approach for estimating construction cost.

However, the application of ANNs in construction

is a relatively new research area (Kim et al., 2004).
Thus the objectives of this study are:
1. to develop an learning model (Artificial Neural Network) for predicting the
accuracy of pre-tender building cost forecasts thereby improve the estimation
of project contingency reserve.
2. to offer a new methodology and tool to complement existing methods used for
improving the accuracy of pre-tender building cost forecasts.

This study is important because the problem of estimate inaccuracies is reflected by

the increasingly large number of projects being completed with cost overrun. If the amount of
inaccuracy in a pre-tender estimate can be predicted, cost advisors would be able to develop
more appropriate contingency reserve for projects. Project owners can be assured of the costs
of their project early in the project development process.

ARTIFICIAL NEURAL NETWORK AND COST ESTIMATION

ANNs are purely data driven models which through training iteratively transition from a
random state to a final model (Hasangholipour and Khodayar, 2010). ANN doesn’t depend
on assumptions about functional form, probability distribution or smoothness (Camargo et al,
2003).
The advantages of using ANN include: It allows the learning from previous project
cost estimates and outcomes (actual project completion cost). It can model a complex set of
relationship between the dependent variables (i.e. output) and the independent variables (i.e.
input variables and in this study the drivers of estimate accuracy. Neural network can also
accommodate multicollinearity in the independent variables.
ANNs are data-driven self-adaptive methods in that there are few a priori assumptions
about the models for problems under study. They learn from examples and capture subtle
functional relationships among the data even if the underlying relationships are unknown or
hard to describe. Thus ANNs are well suited for problems whose solutions require knowledge
that is difficult to specify but for which there are enough data or observations (Zhang et al,
1998).

ANNs can generalize. After learning the data presented to them (a sample), ANNs can often
correctly infer the unseen part of a population even if the sample data contain noisy
information. As forecasting is performed via prediction of future behavior (the unseen part)
from examples of past behavior, it is an ideal application area for neural networks, at least in
principle (Zhang et al, 1998).
At the pretender stage of a project, cost forecasting would depend on limited, noisy
and approximate information. At that stage, it is also difficult to understand the underlying
cost drivers. Also, the relationship between the cost drivers and the cost outcomes could be
significantly nonlinear.
ANN can be used to predict project cost overruns and thereby assist management in
developing an appropriate contingency (Chen and Hartman, 2000). Examples of the
application of ANNs to predict the level of cost overrun/underrun include: Chen and Hartman
(2000) used ANN to predict the final cost of completed oil and gas projects from one
organisation using 19 risk factors as the input data. It was found that 75% of the predicted
final cost aligned with the actual variance i.e. where the ANN model predicted an
overrun/underrun, an overrun/underrun actually occurred. The prediction accuracy of ANN
outperformed multiple linear regression. Chau et al (1997) used 8 key project management
factors to predict the final cost of construction projects. It was found that more than 70% of
the examples did not differ by more than one degree of deviation from the expected.
Gunaydin and Dogan (2004) used 8 design parameters to estimate the square metre cost of
reinforced concrete structure systems in low-rise residential buildings and found that the
ANN provided an average cost estimation accuracy of 93%.The research on the application
of ANN to predict cost performance often compares the accuracy of ANN with multiple
linear regression and in most cases ANN produce more accurate predictions (e.g. Chen &
Hartman, 2000; Sonmez, 2004; Kim at al, 2004; Baccarini (2005)
A single biological neuron is not intelligent. A collection of those neurons is made
intelligent by making cooperate actions. Collection as a network creates a pattern of inputs to
a neural network and processed as a pattern and results as a pattern. The artificial neuron has
been modeled mimicking biological neuron similar way working together to produce
remarkable results. So Artificial Neural Networks (ANN), is a mathematical model that was
developed based on the phenomenon of error minimization. A processing element in ANN,
was arranged as a simple model of biological neuron. ANN learning occurs as given in
Equation (i), which simply represents the cost function of a desired (actual) and ANN output.

