Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Forecasting in Business

Research Using the


ARIMA Box-Jenkins
Methodology

Milagros F. Malaya
Business Management Department, De La Salle University-Manila

Over the years, several techniques have been developed to


deal with diverse forecasting situations, falling broadly into
quantitative and qualitative (or technological) techniques.
Qualitative methods are basically subjective and judgmental
inputs of experts in the field. These are considered to be most
appropriate when historical data and/or scientific expertise are
not available. In some cases, they may even be preferred, as in
very long-range forecasting.
Quantitative methods, on the other hand, can be applied when
the following conditions are present:
1. there is past information;
2. past information is expressed in quantitative form;
3. it can be safely assumed that the past pattern of
relationship will continue until the future.

DLSU Business & Economics Review Volume 12 .~.Vo.l 2000-2001


82 FORECASTING IN BUSINESS RESEARCH

This last condition is referred to as the "assumption of


constancy." Simply put, the underlying premise that explains the
extrapolative nature of all quantitative forecasting methods is
"history will repeat itself."
The two major types of quantitative forecasting models are the
classical (or na"ive, or intuitive) and the modern (or formal) methods.
The first type investigates the horizontal, seasonal trend or cyclical pattern
of the data and uses empirical experience to extrapolate. Although nai"ve
methods are simple to understand and easy to compute, they vary widely
according to the nature of business, type of product, and the forecaster's
background. Formal methods possess the distinct advantages of a solid
statistical foundation (thereby minimizing forecast errors) and
procedures that can be done in a standard and mechanical manner.
This second type includes regression (or causal or econometric)
and time series models (see Figure 1).
Which method of forecasting to use will generally depend on:
(1) purpose or usefulness of the forecasts, including its time frame-
short-, medium- or long-term; (2) the key underlying assumptions;
and (3) the input data and their graphical display.
For causal models, a cause-effect type of relationship is
assumed between the variable to be forecasted and one or more
independent or explanatory variables. Its appeal lies in providing
a better understanding of the environment and the causal factors
that affect it. Time series models, on the other hand, use only the
history of the past values of the variable of interest to explain and
predict its behavior.
The objective is to discover the pattern in the historical data
series and extrapolate that pattern into the future.
Because time series models require data on only one
variable, they can often be used more easily to forecast; whereas
causal models are used with greater success for policy formation
and decision-making.
Another important consideration in selecting the adequate
forecasting method is the data pattern. Data series generally fall
into four distinct patterns:
1. A horizontal pattern exists when data values are found to
have a constant mean over time. Such a data series is
said to be stationary in the mean.
2. A seasonal pattern exists when the data are affected by
seasonal factors (the month or quarter of the year, or day
of the week).

DLSU Business & Economics Review Volume 12 No.1 2000-2001


~ Figure 1. Statistical Models for Time Series
"'c:: .l.
""'
S· _l
•~ Non~Structural or
Time Series Models Structural Econometric Models

i'"
(for short-tlme or (for long·term forecasts)
quick forecasts)
a
3

~

"0' J 1 J I
Moving Average ARIMAor
~ and Exponential Box-Jenkins
ARIMA models Classical Linear
Linear Simultaneous

~ Smoothing Models
with Intervention Regression
Regression with
Autocorrelated
Equation
Models
"' Procedures Errors
"'
9
~

j ~
Decomposition
of Time Series i,
.1 .
s:

"'"'"' Seasonal
..
~

"' Adjustment
"
"'"'
~ ..
.,
84 FORECASTING IN BUSINESS RESEARCH

3. A cyclical pattern exists when data are affected by


economic fluctuations associated with the business cycle.
Whereas a seasonal pattern is by and large of constant
length recurring on a regular periodic basis, cyclical pattern
has a different length and magnitude.
4. A trend pattern exists when the data exhibit an increase
or decrease over time.

Actually, data series will contain a combination of the above


patterns. Different methods use varying ways to identify or, in
certain cases, separate them into components, to arrive at a model
that best fits the data.
This lecture discussion will focus on a specific time series
technique known as the Box-Jenkins or ARIMA methodology.

