Forecasting construction cost escalation

ANDREWN. BLAIR
Blairwood Construction, P.O. Box 3483, Kampala, Uganda, East Africa
AND

LEONARD
M. LYEAND W.J. CAMPBELL
Faculty of Engineering and Applied Science, Memorial University of Newfoundland, St. John's, Nfld.,
Canada A1B 3x5

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

Received July 15, 1992

Revised manuscript accepted December 7, 1992
Escalation can account for a substantial part of construction costs. Therefore forecasts of the amount of escalation
are required for budgetary and bidding purposes. This paper examines methods for forecasting construction escalation
using statistical time series methods. Time series of construction cost indices are used as a proxy of construction cost
escalation. The application of time series methods, their limitations, and their effect on the risk of cost escalation are
demonstrated and evaluated. The analytical methods available are only useful in forecasting for short construction
projects in stable conditions. This is because none of the methods can forecast escalation caused by unpredictable occurrences such as outbreak of war or certain government action. Construction cost escalation remains a risk to be borne
by either the contractor or the owner, or both, depending on the terms of the contract; any logical approach to minimize
the risk is worthwhile.
Key words: construction cost escalation, cost indices, time series forecasting, exponential smoothing. Box-Jenkins
methods, dynamic regression, Statistics Canada.
Une bonne partie des coiits de construction est attribuable a l'escalade de ceux-ci. Par consequent, la prevision de
l'ampleur de cette escalade est necessaire a des fins budgetaires ou d'appel d'offres. Cet article examine des methodes
de prevision de l'escalade des coiits de construction A l'aide de methodes chronologiques statistiques. Les series chronologiques des indices des cofits de construction sont utiliskes comme substitut B l'escalade des coiits de construction.
L'application des mCthodes chronologiques, leurs limites ainsi que leur effet sur le risque d'escalade sont traites et
evalues. Les methodes analytiques existantes ne sont utiles que dans le cas de projets de construction de courte duree
dans des conditions stables. Cela s'explique par le fait qu'aucune methode ne peut prevoir l'escalade causee par des
Cvenements imprkvisibles comme la guerre ou certaines decisions gouvernementales. L'escalade des coiits de construction
est un risque qui doit &treassume par l'entrepreneur, le proprietaire ou les deux, selon les modalites du contrat; toute
methode logique en vue de minimiser ce risque est souhaitable.
Mots clks : escalade des coiits de construction, indice des coiits, prevision chronologique, lissage exponentiel, mkthode
Box-Jenkins, regression dynamique, Statistique Canada.
[Traduit par la redaction]
Can. J. Civ. Eng. 20, 602-612 (1993)

Introduction
Websters dictionary defines the term escalate as "to
gradually increase ... to raise and go up ...." Escalation in
construction costs is the increase in the costs of any construction elements required for original contract works
occurring during construction. The amount included in any
construction cost estimate or construction cost breakdown
to account for escalation in construction costs is an important component of total construction costs.
The financial success of construction projects can be
uncertain and at risk because of the possibility of drastically
changing escalation rates during construction. At the beginning of any given project, there can be a number of different
possible future escalation rates. The use of a n erroneous
escalation rate when estimating construction costs can have
adverse effects o n economic decision making. As a n example, Fig. 1 illustrates the impact of changing escalation rates
o n a hypothetical construction project with a n unescalated
cost of $30 000 000. The cost flow profile of the project is
expected to form the predetermined S-shaped curves shown
NOTE: Written discussion of this paper is welcomed and will be
received by the Editor until December 31, 1993 (address inside
front cover).
Printed in Canada / Imprim6 au Canada

in Fig. 1 (Tham 1980). The project is to be constructed over

a period of 2 years starting 3 months from the date of tender.
Figure 1 indicates that if a n annual escalation rate of 10%
is experienced during construction instead of a prior
estimated rate of 3%, then an additional $3 000 000 expenditure would be incurred. The party who bears the escalation
risk can be devastated by it.
This problem could be minimized or eliminated by inclusion of a mutually agreed "escalation clause" in the contract
documents. This is the practice in many offshore and international construction projects, but unfortunately is not
common in North America.
It is therefore necessary to forecast the amount in
monetary terms of escalation costs that will be incurred
during the execution of a construction project for budgetary
and bidding purposes. T o forecast the amount of cost escalation, one can forecast an applicable escalation rate and apply
this rate to the estimated cash flow. Forecasting the
applicable escalation rate can be achieved by forecasting the
future value of an appropriate cost index. Cost indices are
indicators of the amount of cost escalation. Indices describe
how the cost of a particular construction unit changes with
time. Cost indices are time series because they are generally
produced at regular time intervals. Methods for analyzing

BLAIR ET AL,

Escalati

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

Escalation of 3%

Months from date of tender

FIG. 1. Effect of various escalation rates on construction costs.

