Forecasting elemental building cost percentages using regression analysis and neural network techniques

Early stage project estimates are a key component in business decision making, and generally form the basis of the project's ultimate funding. Their strategic importance has been long recognised, leading to increased research and development in cost modelling. For example, research initiated at The University of Manchester resulted in the production of ProCost - early stage cost estimating software, which has the ability to forecast the total cost of a proposed building in the form of a single figure output. This research project commenced with a nationwide questionnaire survey of current cost modelling and elemental cost estimating practice. One of its major findings was that quantity surveyors are not satisfied with single figure output cost models. Based on this, an investigation into the feasibility of generating an elemental breakdown of the ProCost output was initiated. An investigation into an appropriate elemental output format resulted in the adoption of 17 elements based on the Royal Institution of Chartered Surveyors (RICS) Standard Form of Cost Analysis (SFCA). Models were created for each of these elements using both multiple linear regression and artificial neural network (ANNs) techniques. Initially, data from 120 office buildings were collected and modelled using multiple linear regression analysis. The accuracy of these models, as measured by the mean absolute percentage error (MAPE), ranged from 9.2% to 319.6%. Recognising the deficiency of some of these models, the study proceeded by using ANNs as an alternative modelling method. Accepting the requirement of this method for increased data cases, a second data collection programme was initiated, extending the database to industrial and residential buildings and resulting in a total of 360 projects. ANN produced superior models for the majority of the elements, generating MAPEs from 9.7% to 43%. The final decision support tool presented is a hybrid of these two methods with 5 of the models based on multiple linear regression and 12 on ANN techniques. The mean MAPE of the 17 models is 22.1%. The model compares favourably against previous cost modelling attempts in terms of accuracy, generalisation, sample size, and application spectrum and flexibility. It is anticipated that its utilisation will improve current practice enabling quantity surveyors (cost estimators) to generate quick elemental estimates at an early design stage. Further its elemental character will introduce a crosschecking mechanism into the decision-making process, increasing user confidence in the model's application.