Scribd
Upload
69 views8 pages

Lec4 08

This document provides an overview of seasonal time series analysis and component models. It discusses representing business cycles using seasonal and periodic models. Pure seasonal and multiplicative seasonal time series models are presented, along with their properties. Trend analysis covers deterministic and stochastic trends. The document notes that component models separating a time series into trend, seasonal, and irregular components are not uniquely identifiable from the observed data alone.

Uploaded by

alanpicard2303
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
69 views8 pages

Lec4 08

This document provides an overview of seasonal time series analysis and component models. It discusses representing business cycles using seasonal and periodic models. Pure seasonal and multiplicative seasonal time series models are presented, along with their properties. Trend analysis covers deterministic and stochastic trends. The document notes that component models separating a time series into trend, seasonal, and irregular components are not uniquely identifiable from the observed data alone.

Uploaded by

alanpicard2303
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

Lecture 4: Seasonal Time Series, Trend Analysis & Component Model

Bus 41910, Time Series Analysis, Mr. R. Tsay

Business cycle plays an important role in economics. In time series analysis, business
cycle can be shown in two ways. If the periodicity is fixed, then the cycle can be represented
by a seasonal (or periodic) model. If the periodicity is time-varying, then an AR(2) factor
with complex roots is used. For deterministic function f (.), we say that f (.) is periodic
with a periodicity s if
f (t) = f (t + k s) k = 0, 1, 2,
A typical example of a deterministic periodic function is a trigonometric series, e.g. sin() =
sin( + 2k) or cos() = cos( + 2k). The trigonometric series are sometimes used in
econometrics to model time series with strong seasonality. Of course, seasonal dummy
variables are also used in econometrics to handle strong seasonality.
For stochastic process Zt , we say that it is a seasonal (or periodic) time series with periodicity s if Zt and Zt+ks have the same distribution. Such processes are common in business
and economics. For instance, the series of monthly sales of a department store in the U.S.
tends to peak at December and to be periodic with a periodicity 12.
In what follows, we shall use s to denote periodicity of a seasonal time series. Often s = 4
and 12 are used for quarterly and monthly time series, respectively.
Some examples of seasonal time series:
1. Monthly U.S. Retail and Food Service Sales from January 1992 to August 2008 in
millions of dollars.
2. Electricity consumption of an industrial sector of U.S.
A. Pure seasonal time series
General Model: (B s )Zt = C + (B s )at where C is a constant,
(B s ) = 1 1 B s 2 B 2s P B P s ,

(B s ) = 1 1 B s 2 B 2s Q B Qs

A simple example: Zt = C + (1 B 12 )at . This is a simple seasonal MA model. It is easy


to see that
Invertibility: || < 1.
E(Zt ) = C.
Var(Zt ) = (1 + 2 )a2 .

13.0

Retail and Food Service Sales: 1992.12008.8

12.8
12.6
12.0

12.2

rfs
12.4

1995

2000
tdx

2005

9.6
9.4

log-consu

9.8

10.0

Time Series Plot: Log electric power consumption

1975

1980

1985
year

1990

ACF:

` =

1+2

if ` = 12
if ` =
6 0 or 12.

Another simple example: Zt Zt12 = C + at . This is a simple seasonal AR model. It is


easy to see that
Stationarity: || < 1.
Mean: E(Zt ) =
Var(Zt ) =

c
.
1

1
2.
12 a

ACF:

` =

k for ` = 12k for k = 0, 1,


0 otherwise.

When = 1, the series is non-stationary.


Exercise: Study properties of the seasonal model (1 B 12 )Zt = (1 B 12 )at .
B. Multiplicative seasonal time series
A special, pasimonious class of seasonal time series models that is commonly used in practice
is the multiplicative seasonal model ARIMA(p, d, q)(P, D, Q)s
(B)(B s )(1 B)d (1 B s )D Zt = c + (B)(B s )at
where all zeros of (B), (B s ), (B) and (B s ) lie outside the unit circle. Of course, there
are no common factors between (B)(B s ) and (B)(B s ).
The basic idea of this model is close to the two-way table in analysis of variance in which
the seasonal and regular components are approximately orthogonal. For s = 12, we have
Year Jan.
1
Z1
Z13
2
..
..
.
.

Nov.
Z11
Z23

Feb
Z2
Z14

Dec.
Z12
Z24
..
.

Here the column-effects are the regular serial corrections and the row-effects denote the
annual correlations.
A special model: The airline model
(1 B)(1 B 12 )Zt = (1 B)(1 B 12 )at

where || < 1 and || < 1. This model is the most used seasonal model in practice. It
was proposed by Box and Jenkins (1976) for modeling the well-known monthly series of
airline passengers. It has been shown, Cleveland and Tiao (1976), that the X-11 technique
of seasonal adjustment used by the US government is in fact close to this model.
Let Wt = (1 B)(1 B 12 )Zt , where (1 B) and (1 B 12 ) are usually referred to as the
regular and seasonal difference, respectively. Obviously, Wt = c+(1B)(1B 12 )at
is a multiplicative MA model. It pays to study carefully this seasonal MA model. For
simplicity, assume c = 0.
Mean: E(Wt ) = 0.
Variance: Var(Wt ) = (1 + 2 )(1 + 2 )a2
ACF:

` =

1+2

(1+ )(1+2 )

1+2

(1+2 )(1+2 )

for ` = 0
for ` = 1
for ` = 11
for ` = 12
for ` = 13
otherwise.