and forecasting time series can thus be used to forecast the

rate of escalation of a given construction project (Taylor
and Bowen 1987).
A number of methods are available for forecasting time
series. Many of these methods require a substantial degree
of mathematical dexterity and can be time consuming. In
the past, the parties to a contract were often advised to hire
a consultant to apply these techniques (Stevenson 1984). The
current availability of user friendly computer forecasting
software packages, such as FORECAST PRO (BFS 1988), has
now reduced the amount of mathematical manipulation
necessary for a practitioner. The key requirement in applying
the various forecasting methods using these packages is the
ability to interpret the computer output and to understand
the limitations of the techniques used. The growth of
computer utilization, even among small- and medium-sized
contractors, leads us to believe that the computational techniques will soon be commonplace.
This paper examines the analytical techniques available
to forecast the rate of escalation of construction costs by
forecasting the values of a typical cost index. The cost index
chosen is the "output" index for the prefabricated wood
building industry. The technique for development of this
index is fully documented in Catalogue 62.007, Construction
Price Statistics, published by Statistics Canada. A brief
outline of the theory underlying the various forecasting
methods applicable to construction cost forecasting is given.
Forecasting the values of the typical cost index using
FORECAST PRO is used as an example of the application of
each of the applicable forecasting methods. From these
examples, and from the outline of the underlying theory,
the benefits and limitations of each method are discussed
and a general strategy for choosing between various methods
is given. Finally, the usefulness of the various forecasting
methods to the owner and the contractor of a given construction project is evaluated. Particular attention is given

to whether or not these methods of forecasting future values

of cost indices significantly reduce the risk of financial loss
due to cost escalation.
Forecasting methods
Empirical studies have shown that there is no single best
forecasting method applicable to all situations (Goodrich
1989). To decide on which forecasting method is best for
a given situation, it is necessary to critically examine the
available data. This and an understanding of the fundamentals of the various forecasting procedures are prerequisites
for obtaining realistic forecasts.
Forecasting methods can be classified into three categories: subjective methods, univariate methods, and
multivariate methods (Chatfield 1975).
Subjective methods
Subjective methods are based on human judgement of the
various factors that may have an impact on the required
forecast (Firth 1977). These methods may range from
intuitive and subjective decision made by the decision makers
(Nelson 1973) to highly refined rating schemes that turn
qualitative information into quantitative estimates.
Subjective forecasts are based on judgement, intuition,
commercial knowledge, and any other information the
forecaster deems relevant. A wide range of factors may be
taken into account, depending on the knowledge and the
experience of the forecaster. This makes subjective forecasts
unique to the individual forecaster and therefore not
reproducible.
Subjective methods and intuitive estimates are widely used
in construction estimating and are most useful when there
are insufficient historical data on the appropriate cost index.
Mathematical methods cannot generally be used to make
long range forecasts, that is, forecasts of duration over
2 years (Firth 1977). For such forecasts, subjective methods
have to be used.

CAN. J. CIV. EN(3.

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

604

The forecaster's intuition may often prove to be more

reliable than any mathematical method (Chatfield 1975). As
such, subjective methods can be used as a basis of judging
the accuracy of other methods by comparing the forecasts
obtained using mathematical methods with the forecaster's
intuitive estimates. Since subjective forecasts are not
reproducible, they will not be analyzed and compared to
other methods discussed herein.
Univariate methods
Univariate methods are based on fitting a model to the
historical data of a given time series and extrapolating to
obtain forecasts. The reasons for using univariate time series
methods are provided in Taylor and Bowen (1987). There
are many univariate methods available. They include
extrapolation of trend curves, averaging, stepwise
autoregression, and adaptative filtering. Some of these
simpler techniques have been examined by Taylor and
Bowen (1987) for building price-level forecasting. Extrapolation of trend is inherent in all the other univariate methods.
Exponential smoothing encompasses averaging and is comparable to adaptive filtering (Sullivan and Claycombe 1977).
Stepwise autoregression can be regarded as some form of
the Box-Jenkins method (Granger and Newbold 1977). For
these reasons, an examination of the use of the Box-Jenkins
method and exponential smoothing to forecast construction
cost indices should reveal the benefits and limitations of
using univariate methods to forecast construction cost
escalation.
Exponential smoothing
The most commonly used exponential smoothing methods
are the Holt-Winters family of models (Goodrich 1989).
These model time series use up to three components, representing level, trend, and seasonal influences. Recursive equations are used to obtain smoothed values for the model components. Each smoothed value of any model component is
a weighted average of current and past data with the weights
decreasing exponentially. The Holt-Winters family of
exponential smoothing models can be classified into three
classes, namely, simple exponential smoothing, Holt's twoparameter smoothing, and Winters' three-parameter
smoothing (Goodrich and Stellwagen 1987).
Simple exponential smoothing uses an equation to model
the level of the series of the form:

where a is the level smoothing parameter, Y, is the

observed value of the time series at time t, and L, is the
smoothed level at time t. This equation reduces to the recursive form:
[2]

L,

aYt

(1 - a)Lt-l

The forecasting equation is

where yt(,) is the forecast for lead time m from time t; and
m is the lead time of interest, usually in months.
Holt's two-parameter smoothing uses two equations to
model level and trend. These are given in their recursive form
by

VOL. 20,

1993

where TI is the smoothed trend at time t, y is the trend

smoothing parameter, and other parameters are as previously defined. The forecasting equation is

