What 4,517 projects tell us about schedule overrun
Lessons from my research on predicting schedule overrun with gradient boosting, and how a PMO can use the idea.
By Abdelkarim Sweilem ·
Every project manager has watched a schedule slip slowly. The status stays green for weeks, then turns red in a single steering committee. By then the options are expensive. I wanted to know whether the warning signs were visible earlier in data most PMOs already hold.
What we did
With my co-authors Hala Abdoh and Ahmad Abdoh, I trained a gradient boosting machine (GBM) on a public dataset of 4,517 projects. The model takes five inputs that almost every organization records: planned duration, total cost, team size, a risk factor and client satisfaction. It predicts schedule overrun as a percentage.
What we found
On data the model had not seen, it explained about 83.5% of the variation in overrun (R² of 0.835). Its average error was just under 6 percentage points (MAE 5.946). In plain terms: if a project is heading for a 30% overrun, the model will usually place it somewhere between 24% and 36%. That is precise enough to rank projects by risk and decide where to look first.
Two inputs carried most of the signal: the risk factor and the planned duration. Client satisfaction came next. Team size and cost mattered less than many people expect.
How a PMO can use this
- Use it as an early-warning flag, not a verdict. Run the prediction at each stage gate and investigate the projects that score high. The conversation with the project manager is still where the real diagnosis happens.
- Take the risk score seriously. If the risk factor predicts overrun this strongly, a vague or copy-pasted risk register is a real cost. Better risk data makes better predictions.
- Watch long projects more closely. Longer planned durations carry more exposure. Shorter phases with their own milestones reduce it.
- Start with the data you have. Five fields are enough for a first model. You can add more context later.
Limits
The dataset is public and general. An organization that trains the same approach on its own history will get a model that reflects its own sectors, vendors and habits, which is the step I recommend. The full paper, published in Frontiers in Computer Science and Artificial Intelligence, covers the method and results in detail.