Schedule forecasting model
At a glance
Context: M.S. Data Science capstone project
Problem: Forecasting cost and timeline outcomes for construction projects
Deliverable: A trained model and a Power BI front end for exploring forecasts
Role: Sole author, from problem framing through deployment design
Challenge
Construction schedules are planned to a single date, but real projects finish across a range. A planner needs to know how likely a date is, not just what the plan says.
Approach
- Data. The model was trained on more than 50 completed project schedules exported from Primavera P6.
- Model. Gradient-boosted trees (XGBoost), chosen because they handle mixed project features well and their behaviour can be explained.
- Ranges, not single numbers. Results are reported as P50, P80, and P90: the outcome you would expect half the time, four times in five, and nine times in ten. This is the standard a project controls team can defend.
- Getting it in front of people. Rather than calling the model live, forecasts are scored in advance across a grid of scenarios, stored in a Microsoft Fabric lakehouse, and explored in Power BI with slicers.
Architecture
Completed P6 schedules Feature preparation XGBoost model Pre-scored scenario grid Fabric lakehouse Power BI report with slicers
Why this matters for client work
The same pattern applies outside construction: wherever a team plans to a single number and has history to learn from, a model can put an honest range around it and a report can make that range usable.
Tech stack
- Python and XGBoost
- Microsoft Fabric (lakehouse)
- Power BI
-
Not sure what you need yet? That is what the call is for.
Book a free 30-minute intro call. Tell me what you are trying to answer with your data, and I will tell you honestly whether I can help and what it would take.