Blog

Software estimation that holds up: three-point estimates and PERT

· 5 min read

“How long will it take?” “Eight weeks.” Twelve weeks later, everyone is unhappy, and nobody did anything wrong except compress a range into a single number.

Why single numbers run late

Software work is asymmetric. A task estimated at eight weeks will rarely take five, but it can easily take sixteen: a dependency slips, an integration behaves differently from its documentation, a requirement turns out to mean something else. The possible outcomes have a long tail on the late side. When you ask for one number, people give the most likely value, which sits to the left of the average. Add up many such estimates and the total is late almost by construction.

Three-point estimating

Ask for three numbers instead.

  • Optimistic (O): if everything goes right.
  • Most likely (M): the realistic case.
  • Pessimistic (P): if the known risks materialize.

The PERT estimate is (O + 4M + P) ÷ 6. It weights the most likely value while letting the tail pull the result upward.

A worked example

A payments integration: 6 weeks optimistic, 8 most likely, 16 pessimistic because the partner's sandbox is said to be unreliable. PERT gives (6 + 32 + 16) ÷ 6 = 9 weeks. A rough measure of the spread is (P − O) ÷ 6, here about 1.7 weeks. So: around nine weeks, plausibly between seven and eleven.

The conversation that produces the pessimistic number is often worth more than the formula. Asking “what would make this take sixteen weeks?” surfaces the risks, and each risk named is one you can investigate early.

Check against history

People underestimate their own projects even when they know similar projects overran. It is called the planning fallacy, and awareness does not cure it. Evidence does. Look at your last five comparable projects and compare first estimates with actual durations. If they ran 40% over on average, apply that to the new estimate. A ten-week estimate becomes fourteen.

Presenting a range

  • Give a range with a confidence level: “nine to twelve weeks, and I am about 80% confident in that”.
  • Name the biggest uncertainty and when it will be resolved: “we will know much more after the sandbox spike in week two”.
  • Say which of scope, time and cost is flexible. If all three are fixed, quality will absorb the difference.
  • Commit to re-forecasting at set points, and do it.

Stakeholders push back on ranges because they need a date to plan around. Give them one — the upper end — along with what would have to be true to hit the lower end. Most people handle uncertainty well when they are told about it early.

Estimate less, slice more

Estimates get worse as work gets bigger. Break large projects into pieces of a few days each, deliver them in order of risk, and measure how many you finish per week. After a few weeks, that throughput is a better forecast than any up-front estimate.

Try it — Three-point estimate

PERT estimate

9.0 wk

Standard deviation

1.7 wk

Likely range (±1 SD)

7.3–10.7 wk

Above most likely by

1.0 wk

The band runs from optimistic to pessimistic, the thin line is the most likely value and the thick line is the PERT estimate. A long pessimistic tail pulls the estimate above “most likely” — which is why single-number estimates run late.

Practice this in a lesson

Each engMBA lesson pairs a short explanation with practice questions that grade themselves. Free, with no signup.

See the full curriculum

Keep reading