Neural network-based interval forecasting of construction material prices

Accurate prediction of material costs is essential for proper management and budgeting of construction projects. Material price fluctuation is one of the most important contributors to deviations from the initial estimated cost in construction projects. Traditional machine learning techniques often fail to generate accurate predictive estimates due to high uncertainties associated with material prices. To address this issue, this research proposes an artificial neural network (ANN)-based method to quantify uncertainties through the generation of forecasting intervals. The optimal lower upper bound estimation (optimal LUBE) method is adopted to train ANN to generate intervals directly. The proposed method is used to predict construction material prices in the US for asphalt and steel. It is shown that traditional regression analysis and ANN-based single-point estimates are of limited value for the prediction of material price. In contrast, prediction intervals provide reliable estimates for material prices and they reduce the possibility of project failure due to the inaccuracy of initial estimated costs. The results obtained from three other cost functions are compared to the proposed optimal LUBE cost function to testify the accuracy of the model. The achieved results show that the proposed optimal LUBE cost function presents the most accurate prediction intervals. This study employs a stacking procedure for monitoring and controlling the training process and validation purpose. The proposed interval forecasting method presents a new direction for cost prediction studies and will provide project managers with extra information to manage risks associated with project costs.