Figure 8.2 Explanatory or Causal Relationship
8.6 A critique of the prevailing forecasting methods used for construction demand modelling and prediction
The review o f past works in construction demand modelling revealed that all the studies had adopted
in the studies by Oshobajo and Fellows (1989), and Killingsworth (1990), it only served as a means
o f selecting the independent variables for the regression models. Strictly speaking, the Leading
Indicators technique is intended only to be used to forecast the timing o f turning points and not the
size of the forthcoming downswing or upswing (Ascher, 1978; and Granger and Newbold, 1987).
Hence, it would be misleading to suggest that the classical Leading Indicators technique had been
used in these two studies.
Based on these reviewed studies, it is possible to generalise that the multiple regression technique
is the predominant method employed in the current state o f construction demand modelling and
forecasting research. Although the frequent application o f one particular method should not be
considered as a limitation, the lack o f objectivity in its choice would be. Hence, there is a need to
evaluate the suitability of the multiple regression technique by comparing it with other forecasting
methods.
The main objective o f developing demand models for construction is to facilitate predictions o f
future levels o f demand. Forecasting literature (Makridakis et al., 1982; Lawrence, 1983; Mahmoud, 1984; and Thomas, 1993) have always cited accuracy as the most vital attribute o f a
prediction model, since an important consideration in the decision to use a quantitative model is its accuracy. They have also highhghted that this attribute plays an important role in the selection and
testing o f different forecasting techniques. Empirical studies (Armstrong, 1978; Makridakis and
Hibon, 1979; Slovic, 1972; and Makridakis et al., 1982) have also shown that levels o f forecasting accuracy vary according to the type of method adopted, emphasizing the need to explore different
techniques in order to establish the most accurate one. More often, in construction demand
modelling, when a model is not producing accuracy forecasts, it is attributed to the poor choice or
inadequate use o f modelling variables (Akintoye and Skitmore, 1994), or the presence of
autocorrelated errors (Killingsworth, 1990). The suitability o f the forecasting method was never in
doubt. If it is true that the level of accuracy largely depends on the type o f forecasting method used,
construction demand modelling research to date has failed to focus on an issue fimdamental to its
existence and, that is, to compare and evaluate the forecasting accuracy o f the methods used to model
and predict demand. So far, only Oshobajo and Fellows (1989) have attempted to compare the
accuracy o f two forecasting methods, namely, the pseudo Leading Indicators technique and the
univariate Box-Jenkins approach.
shortcoming o f the regression method is that it forces models into a form whereby they can be treated
by linear least squares during the model estimation stage (Jenkins 1979). This is generally regarded
as inefficient because the linear fimctional assumption may not be capable o f modelling the true
relationship o f the variables. In reahty, relationships are seldom linear, making this functional
assumption even more tenuous. In order to account for non-linearity, some demand models (Tang
et a l , 1990; and Akintoye and Skitmore, 1994) have assumed a log-linear form. Again, the imposition o f a functional form to fit the data would essentially cause the latent relationship o f the
variables to be ignored. Regarding modelling methods that rely on functional assumptions, Newbold
and Bos (1990) indicated that it is quite unrealistic to hope that any simple model and associated
assum ptions can provide a perfectly accurate description o f the complexity that exists in the real
world.
The issues discussed in this section have highlighted two major limitations o f the current state of research in construction demand modelling and forecasting. Firstly, it is felt that the current state
of research has been disadvantaged through not having a formal means o f critically evaluating one
forecasting method relative to another. This is largely caused by a general lack o f interest in
exploring alternative methods of modelling and forecasting construction demand. The focus o f past
studies seems to be on establishing the procedures o f building the model, without taking a step
fiirther to examine the performance of the model. Model building should not be treated as a means
to an end. It should be supported with a formal testing procedure, using accuracy measures as a
basis o f evaluation and comparison. Secondly, it is of the opinion that the prevalent use o f the linear
form in regression models produces demand models that may well fit the sample data but then
exhibit miserable predictive performance. This is because the predictive ability o f the model relies
heavily on the precision o f the functional specification, emphasizing the importance o f this task.
Any inaccuracy in this respect essentially undermines the whole basis o f this predictive model. To
assume that relationships take on simple forms like linear or log-linear is simply ignoring the fact
that complexities exist in the real world. Hence, the practicality o f these demand models is gravely
in doubt.
8.7 Chapter summary
This chapter has provided a broad overview of forecasting, ranging fi'om the general need to forecast
In Section 8.2, two main reasons were cited for the need to forecast. Firstly, forecasts are needed in
order to reduce the degree o f uncertainty in the future; and secondly, accurate predictions o f the
fiiture improve the efficiency o f the decision-making process. Several examples were also given to
justify this need.
A theoretical overview of quantitative forecasting methods was provided in Section 8.3. There are
essentially two main groups o f quantitative forecasting methods, time series and causal methods.
Time-series methods assume explicitly that the underlying pattern o f a data series can be identified
solely on the basis o f its historical data. In contrast, causal methods rely on the cause and effect
relationship between the dependent and independent variables. As such, they are also known as
explanatory models. Regression techniques are most commonly applied to develop causal models.
They range from simple linear regression models to complex econometric models.
Focusing on building causal models using regression methods. Section 8.4 elaborates on the
theoretical aspects o f the regression approach. Basic principles o f the classical linear regression, regression applying linearizing transformation and non-linear regression algorithms were discussed.
It was noted that the possibility o f fitting non-linear models has greatly increased the scope and
flexibility o f regression analysis.
Section 8.5 proceeded to describe several construction demand modelling and forecasting
applications. The review o f past works in demand modelling and forecasting focused on more
recently developed models as they had adopted more sophisticated techniques. Models prior to them
mainly used simple extrapolative methods which did not produce good forecasts. The first study
employed the pseudo Leading Indicators technique to model and predict UK contractors' workload.
The univariate Box-Jenkins approach was adopted as a benchmark method for comparison with the
causal models based on their predictive ability. It was found that the causal model built using
leading indicators could improve forecast accuracy by 5 to 12 per cent. The second study involved
the use o f multiple regression techniques to predict demand on the Thai construction industry.
Factors affecting the demand for residential, non-residential and 'other' construction were used as
independent variables to build the demand models by adopting multiple linear and log-linear
regression analysis. The demand functions were subsequently used to project the growth of the
construction industry for the next five years. The third study formulated a demand forecasting model
for industrial construction using US leading indicators. Multiple regression analysis was also used
The final study adopted leading indicators and regression methods to model and predict UK's private
sector quarterly construction demand. Multiple log-linear regression analysis was applied to develop
demand models for private residential, commercial and industrial construction.
Based on the review o f past works, a critique o f the prevailing forecasting methods used to model
construction demand was provided in Section 8.6. Several issues were highlighted and discussed.
Firstly, it was noted that the multiple regression technique is the predominant method used to
develop construction demand models. Therefore, there is a need to evaluate this method by
comparing its predictive ability with other forecasting methods in order to justify its suitability.
Secondly, it was observed that a linear form was mainly assumed in these models. Otherwise, a log-
linear form was specified for non-linear models. In either case, functional assumptions were used,
ignoring the latent relationship o f the variables. In this respect, it was argued that these models are
impractical because modelling methods that rely on functional assumptions, such as regression
techniques, seldom provide an accurate description o f the complexity o f real-world situations. Finally, it was concluded that construction demand modelling research to date had failed to focus on
the fimdamental issue of comparing and evaluating the forecasting accuracy o f the methods used to model and predict demand. It was attributed to the lack o f a formal means o f critically evaluating
CHAPTER 9
ARTIFICIAL NEURAL NETWORKS: