Research on AI and schedule overrun

The question. If machine learning can predict schedule overrun, which algorithm gives the best balance of accuracy and interpretability?

Method. We compared six regression algorithms on 4,517 software-project records: linear regression, regression tree, random forest, support vector regression, gradient boosting machine and a deep neural network. All used the same predictors, default settings and 10-fold cross-validation.

Findings. Gradient boosting had the lowest errors. The deep neural network came very close and had the highest squared correlation. Random forest ranked third, and support vector regression was weakest under default settings. Risk factor, duration and client satisfaction were again the most consistent predictors. Weighing accuracy and interpretability together, gradient boosting is the strongest practical choice for early-warning systems and portfolio decisions.

The question. Traditional project control relies on expert judgment and static plans, which makes schedule performance hard to predict as technical projects grow more complex.

Method. We trained a gradient boosting machine on a public dataset of 4,517 projects. Inputs were duration, total cost, team size, risk factor and client satisfaction; the target was schedule overrun as a percentage.

Findings. The model explained about 83.5% of the variation in overrun. Risk factor and project duration were the most influential predictors, followed by client satisfaction.