Construction of An Integrated Machine Learning Model for Cost Risk Prediction in Construction Projects

Abstract

Cost overruns are commonly occurred in construction project. Single prediction model always has low accuracy and generalization ability. It is hard to explore complex cost risk factors accurately. This paper develops an ensemble machine learning model based on stacking strategy to solve the problem. Firstly, collect data from 458 construction projects and extract 27 feature variables, including construction period, material price volatility and construction complexity. Then Principle Component Analysis (PCA) reduces dimensionality into 15 critical features. Finally, the base layer applies Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and Support Vector Machine (SVM) for initial prediction. The metalearner applies logistic regression to integrate output of base model. The optimal parameter combination is searched out by grid search based on 10-fold cross validation. On the test set, the model achieves a 6.1 percentage point improvement in accuracy compared to the best-performing GBDT, an 18.8% reduction in RMSE, and a precision increase to 91.2%. Feature importance analysis shows that the average SHAP value of material price volatility is 0.23 across all samples, construction complexity is 0.19, and the number of design changes is 0.15. This integrated model provides an effective predictive tool for construction cost risk management.