An application of deep learning models for project cost and duration forecasting

This research explrores the application of deep learning techniques to improve prediction accuracy of project cost and time at completion as a response to traditional methods limitations due to growing complexity of modern projects. An accurate forecast is key for effective project management, reducing the risk of budget overruns and schedule delays. The proposed multilayer perceptron (MLP)-based deep learning model is compared against benchmark methods which are based on high-performing machine learning model (XGBoost) and traditional Earned Value Management (EVM). This study uses two datasets: Dynamic Scheduling Library (DSLIB), containing 181 projects, and 8 additional projects from Project Portfolio Dataset (Australian) for more diverse context. MAPE (Mean Absolute Percentage Error) serve as key metric to evaluate forecasting accuracy across different project stages: early, mid, and late. The central hypothesis is that well-trained and tuned deep learning architectures can produce more accurate predictions of project cost and time at completion than conventional machine learning approaches, especially during the early project phases when uncertainty is highest. Experimental results showed that while the optimized benchmark model performed well overall, it struggled to make accurate predictions in the early stages of a project, which also applies to EVM calculations. On the other hand, the MLP model consistently delivered comparable or even better forecasts, particularly in the early and mid stages. However, in the later stages, its performance fell slightly behind the benchmark. The aim of the research is to contribute to the expanding body of knowledge on artificial intelligence applications in project management. The potential success of the proposed MLP models could offer enhanced predictive ability, enabling project managers to make more informed project decisions.