UBJ-ARIMA MODELS
George E. P. Box and Gwilyn M. Jenkins are the two people
credited for formalizing the procedure known as ARIMA modeling.
ARIMA stand for Auto Regressive Integrated Moving Average.
Univariate (one variable) Box-Jenkins (UBJ) ARIMA forecasts are
based only on the variable of interest and not on any other data
series (unlike regression). We recall that it is assumed in
traditional regression that the various observations are statistically
independent. However a time series, which describe a sequence
of numerical observations naturally ordered in time, will be
statistically dependent, and therefore regression methods may
not be applied. UBJ ARIMA modeling procedure presupposes that
the data in the series are related and will attempt to determine
the nature of this relationship. This methodology applies to either
discrete or continuous data as long as they are given in equally
spaced or successive time intervals. It is most suited to short-
term forecasting, i.e., up to two or three periods only, since this
type of models puts emphasis on the recent data, rather than the
distant past. It is also highly useful when forecasting involves data
that exhibit seasonal variation. The sample size that is considered
adequate for this type of analysis is about 50, or less if there is no
seasonality.
ARIMA models are classified into:
1. Autoregressive (AR), indicating a relationship with past
observations;

DLSU Business & Economics Review Volume 12 No.] 2000-2001


Milagros F. Malaya 85

2. Moving Average (MA), indicating a relationship with


past random shocks;
3. Mixed Autoregressive-Moving Average (AR-MA).

THE PHILOSOPHY OF BOX-JENKINS


(As compared to Regression Analysis)
In standard regression analysis, the cause-effect relationship
between the interacting variables act as the "black box" process
that generates the observed time series.
In the Box-Jenkins methodology, the starting point is the
observed time series, followed by an examination of its
characteristics, selecting different models or "black boxes" to
generate finally what is termed as "white noise," a purely random
series of numbers.
Table 1. Comparison of Standard Regression Analysis
and Box-Jenkins Methodology
Far standard regression analysis:
1. Specify the causal variables.
2. Use a linear (or other) regression model.
3. Estimate the summary statistics and try other model specifications.
4. Examine the summal)' statistics and try other model specifications.

- -
5. Choose the most desirable model specification {perhaps on the basis of RMSE}

Start here: Explanatory Black Observed


variables box time series

For Box-Jenkins methodology


1. Start with the observed time series.
2. Pass the observed time series through a black box.
3. Examine the time series that results from passage through the black box.
4. If the black box is correctly specified, only white noise should remain.
5. If the remaining series is not white noise, try another black. box.

Start here: Observed


time series
- Black ___.. Purely random
box white noise

CONCEPTS USED IN ARIMA MODELLING

1. Assumption of stationarity
The UBJ-ARIMA method applies only to stationary data series.
A time series is considered stationary if its mean, variance and
autocorrelation function are constant through time. The mean
describes the over-all level of series and is constant if it does not

DLSU Business & Economics Review Volume 12 No.1 2000-2001


86 FORECASTING IN BUSINESS RESEARCH

increase or decrease over time (as seen in a horizontal pattern).


Also, if the data series is stationary, the variance of any subset of
the series is not expected to vary significantly from the variance
of any other subset. Non stationary series are easily transformed
into stationary series by the following operations:
(a) differencing, which is computing successive changes in
the values of the data series, Y, to produce a new series
w, i.e.,
W, = Y, - Y,., t = 2, 3, ... , n
w2 = v2 v,
W, = Y, - Y2
and so on

This operation known as the first differences, generally


transforms the series Y, nonstationary in the mean to
become stationary (W,). If the first difference is not
sufficient, a second differencing (difference of the first
difference) can be done but this is rare.

(b) logarithmic operation to transform a series non stationary


in variance to a stationary series.
If the data series is nonstationary in both the mean and
variance, the logarithmic operation is applied first, followed
by the differencing operation.
The I in ARI MA stands for an integrated series, or one which
has been differenced.

2. Autocorrelation function (ACF)


Autocorrelation coefficients of a time series are computed
to serve two purposes: (a) to identify the pattern in the original
data series (trend, seasonal, or random); (b) to verify if the
residuals are random (test of model adequacy).
The concept of autocorrelation is similar to that of correlation
except that it is applicable to values of the same variable at
different time lags (auto meaning self). Correlation measures
the degree of association between two variables, + 1 being the
perfect direct relation, -1 being the perfect inverse relation, and 0
as no relation. The concept of autocorrelation can be illustrated
by the following example (fable 2) where the original variable sales
(X) is related to another variable X, , which is lagged one period
behind, or to X 2 , which is two periods behind. If the data are

DLSU Business & Economics Review Volume 12 No.1 2000·2001


Milagros F. Malaya 87

stated as monthly figures and the ACF for a time lag of 12 periods,
or a 12-month time lag, is computed to be positive, then a seasonal
pattern of 12 months duration is indicated.