The multiplicative Winters three-parameter smoothing

involves three smoothing parameters for level, trend, and
seasonal effects. The smoothing equations are of the form:

where S, is the smoothed seasonal index at time t, p is the

number of periods in the seasonal cycle, and 6 is the seasonal
index smoothing parameter; other parameters are as previously defined. The forecasting equation is of the form:

s,(,)

is the last available smoothed seasonal index for

where
time t + m.
Simple exponential smoothing is appropriate for data that
fluctuate around a constant or have a slowly changing level
and neither are seasonal nor have any trend. The use of the
Holt's two-parameter model is appropriate for data that
fluctuate about a level that changes with some nearly constant linear trend. The Winters' three-parameter model is
used for data with trend and seasonal effects. The relevant
exponential smoothing equations can be adjusted to represent data that have a damped exponential rather than linear
trend (Goodrich 1989).
All exponential smoothing equations give more weight to
more recent values of data. The larger the values of the
smoothing parameters the more emphasis on recent observations and less on the past. This is intuitively appealing for
forecasting applications.
The smoothing parameters are normally obtained by
either using iterative least squares or a grid search for the
parameters that give the minimum squared error over the
historical data. This calculation process requires a great
number of computations which are normally incorporated
into a computer program.
Exponential smoothing models are robust in that they are
insensitive to changes in the data statistical structure
(Goodrich 1989). No assumptions about the statistical distribution of data are made in exponential smoothing, and
there is therefore no need to analyze the diagnostic statistics
given with most computer programs.
One of the main advantages of using exponential
smoothing is that once the smoothing parameters have been
estimated, only the previous forecast and the most recent
observation have to be stored or are necessary to make a
new forecast. This makes the calculation of a new forecast
computationally very convenient.
Box-Jenkins method
The Box-Jenkins method models time series by making
strong and explicit distributional assumptions about the
underlying data generating process (Box and Jenkins 1976).
The method uses a combination of autoregressive (AR),

605

ET AL.

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

integration (I), and moving average (MA) operations in the

general autoregressive integrated moving average (ARIMA)
model to represent the correlational structure of a univariate
time series.
An autoregressive operation of order p develops a forecast
based on a linear weighted sum of previous data represented
by
[ l l ] P, = ~ , Y , - I+ 42Ylp2 + ... + 4pYf-p + el
where PI is the forecasted value of the series at time t, Y,-i
is the observed value of time series at time t - i, c$~ is the
weighting coefficient of the ith previous period, and el is
the error term at time t. The coefficients are found by
minimizing the sum of squared errors usually using a
nonlinear regression routine.
A moving average operation of order q develops a forecast
which is a function of the previous forecast errors using an
equation of the form:
where 0; is the weighting coefficient for the ith previous
period, and the other terms are as previously defined.
The autoregressive and moving average operations can
only be applied to stationary time series. That is, they can
only be applied to data that have constant mean and standard deviation with time. If a time series is non-stationary,
it has to be transformed to a stationary series by a differencing operation before the AR and MA operations can be
performed. Forecast values have to be transformed back to
the original non-stationary state by the integration (I)
operation.
A three-step procedure of identification, estimation, and
diagnostic checking was originally proposed by Box and
Jenkins (1976) to select a model from the general class of
ARIMA models. This iterative process is depicted in Fig. 2.
The identification process is deciding the best ARIMA ( p ,
d , q) model to fit the data. This means identifying the degree
of differencing d, the AR orderp, and the MA order q. The
estimation process involves statistically estimating the model
parameters. The diagnostic step involves an examination of
the residuals to ensure that the ARIMA modeling assumptions of independence, homoscedasticity, and normality of
the residuals are not violated.
T o use the Box-Jenkins method, the data must have a
strong correlational behavior, and there should be sufficient
data to permit reasonably accurate estimates of the parameters. It is suggested that there should be at least 50 observations for good estimates (Box and Jenkins 1976).
The selected Box-Jenkins model, which satisfies the diagnostic checks, will generally fit the historical data well; the
parameters estimated describe the data on which they are
estimated. These parameters are estimates of unknown
parameters. Therefore when the forecasts using the model
are compared with future data not used in estimating the
model parameters, the fit may not be as good (Abraham
and Ledolter 1983).
Multivariate methods
Choice of the type of multivariate method
Multivariate methods forecast a given time series taking
into account observations of other variables. Generally,
these models use equations developed by regression to
represent the relationship between the dependent or
endogenous variable and the exogenous or explanatory vari-

model to be

I I parametersEstimate
I
in tentatively
Diagnostic checking
(is the model adequate?)

Use model for

forecasting or
control

I
FIG. 2. Stages in the iterative approach to Box-Jenkins model
building (from Box and Jenkins 1976).

ables. Multivariate methods are either single-equation

models or simultaneous-equation models.
In single-equation models, the values of the explanatory
variables determine the value of the dependent value and
the explanatory variables are not influenced by the values
of the dependent variable. Simultaneous-equation models
take into account the simultaneous dependency between the
dependent and explanatory variables.
The construction costs of any given single normal-sized
construction project seldom have any significant influence
on the market forces which cause changes in cost. As such,
single-equation models are most applicable to construction.
Therefore in this paper, only single-equation models will be
discussed.
Single-equation models can be of either nonlinear or linear
specification. A linear specification means that the dependent variable or some transformation of the dependent variable can be expressed as a linear function of the explanatory
variable or some transformation of the explanatory variable.
A model with a linear specification is the most appropriate
to use with construction cost indices because construction
cost components are generally additive.
The multivariate method most applicable to construction
costs is therefore the single-equation linear regression model.
'A normal-sized construction project is defined as a project of
between $2 and $5 million, with a completion time of less than
3 years (R.S. Means Estimating Guides).

