An estimate is a statement about an uncertain quantity, usually duration or cost, produced from evidence and assumptions that ought to travel alongside it. The technique chosen decides how much evidence is needed, how much the estimate costs to produce, and how much confidence it can honestly carry.
There is no best method, only one appropriate to how much is known and to what the decision resting on the number is worth.
The techniques and what each costs to produce
Analogous estimating takes the actual duration or cost of a similar past project and adjusts it for known differences. It is fast, it needs almost no detail, and it can be produced in the meeting where somebody asks for a figure. It is also the least accurate of the family, since everything depends on how similar the analogue really is, so it belongs at initiation and in the charter rather than in a commitment.
Parametric estimating multiplies a measured rate by a quantity, so a known cost per square metre against a floor area gives a cost. It scales, it is repeatable, and it is accurate where the underlying data is good and the work is genuinely proportional to the parameter. Its failure mode is being applied outside the range the data came from, because coordination cost grows faster than the parameter does and a rate drawn from small jobs will not stretch to a large one.
Bottom up estimating decomposes the work, estimates each package, and rolls the results upward. It is the most accurate and by some distance the most expensive, since it cannot begin until decomposition has happened and it takes real time from the people who know the work. It also carries a quiet arithmetic problem, because summing individually padded estimates produces a total padded many times over, which is why the aggregation deserves as much scrutiny as the components.
The formulas in one place
| Technique | Calculation | What it needs | When it fits |
|---|---|---|---|
| Analogous | Adjust the actual figure from a comparable past project | One genuinely comparable project, and somebody who remembers it honestly | Initiation and the charter, where an order of magnitude is what the decision needs |
| Parametric | rate × quantity | A measured rate drawn from work in the same size range | Repeatable work that really is proportional to the parameter |
| Bottom up | Sum of the estimate for every work package | A completed decomposition and time from the people who will do the work | A commitment, a baseline, or any number that will later be defended |
| Three point, triangular | (O + M + P) / 3 | Three values from the estimator | A quick spread where the tails are roughly balanced |
| Three point, beta or PERT | (O + 4M + P) / 6 | The same three values | The usual choice, because overruns run further than underruns |
| Standard deviation | (P - O) / 6 | The optimistic and the pessimistic value | Turning a spread into a confidence range you can quote |
| Activity variance | ((P - O) / 6)² | The same two values | Adding uncertainty along a path, since variances add where deviations do not |
Three point estimating and the beta weighting
A three point estimate replaces one number with an optimistic value, a most likely value and a pessimistic value, which obliges the estimator to describe the spread rather than conceal it.
The triangular average is the plain mean of the three, so optimistic plus most likely plus pessimistic, divided by three. The beta or PERT weighting gives the most likely value four times the weight, so it is optimistic plus four times most likely plus pessimistic, all divided by six.
Running one set of numbers through all of them is the fastest way to see what the weighting actually does. Take an activity with an optimistic value of 6 days, a most likely value of 12 and a pessimistic value of 30.
| Formula | With O 6, M 12 and P 30 | Result | What the answer means |
|---|---|---|---|
Triangular (O + M + P) / 3 | (6 + 12 + 30) / 3 | 16 days | The plain average, dragged upward by one alarming pessimistic figure |
Beta (O + 4M + P) / 6 | (6 + 48 + 30) / 6 | 14 days | The expected duration, with the most likely case given four times the weight |
Standard deviation (P - O) / 6 | (30 - 6) / 6 | 4 days | One deviation either side of 14, so 10 to 18 days at roughly 68 per cent confidence |
| Two deviations | 14 ± 8 | 6 to 22 days | The range worth quoting to a stakeholder, at roughly 95 per cent confidence |
| Activity variance | 4² | 16 | Added along a path, because variances add where standard deviations do not |
The two days between 14 and 16 are the whole argument for the weighting.
The weighting exists because real durations are not symmetric. Work can overrun by a great deal and can only underrun by a little, so the pessimistic tail is longer, and a plain average lets one alarming pessimistic figure drag the expected value too far. The standard deviation, being the pessimistic value minus the optimistic one divided by six, is what turns the spread into a confidence statement.
The long right tail is what the beta weighting is compensating for. Because the pessimistic value sits 18 days above the most likely case while the optimistic one sits only 6 days below it, the plain average lands further from the peak than the weighted one does.
Ranges rather than single numbers
A single number conceals everything a planner needs to know. Saying a task will take twelve days says nothing about whether the honest range is eleven to thirteen or six to thirty, and those two situations call for entirely different decisions.
Publishing a range with the confidence attached keeps the uncertainty where it can be seen. It also protects the estimator, because a number given as a range and later exceeded is a forecast behaving normally, whereas a single number exceeded looks like a broken promise. Estimates are refined as work proceeds and uncertainty falls, and a project still reporting its initiation figures halfway through is presenting a guess as though it were a measurement.
The variance row in the table above is what makes a range for a whole path possible. Variances add along a path where standard deviations do not, so three activities each with a deviation of 4 days give a path variance of 48 and a path deviation of just under 7 days rather than 12. Uncertainty grows more slowly than intuition expects, which is why padding every activity separately overstates the total by a wide margin.
Relative units on adaptive work
Adaptive delivery commonly sizes items relative to one another rather than in hours, because people compare well and predict absolutes badly. Those sizes then combine with observed completion rates to project a date, which grounds the forecast in what the team has actually done rather than in what it intended.
The unit means nothing outside the team that produced it. Comparing one team's numbers against another's has no arithmetic behind it, and an organisation that tracks the comparison has invented a cheap way for a team to look better without delivering anything more.
An estimate from somebody who will not do the work
The people who will perform an activity should produce its estimate, or at least validate it. They know the environment, the tooling and the awkward parts, and none of that is visible to somebody working from a written description.
Ownership matters here as much as accuracy. A number imposed from outside is a target the team had no hand in setting, so nobody defends it when a stakeholder pushes, and nobody feels much obligation to meet it. Imposed estimates also tend to be optimistic, since the person setting them carries none of the consequences of being wrong.