One number is not the outcome
Residual cost, schedule, billing and contractual exposure evolve together. A useful model system therefore estimates more than a binary overrun label: it can include residual EAC, time-to-breach, competing outcome hazards and calibrated uncertainty.
The operator needs to see which range of outcomes is plausible, which assumptions move it, and when the window for action closes.
Model families have different jobs
Hierarchical models can pool evidence across related projects without pretending they are identical. Survival and competing-risk models represent timing and different failure modes. Gradient-boosted or sequence models can capture nonlinear operational signals. Ensembles are justified only when locked evidence shows better decisions after calibration and constraints.
Method names are not proof. A candidate earns influence only if it improves a predeclared decision metric on locked, cutoff-correct evidence.
Uncertainty belongs in allocation
The optimisation layer should consume the model’s uncertainty rather than silently convert it into a single score. Depending on the decision, this may mean conservative bounds, scenario analysis, explicit abstention or prioritising evidence collection.
The final test is operational: did the learned evidence change the selected intervention or capacity allocation in representative cases without breaking hard constraints?