Contingency should have a number behind it
On a great many projects, contingency is set by applying a percentage that has been used before. Ten percent, or fifteen, or whatever the last programme carried. It is a familiar figure and it is arrived at without reference to anything specific about the project it is meant to protect.
The alternative is not complicated. Identify the risks, quantify their range and likelihood, run them against the actual cost and schedule model, and read the contingency off the resulting distribution at whatever confidence level the organisation has decided to hold. The output is a number you can defend in front of a board, and — more usefully — a ranked list of what is actually driving it.
How I run it
Against the real network, not a summary. Schedule risk analysis run on a fifty-line summary of a four-thousand-line programme produces a confident answer to a question nobody asked. The model has to be the model.
Correlation stated explicitly. Treating risks as independent when they are not is the single most common way a quantitative analysis produces a P80 that is comfortably too low. If the same weather, the same labour market or the same supplier drives four durations, that has to be in the model and written down.
Ranges from people who know, captured properly. Elicitation is a skill. Ask an engineer for a worst case and you will usually get a bad-day estimate rather than a genuine tail. The interview technique matters more than the software.
Drivers ranked, not just totals reported. The value of the analysis is rarely the P80 figure. It is the tornado chart underneath it, which tells you the four things worth managing out of a register of two hundred.
Deliverables
- Risk register with named owners, quantified ranges and a live status, not a workshop artefact
- Quantitative cost and schedule risk analysis with the correlation assumptions documented
- Contingency recommendation at the agreed confidence level, with the reasoning attached
- Drawdown plan — how contingency is released, against what, and who approves it
- Scenario cases for the small number of uncertainties that genuinely move the outcome
A note on honesty
A risk analysis that always returns a number the organisation was hoping for is not an analysis. The function only earns its cost if it is occasionally allowed to deliver an unwelcome answer, and if the people commissioning it expect that possibility going in.