Adam O'NeillLast updated 13 min readPart of Quantitative risk analysis (QRA)

What is Capital project estimate confidence level?

A capital project estimate confidence level is the probability that outturn cost will be at or below a stated figure. P50 is the median, with an even chance of being exceeded. P80 carries a one-in-five chance of being exceeded. Both are read from a Monte Carlo distribution of outturn costs, not from a single deterministic estimate.

Source: IPA Cost Estimating Guidance (2021), p.30, now owned by NISTA. Ranges apply to the Anticipated Final Cost (P50-equivalent including uncertainty and contingency). The Cost Estimating Requirements document gives -30% to +100% for SOC by exception.

Business case stageTarget range around the AFCBy-exception rangeTypical AACE estimate classProject definition
Strategic Outline Case (SOC)-20% to +50%-50% to +100%Class 5Less than 5%
Outline Business Case (OBC)-15% to +30%-30% to +50%Class 4 to 3About 30%
Final Business Case (FBC)-10% to +10%-10% to +20%Class 3 to 2More than 60%

What this guide is for

If you sponsor a capital project, sit on an investment committee or approve funding, you will be handed estimates labelled P50, P80, Anticipated Final Cost and optimism bias. The terms are precise and often misused, including in business cases that pass governance. The cost of approving a figure whose confidence position you do not understand is paid later, in overruns and in disputes about what was agreed.

This guide is written for the person approving the figure, not the analyst producing it. For how the percentiles are calculated and which AACE recommended practices apply, see the practitioner companion guide linked at the end.

What a confidence level means on a cost estimate

A confidence level is the probability that outturn cost will be at or below a stated figure. It is read from a Monte Carlo simulation: a cost model run thousands of times with varied inputs, which gives a distribution of possible outturn costs rather than one number. P50 is the figure with a 50% probability of not being exceeded. P80 has 80%. P95 has 95%.

Read it from the funder's side. In the worked example below the P80 is £226m. If you fund at £226m you accept a one-in-five chance of exceeding it. Fund at the P50 of £214m and you accept an even chance. The choice of percentile is a decision about how much overrun risk the funder carries, and the model does not make it for you.

Capital cost distributions are usually right-skewed: the tail of expensive outcomes is longer than the tail of cheap ones. That is why the gap from P50 to P95 is normally wider than the gap from P5 to P50, and why the mean sits above the median.

Why a deterministic estimate sits below the P50

A deterministic estimate is built from quantities and rates, usually most-likely values for each item, and is a single figure. It normally sits below the P50 for one main reason: it leaves out risk events and unquantified uncertainty. The IPA Cost Estimating Guidance (2021), now owned by NISTA (formerly the IPA), makes the point in its definition of the Anticipated Final Cost (AFC), which is the base estimate plus uncertainty plus contingency, presented at the P50-equivalent.

A second, smaller effect comes from adding up most-likely values. The sum of the most-likely values is not the most-likely total, and on a right-skewed model the total tends to sit higher. Treat that as a secondary effect. Missing risk is the primary one. A sponsor who is shown a deterministic figure labelled as the P50 should ask what risk allowance it contains.

The central estimate: P50, the Green Book and the mean

NISTA (formerly the IPA) publishes the Cost Estimating Guidance (2021), which remains the extant document. It says the cost estimate must present the Median Scenario / P-50 equivalent cost. That is the figure to treat as the central estimate in a UK business case that follows IPA guidance.

The Green Book (2026), in its section on expressing uncertainty in quantitative terms, describes the central estimate as the most likely value and treats expected value as an option. The three measures differ. The median (P50) has an even chance of being exceeded. The most likely value is the mode. The mean is the probability-weighted average. On a right-skewed distribution the mean is above the median. In the worked example the P50 is £214m and the mean is £216m, 0.9% higher. Ask which measure the paper calls the central estimate and make sure the same one is used throughout.

