Rapid Pre-construction Forecasting of Schedule Overruns: A Machine Learning Regression Approach

Introduction/Objective This study developed and validated a parsimonious machine-learning regression model that estimates schedule overruns for building construction projects prior to project start. Methods Using a dataset of 60 completed building projects in Lebanon, each with 11 pre-project predictors, an XGBoost regression model was trained and tested on an independent dataset, with repeated cross-validation and hyperparameter tuning. Results Averaged over repeated randomized train/test partitions, the trained and tuned model achieved an R 2 = 79.6% (adjusted R 2 = 74.3%), a mean absolute error of 4.36 ± 0.48 percentage points (95% CI: 3.42–5.30), and a median coverage near the nominal 90% for empirical prediction intervals. Discussion The regression-based approach extends existing delay-risk assessment by estimating delay magnitude rather than only the risk class, while factor-importance analysis identified owner/client performance, legal/permit issues, design quality, and project manager experience as the major drivers of delay. Conclusion The model provides a practical decision-support tool for project managers to estimate the expected magnitude of delay and its uncertainty before construction initiation, requiring only eleven readily obtainable pre-project inputs.