Optimizing Construction Cost Predictions A Comparative Study of Machine Learning Algorithms

Accurate cost forecasting is vital in construction project management, where budget overruns and delays can have significant impacts. Traditional methods like Earned Value Management (EVM) are widely used, but they rely on static, linear assumptions that often fail to capture the complexities of real-world projects. This thesis explores the potential of machine learning (ML) algorithms to improve cost forecasting by addressing the limitations of EVM and offering more dynamic, data-driven predictions during project execution. A comparative analysis of six machine learning models—XGBoost, Extremely Randomized Trees, Random Forest, Support Vector Machine (SVM), Light Gradient Boosting Machine (LightGBM), and K-Nearest Neighbors (KNN)—was conducted using a dataset of 90 real-world construction projects, selected from 181 initial projects. Key project performance metrics, such as Actual Cost (AC), Earned Value (EV), and the Cost Performance Index (CPI), were used as inputs, along with newly introduced features: Project Regularity (RI) and Project Seriality (SP). These static features were introduced to account for non-linear project growth patterns and task structures. The machine learning models were trained on 75% of the data and tested on the remaining 25%, with performance evaluated using Mean Absolute Percentage Error (MAPE) and Normalized Root Mean Squared Error (NRMSE). Results indicated that all ML models significantly outperformed traditional EVM methods, with XGBoost achieving the lowest error rates. The inclusion of RI and SP further enhanced model accuracy, particularly in projects with non-linear progress. Project Regularity (RI) and Project Seriality (SP) were found to be valuable features for improving the predictive power of ML models. RI captured deviations from linear project progression, while SP reflected the structure of tasks, whether serial or parallel. These additional features enabled the models to better account for the dynamic and complex nature of construction projects, leading to more accurate forecasts at various stages of project execution. In conclusion, the study demonstrates that machine learning models offer a superior alternative to traditional cost forecasting methods like EVM. By incorporating dynamic and static project features, ML models provide more precise, adaptive, and reliable cost predictions, helping project managers mitigate risks and make more informed decisions. These findings suggest that further integration of ML in project management practices could lead to improved project outcomes, especially as ML techniques continue to evolve.