Decisions
15 min

Fermi estimation

Decompose an unknown quantity into factors you can bound, so that errors partly cancel and the result lands within an order of magnitude.

Time cost
15 min
Output
A decomposition with a low, central and high estimate.
Steps
5

Use when

  • You need a number now and no data source will produce one.
  • You want to size a branch before deciding whether to research it.
  • Someone has quoted a figure and you want to know whether it is plausible.

Do not use when

  • Precision matters. This gets the order of magnitude, not the number.
  • Real data is available in less time than the estimate would take.

Inputs required

  • The quantity, stated precisely
  • Factors you can bound within a factor of three

Procedure

  1. 01

    State the quantity exactly

    Including units and period. Most Fermi errors are definition errors rather than arithmetic errors.

  2. 02

    Decompose multiplicatively

    Break it into factors that multiply. Population × rate × frequency × amount. Each factor should be one you can bracket.

  3. 03

    Bound each factor

    Give a low and a high that you would bet on, then take the geometric mean. Round aggressively — precision in a factor is wasted.

  4. 04

    Multiply through

    Carry the low path and the high path separately as well as the central one. The spread between them is the honest uncertainty.

  5. 05

    Sanity-check against something known

    Compare the result to a quantity you do know. If it implies something absurd, a factor is wrong by an order of magnitude — find it.

Characteristic failure mode

Correlated errors. Factors chosen from the same optimistic mental model err in the same direction, so they compound instead of cancelling and the range is far too narrow.

Worked example

How many hours a year does a mid-sized office spend in recurring meetings?

  1. 01120 staff × 60% who attend recurring meetings = 72 people.
  2. 02Roughly 4 recurring meetings a week × 45 minutes = 3 hours a week each.
  3. 033 × 46 working weeks × 72 people.

Result

About 10,000 hours a year, with a plausible range of 6,000 to 17,000. Enough to decide whether the question deserves real measurement.

Where to go next

Also cited by
Monte Carlo simulation