Forecasting project Cost Flow using Machine Learning

This paper focuses on forecasting project cost flow using cost curves of completed projects. The paper uses Machine Learning in a time-series framework to model the cost curve as intro­duced by Boussabaine and Kaka (1998) and Cheng and Wu (2009). The research extends on these papers by validating the effectiveness of the methodology using several Machine Learning models. Artificial Neural Network, Support Vector Machine and Random Forest are used on a large data-set of varied projects of Royal Schiphol Group. Subsequently, an out-of-sample com­parison is made based on several forecasting criteria. Overall Artificial Neural Network has the lowest error in forecasting at most stages throughout a project's lifecycle. Conversely, the errors from Support Vector Machine performed relatively poor and show the highest variance. Ran­dom Forest shows decent results but also signs of overfitting. Using Random Forest it is found that the project's budget, duration and performing company division are main cost drivers. The research shows that the Machine Learning framework has better forecasting capability than the logit model.