
∞ d
∆dYt = (−1)k Yt−k
X

k=0 k

Yt is a fractional ARIMA(p,d,q) pro-

cess, called an ARFIMA or FARIMA(p,d,q)
process, if ∆dYt is an ARMA(p,q) pro-
cess.

If −1/2 < d < 1/2, then the process is

stationary. If 0 < d < 1/2, then it is a
long-memory stationary process.

17
Fig 13.10. Samples are simulated from
FARIMA(0,d,0).

d = −0.35 d = −0.35

1.0
2

ACF

0.4
−2 0
x

−0.2

0 500 1000 2000 0 5 10 15 20 25 30

Time Lag

d = 0.35 d = 0.35
0 2 4

0.8
ACF
x

0.4
−4

0.0

0 500 1000 2000 0 5 10 15 20 25 30

Time Lag

d = 0.7 d = 0.7
0.8
−5 0

ACF
x

0.4
0.0
−15

0 500 1000 2000 0 5 10 15 20 25 30

Index Lag

18
Fig 13.11. ACF for simulated
F ARIM A(0, 0.7, 0)

∆ Y
0.7 ∆Y
1.0
0.8
ACF

ACF

0.4
0.4

−0.2
0.0

0 5 15 25 0 5 15 25

Lag Lag

19
Some R functions: simARMA0 in the
longmemo package; cumsum; fracdiff
in the fracdiff package; diffseries

Homework: Exercises 1, 4, 5.

20