δ i p = d i p − y ip .....................................(i)

where
d i p is the desired ith output of the pattern p and y i p is the network’s ith output
received from the neuron system. The error for all patterns in equation (i) is measured using
mean square error (MSE), which is a major performance measurement in ANN learning
shown in equation (ii). Lower MSE indicates higher learning of the set of input pattern.
∑ ∑ (d )
p 2
i − y ip
p i
MSE = .......................(ii)
N *P
where N is number of input data sets and P is number of processing elements.
The learning (training) of ANN model from given inputs and outputs occurs through
the iterations. Equation iii shows the network output y of an ANN calculated from
n elements of an input pattern x through a summation of weighted inputs and a transfer
function.

 n 
y = F  ∑ w i x i + θ  ................................iii)
 i =1 
Where xi denotes ith element of the input pattern x, wi is the weight for the input xi, θ is the
offset and F is a transfer function, which is a smoothing function.

RESEARCH METHOD

Development of Learning ANN Model

Six steps were followed in the development of the learning ANN model for predicting
accuracy of estimates namely:
(1) Defining the ANN output variable – estimate accuracy (or bias)
 (2) Identifying the ANN input variables – the characteristics of a building project that
could influence the accuracy of pre-tender building cost estimates
(3) Data collection on completed building projects
(4) Development of the learning ANN model
(5) Testing the ANN model: i.e. predicting the accuracy of estimates using new data
set.
(6) Evaluating the performance of the learning model and sensitivity analysis.

Defining the output variable – Estimate accuracy (bias)

Skitmore (1991) describes the accuracy of early stage estimate as comprising two aspects,
namely, bias and consistency of the estimate when compared with the contract or accepted
tender price. Bias is concerned with “the average of differences between actual tender price
and forecast” while consistency of estimates is concerned with “the degree of variation
around the average”. Aibinu and Pasco (2008) examined bias of estimate in terms of the
difference between pre-tender forecast and the accepted tender sum expressed as a percentage
of the accepted tender sum. This did not account for uncertainties associated with the
construction process. Thus it has limited application when attempting to prevent project cost
overrun in that upon project award, issue may arise during the construction phase which
could result in project cost overrun. Thus Aibinu and Pasco’s (2008) approach has limited
application that could help ensure project success in the area of cost forecasting and cost
performance. To address this problem, this study defined estimate bias as the difference
between pre-tender forecast and actual completion cost expressed as a percentage of the
actual completion cost.
Thus Percentage cost overestimate or underestimate (estimate error or bias) were
estimated by using the following expression:
pretender cos t estimate − project completion cos t
Estimate bias = x 100
project completion cos t
A positive value of estimate bias implies an overestimation of cost while a negative value
implies an underestimation of cost.

