Predicting Cost Overrun in Construction Projects

Abstract

Two models for predicting cost overrun percentage in construction projects are presented. The first model based on regression analysis. 44 factors that impact cost performance in construction projects gathered from literature. A questionnaire survey was made on construction contractors in Egypt to evaluate the relative importance of these causes from contractors' perspective. Eleven factors were obtained as the most significant causes that lead to cost overrun and these are the independent variables of the proposed model. Data was collected for occurrence of the previous factors on yes/no basis and the corresponding cost overrun percentage (dependent variable) for 30 construction projects and was divided into two sets. The first set contains 20 projects for model building. The results revealed that there is a strong linear relationship between cost overrun percentage and the previous 11 causes that significantly affect cost overrun of projects. These causes are: financial condition of the owner, cash flow of contractor, method of procurement (open tender or selective tender), material cost increase due to inflation, competition at tender stage (aggressive or not), fluctuations in the currency that the payment will be made, project size (small or large), delay in design and approval, risk retained by client for quantity variations, drawings (detailed or not), and inaccurate material estimating. The second set contains 10 projects for validation purposes. The second model is a case based reasoning (CBR) model. CBR method can be an effective means of utilizing knowledge gained from past experience to estimate percentage cost overrun in construction works. Validation of the two models using projects of the second set revealed that regression model has prediction capabilities higher than that of CBR model. Applying the absolute value of standardized coefficient (β) as attribute weight method provides the highest prediction accuracy of cost overrun percentage. Also, feature counting gives results better than the original value of (β).