Reference-class forecasting
Predict from what happened to similar cases rather than from the details of this one, because the details systematically mislead.
- Time cost
- 1 h
- Output
- A class definition, a distribution, and a placed estimate.
- Steps
- 5
Use when
- You are estimating duration, cost or success probability for something that has been done before.
- The inside view is producing a confident, specific, optimistic number.
- A plan has no slack in it and everyone believes it anyway.
Do not use when
- The case is genuinely unprecedented, so no reference class exists.
- You can only find a reference class by stretching the definition until it covers nothing comparable.
Inputs required
- A definition of the class
- Outcome data for the class, even rough
Procedure
- 01
Define the class
What set is this a member of? Be specific enough that membership is checkable, broad enough that there are more than a handful of cases.
- 02
Get the distribution
Not the average — the spread. What happened at the median, the 10th percentile, the 90th. The tail is usually the part that matters.
- 03
Place this case
Where in the distribution does this one sit, and why? The honest answer is usually “near the middle”, because everyone believes their case is above it.
- 04
Start from the class, adjust for the case
Take the class figure as the starting point and adjust for specifics. This is the opposite of the usual order, and the order is the method.
- 05
Cap the adjustment
Limit how far the inside view can move the outside view — say 20%. Without a cap, the adjustment step quietly restores the original optimistic estimate.
Characteristic failure mode
Worked example
A team estimates a six-month build.
- 01Class: internal projects of comparable scope at comparable organisations.
- 02Distribution: median 11 months, 10th percentile 7, 90th percentile 19.
- 03Inside case for being faster: an experienced team. Adjustment capped at 20%.
Result
Estimate moves from 6 months to roughly 9. Nothing about the plan changed — only the starting point did.
Where this disagrees with another method
Fermi builds from the specifics of this case; reference classes deliberately ignore them. Use the reference class when the class exists, and Fermi when it does not.
Where to go next