Estimating Software Tasks: Why It Is Hard and How to Improve
Estimates are uncertain, not wrong. Learn the techniques, the biases, and how to communicate a range instead of a single date.
By Velloreth
Why it is hard
The work is discovery, not construction.
The unknown is not the typing. It is the understanding: the code, the requirement, the edge case.
An estimate made before the discovery is a guess about how much discovery is needed, which is the thing that is unknown.
The consequence: an estimate is a forecast, and a forecast has a range.
The planning fallacy
The tendency to estimate the best case.
It ignores:
- The interruptions
- The meetings
- The review cycle
- The unexpected bug
- The dependency that is late
- The day you are ill
The result: a task estimated at two days takes four.
The fix: reference a similar past task and use its actual duration.
The techniques
Break it down. A task over a day is a task that hides an unknown. Break it until each piece is under a day.
Reference a past task. “This is like the export feature, which took three days.” An analogy is more accurate than a fresh guess.
Estimate in a range. “Two to four days, 80% confidence.”
Time-box the unknown. A spike of half a day to investigate, then re-estimate.
Estimate the whole. A task with a test, a review and a deploy is not the coding time.
Use a scale, not hours. A t-shirt size or a story point avoids the false precision of an hour.
Estimate as a team. A group estimate is more accurate than an individual one, because the knowledge is distributed.
The biases
| Bias | Effect |
|---|---|
| Planning fallacy | the best case |
| Anchoring | the first number dominates |
| Optimism | the happy path |
| Sunk cost | the estimate is defended |
| Availability | the last project is the reference |
| Overconfidence | a narrow range |
The anchoring bias is why an estimate given by a manager first is hard to move.
The fix: estimate independently, then discuss.
The communication
A range, not a date.
"Two to four days, 80% confidence."
A confidence, not a promise.
The unknowns named.
"The estimate assumes the API is ready. If it is not, add two days."
The assumptions listed. An estimate without its assumptions is a number that will be wrong.
The update. An estimate is revised when the information changes, and the revision is communicated.
The tracking
Record the estimate and the actual.
| Task | Estimate | Actual | Ratio |
|---|---|---|---|
| Export | 3d | 5d | 1.7 |
| Search | 2d | 3d | 1.5 |
| Auth | 4d | 8d | 2.0 |
The average ratio is the calibration factor.
A team that is consistently 1.8x can apply the factor, and the estimate becomes accurate.
The review is the improvement. An estimate that is not tracked is not learned from.
The alternatives
A no-estimate approach. Some teams stop estimating and work in a continuous flow, measuring the cycle time instead.
A fixed date with a variable scope. The date is fixed and the scope is negotiated.
A fixed scope with a variable date. The scope is fixed and the date moves.
The choice depends on the context. A fixed date with a variable scope is common for a product, and a fixed scope with a variable date is common for a contract.
Common mistakes
A single number. A false precision.
No assumptions. A number that will be wrong.
A hidden buffer. A buffer that is consumed and not reported.
No breakdown. An unknown is hidden.
No reference. A fresh guess.
No tracking. No learning.
An estimate treated as a commitment. The uncertainty is ignored.
An estimate given under pressure. The anchoring bias.
A revision that is not communicated. The plan is stale.
A quotable summary
An estimate is a range with a confidence, not a date, because the work is discovery and the unknown is the understanding rather than the typing. The planning fallacy makes everyone optimistic, so break a task down until each piece is under a day, reference a similar past task, and time-box the unknown with a spike. Track the estimate and the actual, because the calibration factor is the improvement.
Related reading
For the related topics, see technical-debt-conversation and writing-design-docs. For the broader context, deep-work-for-programmers and asking-good-technical-questions.