Table 2. Example of the Same Variable


with Different Time Lags
t X x, x,
Time Original Variable One Time Lag Two Time Lag
Sales Variable Variable
Constructed Constructed From
' From Sales Sales
1 13 8 15
2 8 15 4
3 15 4 4
4 4 4 12
5 4 12 11
6 12 11 7
7 11 7 14
8 7 14 12
9 14 12
10 12
I
3. Partial autocorrelation function (PACF)
This gives the correlation of stationary observations separated
by lag h adjusting for the effects of the intervening observations.
The estimated PACF, together with the ACF, are used in
determining the ARIMA model that will best fit the data.

The Model Building Procedure


Box and Jenkins propose a practical three stage procedure
for finding a good model, as given in Figure 2.

Preliminary Stage:
The tools used in determining stationarity are
1. historical plot (Y, vs. t): stationary series has no trend
(horizontal pattern) and has constant variability through
time.
2. sample values of stationary series approach 0 fast

A nonstationary series is transformed into a stationary series


by performing the first, or second differences, and/or logarithmic
operation.

DLSU Business & Economics Review Volume 12 No.1 2000-2001


88 FORECASTING IN BUSINESS RESEARCH

Stage 1
The estimated ACF's and PACF's are compared against the
theoretical ACF's and PACF's of the ARIMA Models and the models
which closely resemble the data series are chosen.
For nonseasonal time series, the characteristics of the ACF's
are given in Table 3. The equations of these are shown in Table 4.

Figure 2. The Mode/Building Procedure

I Original Series (Y1 )


I
+
PRELIMINARY STAGE
(Is data set stationary? If not,
transform it to become stationary.)

L
+
stationary series (W 1 ) _]
+
STAGE ONE: ldenUfication
(Choose one or more ARIMA Models


as candidates)

STAGE TWO: Estimation


{Estimate the parameters of the
models chosen in Stage one}

+
STAGE THREE: Diagnostic Checking
(Check the candidate models
for adequacy)

+
I Is model satisfactory? I
+
I YESI
+
Produce Forecasts using the model

DLSU Business & Economics Review Volume 12 No.1 2000·2001


Milagros F. Malaya 89

Table 3. The Characteristics of the


theoretical act's and pact's of ARMA Models

Model ACF PACF

AR (p} Damped exponentials and/or Cuts off at lag p


damped sine waves

MA (q} Cuts off at lag q I Damped exponentials and/or damped


sine waves
ARMA(p, q} Decays exponentially but stays Decays exponentially but stays up for p
up for q lags lags

Table 4. Equations of the General and


Common SpecificARIMA Nonseasonal Models

1. The Autoregressive Model of order p: AR (p)

a)AR(1)

b) AR(2)

2. Moving Average Model of order q: MA (q)

a) MA (1)

b) MA (2)

3. Mixed AR- MA Model of order (p, q)


Wt = E>o-E>,at·t -E>2at-2 - ... -epal-p +0,Wt-t + 02 Wl-2 + ... + 0P WI·P + al

ARMA (1, 1)

a, is white noise, uncorrelated and identically distributed


with mean 0 and constant variance.
0 = represents the AR parameters
= represents the MA parameters
e, = is the deterministic trend component

DLSU Business & Economics Review Volume 12 No. I 2000-2001


90 FORECASTING IN BUSINESS RESEARCH

Stage 2: Estimation
Estimation involves an iterative numerical procedure for
estimating the model of parameters 0; (autoregressive
coefficients) and e,(moving average coefficients) subject to the
inequality conditions found in Table 5

Table 5. Summary of Stationarity and lnvertibility Conditions

Stationarity Conditions lnvertibility Conditions


AR (1) /0,1 < 1 Always invertible
AR (2) 0, + 0, < 1 Always invertible
0, . 0, < 1
! /0j< 1
MA (1) Always stationary /0,/ < 1
MA(2) Always stationary 81 +92 < 1
/0,1 < 1 e 2 - e, < 1
/0 I< 1
ARMA (1, 1) /9,/ < 1

Stage 3: Diagnostic Checking


Estimated parameters are tested for significance using
statistics which are approximately normally distributed.
Diagnostic checking proceeds in a manner similar to residual
analysis in regression, where the Durbin-Watson statistic should
be close to two and residuals fall within the confidence band
approximating 0.
The model is considered adequate in terms of capturing
correlations existing among observations if the sample ACF and
PACF of residuals have values not significantly different from
zero, as indicated by the Box-Pierce or Box-Ljung Chi-square
tests.
Where alternative models have passed diagnostic checking,
the best model is chosen using (1) the coefficient of determination,
R2 , which is a measure of the percentage of variability of W, that
can be explained by the model; the closer it is to 100%, the better;
(2) the mean square error (MSE), which is a gauge of forecast
accuracy and therefore, the smaller, the better; (3) parsimony,
or the use of the smallest number of parameters in the model, an
indication of model simplicity.
The iterative nature of the UBJ methodology is critical in the
analysis since failure of the model to meet the conditions on the

DLSU Business & Economics Review Volume 12 No.1 2000-2001


Milagros F. Malaya 91

estimated parameters and diagnostic checking is a signal for


model reformulation and the procedure is repeated until an
appropriate model is found. The model is then used to build
forecasts.
To summarize, the characteristics of a good ARIMA model
are enumerated as follows:
1. It is parsimonious.
2. It is stationary.
3. It is invertible.
4. It has statistically significant coefficients.
5. It fits available data satisfactorily (i.e., it has passed the
tests of model adequacy).
6. It produces sufficiently accurate forecasts.

