Predicting completed project cost using bidding data

Neural network and regression models have been developed to predict the completed cost of competitively bid highway projects constructed by the New Jersey Department of Transportation. Bid information was studied for inclusion as inputs to the models. Data studied included the low bid, median bid, standard deviation of the bids, expected project duration and the number of bids. A natural log transformation of the data was found to improve the linear relationship between the low bid and completed cost. The stepwise regression procedure was applied, and yielded the best performing predictive model. This regression model used only the natural log of the low bid as independent variable to predict the natural log of the completed cost. Radial basis neural networks were also constructed to predict the final cost. The best performing regression model produced superior predictions to the best performing neural network model. Hybrid models that used a regression model prediction as an input to a neural network were also studied and were found to also produce reasonable predictions. The calculated models produced good predictions of the completed project cost, but were found to be deficient in predicting very large cost increases. Simple models using the natural log of the low bid as input produced the best results. From the analysis it may be concluded that additional information about the variability of the bids submitted does not provide useful information for predicting the final project outcome.