Optimal stopping and the 37% rule

Observe a fixed share of the candidates without choosing, then take the first one better than everything seen so far.

Time cost: 10 min to set up
Output: A look-phase count and a recorded standard.
Steps: 5

Use when

Candidates arrive in sequence and must be accepted or rejected on the spot.
A rejected candidate cannot be recalled.
You know roughly how many candidates there will be.

Do not use when

You can go back to earlier candidates. The rule’s entire structure exists because you cannot, and it is badly suboptimal when you can.
You can judge a candidate against an absolute standard rather than only by comparison. Then use a threshold — satisficing — which does better.
The pool size is unknown, which leaves the 37% undefined.

Inputs required

An expected number of candidates
A rule for comparing any two
No ability to recall a rejection

Procedure

01
Check the assumptions first
No recall, sequential arrival, known pool size, comparison only. If any fails, the rule does not apply — and in most real situations at least one does.
02
Set the look phase
Observe the first 37% — strictly, 1/e — and reject them all, however good. This phase only calibrates the standard.
03
Record the best seen
One number or one name. That is the whole memory the rule needs.
04
Leap
From that point, accept the first candidate better than everything in the look phase.
05
Accept the failure rate
The rule finds the single best candidate about 37% of the time. That is optimal under the assumptions, which means the other 63% is not a mistake — it is the price of no recall.

Characteristic failure mode

Applying it where recall exists. In hiring, in flat-hunting, in most real searches you can go back, or judge against a standard. The rule then rejects good candidates for no reason and its famous number does harm.

Worked example

Twenty flats are available to view over a weekend, each let the same day it is shown.

0137% of 20 is roughly 7. View seven, take none, record the best.
02From the eighth, take the first better than that best.
03In this run, flat 11 clears the bar.

Result

A defensible stopping point in a genuinely no-recall market. In a market where offers stay open for a week, this rule would be the wrong one.

Where this disagrees with another method

vs. Satisficing versus maximising →

The 37% rule uses only comparisons and is optimal when you cannot recall. A written threshold uses outside knowledge and beats it whenever you have any. Most people who quote 37% are in a situation where satisficing would do better.

Where to go next