The P80 is a sensitivity on the central estimate. The funded figure is a separate decision with its own justification. A paper that shows only a budget at P80 hides the central estimate and invites a later dispute about what was approved.

The IPA target ranges at each business case stage

The IPA Cost Estimating Guidance (2021) sets target ranges for the cost estimate around the AFC, and the ranges narrow as the business case matures. They are asymmetric. The table gives them as published in the guidance. They apply to the AFC, which is the P50-equivalent figure with uncertainty and contingency included, and not to the base estimate.

The IPA Cost Estimating Requirements document is a separate publication and gives a different by-exception range at SOC (-30% to +100%). This guide follows the Guidance table. Check which document your gateway reviewers apply.

The ranges describe the spread expected of an estimate at that stage. They are not a pass mark for the P80. Comparing the P80 with the range is a useful sponsor check and is SOMA practice, not an IPA test: if the P80 falls outside the range around the P50, either the project carries unusual risk, which the case should explain, or the model is overstating it, which a peer review should find.

A worked example, with every number tied to its base

A £200m hospital scheme comes to FBC approval. The base estimate is £200m. This is the deterministic build-up with no allowance for risk or uncertainty. The QRA gives the following outturn cost figures.

P50: £214m. This is £14m, or 7.0%, above the base estimate. It is the AFC for the purposes of the IPA range.

Mean: £216m. This is 8.0% above the base and 0.9% above the P50.

P80: £226m. This is £26m, or 13.0%, above the base and £12m, or 5.6%, above the P50.

P95: £234m. This is £34m, or 17.0%, above the base, £20m, or 9.3%, above the P50 and £8m above the P80.

The FBC target range is -10% to +10% of the AFC, so it is applied to the P50: £192.6m to £235.4m. The base of £200m and the P80 of £226m both sit inside it. If the same range were applied to the £200m base, it would run from £180m to £220m and the P80 would appear to fall outside it. That would be the wrong test, because the base is not the AFC.

The funding decision follows. Approving at the base of £200m sits below the P50, so on this model overrun is more likely than not, and the figure does not meet the P50 requirement. Approving at the P50 of £214m accepts an even chance of overrun. Approving at the P80 of £226m provides £26m over the base, of which £14m is the P50 allowance and £12m is the gap between P50 and P80. Approving at the P95 of £234m adds £8m more for a further 15 percentage points of confidence.

The case should state who releases the £12m P50-to-P80 gap and on what trigger. A workable protocol is that the project holds the £14m under its delegated authority. The sponsor or investment committee releases the £12m only against a risk event from the register that has occurred, or an approved change, with each release logged against the driver that caused it.

Optimism bias and QRA contingency cover the same uncertainty

The Green Book (2026), in its section on optimism bias, requires an adjustment for the documented tendency of early estimates to come in low. The adjustment is applied top-down at the outset, using reference-class evidence. Default percentages come from supplementary guidance and departmental versions of it. Check the current table for your sector and year. This guide quotes no default rate.

As risks are identified and quantified, the Green Book says to reduce the optimism bias adjustment in proportion to the level of risk prevented or covered. What remains is the contingency, which the Green Book describes as the remaining optimism bias adjustment. The two methods are alternatives for the same uncertainty at different stages of knowledge. The QRA replaces the top-down uplift as it matures. It does not sit on top of it.

This means a QRA must not be run on a base that has already been uplifted for optimism bias. Doing so charges the same uncertainty twice, once as a percentage and again as modelled risk. Run the QRA on the unadjusted base estimate.

Take the worked example, with an illustrative top-down allowance of 20% of base, which is £40m. At FBC the quantified risks give a P50 allowance of £14m and a P80 allowance of £26m, which is 13.0% of base. The business case should show the modelled allowance at the funded percentile as the contingency, not £40m plus the modelled allowance. Any residual allowance for risks the QRA does not cover is a separate, stated judgement with its own reason.

