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.
What you get out of a QCRA
A good QCRA gives you four things a point estimate cannot:
- An S-curve showing the probability of every possible total cost.
- 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."
- Contingency: how much you need to add to the base estimate to reach the confidence level you choose.
- 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 uncertainty | Risk events | |
|---|---|---|
| What it is | Variation in quantities, rates and prices of known scope | Discrete events that may or may not happen |
| Example | Earthworks could cost $28 to $38 million, most likely $30 million | 35% chance of contaminated soil, costing $2 to $9 million if it occurs |
| How it's modeled | A range (minimum, most likely, maximum) on each estimate line | A probability of occurrence plus an impact range |
| Where it comes from | Estimators and the basis of estimate | The 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
- 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.
- 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.
- 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.
- 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.
- 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.
- Read the results. Build the S-curve, read off P50, P80 and P90, and rank the drivers.
- 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 $120 million. The board's question is simple: what budget do we commit to?
Steps 1 and 2: ranges on the estimate. The estimating team gives a minimum, most likely and maximum for each cost area. Notice the skew: every range reaches further up than down.
| Cost area | Minimum ($M) | Most likely ($M) | Maximum ($M) |
|---|---|---|---|
| Bridge | 40.0 | 42.0 | 55.0 |
| Earthworks and roads | 28.0 | 30.0 | 38.0 |
| Systems and equipment | 14.0 | 15.0 | 18.0 |
| Project management and owner costs | 11.0 | 12.0 | 14.0 |
| Utility relocation | 9.0 | 10.0 | 15.0 |
| Design and engineering | 7.5 | 8.0 | 10.0 |
| Land | 3.0 | 3.0 | 5.0 |
| Total (point estimate) | 120.0 |
Step 3: risk events. Six risks from the register get numbers attached. Together, their expected value is $7.1 million.
| Risk event | Probability | Impact if it occurs ($M, min / most likely / max) |
|---|---|---|
| Extra traffic management needed | 50% | 0.5 / 1.0 / 2.0 |
| Steel price escalation | 40% | 1.0 / 2.5 / 5.0 |
| Contaminated soil found | 35% | 2.0 / 4.0 / 9.0 |
| Permit appeal delays start by six months | 30% | 2.0 / 3.5 / 7.0 |
| Utility owner delays force acceleration | 25% | 1.5 / 3.0 / 6.0 |
| Poor ground at bridge foundations | 20% | 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 level | Total cost ($M) | Contingency above $120M | As % of estimate |
|---|---|---|---|
| Point estimate holds | 120.0 | 0 | 0% (1% chance) |
| Point estimate + flat 10% | 132.0 | 12.0 | 10% (33% chance) |
| P50 | 135.8 | 15.8 | 13% |
| P80 | 143.4 | 23.4 | 20% |
| P90 | 147.6 | 27.6 | 23% |
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.
- 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.
- 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).
- 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.
- 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.
Common mistakes
- Ranges that are too narrow. People anchor on the estimate. Ask for the extremes first and test them: "what would have to happen for this to cost a lot more?"
- Ignoring correlation. Book (1999) showed that with 30 cost items correlated at just 0.2, treating them as independent understates the spread of the total by about 60%. In the Millbrook example, dropping the 0.4 correlation shrinks the standard deviation from $8.4 million to $6.7 million and lowers the P90 by $2.4 million. Our guide on correlation shows the effect in more detail.
- Counting a risk twice, once inside an estimate range and again in the register.
- Adding up expected values as the reserve. The sum of probability times impact is a total that almost never happens; see why the sum of EMVs is not a risk reserve.
- Treating P-values as targets. A P80 budget should be underspent four times out of five. If teams spend up to the budget anyway, the analysis stops meaning anything.
- Running it once. A QCRA isn't a stage-gate document, it's a control tool. The real value is in watching the range narrow over time.
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
- AACE International. Recommended Practice 41R-08: Risk Analysis and Contingency Determination Using Range Estimating.
- AACE International. Recommended Practice 44R-08: Risk Analysis and Contingency Determination Using Expected Value.
- AACE International. Recommended Practice 57R-09: Integrated Cost and Schedule Risk Analysis Using Monte Carlo Simulation of a CPM Model.
- Book, S. A. (1999). Why correlation matters in cost estimating. 32nd Annual DoD Cost Analysis Symposium, Williamsburg, VA.
- Crossrail Learning Legacy. Contingency management.
- Department of Finance, Australian Government. Defining P50 and P80.
- Department of Infrastructure, Australian Government (2023). Cost Estimation Guidance, Guidance Note 3B: Deterministic Contingency, v2.0.
- Flyvbjerg, B., Holm, M. S., and Buhl, S. (2002). Underestimating costs in public works projects: Error or lie? Journal of the American Planning Association, 68(3), 279–295.
- Flyvbjerg, B., Holm, M. K. S., and Buhl, S. L. (2003). How common and how large are cost overruns in transport infrastructure projects? Transport Reviews, 23(1), 71–88.
- Horvat & Partners (2024). Berekening bandbreedte prognose dijkversterkingsopgave, for the Dutch Ministry of Infrastructure and Water Management.
- Kujawski, E. (2002). Why projects often fail even with high cost contingencies. Systems Engineering, 5(2), 151–155. doi:10.1002/sys.10006
- Love, P. E. D., and Ahiaga-Dagbui, D. D. (2018). Debunking fake news in a post-truth era: The plausible untruths of cost underestimation in transport infrastructure projects. Transportation Research Part A, 113, 357–368. doi:10.1016/j.tra.2018.04.019
- Lovallo, D., and Kahneman, D. (2003). Delusions of success: How optimism undermines executives' decisions. Harvard Business Review, 81(7), 56–63.
- NASA (2019). NID 7120.122: Joint Cost and Schedule Confidence Level (JCL) Requirements Updates.