Integrating Machine Learning and Reference Class Forecasting for Construction Risk Contingency Prediction

Projects typically add fixed-percentage risk contingencies, though this method often lacks accuracy. Advances in technology and artificial intelligence (AI) have introduced more reliable methods for forecasting cost and duration. Machine learning (ML) enhances prediction, while reference class forecasting (RCF) offers an outside view to mitigating optimism bias. This study addresses the challenges of contingency estimation in Design-Build (DB) projects by proposing a method that combines ML and RCF using data from 294 DB building projects. To this end, the authors (1) developed ML models to predict final project cost and actual duration, (2) compared different ML algorithms, (3) fine-tuned hyperparameters for accuracy, (4) established a reference class using historical growth data, (5) derived probability distributions for optimism bias uplifts to estimate contingencies, and (6) compared RCF results with ML predictions. Findings revealed that at an 80% confidence level, RCF cost growth is 11.7% while schedule growth is 10.5%.