Skip to content
Find your tool
Guides · 14 min read

Quantitative cost risk analysis: a practitioner's guide

How to turn a single cost estimate into a range of outcomes with probabilities, and how to turn that range into budget decisions. With a complete worked example from a fictional road and bridge project.

By Mike Bouman, Quantil · November 4, 2026

Most project budgets start life as one number. "This will cost $120 million." It sounds firm, it fits on a slide, and it's almost always wrong. Not because the estimators did a bad job, but because a single number can't tell you how likely it's to hold.

Quantitative cost risk analysis (QCRA) turns a single cost estimate into a range of possible outcomes, each with a probability, so you can choose a budget with a known level of confidence. Instead of "$120 million," you get something like: "There's a 50% chance we stay under $136 million and an 80% chance we stay under $143 million." Less comfortable to hear. Much easier to manage.

This guide covers why single numbers let you down, how a QCRA works step by step, and how to turn the results into real budget decisions. The second half works through a complete, fictional example.

Why single-number estimates let you down

The track record isn't great. Flyvbjerg, Holm and Buhl looked at 258 transport infrastructure projects in 20 countries. Nine out of ten went over budget. On average, costs came in 28% higher than estimated: 45% for rail, 34% for bridges and tunnels, 20% for roads (Flyvbjerg et al., 2003).

Why that happens is still debated. Flyvbjerg and his colleagues point to optimism bias and, at times, strategic misrepresentation (Flyvbjerg et al., 2002). Love and Ahiaga-Dagbui (2018) push back hard. They call the "error or lie" framing a false choice and argue that much of the gap comes from how scope develops and how estimates mature. For day-to-day practice, the debate matters less than the part both camps agree on: an early estimate is uncertain, and presenting it as one number hides that from the people who carry the risk.

There's also a simple math problem. Most cost items are skewed. They can overrun by a lot, but they rarely come in far below. So when you add up the most likely value of every item, you don't get the most likely total. You get a number well below where the project will probably land. In the example later on, the most likely values add up to $120 million. The simulated average is $136 million. The $120 million has about a 1% chance of holding.

A practical tip. On many projects the first number on the table becomes the budget, even when everyone knows it’s an early estimate. Then the first risks hit, and each one feels like the project team has failed. But those risks were inside the range all along. Show the range up front, and those moments become expected events instead of bad news.

What you get out of a QCRA

A good QCRA gives you four things a point estimate cannot:

  1. An S-curve showing the probability of every possible total cost.
  2. Confidence levels such as P50, P80 and P90. Each one is the cost you won't exceed with that probability. The Australian Department of Finance puts it well: "P80 isn't a cost plus/minus 20% but instead it's a cost that won't be exceeded 80% of the time."
  3. Contingency: how much you need to add to the base estimate to reach the confidence level you choose.
  4. Drivers: a ranking of the uncertainties and risks that move the total most, so you know where your attention pays off.

Two kinds of uncertainty

Model these separately, because they behave very differently.

Estimate uncertaintyRisk events
What it isVariation in quantities, rates and prices of known scopeDiscrete events that may or may not happen
ExampleEarthworks could cost $28 to $38 million, most likely $30 million35% chance of contaminated soil, costing $2 to $9 million if it occurs
How it's modeledA range (minimum, most likely, maximum) on each estimate lineA probability of occurrence plus an impact range
Where it comes fromEstimators and the basis of estimateThe risk register

AACE International covers both in its recommended practices: range estimating for estimate uncertainty (RP 41R-08), expected value for risk events (RP 44R-08), and integrated cost and schedule analysis with Monte Carlo simulation (RP 57R-09). There's a third layer as well: systemic risk, such as a scope that's still immature. That's hard to capture in a register. The best check is to compare your outcome with similar past projects, an approach known as reference class forecasting (Lovallo and Kahneman, 2003).

How it works, in seven steps

  1. Fix the base. Agree on which estimate, scope, price date and exclusions you're analyzing. Run a QCRA on a moving estimate and you'll get moving answers.
  2. Put ranges on the estimate. Hold a short workshop per cost area. Ask for the minimum and maximum before the most likely value, so people don't anchor. If a range looks suspiciously narrow, challenge it.
  3. Quantify the risk register. Give each material risk a probability and an impact range. Take out anything the estimate ranges already cover, so nothing gets counted twice.
  4. Add correlation. Costs move together. If steel is expensive for the bridge, it's expensive for the barriers too. Leave this out and your result looks far more certain than it really is.
  5. Simulate. Draw random values for every range and risk, add them up, and repeat 10,000 times or more. Each run is one plausible version of the project.
  6. Read the results. Build the S-curve, read off P50, P80 and P90, and rank the drivers.
  7. Decide and govern. Pick the confidence level for the budget, decide who holds the contingency, and agree how it gets released. This is where most of the value is; more on that below.

Worked example: the Millbrook Connector

Illustrative case. The project, names and figures are fictional, built to show the method.

