Abstract
Cost contingency estimation in road projects requires not only quantifying the reserve needed under uncertainty but also identifying the risk sources that justify it, particularly under extreme cost scenarios. This study proposes an explainable framework for estimating and allocating cost contingencies under tail risk by integrating Monte Carlo simulation, Conditional Value at Risk (CVaR), Shapley values, and explainable machine learning. The framework was applied to a Peruvian road project considering 16 risk factors, 50,000 simulation scenarios, and 800 Shapley permutations. For a baseline direct cost of S/312.61 million (USD 81.24 million), the P95 contingency was S/26.88 million (USD 6.99 million), whereas the CVaR95 contingency reached S/30.23 million (USD 7.86 million), highlighting the additional severity captured beyond the percentile threshold. Low productivity, unforeseen interferences, quantity variation, adverse weather events, and freight and transportation jointly accounted for 60.41% of the extreme contingency. Variance contribution, predictive SHAP importance, and CVaR–Shapley attribution showed high ordinal agreement but relevant differences in contribution magnitudes and intermediate rankings. The proposed framework transforms an aggregate tail-risk reserve into risk-specific monetary allocations, providing an auditable basis for prioritizing mitigation measures, strengthening budget justification, and supporting risk-informed planning and financial resilience in road infrastructure projects.