Monte Carlo Analysis: What It Tells You and What It Doesn’t
A sponsor asks when the project will finish and gets a date. What the date conceals is how much of the plan had to go right to produce it. Most single-point schedules are the sum of estimates that each carry a comfortable assumption, and the assumptions compound quietly along the critical path. The finish date that emerges has a probability attached to it whether or not anybody has worked out what that probability is, and on a large plan it is often well under half.
Monte Carlo analysis exists to work it out. It answers one question: given everything in the plan that could vary, how likely is the date or the cost you are about to commit to? The answer arrives as a range with likelihoods attached, and its practical value is that a commitment made at a stated confidence is a commitment you can defend in a room.
Where the spread comes from
A simulation takes the model you already have, the schedule network or the cost breakdown, and replaces its fixed numbers with ranges. The excavation is no longer eighteen days; it is twelve to thirty, most likely eighteen. The licence is no longer forty thousand pounds; it is thirty-eight to fifty-two. The tool then runs the model several thousand times, drawing one value from each range on every pass and recording what comes out of the end. What accumulates is a picture of how the project behaves across thousands of plausible versions of itself.
Simulation belongs to the techniques collected at Section 5 of the PMBOK® Guide, and its distinguishing feature among them is that the uncertainty survives into the answer. An estimate averages a range away and hands you one number to work with. A simulation keeps the range and shows you where your chosen number sits inside it, which is why its output is a confidence level and never a prediction.
Reading the result
The headline output is usually written as a P-number. A P80 date of 14 March means that in eighty runs out of a hundred the project finished on or before 14 March. It does not mean 14 March is likely; the single most probable date sits earlier and carries far less protection. The gap between the P50 and the P80 is the honest size of the uncertainty, and that gap is the figure worth carrying into a conversation about contingency.
The second output is the one project managers underuse. A simulation can rank the uncertain inputs by how much each contributes to the spread, and the ranking usually shows two or three of them accounting for most of it. Shortening a task that contributes nothing to the variance does not move the finish date, however satisfying the meeting about it felt. The ranking tells you where management attention is worth spending.
A refurbishment programme covering four floors of a let office building ran a schedule simulation before committing to a handover date. The planner's deterministic finish was 3 May. The P50 came back at 11 May and the P80 at 2 June, which made for an uncomfortable conversation and a useful one, because the client's lease obligation fell on 15 June. The sensitivity ranking put most of the spread on a single item, the bespoke glazed partitions, whose lead time the supplier had quoted as a range because the glass came from one plant.
The programme manager did two things with that. She released the partition order ahead of the design freeze for the areas already settled, which cut the range at its source, and she ended the fortnightly argument about the plastering sequence, which the ranking showed contributed almost nothing. The rerun P80 came back at 19 May. Nothing in the simulation made the project quicker. It showed her which of the moves available to her would count.
What the simulation cannot see
Input quality is the obvious limit and the least interesting one. A model built on ranges that people invented in a workshop produces a beautifully rendered distribution of invented numbers. Three-point estimates also run systematically optimistic, because estimators build their range around the plan, and the plan is what they believe ought to happen.
Correlation is the limit that does real damage. Simulations draw each input independently unless told otherwise, so the tool assumes that a delay to the mechanical fit-out tells you nothing about the electrical fit-out. On a live site the two often share a subcontractor, a delivery route or a weather window, and when one slips the other slips with it. Independent sampling lets the extremes cancel out, so the distribution comes back far narrower than reality and the P80 sits reassuringly close to the P50. Where inputs move together, say so in the model or discount what it gives you.
Unmodelled risk is the third limit and the quietest. A simulation varies the things you put into it, and it has nothing to say about the planning authority reopening a condition, the anchor tenant changing their requirement, or asbestos appearing behind a riser. Those belong in the risk register, with contingency decided separately, and a team treating the P80 as cover for everything has quietly turned a schedule tool into a promise.
Whether to run one at all comes down to what is at stake and what you already know. Where a plan carries real dependency structure, where a contingency figure is being negotiated, or where a committed date carries a penalty, a simulation earns its afternoon. Where the plan is short, the dependencies are few, or nobody can defend a single range, it produces confident-looking output that adds nothing to the decision. Recognising which of those situations you are in before the model gets built is among the judgements a structured PMP exam preparation course counts as knowing the technique at all.
For a PMP® candidate, the distinction to hold is between a range and a forecast. A simulation produces a range with likelihoods attached, and any answer treating its output as the date the project will finish has misread what it is. It helps to ask, whenever a confidence figure appears in front of you, what was varied to produce it and what was held fixed.
On a live project the most valuable half-hour is often the argument about the ranges that precedes the run. Asking a trade contractor for a best and worst case, and for the reasoning behind each, surfaces assumptions a single number hides completely. Teams that simulate once and never return to it get a slide for the board pack. Teams that rebuild the ranges at each stage boundary get an early warning system, because a spread that widens between runs is a signal in its own right.
Interested in going further?
Knowing when a confidence figure deserves weight, and when it is a tidy number resting on ranges nobody can defend, takes exposure to both kinds. Omega's PMP® Exam Preparation covers simulation alongside the estimating decisions that feed it, in the predictive and hybrid settings where the two meet.
The PMBOK® Guide Eighth Edition sets simulation alongside the estimating and forecasting methods it depends on.
Ad · Amazon affiliate link.
References
A182: Multicriteria Decision Analysis Explained
A161: Expected Monetary Value Explained
A180: Facilitation Techniques for Difficult Project Conversations
A184: Trend Analysis vs Variance Analysis
A162: Decision Trees Without the Headache
PMP and PMBOK are registered marks of the Project Management Institute, Inc.
