BuzbyTimeRead the requirements guide

Reading the numbers · 2.3

Estimates against actuals

Estimates against actuals. What decides it, what it costs, and what usually goes wrong. For a practical comparison point, see Monitask's freelancer time tracking software.

Two numbers answering different questionsThe distinction

An estimate is a forecast made under uncertainty. An actual is a record of what happened. Comparing them tells you about your forecasting, which is useful, and tells you nothing about the person who did the work.

Nearly every organisation that starts comparing them ends up using the variance as a judgement about people, at which point the estimates stop being forecasts. For broader background, see Microsoft.

Why estimates inflate the moment they are scored

Somebody assessed on hitting their estimate has one reliable strategy available, and it is not estimating better. Padding is rational, it is undetectable from outside, and it makes every forecast built on those estimates worse.

The organisation then observes that delivery has slowed, which it has not: the estimates moved.

Underestimation is systematic

People underestimate how long work will take, consistently, and the bias survives experience. It is one of the more robust findings about planning and it has a practical consequence: exhortation does not fix it because it is not a discipline problem.

What works is a correction derived from your own history. If your team's estimates have run at seventy per cent of actuals for a year, plan against that ratio rather than against the estimates, and stop asking anybody to try harder.

Where to compare them

  • At team level, over quarters, to derive the correction factor.
  • At project level, retrospectively, to find which kinds of work are consistently misjudged.
  • Never at individual level, for any purpose.

The second is the most useful and the least done. Estimation error is rarely uniform: it clusters in particular kinds of work, usually the ones with a dependency on somebody outside the team.

Re-estimating

An estimate made before anything was known is worth less than one made a third of the way in, and both are worth recording. A team that re-estimates at a defined point produces a much better planning input and a visible record of how uncertainty resolved.

What it must not become is a negotiation in which the re-estimate is challenged, because that returns the whole exercise to the first problem.

What the variance is actually telling you

Usually one of four things: the work was different from what was described, a dependency was late, the estimate was made by somebody who would not be doing the work, or the scope changed and nobody recorded it.

Only the third is about estimating. The other three are about how work arrives, which is a more useful thing to have discovered.

Estimating in ranges

A single number invites treatment as a commitment. A range, with the assumptions that would move it to either end, communicates what an estimate actually is and survives contact with a planning meeting better.

Where an organisation insists on one number, supply the upper end and say so.

Who estimates

The person who will do the work, or somebody who has done it. Estimates produced by whoever is selling the work are forecasts of what the client will accept, and comparing them with actuals measures the sales conversation.

Keeping the history

Original estimate, re-estimate, actual, and what changed. Four fields, kept for a year, and they produce a correction factor specific to your work that no benchmark can supply.

Estimates for clients are a different animal

An external estimate is a commercial commitment with a negotiation attached. Comparing it with actuals measures pricing and scope control rather than forecasting, and mixing the two analyses produces conclusions about neither.

Keep the internal planning estimate and the client-facing figure as separate fields. Most systems support this and most organisations use one field for both.

Where to start

Compute your team's ratio of estimate to actual over the last year, at team level. One number, an hour's work, and it improves every forecast you make afterwards.

The estimate nobody records

How long somebody thought a task would take at the moment they started, as distinct from what was in the plan weeks earlier. It costs a field and it is the estimate most closely related to what actually happens.

When actuals should change an estimate

Not during the work, if the estimate is being used to judge. Afterwards, always: a library of what similar work actually took is the best forecasting input available and it accumulates for free.

The forecast that actually helps

Not a single date. A range, derived from your own history of similar work, presented with the assumption that would move it. That is buildable from a year of records and it is more useful than any estimating technique applied to a blank page.

A short checklist

Estimate in ranges. Record the original and the re-estimate. Compare at team level only. Derive a correction factor. Never put variance in anybody's objectives.

One sentence to carry

An estimate scored is an estimate padded, and the padding is invisible to everybody including the person doing it.

Where teams get this right

The ones that treat estimating as a forecasting skill owned by the team, reviewed by the team, and reported to nobody as a performance figure. It sounds like an absence of accountability and it produces the accurate forecasts everybody wanted.

Historical data as the estimate

For repeatable work, the best estimate is what similar work actually took, taken from your own records rather than produced by anybody. It requires only that the categories are stable, which is the argument the entry on categories makes for a different reason.

Also in reading the numbers