PERT

Goal

Three-point estimation asks whoever knows the work for the three bounds between which they place their estimate, and it derives two numbers from them: an expected value and a measure of its spread. It thus makes visible what a single number hides: giving one number carries a lie by omission, because it says nothing about its own reliability, and two tasks both quoted at six days can cover opposite realities, one almost certain, the other liable to double.

The decision it supports is a commitment made under uncertainty. Setting a delivery date, reserving a budget, comparing two options one of which is riskier than the other at equal expectation, all of this requires knowing the average cost of a thing and the margin around it. PERT supplies that margin in a quantified and combinable form. The expected values of several tasks add up into a project estimate, and their spreads compose into an overall confidence interval. A manager can then state "the project runs nineteen days, and there is about a nine-in-ten chance it lands between fifteen and twenty-three days", a claim of an entirely different value from a bare "nineteen days" for anyone who has to commit to it.

The deliverable is a table: one row per component, the three elicited values, the expected value and the standard deviation that follow from them, then a summary row carrying the project estimate and its confidence interval. PERT is one of the methods in the estimation family, the one that fills this table at the finest elements of a breakdown; the total it produces then reads as a range with a probability attached, where a single-point estimate leaves only the choice between excessive caution and blind optimism.

Usage

When to use it

• Uncertainty to quantify: the decision needs a margin in numbers around the estimate.
• A project-level confidence interval: several components to add into an overall probabilistic range, which only a combinable spread allows.
• Two options at comparable expectation, different risk: PERT exposes which is the tighter, a distinction a single number erases.
• A breakdown already available: the elements to estimate exist at a fine level, PERT estimates them at three points then sums them.
• Experts able to name three credible values: an optimistic, a most likely and a pessimistic value grounded in experience or an analogue.

When not to use it

• Too little information for three credible values: at early strategic framing, prefer the rough order of magnitude (ROM), a single figure with an openly wide range, then return to PERT once the ground is known.
• Repetitive units with a calibrated model: when history feeds a per-unit cost or effort model, parametric estimation is faster and often tighter than eliciting three points each time.
• No shared quantitative basis: expert disagreement that no formula settles, first seek consensus through the Delphi method, whose converged values can then feed a PERT calculation.

Description

The three values

The technique begins by eliciting, for each component to be estimated, three durations or three costs from the person who will do the work. The optimistic value (O) is the one reached if everything runs without a hitch, without being a miracle for all that. The most likely value (M, for most likely) is the one the estimator would keep if they had to give only one, the mode of their belief. The pessimistic value (P) is the one reached if the reasonably foreseeable difficulties occur. The care taken over P decides the quality of everything else, because it carries most of the spread. A P set as "a bad day" rather than a genuine adverse scenario artificially compresses the uncertainty and gives a falsely reassuring confidence interval.

Weighting the expected value

PERT reduces the three values to an expected value, written tE, by a weighted average:

tE = (O + 4M + P) / 6

The most likely scenario weighs four times as much as each extreme. This weighting approaches the mean of a beta distribution tuned so that its bounds are O and P and its mode M. The beta has the property that matters for estimation: it can be asymmetric. When the pessimistic tail is long, say an optimistic value at three days, a mode at five, a pessimistic value at thirteen, its mean shifts beyond the mode, pulled toward the adverse side. The expected value is then six where the most likely scenario was five, and that one-day gap is the risk the tail places on the component. A single-point estimate set on the mode would have ignored it.

This formula is commonly confused with another. The confusion is a substantive error. The triangular mean, (O + M + P) / 3, gives equal weight to the three values. It describes a triangular distribution, not a beta, and suits cases where the history is too thin to justify the beta weighting or where the mode is not more likely than the average of the three. The PMBOK Guide, 8th edition, writes that formula under the name multipoint estimating; the beta weighting stays with the 1959 paper. The two formulas coexist in practice as in the textbooks and are often cited one for the other. PERT is the four-weighted version, (O + 4M + P) / 6; calling the other one "PERT" is a misreading that silently changes every estimate.

The spread: standard deviation and variance

The same triplet supplies a measure of the uncertainty. The standard deviation of the component is the range divided by six:

σ = (P − O) / 6    and so    σ² = ((P − O) / 6)²

The division by six has the same origin as the weighting: the range P − O stands in for the six standard deviations that, in the assumed distribution, separate the low bound from the high bound, since most of a distribution spreads over roughly three standard deviations on either side of the centre. A component with a small range is therefore tight, a component with a wide range is uncertain, and two components with the same expected value can have standard deviations with nothing in common. That is the reading PERT makes possible and the single number forbids.

Rolling up to the project level

Expected values add up simply: the project estimate is the sum of its components' tE. Spreads, however, do not add the same way. It is the variances that sum, and the project standard deviation is the root of that sum:

σ_project = √(Σ σ²)

Adding the standard deviations directly is the most frequent mechanical error and it overstates the spread. The root of the sum of the variances is the right operation, and it rests on an independence assumption. The formula is valid only if the components' uncertainties are uncorrelated, that is if no common risk factor links them. If three components depend on the same supplier, the same critical resource or the same regulatory constraint, a delay on one accompanies a delay on the others, and the root of the sum of the variances understates the project's real risk because it treats as independent uncertainties that move together. The assumption must be stated next to the result. Where it does not hold, the remedy is to flag it qualitatively or to turn to a simulation that models the correlation.