Selection of Input variables: Factors affecting the accuracy of pre-tender building cost
estimates
According to Seo et al (2002) attributes used as inputs for ANN model may be derived from
the literature. They must be meaningful to the estimator and the design team at the pre-tender
stage hence should consist of attributes that are known during that stage. It is also useful that
the attributes should, as much as possible, be high-level project characteristics. An overview
of previous studies suggests that a large number of variables may contribute to the accuracy
of pre-tender cost estimate. Gunner and Skitmore (1999a) reviewed previous studies and
summarised the factors as follows: building function, type of contract, conditions of contract,
contract sum, price intensity, contract period, number of bidders, good/bad years,
procurement basis, project sector (public, private or joint), number of priced items and
number of drawings. Gunner and Skitmore (1999a) analysed the estimates of 181 projects in
Singapore. They found that a majority of the factors influenced the accuracy of estimates.
Using data from 42 projects in Singapore Ling and Boo (2001) found similar results when
they compared five variables against Gunner and Skitmore’s (1999a) work. Skitmore and
Picken (2000) studied the effect that four independent factors (building type, project size,
project sector and year) had on estimating accuracy. They tested the four factors using data
from 217 projects in the United States of America. They found that bias in the estimate of the
projects is influenced by project size and year, while consistency in the estimates is
influenced by project type, size and year. In a study of 67 process industry construction
projects around the world, Trost and Oberlender (2003) identified 45 factors contributing to
the accuracy of early stage estimates. They summarized the factors into 11 orthogonal
elements. Of the 11 factors, the five most important include: process design, team experience
and cost information, time allowed to prepare estimates, site requirements, and bidding and
labour climate. All these studies suggest that there are a large number of variables that may
substantially influence the accuracy of an early stage estimate.
According to Gunner (1997) the factors influencing accuracy of estimates are inter
correlated so that the true bias of one factor could be masked by one or more factors. For
example, Gunner and Skitmore (1999b) theorise that “Price Intensity alone is both necessary
and sufficient to account for systematic bias (inaccuracy) in building price forecasting”. Price
intensity is the total cost of a building divided by the gross floor area. Price intensity theory
states that buildings with low unit rates (cost/m2 gross floor area) would tend to be
overestimated, while those with high unit rates would tend to be underestimated. In a study of
89 construction projects in Hong Kong, Skitmore and Drew (2003) support the price intensity
theory. In another study, Skitmore and Picken (2000) using data from 217 projects in the
United States found that ‘year’ was the underlying variable responsible for the bias and
inconsistency in cost estimates, after partialling out confounding effects of the four factors
put forward. The finding contrasts Gunner and Skitmore’s (1999b) ‘price intensity’ theory.
However, their result supports Gunner’s (1997) theory which states that intercorrelations
among variables cause confounding effects. It also supports Gunner and Skitmore (1999a) in
their suggestion that a single underlying variable is the cause of bias and consistency seen in
estimates.
In their study of 56 projects in Australia, Aibinu and Pasco (2009) found no evidence
to support the price intensity theory. The study defined bias as cost forecasts compared with
accepted tender sum. Based on regression analysis, the study discovered that the project size
is the only factor that significantly influenced the accuracy of the estimates (bias) of the
projects studied. However, project size explained only 29% of the changes in estimate bias
with 71% bias unaccounted for. Thien (2008) conducted a regression analysis of the factors
that influenced estimate bias on 100 projects in Malaysia. The study defined bias as the cost
forecasts compared with final completion cost. Thien (2008) work shows that 7% of the bias
in the estimates of the projects can be explained by project type with commercial project
tending to significantly suffer most from estimate bias. The study however has 93% of bias
unaccounted for.
It appears that in Aibinu and Pasco (2008) and Thien (2008) the regression model was
unable to detect the subtle and non-linear relationship between the factors influencing bias
and the bias observed; and may have been responsible for the low explanatory power of their
models. The use of ANN should produce a better predictive model. In this study, rather than
selecting the significant factors discovered by previous studies, we use nine high level project
attributes (Table 1) identified from the literature as input variable in the training of the
learning ANN model.
Table 1 Project characteristics used in training the learning model (input variables)
Project Characteristics Unit Type of information Descriptors
Gross Floor Area (GFA) m² Quantitative n.a
Principal structural material No unit Categorical 1- steel
2 - concrete
Procurement route No unit Categorical 1- traditional
2- design and construct
Type of work No unit Categorical 1- residential
2 - commercial
3 - office
Location No unit Categorical 1 - central business district
2 - metropolitan
3- regional
Sector No unit Categorical 1 - private sector
2- public sector
Estimating method No unit Categorical 1- superficial method
2 - approximate quantities
Number of storey No unit Categorical 1 – one to two storey(s)
2- three to seven storeys
3 - eight storeys and above
Estimated Sum Cost/ m² Quantitative n.a