Picture a regional transport authority planning the Millbrook Connector: 8 km of road widening with a new river bridge. The estimate adds up to $120 million. The board's question is simple: what budget do we commit to?

Steps 1 and 2: ranges on the estimate. The estimating team gives a minimum, most likely and maximum for each cost area. Notice the skew: every range reaches further up than down.

Cost areaMinimum ($M)Most likely ($M)Maximum ($M)
Bridge40.042.055.0
Earthworks and roads28.030.038.0
Systems and equipment14.015.018.0
Project management and owner costs11.012.014.0
Utility relocation9.010.015.0
Design and engineering7.58.010.0
Land3.03.05.0
Total (point estimate)120.0

Step 3: risk events. Six risks from the register get numbers attached. Together, their expected value is $7.1 million.

Risk eventProbabilityImpact if it occurs ($M, min / most likely / max)
Extra traffic management needed50%0.5 / 1.0 / 2.0
Steel price escalation40%1.0 / 2.5 / 5.0
Contaminated soil found35%2.0 / 4.0 / 9.0
Permit appeal delays start by six months30%2.0 / 3.5 / 7.0
Utility owner delays force acceleration25%1.5 / 3.0 / 6.0
Poor ground at bridge foundations20%3.0 / 6.0 / 14.0

Steps 4 and 5: correlation and simulation. The cost areas are correlated at 0.4, a common choice for cost estimates when there's no better data. The model uses triangular distributions and runs 100,000 iterations.

Step 6: the results.

Confidence levelTotal cost ($M)Contingency above $120MAs % of estimate
Point estimate holds120.000% (1% chance)
Point estimate + flat 10%132.012.010% (33% chance)
P50135.815.813%
P80143.423.420%
P90147.627.623%

A few things jump out. The $120 million estimate has roughly a 1% chance of holding. The classic flat 10% allowance gives you about a one-in-three chance of staying on budget. And most of the gap doesn't even come from the risk register. The skewed ranges alone push the average from $120 million to $129 million; the risk events add about $7 million on top.

The biggest drivers are the bridge and earthworks ranges, followed by the risk of poor ground at the bridge foundations. That's useful to know: it tells the team an extra site investigation would probably pay for itself.

From numbers to decisions

The numbers support four decisions. None of them is a calculation. Each one is a choice about how much risk you're willing to carry and how you want to stay in control.

  1. Which confidence level becomes the budget? Organizations choose differently. The Australian Government requires both a P50 and a P90 estimate for projects above $25 million that seek Commonwealth funding (Department of Infrastructure, 2023). NASA budgets its programs and projects at a 70% joint cost and schedule confidence level (NASA, 2019). In the Netherlands, the forecast for the national flood defense program was reported as a P50 of €18.4 billion, with a range of €14.3 billion (P15) to €22.5 billion (P85) (Horvat & Partners, 2024). P50, P80 and P90 explained goes deeper into the choice.
  2. Who holds the contingency? Kujawski (2002) shows that projects can overrun even with plenty of contingency if you split it across sub-projects up front. A pooled reserve, released to risks as they happen, works better. Crossrail held contingency at several levels, from the board down to project managers, each with formal delegated authority (Crossrail Learning Legacy).
  3. When can contingency be released? Agree the rules before you need them: what counts as a valid claim, how much each level can approve, and what gets reported.
  4. When do we run it again? At every stage gate, after major contract awards, and whenever one of the top five drivers changes. Risks that have passed should hand their contingency back to the program.
A practical tip. Take this to the steering group on a single page: three numbers (P50, the proposed budget level and a high case), the top three drivers, and the four decisions written as proposals they can approve. A useful rule of thumb: qualitative where you can, quantitative where you must. For day-to-day control, a sharp risk register is usually enough. But as soon as a decision depends on a budget, a reserve or a completion date, show the uncertainty in probabilities and amounts. The discussion nearly always ends up at the same question, who holds the contingency, so agree the layers and approval limits before the first claim comes in.

Common mistakes

Is it worth the effort?

If an overrun would force a new funding decision, or if several projects share one reserve, almost always. For small, repeatable work with good historical data, a factor-based contingency is often enough. The Australian guidance draws the line at $25 million. Where your line sits depends on what a surprise would cost your organization.

Frequently asked questions

Is Monte Carlo better than a percentage contingency?

It's more honest. A percentage tells you nothing about your chance of staying on budget. A simulation tells you that chance, and what is driving it. See why a percentage on top is not a risk budget.

How many iterations do I need?

For stable P80 and P90 values, 10,000 is usually plenty. Run it twice. If the percentiles shift by more than a fraction of a percent, add more.

Do I need special software?

No. You can run a simulation in a spreadsheet; how to calculate contingency with Monte Carlo shows the formulas. Dedicated tools save time on correlation, reporting and audit trails.

Can I combine cost and schedule risk?

Yes, and you often should. Delays cost money through escalation and time-related costs. AACE RP 57R-09 describes an integrated approach.

Quantil's QRCIM runs this full analysis in Excel, from estimate ranges and risk register to S-curve, drivers and board report. Try the free demo.

References