You are on page 1of 2

354 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

AUTOREGRESSIVE MOVING AVERAGE MODELs


A model with autoregressive terms can be combined with a
model ha.
erage terms to get a "mixed" autoregresSive-moving
average model. It is conv
use the notation
ARMA(p. q), where p is the order of the autoregressive
the order of the moving average pa
part, to represent these models. An AD
model has the general form

Y=h+
ARMA
+-t. +0, te,0-12-
ARMA(p, q) models can describe a wide variety of behaviors for
series' Forecasts
generated by an ARMA(P, q) model will depend on current stationarvtime
values of the response and nae
Yas well as current and past values of the errors
Figure 9.4 shows the equation of an
of the theoretical ARMA(1,1) model and the (residuals)
autocorrelations and partial autocorrelations. In possible behavio
autocorrelations and partial autocorrelations die this case, both the
out, neither cuts off
SUMMARY
Theautocorrelation and
average processes can be partial autocorrelation patterns for
summarized as follows: autoregressive-moving
Autocorrelations
MA() Partial autocorelation
Cut off after the order
4 of the process Dieout
AR(P) Die out
Cut off after the
order
ARMA(P.) Dieoout Pof the process
Die out
The numbers of
ARMA model are autoregressive and moving average terms
determined from
partial autocorrelations the
patterns of the sample p(orders
q) in an and
in later section of this
a and the values of the model autocorrelations and
chapter. In practice, the valuesselection criteria that are discussed
of p and q each
Implementing the
Model-Building rarely exceed 2.
As
shown in Figure 9.1, the Strategy
strategy that consists of Box-Jenkins approach uses an
model coefficients selecting an initial model iterative
(parameter (model identification),model-building
ing). If necessary, the initial estimation), and estimating the
analyzing the residuals (model
residuals indicate no further model is modified and the check-
can be used
for modification is necessary. Atprocess is repeated until the
The forecasting.
steps in the
this point, the fitted model;
model-building strategy examined in some detail
are
STEP 1: MODEL next.
1. The first IDENTIFICATION
step in model
ary, that is, whether the identification is to determine
to look at a time series whether the series is
plot of the series alongappears
with the
to vary about
a fixed station
level. It is useful
Notice when q=0, the sample autocorrelation
when p=0, the ARMA (p.0) model function. A
ARMA (0q) model is a pure reduces
to a
pure
autoregressive
moving average model model
of order q.
of order p.
Similarly,
CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 355S
nonstationary time series is indicated if the series appears to grow or decline over
time and the sample autocorrelations fail to die out rapidly. The time series' pic
tured in Figure 8.2 are nonstationary, and the pattern of sample autocorrelations
shown in Figure 8.3 is typical for a nonstationary series. (See p. 297.)
If the series is not stationary, it can often be converted to a stationary series by differ-
encing.
model isThat
thenis,specified
the original
for series
the differenced
is replacedseries.
by a In
series
effect,
of differences.
the analyst An
is modeling
ARMA?

changes rather than levels.


over time, but
As an example, suppose the original series Y, is generally increasing
the first differences AY = Y,- Y- vary about a fixed lev
It may be appropriate to
.

model the stationary differences using an ARMA model of,say, orderp=1and q=1.
In this case the model is

AY,=%AY-1t6,-0E-1
OI
(-Y-)=,(--Y-+6,-04,
the differences before stationary
In some cases, it may be necessary to difference
done twice and the stationary data are
data are obtained. Simple differencing is
A=A(AY) =A(,--Y-9=Y-27,-+Y-
about a fixed
the data indicates the series varies
Differencing is done until a plot of of differences
die out fairly rapidly. The number
level, and the sample autocorrelations
denoted by d.
required to achieve stationarity is moving aver-
Models for nonstationary series are called autoregressive integrated
order of the autore-
age models and denoted by
ARIMA(pdq).° Here p indicates the the order of the
indicates
amount of differencing, and q
gressive part, d indicates the series is stationary, d 0 and the
= ARIMA models
moving average part. If the original from this point on, the ARIMA(p,dg)
no-
reduce to the ARMA models. Consequently,
tation is used to indicate models for both stationary
(d 0) and nonstationary (d>0)
=

time series. series can


involve differences, forecasts for the original
Although ARIMA models
from the fitted model.
always be computed directly of
the analyst must identify the form
2. Once a stationary series has been obtained,
the model to be used.

This step is accomplished by comparing the


áutocorrelations and partial autocor
and partial auto-
to the theoretical autocorrelations
relations computed from the data correlations for some of
models. The theoretical
correlations for the various ARIMA
are shown in Figures 9.2,9.3,
and 9.4 for help in se-
the more common ARIMA models
model.
lecting an appropriate set of autocorrelations and partial
autocorrela-
Each ARIMA model has a unique values to one
be able to match the corresponding sample
tions, and the analyst should
of the theoretical patterns. model from
an appropriate ARIMA
There may be some ambiguity in determining autocorrelations. Thus, the initial
autocorrelations and partial
the patterms of the sample

differenced series, the constant term may not be required.


sWhen an ARMA model is used for a to get the original series
differences must be summed (or integrated)
The term ntegrated means the

You might also like