606

CAN. J.

CIV. ENG. VOL. 20.

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

1993

dynamic features of autoregression and the effects of

explanatory variables. In dynamic regression, the dynamic
portion of the model, that is, the lagged dependent variables
and the autoregressed error terms, must be determined term
by term by hypothesis testing. The specification of the
explanatory variables must be such that all the necessary
explanatory variables are included. It should further be
verified that all variables included in the model are statistically significant and that underlying assumptions are not
violated.
Dynamic regression is generally used when the data available are long enough or stable enough to support a correlational model and the explanatory variables result in a definite
increase in accuracy. The various measures of accuracy will
be given in the examples that follow.

Attributes of regression models

Regression models have the advantage of being amenable
to the investigation of various "what-if" scenarios. This is
appealing because one can tell the influence of a change in
an explanatory variable of interest, whether known, foreseen, or probable.
One shortcoming of regression models is that if some
100 0
crucial explanatory variable has not been varying in the past,
1 7 1 7 1 7 1 7 1 7 1 7 1 7
it is not possible to include it in the model. Thus the effect
1 1983 1 19B4 1 1985 1986 1987 1988 '198g1 of a change in such a variable cannot be assessed.
problems are encountered in the use of regression models
MonthIYear .
in cases where the values of the explanatory variables for
FIG. 3. Plot of the prefabricated wooden building industry
use in obtaining forecasts are not known or required to be
index (PWBII).
forecast. This will occur when there is no lag relationship
between the explanatory and dependent variables or when
This regression model is of the form (Pindyck and Rubinfeld
the lag is not sufficient.
1976):
Application of forecasting techniques
From the above discussion of the various forecasting
where Y, is the dependent variable at time t, Pi are the coefmethods, it would appear that statistical expertise would be
ficients of Xi,, Xi, is the observed value of the ith
required to apply these methods. Fortunately, available comexplanatory variable at time t, and e, is the error term at
puter software reduces the application process to the intertime t.
pretation of the computer output and running various trials
;o
obtain the best forecasting model. One such advanced
Requirements for the use of regression models
statistical forecasting package is FORECAST PRO (BFS 1988)
The single-equation linear regression model assumes that
developed by Business Forecast Systems, Inc., which is used
the residuals are normally distributed random variables with
in the examples that follow.
a mean of zero and a constant variance. It is also required
The data used consist of 81 monthly values of the
that the explanatory variables are linearly related to the
prefabricated wooden building industry index (PWBII)
dependent variable (or can be transformed into some linear
published by Statistics Canada for the period January 1983
relation), and that explanatory variables are not collinear,
to September 1989. This is an output index based on industhat is, they are not correlated to one another.
try selling price for completed structures. The authors feel
Building regression models requires theoretically plausible
that this index could not be materially changed by new techexplanatory variables and sufficient historical data to estinology or code changes in the brief period (6 years) conmate the model parameters. The models have to be tested
sidered. A time plot of the data used is shown in Fig. 3. The
through the examination of various statistical diagnostics
first 72 observations are initially used for fitting the various
to ensure that the assumptions on which the models are
models and the last 9 observations are used to compare the
based hold.
forecasts obtained using various methods. From a visual
inspection, it can be seen that the series is non-stationary,
Improvement to the linear regression model
has an upward trend, and appears to be seasonal in the last
The true relationship between the explanatory variables
half of the data. For such data, both Winters threeand the dependent variable is rarely known. Use is therefore
parameter exponential smoothing and Box-Jenkins models
made of empirical evidence to develop an approximate relaare applicable. Dynamic regression can also be applied to
tionship. Thus, the explanatory variables may not suffiobtain forecasts if suitable explanatory variables can be
ciently account for the variation in the dependent variable,
found. All these methods will be used and the best method
in which case the use of dynamic regression may improve
will be chosen from the resulting forecasts. The use of anathe model.
The term dynamic regression is adopted to represent
lytical methods to forecast a construction cost index will thus
multivariate models that combine the time series oriented
be demonstrated by these examples.

BLAIR ET

AL.

607

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

TABLE1. Computer output for exponential smoothing model

Historical fit of exponential smoothing
model
Dependent variable:
R-squared:
Adjusted R-squared:
Standard forecast error:
F-statistic:
Durbin-Watson:
Ljung-Box:
Standardized AIC:
Standardized BIC:
Three-parameter Winters
smoothing parameter values
Level:
Trend:
Seasonal:

