Late projects are late by a multiple and not by a margin
💎CORE ARGUMENT
A task estimated at two days can finish in one. It cannot finish in minus three. It can, however, take twenty. The possible outcomes are bounded on the good side and open on the bad side, and that lopsided shape has a consequence I managed to ignore for ten years: lateness scales with the size of the thing, so it is a ratio and not an amount.
When I say about two weeks I am usually naming the outcome I think most likely, or the middle one. For a lopsided distribution the average sits well to the right of both, because the rare long overruns weigh so much. Add up twenty tasks estimated at their middles and the total is not the middle of the project. It is an optimistic number in a careful number's clothes.
This is why a fixed buffer fails. Add a week for safety treats the overrun as a constant, as though a three-day task and a three-month project would both slip by the same few days. The honest pad is a multiplier, and it should be larger where I know less. A familiar change might deserve one and a half. An integration with a system nobody on the team has touched might deserve three, and I should say so out loud.
I wrote the first version of this in the low days after recommending the slip. The launch had been planned with a margin. It needed a factor.
Two things I do now. I give a range whose upper end is a multiple of the lower. And I ask of any plan which single item could run to five times its estimate, because that item is the schedule.
🧱BUILT FROM
🌍WHY IT MATTERS
I pad by ratio, and I name the ratio. Estimates leave my desk as a range, low to high, and the conversation is about which item owns the high end.
Still open. What is my multiplier? I have one launch measured properly and a handful of remembered ones. Until I have logged estimate against actual for a year, one and a half and three are opinions.
☑️NEXT STEPS
- Log estimate against actual for every piece of work over a week long
- Write the build estimate in the January memo as a range with its ratio named
- Work out my own multiplier once a year of the log exists
An essay arguing that estimates are medians of a skewed distribution, so the average outcome lies far to the right of the number people quote.
The same mistake seen from the planner's chair is 💎 My launch estimates ignore every launch I have seen. Where the long tail comes from is 💎 Reality has more detail than any plan can hold.
Durations cannot be negative and have no upper limit, so their distribution is skewed to the right. A typical-case estimate is close to the median; the mean is higher, and a sum of medians underestimates the sum. The error is proportional to the estimate.
A multiplier is an invitation. Announce that everything is doubled and the first estimates shrink to compensate, or the work expands to fill the doubled time. A known pad stops being a pad.
It goes into the year's list of what I got wrong, and it decides how the build estimate in the January memo will be written.