Rethinking construction cost overruns: an artificial neural network approach to construction cost estimation
Ahiaga-Dagbui, Dominic Doe
MetadataShow full item record
The main concern of a construction client is to procure a facility that is able to meet its functional requirements, of the required quality, and delivered within an acceptable budget and timeframe. The cost aspect of these key performance indicators usually ranks highest. In spite of the importance of cost estimation, it is undeniably neither simple nor straightforward because of the lack of information in the early stages of the project. Construction projects therefore have routinely overrun their estimates. Cost overrun has been attributed to a number of sources including technical error in design, managerial incompetence, risk and uncertainty, suspicions of foul play and even corruption. Furthermore, even though it is accepted that factors such as tendering method, location of project, procurement method or size of project have an effect on likely final cost of a project, it is difficult to establish their measured financial impact. Estimators thus have to rely largely on experience and intuition when preparing initial estimates, often neglecting most of these factors in the final cost build-up. The decision-to-build for most projects is therefore largely based on unrealistic estimates that would inevitably be exceeded. The main aim of this research is to re-examine the sources of cost overrun on construction projects and to develop final cost estimation models that could help in reaching more reliable final cost estimates at the tendering stage of the project. The research identified two predominant schools of thought on the sources of overruns – referred to here as the PsychoStrategists and Evolution Theorists. Another finding was that there is no unanimity on the reference point from which cost performance could be assessed, leading to a large disparity in the size of overruns reported. Another misunderstanding relates to the term “cost overrun” itself. The experimental part of the research, conducted in collaboration with two industry partners, used a combination of non-parametric bootstrapping and ensemble modelling with artificial neural networks to develop final project cost models based on about 1,600 water infrastructure projects. 92% of the validation predictions were within ±10% of the actual final cost of the project. The models will be particularly useful at the pre-contract stage as they will provide a benchmark for evaluating submitted tenders and also allow the quick generation of various alternative solutions for a construction project using what-if scenarios. The original contribution of the study is a fresh thinking of construction “cost overruns”, now proposed to be more appropriately known as “cost growth” based on a synthesises of the two schools of thought into a conceptual model. The second contribution is the development of novel models of construction cost estimation utilising artificial neural networks coupled with bootstrapping and ensemble modelling.