Predicting construction cost overruns using text mining, numerical data and ensemble classifiers

• Some words and phrases are associated with the level of project cost increase.
• Text mining and SVD algorithms were used to convert text to numeric data.
• The text information was combined with numerical data describing the project.
• The combined data were used to predict cost overruns using data mining algorithms.
• Best results were obtained using the stacking ensemble method.

This paper discusses how text describing a construction project can be combined with numerical data to produce a prediction of the level of cost overrun using data mining classification algorithms. Modeling results found that a stacking model that combined the results from several classifiers produced the best results. The stacking ensemble model had an average accuracy of 43.72% for five model runs. The model performed best in predicting projects completed with large cost overruns and projects near the original low bid amount. It was found that a stacking model that used only numerical data produced predictions with lower precision and recall. A potential application of this research is as an aid in budgeting sufficient funds to complete a construction project. Additionally, during the planning stages of a project the research can be used to identify a project that requires increased scrutiny during construction to avoid cost overruns.