PWBII
0.992
0.992
0.710986
3020.771 (1.OOO)
1.784
14.733 (0.744)
0.725630
0.760876

Measures of goodness-of- fit and diagnostic statistics

T o aid the selection of the best forecasting model, some
form of objective statistical criteria are required. The
statistical criteria fall into two categories: goodness-of-fit
statistics and diagnostic statistics. The goodness-of-fit
statistics used in FORECAST PRO are the R-squared, adjusted
R-squared, and the standard forecast error. The R-squared
statistic indicates the percentage of variance that is explained
by the model. An R-squared of l .OO (100%) would indicate
a perfect fit of the model to the historical data. The adjusted
R-squared statistic adjusts the R-squared value to reflect the
number of parameters used in the model. For univariate time
series models, the adjusted R-squared may be the same as
the R-squared value. Another indication of model fit is the
standard forecast error which is calculated by taking the
square root of the average of the squared one-step ahead
forecast error. The smaller the standard forecast error the
better the fit. A standard forecast error of 0.0 would indicate
perfect fit.
The diagnostic statistics used in FORECAST PRO include the
F-statistic, Durbin-Watson statistic, Ljung-Box statistic, the
Akaike information criterion (AIC), and the Bayesian information criterion (BIC). The F-statistic which is the result
of a F-test indicates whether the null hypothesis that the
actual model parameters are 0 holds. The percentage value
of the test statistic is given in parentheses in the output
tables. A percentage value greater than 0.95 would indicate
that the actual model parameters are statistically different
from 0 at the 95% confidence level. The Durbin-Watson
statistic and the Ljung-Box statistic test whether or not the
residuals are autocorrelated. Interpretation of the DurbinWatson and the Ljung-Box statistics requires reference to
statistical tables. The Durbin-Watson statistic indicates the
acceptance of the null hypothesis that there is no serial
correlation in the first lag. A value close to 2 is desirable.
While the Durbin-Watson test checks for first-lag autocorrelations, the Ljung-Box test checks for autocorrelations
up to the maximum of lag L. The percentage value of the
Ljung-Box test statistic is given in parentheses in the output
tables. A percentage value less than 0.95 would indicate that
the residuals are not correlated. The two tests should be used
as complements of each other (Box and Jenkins 1976). The
AIC and BIC statistics evaluate how well a given model will
perform relative to another. They reward goodness of fit
to historical data and penalize model complexity. Based on

Month /Year
A

Actual Index Values

Winters f i t

Frc. 4. Forecast comparison: Winters three-parameter

exponential smoothing model.

empirical research, the model with the lowest AIC or BIC

will generally be the most accurate (Goodrich and Stellwagen
1987).

Exponential smoothing modelling

Winters three-parameter exponential smoothing model
was fitted t o the data using FORECAST PRO. The results are
shown in Table 1. Since no statistical distribution assumptions have been made about the data, it is not necessary to
closely scrutinize all the diagnostic statistics produced by the
software package.
The smoothing parameter values are obtained by
FORECAST PRO using an iterative search method to minimize
the squared errors over the historical data. The computerized
iterative search employs the simplex method of nonlinear
optimization.
Examining the exponential smoothing parameters reveals
that the seasonal parameter value is close to 0.5, indicating
that the best forecast for the next seasonal effect is half the
last seasonal effect and a weighted average of preceeding
seasonal effects. The small trend parameter value of 0.05
indicates that the smoothing model has a long memory of
trend, and distant trends have a n effect on the forecasted
trend component. The large value of the level parameter
indicates that the model is highly adaptive to the last
observed level of the series.
Figure 4 shows the plot of the fitted and forecasted values
as compared to the actual figures. Though the model gives
a visually good fit to the historical data, it does not forecast
the turning point after the last fitted value. A 1-month lag
between the fitted and actual values is apparent. This shows
the strong influence of the previously observed level.

CAN. J.

608

CIV. ENG.

VOL. 20, 1993

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

TABLE
2. Computer output for Box-Jenkins model
Historical fit of Box-Jenkins model
Dependent variable:
R-squared:
Adjusted R-squared:
Standard forecast error:
F-statistic:
Durbin-Watson:
Ljung-Box:
Standardized AIC:
Standardized BIC:
Box-Jenkins model parameters
B[l]:
Simple difference:

PWBII
0.991
0.990
0.763177
7298.05 1 (1 .OOO)
2.066
11.454 (0.510)
0.768608
0.781052

TABLE3. Computer output for dynamic regression model:

initial trial
Historical fit of dynamic regression model
Dependent variable:
R-squared:
Adjusted R-squared:
Standard forecast error:
F-statistic:
Durbin-Watson:
Ljung-Box:
Standardized AIC:
Standardized BIC:

PWBII
0.965
0.962
1.564319
303.379 (1 .OOO)
0.557
120.896 (1 .OOO)
1.627879
1.789863

Variable Coefficient Standard error T-statistic Probability

UWRI
CBMPI
CBLR
S&PMPI
AM1
Constant

0.361754
0.351710
0.064706
0.000548
0.120562
7.109199

0.073729
0.137117
0.044185
0.039341
0.129265
5.230716

4.907
2.565
1.464
0.014
0.933
1.359

1.OOO
0.990
0.857
0.011
0.649
0.826

Box- Jenkins modelling

Without software packages like FORECAST PRO, considerable skill and judgement is required to identify a suitable
Box-Jenkins model to fit the data. The originally proposed
method of identification through an examination of autocorrelations and partial autocorrelations requires specialist
knowledge (Firth 1977). This specialist knowledge would
have to be provided by professional statisticians. FORECAST
PRO identifies an appropriate model automatically by the
BIC statistic.
The use of these criteria in a computer package relieves
the user from the lengthy identification, estimation, and
diagnostic steps in the Box-Jenkins methodology. The
results of an application to the PWBII are shown in Table 2.
Because of the strict distributional assumptions in the BoxJenkins model, an examination of the diagnostic statistics
is required.
The R-squared statistic of 0.991 indicates that 99% of the
variation of the index is explained by the model. This shows
that the model fits the historical data very well. The
F-statistic is highly significant (as indicated by the number
1.000 in parentheses). This rejects the null hypothesis that
the actual model parameters are 0.
The Ljung-Box statistic is not significant at both the 95%
confidence level and the 99% confidence level. There is
therefore no evidence of serial correlation in the first several
lags of the residuals.