Data collection and processing

Data from 100 construction projects completed over ten years (1999 – 2007) were collected
from the office of a quantity surveying firm in Malaysia. Information obtained in respect of
each project include: project size (pre-tender cost estimate, final completion cost, the number
of floors, and gross floor area - GFA) and other information relating to the input variables
(drivers of estimate accuracy) – see Table 1. The 100 projects were identified by a systematic
sampling process. Projects that were not suitable for analysis because they were discontinued
at the feasibility stage were discarded. The researchers had first hand access to all data, such
as estimating report, tender evaluation report, and final accounts. Table 2 shows the profile of
the 100 projects data set obtained including the mean estimate bias, standard deviation and
the coefficient of variation. The standard deviation and the coefficient of variation (CV) were
determined for projects in the different categories of the 9 factors used in the model training.
Coefficient of variation is a measure of predictability of estimate bias. Large coefficient of
variation implies that estimate bias is volatile and unpredictable. Standard deviation (S) was
computed for the projects using the expression:

S= Σ( x − x ) 2
n
Where: x = estimate bias; x = mean estimate bias
n = number of projects
Thereafter, the consistency in the estimates was determined by calculating the coefficient of
variation (CV) as follows:
s tan dard deviation
CV = x 100
mean estimate error
By visual inspection of Table 2, the estimates for the 100 projects are somewhat inconsistent
with coefficient of variation ranging from 11.77% to 21.17% across the various project
categories. It is assumed that a double digit coefficient of variation is large. Thus the risk of
estimation bias is not small. It also suggests that firms have little control over the propensity
that estimates would be biased.

Table 2: Profile of projects according to factor category

Project factors Number of Mean Error Standard Coefficient

Projects (%) Deviation of Variation(%)
(Estimate (Estimate
Bias) (%) Consistency)
Estimated Sum
(RM)
1 - 5,000,000 27 2.41% 20.46% 19.98%
5,000,001 -
10,000,000 8 0.63% 18.15% 18.04%
10,000,000 65 0.51% 13.77% 13.70%
Gross Floor Area
(m²)
1 - 3,000 27 2.33% 21.66% 21.17%
3,001 - 10,000 18 - 4.83% 12.99% 13.65%
above 10,000 55 2.31% 13.39% 13.09%
Number of Storeys
1-2 storeys 48 1.52% 16.84% 16.59%
3-7 storeys 26 3.58% 17.99% 17.37%
8+ storeys 26 -1.59% 13.35% 13.57%
Location
CBD 39 -0.31% 18.70% 18.76%
Metropolitan 49 1.55% 14.83% 14.60%
Regional 12 3.25% 12.15% 11.77%
Procurement Route
Traditional (Lump
Sum) 59 0.53% 16.62% 16.53%
Design & Construct 41 1.76% 15.61% 15.34%
Principal Structural Material
Steel 39 1.15% 16.84% 16.65%
Concrete 61 0.93% 15.82% 15.67%
Estimating Method
Superficial 64 0.20% 16.11% 16.08%
Approximate
Quantities 36 2.47% 16.33% 15.94%
Sectors
Private 59 0.97% 15.10% 14.95%
Public 41 1.12% 17.72% 17.52%
Project Type
Residential 50 3.94% 12.57% 12.09%
Commercial 32 -5.63% 19.34% 20.49%
Office 18 4.78% 15.58% 14.87%
The ANN Architecture and Training
Using 85 out of the 100 data sets, different Networks from different topologies of ANN
model were trained and optimized, by changing the network parameters in order to obtain a
good relationship with better correlation coefficients and performance. The ANN models that
were trained include Generalized Feed Forward (GFF), Jordan and Elman Networks, Radial
Basis Function (RBF) and Self Organizing Feature Maps.
Among the abovementioned networks, Feed Forward RBF was found to be the best topology
that provided significant learning for the aforementioned relationships (Figure 1). The RBF
has been constructed using the mathematical function (Neurosolutions 4.32, Lefebvre et al.,
2003) in a hidden layer with appropriate number of processing elements. Equation (iv)
shows the mathematical form of RBF.
 −1 p 
G ( x ; x i ) = exp  2 ∑
( x k − x ik ) 2  ……………….(iv)
 2σ i k = 1 
where G is multivariate Gaussian function, σi is variance of p data points, xi is mean at ith
node. Figure (1) is a schematic diagram of the RBF network, shows input layer with two
inputs, RBF layer, four hidden layers and output layer with single output.

