Data-Driven Approaches for Forecasting Cost Overruns in Infrastructure Projects

Purpose: The aim of the study was to assess sustainable development depends on the infrastructure projects that affect the long-term economic, social, and environmental impacts. Cost overruns, however, tend to occur in construction projects. Methodology: This research study uses the Build Bridges, Not Walls dataset available in the U.S. National Bridge Inventory, which contains over 600,000 bridge records, and builds a predictive model to estimate the number of infrastructure cost overruns. To begin with, it involved extensive preprocessing, including imputing missing values, scaling the features, and analyzing their correlations. The model was assessed using R2, RMSE, MAE, MAPE, and SHAP interpretability indicators. Findings: The XGBoost model performed very well on the prediction accuracy metric, achieving an R2 of 95.67%, exceeding the ANN benchmark (R2 = 77.3%). The analysis of the Shapley values showed that landform and location were the most influential features, with construction type being a minor factor. Additionally, the strength of XGBoost was further confirmed through evaluations of RMSE, loss curves, and feature-wise performance plots. Implications to Theory, Practice and Policy: The overall conclusion from the study is that applying ensemble learning methods especially XGBoost was the right approach for plausible costing of large-scale infrastructure projects, leading to better planning and decision-making.