1 7 1 7 1 7 1 7 1 7 1 7 1 7
11983 1 1984 1 1985 1 1986 1 1987 1 1988 119891

Month /Year

-Actual index Values

-t-

Box-Jenkins f i t

FIG.5. Forecast comparison: Box-Jenkins model.

The autocorrelations of the residuals were examined using
and were found to exhibit no systematic
pattern. They were also small in magnitude, being less than
2 times the standard error.
A comparison of the AIC and BIC for the Box-Jenkins
and exponential smoothening models indicates that on the
basis of these criteria the exponential smoothing model is
vrobablv more accurate. However, this is not conclusive
becausethe most important criteridn in forecasting is how
good the forecasted values fit the actual values and not the
goodness of fit to historical data.
Figure 5 shows the plot of the actual, fitted, and forecast
values. The model gives a visually good fit to the historical
data, but does not forecast the turning point in the data right
after the last fitted value.
FORECAST PRO

Dynamic regression modelling

Dynamic regression requires additional variables which
are correlational significant and theoretically plausible to
explain the dependent variable. In this example, various
indices published in Statistics Canada Catalogue 62-007 were
used as possible explanatory variables to forecast the
PWBII. These include the union wage rate index (UWRI),
the construction building materials price index (CBMPI),
the commercial bank lending rate index (CBLR), the sawmill
and planing mill products index (S&PMPI), and the architectural materials index (AMI). These variables were plotted
and their plots compared with the PWBII. Possible lag relationships and possible transforms were sought through a
visual inspection of the plots.
Table 3 shows the results of the initial trial. The value
of the R-squared statistic is quite high (97%). This is probably due to trying to develop a model using a large number
of explanatory variables. Each additional explanatory variable can only increase the R-squared value, but may not
improve the model's forecasting accuracy.

BLAIR ET

AL.

609

TABLE4. Computer output for dynamic regression model: final

trial

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

Historical fit of dynamic regression model

Dependant variable:
R-squared:
Adjusted R-squared:
Standard forecast error:
F-statistic:
Durbin-Watson:
Ljung-Box:
Standardized AIC:
Standardized BIC:
Variable

PWBII
0.992
0.992
0.658523
2456.983 (1 .OOO)
2.032
14.390 (0.724)
0.674757
0.71 1028

Coefficient Standard error T-statistic Probability

UWRI[- 51 0.082996
AMI[- 121 0.111773
PWBII[- I] 0.797042

0.033897
0.031417
0.059730

2.448
3.558
13.344

0.986
1.000
1.000

The Durbin-Watson and Ljung-Box statistics indicate

problems of serial correlation. This means the model has
to be adjusted because the assumptions on which it is based
do not hold.
The coefficients of four of the explanatory variables are
not significant at the 95% confidence level, implying that
these variables should be omitted, transformed, or lagged.
Some of these explanatory variables may appear not to be
statistically significant because of multicollinearity, that is,
these explanatory variables may be correlated with one
another. FORECAST PRO, however, does not provide tests for
multicollinearity.
Various trials using lagged dependent and transformed
explanatory variables were performed until a near optimum
regression model was obtained as one possible solution.
There are a number of available selection strategies that can
be used to empirically construct the required regression
model. For this example, the backward elimination procedure, whereby the selection process started with the largest
possible model, was used. Other selection strategies include
forward selection and stepwise regression (Abraham and
Ledolter 1983). The choice of the search strategy to use
depends on the forecaster's individual preference, since all
methods produce a regression model with the necessary
statistical attributes. The result of the backward elimination
selection strategy is given in Table 4. The explanatory variables found to give the statistically satisfactory model shown
in Table 4 are the following:
UWRI[- 51, the UWRI lagged 5 months; AMI[- 121,
the AM1 lagged 12 months; and PWBII[ - 11, the dependent
variable lagged 1 month.
This model implies that the value of the PWBII at any
given time can be forecast using values of the AM1
12 months before, the UWRI 5 months before, and the
PWBII 1 month before. These results seem to indicate the
time periods for the input variables to affect the output
variable, and are not unrealistic given the time for material
inventory to work through.
The signs of the coefficients of the explanatory variables
are all positive. This agrees with a priori expectations that
the PWBII will increase with increases in the UWRI and the
AMI. In this model there is no problem of serial correlation,
as evidenced by the Durbin-Watson and Ljung-Box
statistics. All independent variables are significant at the
95% confidence level. This model explains 99% of the varia-

