PM Planning

Project Timeline Estimation: PERT, Contingency, Example

By Vact Published · Updated

A defensible timeline estimate is a range with a stated confidence, not a single date. The standard way to build one is the three-point (PERT) estimate: for each activity, collect an optimistic time (O), a most likely time (M) and a pessimistic time (P), then compute the expected duration as (O + 4M + P) / 6 and the standard deviation as (P - O) / 6. Add the expected durations along the controlling path, add the variances (the squared standard deviations) along the same path, and read off the date that matches the confidence the sponsor needs. The difference between that date and the plain schedule is the contingency, and it is shown separately from the work.

The method comes from the US Navy’s Polaris Fleet Ballistic Missile program: Malcolm, Roseboom, Clark and Fazar described it in Operations Research in October 1959 as a technique for measuring and controlling development progress, with time estimates obtained from responsible technical staff and then expressed in probability terms. PERT works alongside other common ways to set an activity’s duration, and each has its own weak point:

MethodUse it whenWhat it needsMain weakness
Three-point (PERT)One task has a real range of outcomesO, M and P for each task, with reasonsInherits any optimism in O, M and P
AnalogousA similar task or project is already finishedActual durations from that workOnly as good as the match
ParametricDuration scales with a measurable quantityA history showing the relationshipNeeds enough data to show it
Expert judgmentNothing comparable existsThe person who will do the workCannot be audited without a stated basis

Try the four-activity timeline calculator

The network is A, followed by parallel activities B and C, followed by D. D can begin only when both B and C finish. Enter durations in the same working-day unit. The calculation assumes B and C have the resources to run in parallel and uses no holidays or imposed dates. Add contingency as a separate number of working days.

With the example values, A-B-D takes 9 working days and A-C-D takes 7. B controls the finish.

The calculator runs in your browser and does not save or submit the entries. The critical-path guide works through a full forward and backward pass, which finds the controlling path and each activity’s float. The calculator takes one number per activity, so enter the most likely values to see the plain schedule, then the expected values from the PERT table below (rounded to its half-day step) to see the effect of uncertainty.

Three-point (PERT) estimates: formula and standard deviation

For each activity, collect three durations in working days, each with its reason:

  • O, optimistic: the best case if the main uncertainties resolve favorably.
  • M, most likely: the single most probable duration under normal conditions.
  • P, pessimistic: the worst case if the main uncertainties resolve unfavorably.

NASA’s Schedule Management Handbook (NASA/SP-2010-3403, Revision 1, effective 30 January 2020) recommends collecting these three values with “supporting rationale for each value.” It adds that where a task is very well defined and historically substantiated, such as an off-the-shelf procurement or standardized testing, the three values may be the same or nearly so, and an uncertainty estimate may not be necessary. It also says the person assigned the task must keep ownership of its schedule, including review and approval of its durations, and that durations should not be “padded in order to keep a hidden cushion, reduced to be unrealistically optimistic, or arbitrarily cut by management.”

The classic PERT formulas, as presented in Carnegie Mellon’s construction scheduling text, are:

QuantityFormulaBasis
Expected durationE = (O + 4M + P) / 6Mean of a beta distribution with a restricted shape
Standard deviationSD = (P - O) / 6Range treated as six standard deviations
VarianceSD squared, or (P - O)^2 / 36Added along a path, if durations are independent

These formulas treat O and P as absolute limits. Because absolute limits are hard to estimate from historical data, the same text describes a common practice of setting O and P at the 5th and 95th percentiles instead, so that 90% of outcomes fall between them. The expected duration is unchanged, but the variance divisor becomes 10 instead of 36, so the same O and P give a variance 3.6 times as large. Whichever reading a team uses, it should state it, because the choice changes every probability that follows.

Worked example on the calculator’s network

The most likely values below are the calculator’s defaults, so the plain schedule is A + B + D = 2 + 5 + 2 = 9 working days. The optimistic and pessimistic values are invented for the illustration: A and D are routine, while B and C each carry an unresolved risk, such as an unproven integration, that could add seven working days.

TaskOMPExpectedSDVariance
A: preparation1252.330.670.44
B: first parallel35125.831.502.25
C: second parallel23104.001.331.78
D: final1252.330.670.44

Path A-B-D has an expected duration of 2.33 + 5.83 + 2.33 = 10.5 days and a variance of 0.444 + 2.25 + 0.444 = 3.14, so its standard deviation is the square root, 1.77 days. Path A-C-D has an expected 8.67 days and a standard deviation of 1.63 days. The expected finish is 10.5 days, one and a half days beyond the 9-day plan, because every pessimistic tail is longer than its optimistic one.

From path variance to a date

Treat the controlling path as roughly normal, as the Carnegie Mellon text does in its own example. The probability of finishing by a target day T is the standard normal cumulative probability of (T - 10.5) / 1.77.

