Artificial Intelligence and Machine Learning Applications for Predicting Cost Overruns in Construction Projects: A Systematic Literature Review

Abstract

Cost overruns remain a persistent challenge in construction projects, often undermining financial performance and stakeholder trust. While traditional estimation methods have proven insufficient, artificial intelligence and machine learning offer promising predictive capabilities. This systematic literature review aims to synthesize and critically evaluate the state of research on AI and ML applications for forecasting cost overruns in construction. Our objective was to identify prevailing models, their predictive accuracy, and the contextual factors that influence their performance across different project types and phases. We conducted a comprehensive search of peer-reviewed literature following the PRISMA guidelines, screening studies that employed supervised learning, ensemble methods, or deep learning architectures for cost-related predictions. The methodology involved a structured extraction of model characteristics, data sources, feature engineering approaches, and reported performance metrics. The results reveal that ensemble methods such as random forests and gradient boosting consistently outperform single classifiers, often achieving mean absolute percentage errors below 10% when trained on historical project data. Neural networks and support vector machines also show strong performance but require larger datasets and more careful tuning. Importantly, the integration of AI with building information modeling and digital twin technologies has emerged as a particularly effective strategy, enabling real-time cost trajectory adjustments. However, we found that most studies focus on commercial and residential projects, with limited validation for infrastructure or green building contexts. Furthermore, many models lack rigorous cross-validation or external testing, raising questions about generalizability. The conclusion is that AI and ML demonstrably improve cost overrun prediction accuracy compared to conventional methods, yet adoption remains hampered by data scarcity, model interpretability issues, and insufficient industry validation. This review provides a synthesized evidence base to guide future research toward more robust, scalable, and practically deployable predictive systems. We therefore recommend prioritizing hybrid models that combine domain-specific rules with data-driven learning, alongside strategies for leveraging limited project data through transfer learning or synthetic data augmentation.