Figure 1: RBF network trained – Four hidden layers with 100, 80, 20 and 4 PEs in each
layer respectively. Hidden layer 0 is RBF.

The network found has been consisted of four hidden layers including RBF layer and GFF
layers. No of processing elements were 100, 80, 20 and 4 for 9 inputs that produced ‘estimate
bias’ single output. Gaussian transfer function was used in the aforementioned RBF layer
Tanh transfer functions were used in all other GFF layers.
Input data was normalized and arranged in three sets training, cross validation and
testing. The training dataset was used to train networks whilst cross validation data set is
evaluating the training. Once a network is learnt, the test data set was used to forecast
‘estimate bias’ from the network and comparison could be made against actual. Normalized
data set is useful in training as each parameter is mapped into a radius of 1 in order to setup a
unique boundary.

Results
In this study, out of more than 60 different Networks that were trained and optimized using
85 training and cross validation data sets, Figure 2 and Table 3 shows the performance of the
best Network found.

Figure 2: Performance of the best network

Table 3: Performance of the best network

Table 4: Network training and validation

Figure 3 shows the actual and predicated values of Estimate bias versus estimated sum for the
entire data set. Similarly Figure 4 shows the actual and predicated values of estimate bias
versus gross floor area. The correlation coefficient stands at 90%.
Figure 3: Comparison between Estimate Bias, actual and network forecast vs Estimated Sum
– 90% correlation coefficient

Model Validation and Performance of ANN

Figure 4 shows the trained network forecast progressively on test data set. The major
performance measure used in the training was Mean Squared Error (MSE) which is 0.002
(Table 5). Based on the normalized input dataset the correlation coefficient was found to be
73% while the maximum absolute error is 0.10 (Table 5). This means that in 73% of the test
cases, the predicted estimate bias did not differ by more than 10% from the expected. Based
on this performance measures the trained network is suitable for forecasting estimate bias and
can be extended to forecast actual cost when estimated cost is given.
 Figure 4: Network forecast comparison – Estimate bias actual vs. Network prediction

Table 5: Model performance measures

Contribution of Input Parameters

Some input variables are more effective than others in the estimate. Analysis in this regard is
shown in Figure 6 and percentages of contributions are tabulated in Table 6 arranged from
highest to the lowest.
 Figure 5: Input contribution chart

Table 6: Input contribution table in percentage

Table 6 shows that ‘Type of work’ contributed the highest to estate accuracy whilst ‘GFA’ is
the lowest.
Conclusion
At the pretender stage, little information is often available about projects thereby making
estimates of contingency allowance difficult. If quantity surveyors can accurately predict the
inaccuracies in their estimate they can include contingency allowance to cover such
inaccuracies. Learning model can facilitate the prediction of estimate inaccuracies because
the modeling assumptions are less rigid when compared to regression modeling. Using ANN
models can offer an efficient method of predicting cost at the pretender stage. This study
applied neutral network approach for estimating the accuracies in pretender building cost
estimates. The trained ANN model can be used as a decision making tool by cost advisors
when forecasting building cost at the pretender stage. The model can be queried with the
characteristics of a new project in order to quickly predict the error in the estimate of the
project. The predicted error represents the additional contingency reserve that must be set
aside for the project to cater for cost overruns. The model can also be extended to forecast
actual cost of a project when the estimated cost is known. Further study is needed with larger
sample size to improve on the prediction power of the model. Future study will also seek the
integration of artificial intelligence and probability estimating. Rather than obtain point
estimate, the probability estimating model with artificial intelligent will produce a minimum
estimate, most likely estimate and a worst scenario estimate of contingency allowance. This
should provide more robust information for decision making at the design stage.