APPLICATIONS OF THE BJ MODELLING PROCEDURE

Example 1. Jollibee Stock Prices


The graph of the daily closing prices of Jollibee Stocks from
January 4 to September 30, 1999, shows a series nonstationary
in the mean due to the presence of an increasing/decreasing
trend. Also, the sample ACF and PACF computed for 44 lags
decay very slowly. The Box-Pierce Q statistic of 3056.252 is
extremely high and significant as compared to the x2 value of 55.8 at
40 degrees of freedom, a = .05.
The series is successfully transformed into a stationary one by
computing the first difference: its graph is shown as having a
horizontal pattern. The correlogram (which is the graphical display
of the estimated ACF and PAC F) of the differenced series does not
have any spikes lying outside the confidence band and yields an
insignificant Q statistic (=16.157). Thus, the difference of Y, is found
to be an adequate representation since only white noise remains in
the correlogram. However, we shall choose to express it in an
alternative form to generate the other tests of model adequacy and
for forecasting purposes.
An AR (1) model equation relates Y, to its immediate past
observation Y,., plus a random error component, or

Y, = 6, v,., + a,
Where ~ is the autoregressive parameter of lag 1. Where
~. is not significantly different from one, then the AR (1) model is

DLSU Business & Economics Review Volume 12 No.1 2000-2001


92 FORECASTING IN BUSINESS RESEARCH

identical to the first difference of Y,, that is,


W, = Y,- Y,_,

Applying this to the Jollibee d<tta series, 6, (=.981 0435) is


tested to be significantly different from zero but not different from
one. Adding the constant c to the model is an improvement since
it is significant and R2 increases. Both the Durbin-Watson and
the MSE values are acceptable.
The final model equation is Y, = 18.411184 + Y,_, +a, which
follows the random walk model for stock prices series.
With the model based on periods 1 to 187, the forecast is
developed for periods 188 to 190. The forecast as compared to
the actual Jollibee stock prices stock prices for the periods 188
to 190 yields an average forecast error of 0.5%.

Example 2. ABSCBN Stock Prices


The graph of the daily closing prices of the ABS-CBN stock
prices from January 4 to September 30, 1999, shows a definitely
increasing trend, with the correlogram of the sample ACF and
PACF as dampening slowly. Hence, the series is said to be
nonstationary in the mean. Computing the first difference stabilizes
the mean as seen in the horizontal pattern of its graph, but
nonstationarity due to a changing variance is indicated by
inordinately high peak towards the tail end of the graph.
Where nonstationarity results from both the changing mean
and variance, the logarithmic transformation is first applied to
stabilize the variance, followed by the differencing operation to
stabilize the mean.
We search for a model on the differenced log data series.
An AR (1) or MA (1) yield an extremely low R2 (only 3%), indicating
that the successive transformations (logarithm and differencing)
may have produced adverse effects. To verify this speculation, only
one operation (logarithm) is retained and an MA (1) model is tried.
R2 immediately jumps to 69%, supporting the contention. The model
is further improved by applying the same technique used in the
Jollibee series of substituting an AR (1) in lieu of the differencing
operation. Again, we take note that the AR (1) coefficient is not
different from one. R2 rises to 98.5% but the constant term of the
model is dropped due to nonsignificance. Values of the Durbin-
Watson and MSE are within normal levels. The residuals of the
ACF and PACF yield a Q (=22.188) that is no longer significant

DLSU Business & Economics Review Volume 12 No.1 2000-2001


Milagros F. Malaya 93

The model is developed for periods 1 to 186 while the forecast


(restored to its original form by getting the exponential) is obtained
for periods 187 to 189. The model equation can then be expressed
as
v, = v,_, + a,_, + a,
which is just a variation of the random walk model.
The actual and forecast ABS-CBN stock prices give an
average forecast error of 1.55%.