Target finishz-scoreChance of finishing by then
9 working days (the plain schedule)-0.85about 20%
10.5 working days (the expected value)050%
12 working days+0.85about 80%

An 80% commitment therefore needs 12 working days, which is 3 days of contingency above the plain 9-day schedule. The other path is nowhere near binding at that date: A-C-D has roughly a 98% chance of finishing within 12 days. That check matters because the Carnegie Mellon text warns that PERT “focuses upon a single critical path” and so “typically underestimates the actual project duration,” since a path with a shorter expected time can become the longest on a given run. It also warns that positively correlated durations make PERT underestimate the variance, producing over-optimistic expectations of meeting a deadline. In its nine-activity example, 500 simulated runs averaged 30.9 days against a PERT figure of 30.0, about 3% higher; with a correlation of 0.5 between each pair of four activities, the simulated mean rose to 36.5 days, about 20% above PERT.

A simulation shows the single-path effect on this network. Our own 200,000-run Monte Carlo of the four tasks draws each duration independently from the beta-PERT distribution, the beta shape with parameters 1 + 4(M - O) / (P - O) and 1 + 4(P - M) / (P - O), whose mean is exactly (O + 4M + P) / 6. It gives a mean of 10.7 days and an 80th percentile of 12.2 days, close to the hand calculation; the extra 0.2 days comes from the roughly 18% of runs in which C outlasts B. Drawing the same three-point estimates from triangular distributions, which GAO’s Schedule Assessment Guide uses for its house-construction example, raises the mean to 12.4 days and the 80th percentile to 14.3, because a triangular distribution’s mean is (O + M + P) / 3, which gives the long upper tails more weight. The optimistic, most likely and pessimistic values are the same in both runs. The shape assumed between them is not, so state the distribution next to any percentile.

Rounded to the calculator’s half-day step, the expected values are A = 2.5, B = 6, C = 4 and D = 2.5, which give 11 days. Entering contingency of 1 day brings that to 12.

Analogous and parametric estimates

Both methods start from history instead of a forecast made from scratch. The NASA handbook describes them as two of several ways to derive or check a duration:

  • Analogy: the actual duration of a similar activity is used as the basis for the new one, “often adjusted for differences in complexity,” with careful thought given to choosing the analogy.
  • Parametric analysis: a “mathematical relationship that defines schedule as a function of one or more parameters.” NASA calls these Schedule Estimating Relationships and compares them with the Cost Estimating Relationships used for budgets. They apply only if the driver shows a measurable relationship with duration, and the handbook admits that getting data good enough to quantify it “can be difficult.”

An illustrative case: the last help-center launch took 20 working days for 12 articles, or about 1.7 days per article including review. If the next one has 18 articles, the analogy gives roughly 30 working days, to be adjusted for any difference in reviewers or tooling. A parametric version would use several past launches to fit a fixed setup time plus a per-article rate, and would be trusted only once those launches showed a consistent relationship. Neither replaces a range. The analogy is a most likely value for M, and the spread among past launches is a defensible basis for O and P.

For the underlying scope list, build the estimate from a work breakdown structure so that each estimated activity has a deliverable and an owner.

The planning fallacy: why the most likely value runs short

Estimates from the people doing the work are biased in a known direction: they run short. Buehler, Griffin and Ross, writing in the Journal of Personality and Social Psychology in 1994 (DOI 10.1037/0022-3514.67.3.366), describe the planning fallacy, a term they credit to Kahneman and Tversky (1979), as “the tendency to hold a confident belief that one’s own project will proceed as planned, even while knowing that the vast majority of similar projects have run late.” In their first study, 37 psychology students in the final semester of the Honors Thesis course at the University of Waterloo predicted when they would submit their thesis, giving a best estimate and predictions for the best and worst cases. Four had still not submitted two semesters later, so the means in the table cover the other 33.

Prediction asked forMean predicted daysShare who finished by that time
”Everything went as well as it possibly could”27.410.8%
Best estimate33.929.7%
“Everything went as poorly as it possibly could”48.648.7%
Actual mean completion55.5

Fewer than a third finished within their own best estimate, and fewer than half finished within a worst case they had been asked to imagine. The authors traced the bias to how people predict: they focus on plan-based scenarios for the task at hand rather than on relevant past experiences. In their fourth study, the optimistic bias was eliminated for participants who were instructed to connect relevant past experiences with their predictions. The paper opens with the transcontinental railway promised to British Columbia by 1881, whose last spike was driven in 1885, and with the Sydney Opera House, estimated in 1957 at $7 million for completion in early 1963 and opened in a scaled-down form in 1973 at a cost of $102 million.

