Construction projects are significantly impacted by uncertainties, leading to time and cost overruns. Cost overruns pose a significant threat to the construction industry’s profitability. The potential benefits of recent advancements in machine learning (ML) models have not yet been fully utilized against such chronic threats. This research aims to lay the groundwork for the potential application of advanced ML techniques in predicting cost overruns by employing a broad set of ML algorithms.
The features used to predict the cost overrun ratio in construction projects were extracted from relevant studies and available field data, resulting in 12 key features. The principle of “let the data speak for itself” was applied to this study, which employed atypical tools, specifically ML methodologies, to a dataset of 836 public projects. The prediction models were developed using 15 ML regression algorithms and then further evaluated and cross-validated.
The CatBoost model demonstrated superior predictive accuracy on the test set (R2 = 0.883), followed closely by Stacking Regressor (R2 = 0.881). Other models with high accuracy included Voting Regressor (R2 = 0.867), XGBoost Regressor (R2 = 0.844), Gradient Boosting Regressor (R2 = 0.833), LGBM Regressor (R2 = 0.813) and Random Forest Regressor (R2 = 0.802). The predictive model identified three key factors in forecasting cost overruns: variation orders, which had the highest feature importance at 41.16%, followed by excessive quantities at 21.86% and budgeted costs at 20.96%.
The findings hold substantial implications for research, practice and society. The study validates the effectiveness of machine learning algorithms in forecasting cost overruns in construction projects through a comparative analysis of various ML algorithms. The investigation underscores the efficacy of the CatBoost algorithm and advocates for additional inquiry into machine learning applications within this domain. The CatBoost model, recognized as a top-performing machine learning model, serves as an effective and dependable tool for practitioners and project planners to predict cost overruns in construction projects. This promotes the advancement of data-informed cost estimation and management approaches, resulting in enhanced decision-making and minimized project risks. From a societal perspective, the capacity for accurate overrun predictions will guarantee the achievement of the desired project baseline. Consequently, the implementation of the research findings will lead to a reduction in overruns in construction projects, particularly within the government sector, thereby refining construction project management practices. This leads to enhanced infrastructure development and may result in reduced project costs for consumers. Furthermore, employing machine learning models for cost prediction can enhance transparency in construction projects.
The application of advanced predictive methodologies, such as the CatBoost algorithm, in the construction sector offers actionable insights since industry practitioners will be able to use the developed models to improve business practices. This research presents an innovative method for forecasting construction cost overruns by thoroughly evaluating 15 ML models. To the best of the authors’ knowledge, this study uniquely examines CatBoost for cost overrun prediction in construction projects while utilizing a more extensive set of data from 836 public construction projects.