References
Hasangholipour, T. and Khodayar, F. (2010) A Novel Optimized Neural Network Model for
Cost Estimation Using Algorithm . Journal of Applied Science. Vol 10, Issue 6 , pp. 512 –
516.

Camargo, M.B., Rabenasolo, J.M., Castelain, J.M, and Jolly-Desodt, A.M. (2003)
Application of the parametric cost estimation in the textile supply chain. Journal of Textile
Apparel Technology Management. Vol 3, pp 1-12.

Aibinu, A.A and Jagboro, G.O. (2002) The effects of construction delays on project delivery
in Nigerian construction industry. International journal of Project management, 20, 593 –
599.

Aibinu, A.A. and Pasco, T. (2008) The Accuracy of Pre-tender Building Cost Estimate in
Australia. Construction Management and Economics, Vol. 26, 12, pp. 1257 – 1269.

Khan, S., Dassanayake, D., and Rana, T. (2005). Ocean based water allocation forecasts
using an artificial intelligence approach. in: MODSIM 2005 International Congress on
Modelling and Simulation Advances and Applications for Management and Decision Making,
12-15 December 2005, Melbourne: Modelling and Simulation Society of Australia and New
Zealand: 1667-1673.
http://www.mssanz.org.au/modsim05/papers.htm

Khan, S., Rana, T., Dassanayake, D., Abbas, A., Blackwell, J., Akbar, S., and Gabriel, H.
(2008). Spatially distributed assessment of channel seepage using geophysics and artificial
intelligence. Irrigation and Drainage, (Published online 13 Aug 2008. doi: 10.1002/ird.415):
[np].
http://www3.interscience.wiley.com/journal/121377494/abstract?CRETRY=1&SRETRY=0

Khan, S., Robinson, D., Beddek, R., Wang, B., Dassanayake, D., and Rana, T. (2004).
Hydro-climatic and economic evaluation of seasonal climate forecasts for risk based
irrigation management.CSIRO Land and Water Technical Report 5/045/04. Griffith: CSIRO
Land and Water. http://www.clw.csiro.au/publications/technical2004/tr5-04.pdf

Zhang, G. Eddy, B. and Hu Y.M. (1998) Forecasting with artificial neural networks:: The
state of the art, International Journal of Forecasting, Vol. 14, 1, pp.35 – 62.

Baccarini, D. 2005. Estimating Project Cost Contingency - Beyond the 10% syndrome.
Australian Institute of Project Management National Conference, 09/11/2005. Victoria:
Australian Institute of Project Management.

Garza, J. and Rouhana K. (1995) Neural network versus parameter-based application. Cost
Engineering, 37,25, pp. 14-18.

Kim, G.H.., An, S., Kang, K (2004) Comparison of Construction Cost Estimating
Models Based on Regression Analysis, Neural Networks, and Case-Based Reasoning.
Journal of Building and Environment. Vol. 39, pp 1235-1242.

Baccarini, D. (2004) Accuracy in Estimating project cost construction contingency – a

statistical analysis. RICS COBRA. 7-8 September Leeds metropolitan University, UK .

AIQS (2003), AIQS Website: www.aiqs.com.au. Canberra.

Akintoye, A. and Fitzgerald, E. (2000). "A survey of current cost estimating practices in the
UK." Construction Management and Economics 18: 161 - 172.

Bromilow, F.J., Hinds, M.F. and Moody, N.F. (1988). The Time and Cost Performance of
Building Contracts 1976-1986. The Australia Institute of Quantity Surveyors.

Brown, M.B. and Forsythe, A.B. (1974). Journal of the American Statistical Association, 69,
364–367.

Cheong, P.F. (1991) Accuracy of design stage cost estimating, MSc Dissertation, National
University of Singapore.