Two consequences for a PERT estimate follow:

  • A range does not fix bias in the inputs. Feeding the study’s group means into the formula (O = 27.4, M = 33.9, P = 48.6) gives an expected 35.3 days and a standard deviation of 3.5, while the actual mean was 55.5, about 5.7 standard deviations higher. The stated confidence would look precise and be wrong. This is an arithmetic illustration on group averages, not a result from the paper.
  • The fix is evidence, not a wider range. Asking for a worst case moved the prediction only partway toward reality: the mean pessimistic prediction, 48.6 days, still fell 6.9 days short of the actual mean. The remedy that worked in the fourth study was connecting the prediction to relevant past experience, which is what the analogous method and an actuals log do. GAO’s guide makes the related point that activity durations and logic “may be overly optimistic if there is pressure from the customer or instruction from management to finish earlier than the unbiased duration estimates imply.”

Record each estimate with its basis, then compare it with the actual when the task closes. After a few weeks the log shows the team’s own ratio of actual to most likely, a correction for M drawn from the team’s own history. The resource-allocation guide covers the availability side of the same problem, since a duration that assumes a person is free is wrong when the person is not.

Contingency vs management reserve

The buffer between the planned finish and the committed date has several names, and sources disagree on them, so fix the vocabulary before the first status meeting.

SourceTermMeaning
GAO Schedule Assessment Guide (2015)Schedule contingency, also “margin or a reserve”Extra time in the baseline for known and quantified risks and uncertainty, between the planned finish and the committed date
NASA Schedule Management Handbook (2020)Schedule marginA separately planned quantity of working days above the planned work duration, for risks and uncertainties, with no scope or budget
NASA Schedule Management Handbook (2020)ContingencyNon-working days or times, such as holidays, weekends or extra shifts, that could be used to overcome performance delays
NASA Schedule Management Handbook (2020)Management ReserveBudget withheld for management control, typically at project or contract level, for unforeseen and unplanned events; outside the performance baseline, with no scope assigned

The two “contingency” rows mean different things: GAO’s is the time buffer itself, while NASA’s is non-working time that can be pressed into service. NASA prefers “margin” for the time buffer and notes that NPR 7120.5 treats “reserve” as an obsolete term, using “schedule margin” for schedule and Unallocated Future Expenses (UFE) for funding.

In the GAO guide, the contingency is sized by a schedule risk analysis. It compares the deterministic finish with the result at a chosen level of certainty. “An organization may want to adopt an 80 percent chance that its program will finish on time or earlier,” and the contingency is the gap between that date and the schedule’s own finish. In the worked example that gap is 3 days: the 80% date is day 12 against a 9-day plan. Some organizations use the 65th or 55th percentile instead; the guide warns that because schedule distributions tend to be right-skewed, with a greater tendency to finish late than early, those percentiles are not as certain as the 80th and may expose the program to overruns.

Three working rules follow from the two documents:

  1. Size it from the risk, then show it as its own line. GAO prefers one contingency activity just before the finish milestone. Apportioning contingency to specific activities in advance is not recommended, it says, because the risks that will actually occur, and the size of their effects, are not known beforehand; dispersed contingency is also less visible and may be harder to track. The Carnegie Mellon text records the other practice, an allowance inside each activity, such as scheduling an activity with an expected duration of 2 days for 2.2 days, a ten percent contingency.
  2. Put it under change control. GAO says contingency is held by the program manager and, when allocated to others, should follow a formal change process so that its use is “transparent and traceable.” NASA treats schedule margin the same way, as a control parameter owned by the project manager that “can only be released through a formal process.”
  3. Keep the time buffer and the money reserve apart. NASA’s schedule margin is time, and its margin activities carry no budget. Its management reserve is budget, withheld for management control rather than for any specific task, typically held at project or contract level for unforeseen and unplanned events, and not to be used to cover past performance variances.

Tie each material risk to an owner and a trigger in the risk register, so the team knows when contingency is being spent and why.

Separate effort from duration before applying any formula

Effort is work, in person-hours. Duration is elapsed working time on the calendar. A task of 12 hours of effort with three available hours a day takes four working days if the work can be split that way, and a required approval that arrives only on Friday sets its own constraint. Do three-point estimates on durations that already include the calendar.

Resource conflicts can outweigh any estimate range. In the calculator’s example, with separate people for B and C the plain schedule is 9 days. If one person must do B and then C, it is 2 + 5 + 3 + 2 = 12 days before any contingency, because the activities no longer overlap. With the PERT expected values the sequential path is 2.33 + 5.83 + 4 + 2.33 = 14.5 days. Resolve this with a capacity check, not a larger buffer.

Update the forecast as work finishes

Replace each estimate with its actual on completion and re-estimate what remains, keeping the approved baseline visible so movement can be explained. Recheck which path controls after every change. In the calculator, raising C from 3 to 6 days makes A-C-D the longer path at 10 days, and attention on B alone would miss it. Report the next milestone, its forecast, and how much contingency is left.

Sources