Estimating From Your Own Data
The most valuable thing a record produces: a reference class of comparable past work, which beats any estimate made by imagining the task.
Estimating by thinking about a task is unreliable in a known direction. Estimating by looking up three similar past tasks is not. For teams applying “Estimating From Your Own Data” to client or project work, weekly timesheet template provides an operational record to compare with estimates; the useful conversation is about scope and delivery, not activity as a substitute for results.
Why the reference class works
It bypasses imagination entirely.
Past actuals already include the interruptions, the waiting, the rework and the last ten percent — all the things an estimate omits by construction.
And it requires no correction factor, because the correction is already embedded in the historical number.
This is the strongest practical argument for keeping any record at all.
Building one
For each completed piece of work: what it was, how long it actually took, and two or three words about what made it unusual.
Three lines. That is the whole system.
After a dozen entries you can estimate; after fifty you can estimate well.
Keep it separate from the daily record so it survives changes of method and tool.
Using it
Find three past items most like the new one.
Take the range, not the average: "between four and seven days, most likely five" is a more honest statement and a more useful one.
Adjust for the stated differences, which is where the two or three words earn their place.
And quote the range, because a single number invites being held to its lower end.
When there is no comparable past work
Decompose until the parts are comparable. You may not have built this thing before, but you have written a specification, done a review and fixed the thing that did not work.
Where nothing is comparable, say so and price the uncertainty rather than producing a confident number.
And treat the first instance as an investment in the reference class, which makes the second estimate real.
The optimism that survives anyway
People with good records still shade estimates downward under pressure to quote low.
Which is a commercial decision rather than an estimation error, and should be made knowingly.
Write the honest number first, then decide what to quote. Conflating those two is how the record stops being trusted internally.
What to record about outcomes
Whether the estimate held, as a yes or no.
What the overrun consisted of, which after a while shows a pattern: for most people it is one or two recurring categories rather than general slippage.
And the elapsed time as well as the effort, because for anything with dependencies those are different numbers and clients usually care about the first.
The reference class for other people
Estimating work you will delegate needs their history, not yours.
In its absence, assume worse than your own by a meaningful margin, which is uncomfortable and consistently borne out.
What to check
Do you have any record of how long past work actually took?
Could you find three comparable items for the next thing you estimate?
Do you quote a range or a number?
And when your estimate last overran, did you record what the overrun consisted of?
Quote a range
Between four and seven days, most likely five, is more honest and more useful than a single number.
A single figure invites being held to its lower end, and the range is what your reference class actually supports.
Worth keeping in mind
Record what the overrun consisted of, not just that it happened. After a while the pattern shows one or two recurring categories rather than general slippage, and those are addressable.
When there is no comparable work
Decompose until the parts are comparable: you may not have built this before, but you have written a specification, done a review and fixed the thing that did not work.
Where nothing is comparable, say so and price the uncertainty rather than producing a confident number, and treat the first instance as an investment in the reference class.
In short
Record the outcome as well as the estimate: whether it held, what the overrun consisted of, and elapsed time separately from effort.
For anything with dependencies those are different numbers and clients usually care about the first. For a complementary perspective on work, measurement or planning, consult Notion.