The Green Book also says contingency should typically not be allocated directly to the project, because government self-insures. A request to fund at P80 therefore needs its own justification, for example a funder's risk appetite or a requirement set by the sponsoring department. The Green Book does not require P80. Its percentile example is a reference-class device: choose the road at the 90th percentile of a list of comparable roads, one exceeded by only one in ten similar projects. It is not a Monte Carlo output. The Green Book directs high-cost, high-risk proposals to Monte Carlo analysis and cautions that a P90 reference value can understate tail risk.

Choosing the funded percentile

P80 is common practice for the upper-bound sensitivity on UK public capital programmes. It is a convention, not a rule, and no document cited in this guide mandates it. The Green Book does not say P80, so "P80 because the Green Book says so" is not a defensible reason. "P80 because the funding body has set that risk appetite" is.

The right percentile depends on the consequence of overrun, the funder's appetite for it and the price of extra certainty. In the worked example, moving from P80 to P95 costs £8m for 15 additional percentage points of confidence. Moving from P50 to P80 costs £12m for 30. Those marginal costs are what the sponsor is buying, and the paper should show them.

A P50 funding position is legitimate where a funder knowingly accepts a one-in-two chance of overrun in return for lower capital commitment. It must be stated as such. The failure mode is funding at P50 while describing the position as P80 confidence.

Whether to fund above P80 is a matter for the funder's own rules. This guide cites no sector regime that sets P95, so check the rules that apply to your programme. A portfolio sponsor should note that diversification across projects puts the portfolio P80 below the sum of the individual project P80s, so adding up project P80s overstates the portfolio position.

Six questions to put to the QRA

These are the six questions to ask before approving any QRA-based figure. The decision section below refers back to them.

One - is the risk register specific to this programme? A register copied from a template gives generic numbers. Ask for the risks most specific to this scheme and the evidence behind each range. If the team cannot name them, the analysis is generic.

Two - how was correlation handled? Zero correlation, the default in many tools, usually understates the spread because risks share causes such as ground conditions, market capacity or one contractor. Ask to see the correlation assumptions and the pairs that matter most.

Three - does the output distribution make sense? A right-skewed shape is the normal pattern, but shape alone proves nothing. A narrow output usually means input ranges were set too tight, which is optimism, or that correlation was left out. Ask for the histogram and the input ranges, and compare the ranges with benchmark data or with outturns on similar schemes.

Four - what drives the result? The tornado chart or sensitivity table should show a handful of dominant drivers, typically three to six as a SOMA rule of thumb. A flat chart where every input contributes equally suggests a generic model. Ask for the top drivers by name and the mitigation planned for each.

Five - what has changed since the last run? Tie currency to events, not to a calendar: a change control, a re-baseline, a stage gate, a major procurement outcome or a significant risk occurring. If any of these has happened since the last run, the figures are out of date.

Six - who has reviewed it independently? A QRA reviewed by someone outside the project team carries more weight at gateway than one that has not. Ask for the reviewer, the date and the findings.

The decision in front of you

Before approving, you should be able to state six things in writing: the P50 central estimate and the measure used; the funded percentile and why it was chosen; the target range for the stage and where the P50 and P80 sit against it; whether optimism bias was applied and how it was reduced as risks were quantified; who holds and who releases the contingency; and the answers to the six questions above.

If any of these is missing, send the case back for the evidence before approving. Approving against an incomplete confidence picture is what produces disputes later about what was agreed at sanction.

On a pre-sanction QRA review we ask for the risk register, the correlation assumptions, the histogram and the sensitivity ranking, then test the result against wider benchmark ranges. We report whether the P50 and P80 move and which drivers move them, so the sponsor has the six answers before the paper is signed.

Strengthening your QRA function?

SOMA runs QSRA, QCRA and integrated risk analysis to AACE recommended practice: workshop facilitation, calibrated three-point inputs, risk-driver Monte Carlo and a report written so a board can act on the number. Send us your scope and we will tell you what it needs.