7 1 7 1 7 1 7 1 7 1 7
1984 1 1985 1 1986 1 1987 1 1988 119891

Month /Year

Actual index Values

Regression f i t

FIG. 6. Forecast comparison: dynamic regression.

tion in the PWBII, as indicated by the R-squared statistic,

and therefore fits the historical data well. Figure 6 shows
the plot of the fitted and forecast values as compared with
the actual values. The regression model appears to attempt
to account for the turning point in the data after the last
fitted value. This example demonstrates some of the problems of using regression models. Though, from theoretical
considerations, interest rates affect construction costs, the
CBLR was not found statistically significant enough to be
used in the model to forecast or explain the variation in the
PWBII. Thus factors that are known to affect construction
costs may not be found useful in forecasting by regression
models even when data on the variation of these factors
exist.
Choice of the forecasting model
In order to decide which of the three fitted models would
be best to use for forecasting the PWBII, the following
criteria are used:
1. The accuracy of the forecasts and not the historical fit
to the data will be considered.
2. A more complex model will only be recommended if it
undoubtedly improves the forecasting accuracy.
Various methods of measuring forecast accuracy exist.
Mahmoud (1984) surveys the following: the mean square
error, the mean percentage error, the mean absolute percentage error, Theil's U-statistic, the root mean square error,
the mean error, the mean absolute deviation, turning points,
and hits and misses.
In forecasting construction cost escalation, it is most convenient to use a measure of accuracy consistent with the
normal measure of accuracy used in construction cost
estimation. The measure of accuracy used in construction
cost estimation is normally the percentage error of the estimate. Of particular interest to the cost estimator would also
be the turning points in the time series where the rate of cost

CAN. J. CIV.

ENG. VOL. 20,

1993

TABLE
5. Mean absolute percentage error of various forecast models and scenarios*

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

Period

Actual
PWBII

Jan. 1989
Feb. 1989
Mar. 1989
Apr. 1989
May 1989
June 1989
July 1992
Aug. 1992
Sept. 1992

Exponential
smoothing
forecast

Absolute
error

133.40
133.43
133.90
134.48
134.81
135.04
135.67
136.01
136.36

2.00
2.47
2.00
1.12
0.69
0.56
0.07
0.41
0.76

Box-Jenkins
forecast

Absolute
error

Regression
forecast

Absolute
error

*Nine-month forecast beginning with January 1989.

t ~ . ~ . ~mean
. ~ absolute
. ,
percentage error.

TABLE
6. Mean absolute percentage error of various forecast models and scenarios (continued)*

Period

Actual
PWBlI

Exponential
smoothing
forecast

Absolute
error

Box-Jenkins
forecast

Absolute
error

Regression
forecast

Absolute
error

July 1988
Aug. 1988
Sept. 1988
Oct. 1988
Nov. 1988
Dec. 1988
Jan. 1989
Feb. 1989
Mar. 1989

M.A.P.E.
*Nine-month forecast beginning with July 1988.
t ~ . ~ . ~mean
. ~ absolute
. ,
percentage error.

TABLE
7. Mean absolute percentage error of various forecast models and scenarios (continued)*

Period

Actual
PWBlI

Exponential
smoothing
forecast

Absolute
error

Box-Jenkins
forecast

Absolute
error

Regression
forecast

Absolute
error

Oct. 1988
Nov. 1988
Dec. 1988
Jan. 1989
Feb. 1989
Mar. 1989
Apr. 1989
May 1989
June 1989

*Nine-month forecast beginning with October 1988.

t ~ . ~ . ~mean
. ~ absolute
. ,
percentage error.

escalation changes. As such, to measure the accuracy of

forecasts, the mean absolute percentage error and the precision with which a method forecasts turning points is
examined.
In terms of complexity, exponential smoothing is the
simplest method to apply. The other methods should
therefore be tested to examine whether they are undoubtedly

more accurate and forecast the turning points much better

than exponential smoothing.
To guard against spurious accuracy, three forecasting
scenarios were used. These scenarios were obtained by fitting
the different models to the data and forecasting 9 months
ahead but using three different last fitted values. The three
forecasting scenarios are forecasting 9 months ahead begin-

61 1

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

BLAIR ET AL.

