Stop Reporting. Start Forecasting.
A single completion date is a claim with the uncertainty deleted. What P10/P50/P90 forecasting actually requires from your project data — and what it gives back.
Ask a project director when a project will finish and you get a date. Ask how confident they are and you get a pause.
The pause is the honest part. The date is a point estimate extracted from a distribution nobody wrote down.
What a single date hides
A baseline schedule is a model of a future that has already stopped being true. Every progress update is a small correction applied to a structure that assumed the corrections would not be needed. By the time variance is large enough to show up in a monthly report, the decision window that could have absorbed it has closed.
This is the core failure of reporting-driven project control: it is a description of the past presented at the cadence of the past.
What a forecast requires
Probabilistic forecasting is not a more expensive way to produce the same date. It requires different inputs:
- Reconciled actuals, not reported actuals. Percent-complete claimed in a progress meeting and percent-complete evidenced by inspection records are different numbers. The gap between them is itself a signal.
- Duration distributions, not duration estimates. How long this crew has historically taken on this activity type, on this project, under these conditions.
- Explicit driver modelling. Weather, crane availability, delivery lead time, approval latency, rework rate. Each with its own uncertainty.
- Dependency structure that reflects reality. Including the soft logic that exists in the superintendent's head and nowhere in the schedule file.
Given those, the output is not a date. It is a range with a shape: a P10 optimistic case, a P50 most-likely case, a P90 case that governs how much contingency you actually need.
What it gives back
Three things a static date cannot.
Lead time. A forecast degrades gradually and visibly. You see the P90 case walk outward weeks before the P50 case moves at all. That interval is your decision window.
Attribution. When the range widens, you can ask which driver widened it. "The forecast slipped" is not actionable. "The forecast slipped because connection-plate lead time moved from 12 to 19 days and three downstream activities share that constraint" is.
Comparability of options. Recovery scenarios stop being arguments and start being distributions you can put side by side. Resequencing Zone C and adding a second crane are no longer competing opinions — they are two curves, with costs attached.
The discipline it demands
Forecasting is unforgiving in a way reporting is not. A report can be optimistic without consequence. A forecast that was optimistic is visibly wrong three weeks later, and the record of it persists.
That is the point. A forecasting system that nobody can audit is just a dashboard with more decimal places. The forecast has to carry its evidence, or it does not deserve the confidence it displays.