Under the independence assumption, the sum of a sufficient number of components tends toward a normal distribution even if each follows a beta, which allows the confidence interval to be read on the normal law. A project expected value plus or minus 1.645 times its standard deviation covers about 90 % of outcomes; plus or minus 1.96 times, about 95 %. This band is PERT's deliverable. A last pitfall lodges in its presentation: a tE shown to two decimals on inputs that are expert judgements displays a precision the data do not carry. The table keeps the decimals so that the calculation can be rechecked to the digit; the estimate one commits to, for its part, is stated as a rounded range, fifteen to twenty-three days. The range tells the truth the decimal disguises.

AI considerations

The calculation calls for no artificial intelligence: the weighted average, the variance and the root of their sum fit in a few spreadsheet cells. A model's useful contribution lies upstream of the three values and downstream of the total.

Upstream, a language model helps to ground the three points. Asked about the history of comparable tasks, the closed tickets of an earlier project, the durations actually observed on analogous work, it proposes an optimistic, a most likely and a pessimistic value with support that an expert corrects, which beats a triplet set from memory. It can also produce, from a single set of components, the three readings a decision needs, the expected value, the standard deviation and the confidence interval. It is downstream that it gives the most, on the independence assumption: a Monte Carlo simulation, which an assistant sets up quickly, replays the project thousands of times letting the correlated components vary together and gives the real interval where the root of the sum of the variances gives only an approximation valid under independence.

What the machine cannot supply lies in the inputs and the assumptions. The three values are an expert judgement about this specific work, with its own hazards, and a model queried without that context produces plausible, groundless numbers. The pessimistic value above all demands knowing what can go wrong on this task, field knowledge no general model holds. And the finding of independence or correlation is a business reading: it is the analyst who knows that three components pass through the same integrator, information that follows from no table of numbers and yet decides whether the consolidated standard deviation is honest.

Examples

A French-speaking Swiss SME estimates the effort for an overhaul of its payroll system, which must correctly compute AVS and LPP contributions before go-live. The work is broken into four packages, each estimated at three points in person-days.

Data migration and the AVS/LPP interface have the same expected value, six days, and a single number would show them identical. Their variances separate them nonetheless: 2.778 against 0.444, a standard deviation of 1.67 days against 0.67. The migration is the real schedule risk, the interface is all but settled, and only three-point estimation surfaces that difference where a single six would hide it. At the project level, the expected values sum to nineteen days; the variances sum to 6.111, whose root gives a project standard deviation of 2.47 days, provided the four packages are estimated independently, an assumption to check in this case since a single integrator could link several packages. The roughly 90 % confidence interval is 19 ± 1.645 × 2.47, a range of about fifteen to twenty-three person-days. Applied to a consultant day rate of CHF 1'350, the expectation comes to about CHF 25'650 and the range to about CHF 20'250 to CHF 31'050.

Visualisations

The technique produces two objects, and each calls for its own medium. The first is the estimating table: made of rows and columns, it is the deliverable itself, and its columns are what a reviewer recomputes to check the estimate. It stays in HTML, selectable and recomputable, rather than frozen into an image.

The second is the logic of the formula, which is better seen drawn than tabulated. The beta distribution curve of one package, the data migration, carries its three values on the axis, the mode at the peak and the expected value marked to the right of that peak, pulled by the long pessimistic tail. It shows at a glance why the expected value exceeds the most likely scenario and why the range measures the spread. The curve reveals the mechanics of the weighting, the table carries the reproducible calculation.

Cost

Tools

The spreadsheet is the honest choice and hard to beat. The three formulas sit in a few cells, the total and the interval recompute at every change of a value, and sensitivity analysis, replaying the estimate under revised triplets, happens in the same workbook. For a project of a few dozen components, nothing more is needed.

Project-planning tools often support three-point estimation natively: you enter O, M and P on each task and the tool computes the expected value and propagates the durations through the schedule, which avoids keeping a spreadsheet alongside the plan. Where the independence assumption does not hold, Monte Carlo simulation modules or plug-ins take over: they draw thousands of scenarios letting the correlated components vary together and give the real confidence interval, which the root of the sum of the variances can give only under independence. This extra tooling is justified only when the correlations are real and the stakes of the estimate high; below that, it adds a licence without providing anything the spreadsheet does not already do.

Sources

• IIBA, A Guide to the Business Analysis Body of Knowledge (BABOK Guide) v3, §10.19 Estimation: the placement of PERT among the estimation methods, the weighted average that grants four times more weight to the most likely scenario and the tie between estimation and the confidence interval, the variance and the standard deviation as measures of precision. The BABOK describes these methods and does not prescribe a variance formula, which comes from the 1959 founding paper.
• D. G. Malcolm, J. H. Roseboom, C. E. Clark and W. Fazar, "Application of a Technique for Research and Development Program Evaluation", Operations Research, 7(5), 1959, pp. 646-669: the founding paper of PERT, developed at the US Navy's Special Projects Office for the Polaris programme, which sets out the beta-distribution approximation and the derivation of the expected value and the variance from the three estimates.
• PMI, A Guide to the Project Management Body of Knowledge (PMBOK Guide), 8th edition, multipoint estimating: the PMBOK's own name for the technique and its definition, an average or weighted average of the optimistic, most likely and pessimistic estimates where an activity estimate is uncertain, together with its use in estimating durations and costs. The one formula it writes out is the triangular mean tE = (tO + tM + tP) / 3, given as "one commonly used formula"; the four-weighting and the standard deviation (P − O) / 6 come from the 1959 paper.