Stockholders and investors alike can take note that both


Jollibee and ABS-CBN Stock prices exhibit the "random walk"
phenomenon, which means that, barring direct intervention, stock
prices are not affected by any long-term influence from external
factors. Since the significant effect on today's price is only
yesterday's price, stock prices react immediately and more
impulsively to market forces and other events with economic
impact, political scenarios, peace and order situations, etc. That
is, such effects are only short-term, and Jollibee and ABS-CBN,
being business leaders in their respective industries, are able to
filter well the external disturbances affecting their organizations.
Hence, factors that affect significantly changes in stock prices are
internal to the firm and its line of products/services. Finally, for Jollibee,
a trend pattern exists which persistently pulls its prices upward, an
evident market sentiment of a continuing confidence in Jollibee as
one of the fastest growing fastfood chains in the country.

CONCLUSION
Permit this researcher to write a personal reaction as a fitting
conclusion to this paper.
The writer understands the economist's seeming obsession
with regression as a tool for analysis because of his concern for
causality. On the other hand, statisticians are wary of regression.
There are too many assumptions involved in regression.
Secondly, these assumptions are highly restrictive. In some
cases, just testing the assumptions can constitute a study in itself.
And the transformations to remedy the violations are just as equally
confusing. Additional assumptions are made, which to a large
extent remain unverified. For many, the final test is simply the
adequacy of the model in relating to actuality, or just how well the
researcher can defend any inconsistency or deviation.

DLSU Business & Economics Review Volume 12 No.1 2000-2001


94 FORECASTING IN BUSINESS RESEARCH

On the other hand, time series analysis, specifically the Box-


Jenkins methodology, contains no such restrictions. Over other
estimating procedures, Box-Jenkins methodology has the
following advantages: (1) B-J is based on the ARIMA models
which are well substantiated by classical probability theory and
mathematical statistics; (2) there are several choices for
modeling since ARIMA is a family of models; (3) the appropriate
ARIMA model produces optimal forecasts, that is, the least
forecast error variance among all linear models.
It is however not very popular probably due to the following
reasons:
1. The statistical concepts of ARI MA models are highly
theoretical.
2. The construction of the proper BJ models may require more
experience and computer time than some methods.
3. While it is true that regression is more meaningful, with
the interpretation given to the regression coefficients and
R2 , ARIMA models may not be as satisfying, but certainly
present a description of the underlying structure of the
movement of the time series. It is unfortunate that for
many researches using ARIMA models, the primary object
is forecasting and virtually no attempt is made to explain
the model equation. Parameter estimates of ARIMA give
the weights, or degree of importance, of the effect of the
past observation or residual on the current observation.
The model equation also shows the length of period of
persistence of such effects and seasonality of behavior, if
any. The constant term, if significant, indicates that the
series has a trend or is moving towards a specific
direction.
4. Possibly from the economist's point of view, it has limited
applications. I do believe that ARIMA modeling can serve
as a preliminary data analysis since there are no restrictive
assumptions. ARIMA modeling, as an "exploration into
the unknown", may yield valuable clues so that further
methods may be devised to reveal specific details. This
researcher also suggest ARIMA modeling as an
independent technique for data analysis to support or
challenge a contention or claim. For instance, an
econometric model and an ARIMA model may be

DLSU Business & Economics Review Volume 12 No.1 2000-2001


Milagros F. Malaya 95

developed simultaneously and their forecasting abilities


compared. Finally, there are situations where the factors
that affect the dependent variable interact, or are difficult
to segregate or quantify. Probably, the only way to analyze
the behavior of the dependent variable is by time series
modeling.

Having stated this writer's position, it is hereby clarified that


there is not intention to challenge the place of regression in the
heart of the economist. No amount of argument can ever do that.
It is hoped that this paper may be instrumental in changing the
attitude of a great number who have reserved for ARIMA modeling
just a berth for mental and computer gymnastics.

Notes:
1. Tables 1 and 2 are taken from Wilson and Keating.
2. Tables 3, 4, 5 are taken from Pankratz.

BIBLIOGRAPHY

Makridakis, S. and Whellwright, S. (ed). The Handbook of


Forecasting, New York: John Wiley & Sons, 1982.

- - - - - - · - - - . , . - - · Forecasting: Methods and


Applications. New York: John Wiley and Sons, 1978.

---:-::--.-,..-..,...,--:-:--::::--.,-· Univariate and M uItiva ria te


Methods. Holden Day, Inc, 1978.

Pankratz, A. Forecasting with Univariate Box-Jenkins Models.


New York: John Wiley and Sons, 1983.

Wilson, J. H. and Keating, B. Business Forecasting. Illinois:


Richard D. Irwin, Inc., 1994.

DLSU Business & Economics Review Volume 12 No.1 2000-2001

You might also like