ning with January 1989, July 1988, and October 1988. The
various forecasts are given in Tables 5-7. These tables show
that Winters three-parameter exponential smoothing yields
superior forecasts in two of the three forecasting scenarios
and the regression model gives a superior forecast in the
other.
Exponential smoothing is the simplest of the three types
of forecasting models. From the evidence in Tables 5-7,
none of the more complex models give undoubtedly better
forecasts of the PWBII than the Winters three-parameter
model. Based on the previously stipulated criteria, exponential smoothing would be the best method to use in these circumstances to forecast the PWBII.
It should be noted that due to the specification of the
regression model, forecasts of the PWBII can only be made
5 months ahead without having to first obtain forecasts of
the UWRI. In obtaining the 9-month forecasts given in
Tables 5-7, actual values for the UWRI were used up to
4 months ahead. This gave the regression forecast for the
last 4 months of the 9-month forecast period unrealistic
accuracy. Nonetheless, the regression forecasts were less
accurate than those from exponential smoothing in two of
the three scenarios. It is therefore not necessary to investigate the accuracy of the regression model when 4-month
forecasts instead of actual values of the UWRI are used.
Nevertheless, the regression model could still be useful
in some circumstances, such as in the case when the
forecaster knew of ongoing labour union negotiations which
were likely to increase the UWRI by some estimated amount.
The regression model could be used to forecast the effect
of this increase on the PWBII. This illustrates one of the
benefits of developing a regression model.
Conclusion
Forecasting construction cost escalation is important
because escalation accounts for a substantial part of the costs
of many construction projects. With the availability of user
friendly forecasting software, many complex statistical
forecasting techniques can now be used to forecast construction cost escalation. This can be done provided the practitioner can interpret the results produced by these software.
Univariate time series are based on the assumption that
existing patterns in the data will continue. Therefore they
cannot usually predict turning points. They are not recommended if there is reason to believe existing conditions will
change dramatically.
Multivariate forecast methods are dependent on the
accuracy of the explanatory variables used in the forecasts.
One of the main difficulties in their use is the identification
of statistically significant explanatory variables. The accuracy of the multivariate forecasts produced depends on the
accuracy of the explanatory variables used to make the
forecasts.
The analytical forecasting techniques discussed herein are
only valid for short-term forecasting, generally less than
1 year ahead. No analytical forecasting technique reviewed
is capable of long-term forecasting of cost escalation.
Being able to give quantitative forecasts of escalation does
not eliminate the risk caused by cost escalation. This is so
because none of the discussed techniques can forecast escalation caused by unpredictable occurrences, which include
major events like the outbreak of war or government action.
The analytical methods available are only useful in fore-

casting for short construction projects in stable conditions.

Construction cost escalation, with or without the use of these
forecasting methods, still remains a risk to be borne by either
the contractor or the owner, or both, depending on the terms
of the construction contract.
Acknowledgement
The authors would like to thank one of the anonymous
referees for providing the reference for Taylor and Bowen
(1987), and for providing insightful comments and
suggestions.
Abraham, B., and Ledolter, J. 1983. Statistical methods for
forecasting. John Wiley and Sons, New York, N.Y.
BFS. 1988. FORECAST PRO. Business Forecast Systems, Inc.,
Belmont, Mass.
Box, G.E.P., and Jenkins, G.M. 1976. Time series analysis for
forecasting and control. Holden Day, San Francisco, Calif.
Chatfield, C. 1975. The analysis of time series: theory and practice.
Chapman and Hall, London, United Kingdom.
Firth, M. 1977. Forecasting methods in business and management.
Edward Arnold (Publishers) Ltd., London, United Kingdom.
Goodrich, R.L. 1989. Applied statistical forecasting. Business
Forecast Systems, Inc., Belmont, Mass.
Goodrich, R.L., and Stellwagen, E.A. 1987. Forecast Pro statistical
reference. Business Forecast Systems, Inc., Belmont, Mass.
Granger, C. W.T., and Newbold, P. 1977. Forecasting economic
time series. Academic Press, New York, N.Y.
Mahmoud, E. 1984. Accuracy in forecasting: a survey. Journal
of Forecasting. 3: 139-159.
Nelson, C.R. 1973. Applied time series analysis for managerial
forecasting. Holden-Day Inc., San Francisco, Calif.
Pindyck, R.S., and Rubinfeld, D.L. 1976. Econometric models
and economic forecasts. 2nd ed. McGraw-Hill Book Company,
New York, N.Y.
Stevenson, J. J. 1984. Escalation modelling, forecasting and tracking. Transactions of the American Association of Cost Engineers,
28th Annual Meeting, Montreal, Que., pp. F.3.1-F.3.7.
Sullivan, W.G., and Claycombe, W.W. 1977. Fundamentals of
forecasting. Reston Publishing Co. Inc., Reston, Va.
Taylor, R.G., and Bowen, P.A. 1987. Building price-level
forecasting: an examination of techniques and applications.
Construction Management and Economics, 5: 21-44.
Tham, T.B. 1980. Fast relief for the pain of contract cash flows.
Transactions of the American Association of Cost Engineers,
24th Annual Meeting, Washington, D.C., pp. A.2.1-A.2.5.

AIC
AM1
AR
ARIM A
BIC
CBLR
CBMPI
d
el

I
L,
MA
P
Sl

S&PMPI
T1
UWRI

List of symbols and abbreviations

Akaike information criterion
architectural materials index
autoregressive
autoregressive integrated moving average
Bayesian information criterion
commercial bank lending rate index
construction building materials price index
degree of differencing
error term at time t
integration
smoothed level at time t
moving average
number of periods in the seasonal cycle
smoothed seasonal index at time t
saw mill and planing mill products index
the smoothed trend at time t
union wage rate index

612

YI
Y/(m)

YI

Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by 142.243.254.209 on 01/13/15

For personal use only.

YI-i
xi,

CAN. J . CIV. ENG. VOL. 20, 1993

observed value of time series at time t

forecast for lead time m from time t
forecasted value of time series at time t
observed value of time series at time t - i
observed value of the ith explanatory variable
at time t

01

Pi

Y
4;

O4

level smoothing parameter

coefficient of Xi
seasonal index smoothing parameter
trend smoothing parameter i
weighting coefficient of the ith previous period
weighting coefficient for the 4th previous period