Simulation

Risk technique · See it on the map

Guide only — A Monte Carlo exists for schedule forecasting; it is not wired to any risk output.

Running a risk model many times with varying inputs to see the range of likely outcomes.

When to use it

When you need a probabilistic answer to 'how likely are we to finish by date X or under budget Y' — one that accounts for many uncertain inputs combining, including correlations between them (two tasks sharing the same delayed vendor, for instance) that a simple sum of point estimates can't represent.

When to avoid it

When you don't actually have distributions for the inputs — a simulation run on made-up ranges produces a precise-looking output from imprecise input, which is worse than an honest three-point estimate that admits it's rough. Also skip it when the answer doesn't need to be probabilistic at all; a straightforward reserve calculation is often enough.

Steps

1. Build a model of the project's cost or schedule logic — the tasks, dependencies, and cost items the simulation will vary.
2. Assign a probability distribution to each uncertain input (see representations of uncertainty) rather than a single point estimate.
3. Identify and encode correlations between inputs where they exist — inputs that move together (shared vendor, shared team) will understate risk if treated as independent.
4. Choose an iteration count large enough for the output distribution to stabilize, and run the simulation.
5. Read the output as a distribution — a P50, P80, P90 completion date or cost — not as a single answer, and decide what confidence level the commitment you make should be pinned to.

What it produces

• A probability distribution of possible project outcomes (cost, duration, or another modeled quantity), summarized as percentile figures such as P50/P80/P90.
• A defensible basis for setting a contingency reserve or a committed date at a stated confidence level rather than a guess.

Common pitfalls

• Running the simulation on fabricated distributions and reporting the output with the same confidence you'd give a result built on real historical data — the output is only as honest as the inputs.
• Ignoring correlation between inputs, which systematically understates the tails of the outcome distribution.
• Confusing this with what Driftless actually has today: the schedule forecaster in ``driftless.calc.forecast`` runs a bootstrap simulation over past sprint velocity to forecast a completion date band. That is a real Monte Carlo method, but it samples historical velocity, not a risk model with cost/schedule distributions and correlations across the register. Driftless does not run a risk simulation of the kind described here, and there is no risk-model input (distributions, correlations, iteration count) anywhere in the product to run one from. A reader who wants this technique needs a separate quantitative risk analysis tool; this entry explains what the technique is and requires, not something you can launch from here.

Worked example

A construction estimator builds a cost model with twelve line items, each given a triangular distribution instead of a single figure, with steel and concrete costs correlated because both depend on the same regional supply conditions. Ten thousand iterations produce a cost distribution whose P80 figure — not the P50 — becomes the number the client contract is priced against, because the client wants 80% confidence of not exceeding the price.

Source

• PMBOK-6 §11.4.2.5

Where it comes from: this technique is named by the PMBOK Guide, 6th edition.