Note that 11 = 13 6= 0, which can be regarded as an interaction between the


regular and seasonal correlations. Also, the seasonal factor does not affect the regular
correlation, neither the regular factor affects the seasonal correlation.
Exercise: Study the ACF of the series Wt = (1 1 B 2 B 2 )(1 B 12 )at .
Exercise: Study the ACF of the series Wt = (1 B)(1 B 4 )at and Rt = (1 B 4 )at .
C. Non-multiplicative seasonal model
In some applications, a non-multiplicative model might be suitable. A simple example of
the model is
Zt = (1 B B 12 )at
The ACF of this series is (for ` > 0)

` =

1+2 +2

1+2 +2

1+2 +2

for ` = 1
for ` = 11
for ` = 12
otherwise.

Notice that the difference between this and that of multiplicative model. The ACF structure also highlights the parsimony of the multiplicative model as both models use two
parameters, yet the multiplicative model covers serial correlation at lag 13.
5

Exercise: Study the properties of the model


Zt Zt12 = (1 B)at
where || < 1. This model is also useful in practice.
D. The simplifying operator
The seasonal difference (1 B 12 ) can be factorized as

(1 B 12 ) = (1 B)(1 3B + B 2 )(1 B + B 2 )(1 + B + B 2 )(1 + 3B + B 2 )(1 + B)(1 + B 2 )


All 12 zeros of this polynomial are on the unit circle. Each factor produces different response
function. The overall pattern, however, has a period of 12. The factor
(1 + B + B 2 + + B 11 )
represents an average of 11 consecutive observations. It can be used as a filter to remove
the seasonality in a monthly time series. The factor (1 B) is not included, because it
corresponds to a trend.
Similar comments apply to (1 B 4 ) and (1 B s ) in general.
E. Trend Analysis
By and large, two types of trend are commonly used in business and economic time series
analysis, namely deterministic and stochastic trends.
Deterministic trend:
Linear trend: Zt = 0 + t + Xt , where Xt is a stationary time series, e.g. a white
noise series.
Exponential trend: ln(Zt ) = 0 + t + Xt .
Cyclical trend: Zt = r cos(t + ) + Xt , where r is amplitude, is the frequency with
period 2
, and denotes the phase shift. More generally,

Zt =

k
X

ri cos(i t + i ) + Xt =

i=1

k
X

[Ai cos(i t) + Bi sin(i t)] + Xt

i=1

where Ai = ri cos(i ) and Bi = ri sin(i ).


Stochastic trend:
Linear trend: Zt = t + at , where t = t1 + t with {t } a white noise series
independent of at .
Quadratic trend: Zt = t + at , where (1 B)2 t = t .
6

Seasonal trend: Zt = t + at , where (1 B s )t = t .


The deterministic trend can be regarded as a special case ofstochastic trend. For instance,
if i = 2i
for i = 1, 2, , 6, then we have cos(i ) = 1, 23 , 21 , 0, 21 , 1. Therefore, by
12
apply (1 B 12 ) to the general cyclical trend model we have
6
X

(1 B 12 )[

ri cos(i t + i )] = 0

i=1

and
(1 B 12 )Zt = (1 B 12 )Xt .
This latter equation points out an important fact that is commonly overlooked by data
analysts. The model seems to indicate that there is a common factor (1 B 12 ) on both
sides of the equation, implying that one might say that Zt = Xt . However, this is only part
of the picture, as we know that the origial time series Zt is Xt plus some cyclical trend.
Thus, the correct cancellation formula is
Zt = f (t, 12) + Xt
where f (t) is a deterministic function of period 12.
F. Component Models
There is a growing literature in considering component models in time series literature.
The component model has a long history, it is basically assume that
Zt = Tt + St + Rt
where Tt , St , Rt are respectively the trend, seasonal and irregular components of
Zt . The three components are assumed to be independent. The common approach to
component model is the structural model, e.g. Harvey (1990), which assumes a particalar
model for each of the three components, then estimate the parameters involved by maximum
likelihood method.
The idea of such a component model is appealing. However, one must use the model with
care. Why? Basically, the model is not identifiable. In other words, there are infinite
many ways to decompose a time series into the three components.
A simpel example is in order. Consider the ARIMA(0,1,1) model
(1 B)Zt = (1 B)at .
This is a model we can build from data. However, this model may arise from many sources.
Case 1: Write Zt = Tt + bt where Tt = Tt1 + et and {et } and {bt } are independent white
noise series. Then, we have
(1 + 2 )a2 = e2 + 2b2
7

and a2 = b2 .

Thus, and a2 are determined by b2 and e2 .


Case 2: Write Zt = Tt + bt where Tt = Tt1 + t t1 with {t } a white noise independent
of {bt }. In this case, it is easily seen that and a2 are determined by
(1 + 2 )a2 = (1 + 2 )e2 + 2b2

and a2 = e2 + b2 .

Thus, given models for the component Tt and bt , we can determine and a2 . On the other
hand, given and a2 , there is no way we can determine which case is the true underlying
model. In practice, only Zt is available (observable), implying that we can only CHECK
the model for Zt . Therefore, the identifiability problem arises.
One can resolve the identifiability problem if he/she is willing to add certain conditions.
For example, in the above instance, one may require that Tt is a random walk. Then, case
1 is the solution. Do not overlook this identifiability problem if you make inference about
the components.
In summary, the component model is suitable for forecasting. To use it to make inference
on components, one must understand the assumption used to obtain the decomposition
and the fact that the decomposition obtained is only one of many possible decompositions.

You might also like