The ability to accurately predict the duration and final costs of a project is essential to implement timely corrective actions during its execution. Estimates at Completion (EAC) are the key tool to support project control. In recent years, the integration of Machine Learning (ML) techniques with traditional methodologies such as Earned Value Management (EVM) and Earned Schedule (ES) has improved the accuracy of forecasting. However, many studies have focused exclusively on static data, neglecting both the influence of dynamic data and topological indicators and the issues of underfitting and overfitting that compromise model robustness. This aims to address both issues by proposing a structured ML pipeline, including automated preprocessing, feature engineering, and cross-validation procedures, with the goal of improving the generalizability of predictive models.