Improving Cost Contingency Estimation in Infrastructure Projects with Artificial Neural Networks and a Complexity Index

Machine learning (ML) algorithms have been developed for cost performance prediction in the form of single-point estimates where they provide only a definitive value. This approach, however, often overlooks the vital influence project complexity exerts on estimation accuracy. This study addresses this limitation by presenting ML models that include interval predictions and integrating a complexity index that accounts for project size and duration. Utilizing a database of 122 infrastructure projects from public works departments totaling HKD 5465 billion (equivalent to USD 701 billion), this study involved training and evaluating seven ML algorithms. Artificial neural networks (ANNs) were identified as the most effective, and the complexity index integration increased the R2 for ANN-based single-point estimation from 0.808 to 0.889. In addition, methods such as bootstrapping and Monte Carlo dropout were employed for interval predictions, which resulted in significant improvements in prediction accuracy when the complexity index was incorporated. These findings not only advance the theoretical framework of ML algorithms for cost contingency prediction by implementing interval predictions but also provide practitioners with improved ML-based tools for more accurate infrastructure project cost performance predictions.