Conover, W. J., Johnson, M. E., et al (1981). "A Comparative Study of Tests for
Homogeneity of Variances, with Applications to the Outer Continental Shelf Bidding Data."
Technometrics 23(4): 351 - 361.

Fleiss, J. L. (1971). "Measuring nominal scale agreement among many raters." Psychological
Bulletin 76(5): 378 - 382.

Gunner, J. (1997). Accuracy of construction pre-bid forecasts, PhD thesis, University of

Salford.
Gunner, J. and Skitmore, M. (1999a). "Comparative analysis of pre-bid forecasting of
building prices based on Singapore data." Construction Management and Economics 17: 635
- 646.

Gunner, J. and Skitmore, M. (1999b). "Pre-bid building price forecasting accuracy: price
intensity theory." Engineering, Construction and Architechural Management 6(3): 267 - 275.

Kometa, S. T., Olomolaiye, P. O., et al (1994). "Attributes of UK construction clients

influencing project consultants' performance." Construction Management & Economics
12(5): 433 – 443.

Landis, J. and Koch, G. (1977). "The measurement of observer agreement for categorical
data." Biometrics 33(1): 159 - 174.

Levine, D.M., Stephan, D. Krehbiel, T.C., and Berenson, M.L. (2005) Statistics for
Managers, 4th edition. Prentice hall: New Jersey. Page 429.

Ling, Y. Y. and Boo, J. H. S. (2001). "Improving the accuracy of approximate estimates of

building estimates." Building Research & Information 29(4): 312 - 318.

Lowe, D. and Skitmore, R. M. (2001). "Human effects in construction contract price

forecasting: experience and experimental learning styles." Journal of Construction
Engineering and Management 127(6): 485 - 493.

McCaffer, R. (1976) Contractor's bidding behaviour and tender price prediction, PhD thesis,
Loughborough University of Technology, UK.

Mills, A. (1997) Analysis of the accuracy of building price forecasts. The Australian Institute
of Quantity Surveyors, 1 (2), 31-38.

Morrison, N. (1984) The accuracy of quantity surveyors' cost estimating. Construction

Management and Economics, 2, 57 - 75.

Professions Australia’s Education Committee (2005) Skill Mapping: Assessing Australia's

Longer Term Requirements for Professional Skills. A Discussion paper.

Rawlinsons (2007), Australian Construction Handbook 2007. Perth

Rawlinsons (2006), Australian Construction Handbook 2006. Perth

Serpell, A. F. (2005). "Improving Conceptual Cost Estimating Performance." AACE

International Transactions, EST.13: EST.13.1 - 13.6.

Skitmore, M. (1991), Early Stage Construction Price Forecasting. London.

Skitmore, M. and Drew, D. (2003). "The analysis of pre-tender building price forecasting
performance: a case study." Engineering, Construction and Architechural Management
10(1): 36 - 42.
Skitmore, R. M. and Picken, D. (2000). "The accuracy of pre-tender building price forecasts:
an analysis of USA data." Australian Institute of Quantity Surveyors Refereed Journal 4(1):
33 - 39.

Smith, P. (1994) Trends in the Australian Quantity Surveying Profession. A paper presented
at the 4th World Congress on Cost Engineering, Project Management and Quantity
Surveying, and 18th International Cost Engineering Congress, 17-21 April, 2004, Cape Town,
South Africa, by International Cost Engineering Council.

Thien V.C (2008) Improving the accuracy of preliminary building cost estimate in Malaysia.
A dissertation submitted to the Faculty of Architecture Building and Planning, the University
of Melbourne.

The Australians Department of Employment and Workplace Relations (DEWR) Skilled

Occupation Lists (2008) available at http://www.immi.gov.au/allforms/pdf/1121i.pdf. and
accessed on 30th May 2008.

Trost, S. M. and Oberlender, G. D. (2003). "Predicting Accuracy of Early Cost Estimates

Using Factor Analysis and Multivariate Analysis." Journal of Construction Engineering and
Management 129(2): 198 - 204.