Monte Carlo Simulation for Cost Estimates: Modeling Risk and Determining Contingency
Monte Carlo simulation is a probabilistic modelling technique used in cost estimating to simulate thousands of project scenarios and determine the probability of different cost outcomes.
Every capital cost estimate is a distribution pretending to be a number. The line item that reads $48.2 million for structural steel is shorthand for a range of plausible outcomes driven by commodity prices, fabrication productivity, design growth, and schedule slip. Deterministic estimating collapses that range to a point, then bolts on a percentage contingency to cover what the point cannot see. On small, well-understood work that shortcut is defensible. On a multi-hundred-million-dollar project with long lead times and correlated risks, it hides exactly the information sponsors need in order to fund the project responsibly.
Monte Carlo simulation is the standard technique for restoring that information. By running thousands of randomized scenarios across defined probability distributions and a project risk register, it produces a cost probability curve rather than a single answer. This article covers what the method actually does, the three inputs it depends on, how to read P50 and P80 outputs, a worked $300M highway example, and the governance conditions under which the numbers are worth trusting.
Why Probabilistic Estimating Replaces Percentage Contingency
Deterministic estimates with a flat percentage uplift treat contingency as a blanket — typically 10 to 20 percent on a Class 3 estimate — regardless of which risks actually sit behind the number. The uplift covers nothing specific, so it cannot be defended line by line, drawn down against identified events, or benchmarked against similar projects. When it runs out, nobody can explain whether the cause was poor productivity, unknown scope, or under-provisioning from day one.
Probabilistic estimating replaces that blanket with a model. Each significant cost element and each entry on the risk register carries a distribution, and the simulation aggregates them into a total-cost distribution that shows the probability of the project landing at any particular number. Contingency becomes the distance between the deterministic base and whatever confidence level the sponsor chooses to fund — usually P50 or P80 — and every dollar of that contingency can be traced back to specific uncertainties in the model.
This shift is why Monte Carlo has become routine on Class 3 and Class 2 estimates for large infrastructure, mining, and energy projects. The method does not make the underlying uncertainty smaller; it makes it explicit, priced, and debatable.
The Three Inputs the Model Depends On
A Monte Carlo cost model is only as good as the three inputs feeding it. The first is a credible deterministic base estimate, built bottom-up, parametrically, or from benchmark data, representing the expected cost of the defined scope before any risk loading. Weak base estimates produce confident-looking distributions around the wrong mean — probabilistic dressing on a deterministic error.
The second is a project risk register. Each entry carries a probability of occurrence and a cost impact range. Typical entries on an infrastructure project include geotechnical uncertainty driving rock excavation, scope growth during detailed engineering, commodity price volatility on steel and concrete, labour productivity variance against plan, and weather or permitting delays. The register is where subject-matter expertise enters the model, and its quality almost entirely determines whether the output is signal or decoration.
The third input is a set of probability distributions assigned to the uncertain variables. The three most common choices, and the conditions under which each is appropriate, are summarized below.
| Distribution | When to use it | Typical cost application |
|---|---|---|
| Triangular | Minimum, most likely, and maximum are known from expert judgement | Line items with bounded expert estimates — productivity factors, unit rates |
| Normal | Uncertainty is symmetrical around the expected value | Well-sampled quantities, labour hours with stable history |
| Lognormal | Upside tail is longer than the downside — overruns larger than savings | Commodity prices, schedule-driven costs, claims exposure |
Correlation between variables matters almost as much as the distributions themselves. Steel price and concrete price tend to move together; productivity across trades often correlates with overall site conditions. Ignoring these correlations understates the tail of the total-cost distribution, producing a P80 that looks reassuring but is structurally optimistic.
How the Simulation Produces a Cost Distribution
The mechanics are straightforward. The model takes the base estimate, samples one value from each uncertain variable’s distribution according to its probability weights, applies the sampled risk impacts to the base, and records a single total-project-cost scenario. It then repeats that process thousands of times — ten thousand iterations is typical — producing a large population of plausible total-cost outcomes.
Those outcomes are then assembled into a histogram and, more usefully, into a cumulative probability curve. The curve plots total project cost on the x-axis against the probability that the final cost will not exceed that value on the y-axis. The P50 is the point where there is a fifty percent chance the final cost lands at or below that number, and a fifty percent chance it lands above. The P80 is the same logic at the eighty percent confidence level and is commonly used as a funding target in public infrastructure and regulated sectors because it provides meaningful downside protection without pricing in extreme tails.
Worked Example: A $300M Highway Project
Consider a highway construction project with a deterministic base estimate of $300M built bottom-up from quantities and unit rates. The risk register highlights three material uncertainties: geotechnical conditions that may require unplanned rock excavation, steel price volatility during the procurement window, and weather-related schedule delays in the second construction season. Each is assigned a probability of occurrence and a cost-impact distribution — triangular for the geotechnical and schedule risks, lognormal for the steel exposure to capture the upside tail.
Running ten thousand iterations produces a cost distribution with a P50 of $330M and a P80 of $360M. The interpretation is specific. The median outcome, assuming the risk model is correct, is $330M — a $30M gap between the deterministic base and the most likely total cost, reflecting expected but not-yet-priced risk. Funding the project at $360M gives an eighty percent confidence that final cost will not exceed budget, with $60M of risk-based contingency covering the identified uncertainties at that confidence level.
The sponsor’s decision is then explicit: fund at P50 and accept a roughly fifty-fifty chance of an overrun request, or fund at P80 and carry a larger but defensible contingency. Either choice is informed. The deterministic $300M number, without this analysis, allows neither.
Where Monte Carlo Helps and Where It Misleads
The method’s strengths are real. It forces risks to be quantified rather than narrated, ties contingency to specific uncertainties instead of a round percentage, and gives executives a risk profile they can compare across competing investments. Projects that adopt probabilistic estimating and maintain the underlying risk register through execution tend to produce more reliable final-cost outturns and fewer mid-project funding requests.
The failure modes are equally real. Garbage-in-garbage-out is the first: a poorly populated risk register or optimistically narrow distributions produce a confident-looking P80 that is nowhere near reality. The second is false precision — reporting P80 to the nearest million when the underlying impact ranges were expert guesses with order-of-magnitude error bars. The third is neglected correlation, which systematically understates tail risk. The fourth is using Monte Carlo to replace judgement rather than to structure it; the model cannot tell you which risks to include, only what the consequences look like once you have.
Practical safeguards are straightforward: build the risk register with the project delivery team rather than in a controls silo, calibrate distributions against historical project data where possible, model correlations explicitly rather than ignoring them, and integrate cost and schedule risk so that delay-driven cost impacts flow through the same model. Validated inputs are worth more than additional iterations.
Key Takeaways
- Monte Carlo simulation converts a deterministic cost estimate into a probability distribution by sampling thousands of scenarios across defined risk and variable distributions.
- The three inputs that matter are a credible base estimate, a disciplined risk register with probabilities and impact ranges, and appropriate distributions (triangular, normal, lognormal) with explicit correlations.
- P50 is the median outcome; P80 is a commonly used funding target that provides eighty percent confidence the final cost will not exceed budget.
- Contingency determined probabilistically is traceable to specific risks and defensible to sponsors, unlike a flat percentage uplift.
- The method’s output is only as credible as the risk register and distributions behind it — validate inputs before trusting iteration count.