Predicting project cost overrun levels in bidding stage using ensemble learning

Abstract

Predicting project cost overruns in the bidding stage has undergone significant changes with the application of state-of-the-art techniques. Both modeling techniques and domain knowledge should be integrated to enhance predictions of cost performance. This study developed an ensemble-learning classification model to predict the expected cost-overrun levels of public projects and derive explanatory factors and key predictors. A database of 234 public-sector projects in South Korea was used, including project characteristics (i.e., project delivery method, project types, cost, and schedule) in combination with bidding characteristics (i.e., award method, number of bidders, bid to estimate ratio, number of joint ventures). The results yielded an average accuracy of 61.41% for five model runs. Furthermore, information on the project type being constructed is an important contributor to prediction accuracy. Results of the model enable project owners and managers to screen projects that are expected to incur excessive cost overruns and to anticipate budget loss during the bidding stage and before contracts are finalized.