A Machine Learning-Driven Framework for Sector-Specific Construction Cost Overrun Prediction and Mitigation

The construction industry is essential for infrastructure development in residential, commercial, and industrial sectors. Despite careful planning, cost overruns persist due to factors like imprecise estimations, inadequate risk mitigation, inflation, and site conditions. This study proposes a data-driven cost-management framework by intertwining statistical analysis with machine learning to anticipate and mitigate financial deviations. A structured Likert-scale questionnaire designed through an extensive literature review and the factors thus found influential were categorized into five major categories, each encompassing three vital factors causing cost overruns, questionnaires were used to collect 70 responses from diverse stakeholders (engineers, contractors, and owners) across all sectors. Among the various models tested, Random Forest Regression outperformed all others, achieving R2 scores of 0.8001 (Overall), 0.8715 (Residential), 0.8715 (Commercial), and 0.7990 (Industrial & Heavy). Comparatively, XGBoost yielded [0.7901, 0.8615, 0.8615, 0.7890], CatBoost [0.7851, 0.8565, 0.8565, 0.7840], and MLP regressors [0.7801, 0.8515, 0.8515, 0.7790] for overall, residential, commercial, industrial, and heavy, respectively. whereas classical models, such as Ridge and Linear Regression, trailed behind. The strength of Random Forest lies in capturing nonlinear interactions within perceptual data, while enabling interpretability through SHAP and feature importance analysis. Although prior studies have employed machine learning, the novelty of this research lies in its sector-specific, stakeholder-informed real-time data approach, offering actionable insights for targeted risk mitigation, cost control, and effective execution, bridging a critical gap in the construction cost overrun literature and also posing a major contribution of the study. The novelty of this work lies in its sector-specific (residential, commercial, industrial/heavy), stakeholder-informed real-time perceptual data approach combined with explainable AI (SHAP and PCA), filling critical gaps in generic, single-sector, or archival-data-focused models prevalent in prior research.