Interpretable Gradient-Boosted Framework for Highway Construction Cost Estimation

Abstract

Estimation of highway construction cost at the early stage is an essential element for budgeting, scope definition, and risk management; but the traditional methods for cost estimation are based on detailed bills of quantities (BoQ) which are not available at early-stage planning. We propose an explainable machine learning approach for BoQ-based CPM prediction and cost breakdown to the work-item level. With a dataset of highway projects from Iraq (N=170, 2009–2023), we construct CPM and item shares, apply log transforms to address skewness, and estimate models with leave-one-year-out cross-validation to mimic forward-in-time generalization. We compare a baseline shallow ridge regression to gradient-boosted trees (LSBoost). For explanation analysis, we use partial dependence and Shapley/importance to explain the marginal effects of length vs. year. Results indicate that boosted trees improve CPM prediction compared to linear baselines and remain interpretable: CPM decreases with project length (economies of scale) and shows time shifts consistent with market/inflation effects. Cost-share modeling suggests that paving is the primary cost, with systematic variation over length and time. The provided software framework provides uncertainty-aware, comprehensible predictions applicable to construction management and infrastructure planning. To improve transferability between regions, future work will incorporate official price deflators and geometric/traffic descriptors.