Tool/Decision Theory/No. 0696
Optimal Stopping
Optimal stopping is a mathematical framework for deciding when to end a search or process by comparing the value of stopping with the expected value of continuing. Rooted in sequential analysis, it accounts for search costs, available information and the risk of losing options.
- Evidence
- Well established
- Read
- 7 min
- Links
- 11 connections
01You've seen this when…
- in life
A used bike fits you and falls within your budget. You spend another evening comparing listings, then discover the seller has accepted someone else’s offer.
- at work
A strong candidate needs an answer by Friday. Your team wants to interview three more people next week, but waiting means the current candidate may take another job.
- out in the world
A city requests another round of supplier quotes for an urgent repair. The quotes might get cheaper, but each day the repair waits adds another day of disruption.
02The idea
After enough searching, the problem changes. Your focus shifts from whether a better option exists somewhere to whether trying to find it is worth what you might spend or lose.
Optimal stopping treats that as a decision in its own right. At each stage, compare the value of stopping now with the value of continuing and making another decision after you learn more. Continuing can improve your choice. It can also cost time and money while you risk losing an acceptable offer.
The stopping rule depends on the setting. Can you return to an earlier option, and how many opportunities remain? Some searches aim to find the single best candidate, while others aim to meet a minimum standard or get the best result after costs. Different answers produce different rules.
The famous 37% rule belongs to the secretary problem, one narrow version. The number of candidates is known, and they arrive in random order. You can judge their relative rank. Each rejection is final, and your goal is the highest probability of selecting the very best. For a large pool, the rule is to reject roughly the first 37%, then take the next candidate who beats everyone seen so far. Even under those assumptions, it selects the best only about 37% of the time.
The general lesson is not to spend 37% of your life browsing. It is to choose a stopping rule that fits the opportunities you actually have. Value of information asks what more knowledge is worth; optimal stopping adds the question of when to stop obtaining it.
03How to use it
- Specify what a successful choice means. Write down the outcome you care about: lowest total cost, a qualified hire, a safe decision, or the greatest chance of picking the best. If several options would serve you equally well, finding the absolute best may not be worth much.
- Map the search rules. Check whether the search has a deadline, then record which offers can expire and which rejected options you can revisit. A shortlist you can return to is different from a sequence of take-it-or-leave-it offers. Keeping an offer open has option value.
- Count the cost of continuing. Include fees, time, delayed benefits, and the chance of losing your current option. Time has an opportunity cost, even when browsing itself is free. Exclude hours already spent from the decision to spend another hour.
- Estimate what further search could deliver. Consider both the size of a plausible improvement and its chance. Use ranges when you lack reliable probabilities. If another search costs $50 and the likely improvement is only $10, precision is unnecessary. If reasonable estimates reverse the answer, your uncertainty matters.
- Choose a rule before the next attractive option appears. Set a minimum acceptable result and a point at which you will choose among acceptable options. Or use a value-based rule: continue only while a feasible further search looks worth its cost. Update when new evidence changes the calculation. If committing merely feels uncomfortable, stick with the rule.
- Allow stopping without buying or choosing. Ending a search can mean declining an unsafe product, an unqualified candidate, or an unaffordable home. Sometimes the best available action is to decline them all.
In formal models, this comparison accounts for the future decisions that searching makes possible. That is the connection to dynamic programming and Bellman’s principle. In everyday use, a rough comparison of feasible search plans is often more useful than pretending to have exact probabilities.
04A worked example
Consider an illustrative case. A freelancer needs to order a replacement laptop before today’s shipping cutoff. A suitable machine costs $700, and the shop will hold it until the cutoff. There is time for one final hour of searching.
The freelancer estimates a one-in-four chance of finding an equally suitable machine for $640. Otherwise, the search produces nothing better. That hour would displace $30 of paid work. These are planning assumptions, not measured probabilities.
What it looks like Another hour could save $60, so stopping seems premature. Four hours of earlier research also make it tempting to keep going until that effort produces a bargain.
What’s actually going on The additional search has an expected saving of $15: a quarter chance of saving $60. It costs $30 in displaced work. With only this final search opportunity remaining, and the current offer protected, stopping has the higher expected net value. The four earlier hours change neither option; letting them drive the decision would be a sunk cost fallacy.
What made it work The freelancer compares the remaining alternatives, checks that the machines really are equivalent, and includes the cost of time. If the chance of saving money or the size of the saving were much greater, continuing could be the right answer.
05When to reach for it
06When it misleads
- The famous percentage solves the wrong problem. Candidates in practice may arrive in a nonrandom order. You may be able to recall them, or care more about avoiding a bad hire than finding the single best. Those changes can overturn the 37% rule.
- Your estimates are guesses wearing numbers. An invented probability gives even a precise calculation an unreliable foundation. Test a range of plausible assumptions, especially when the decision is expensive.
- The next step is only part of the search plan. One low-value step might unlock a valuable new source of options. Comparing only the next click can miss that benefit; compare realistic paths of continued search.
- Safety gets reduced to average value. Completing required inspections and avoiding catastrophic downside take priority over a small expected benefit. Put safety and legal constraints into the decision before optimizing within them.
- Stopping becomes a synonym for settling. Satisficing means accepting an option that meets a standard. It can be a sensible stopping policy, and whether it is optimal depends on the setting. Sometimes searching longer improves the outcome after costs; sometimes abandoning the search is better.
07Roots
During World War II, Abraham Wald worked with the Statistical Research Group at Columbia University. Wartime testing created a practical problem: how much evidence was enough? Inspecting more items could improve a judgment about production quality, but testing consumed scarce time and resources.
The conventional approach fixed the sample size beforehand. Wald developed sequential methods that examined evidence as it arrived. His sequential probability ratio test used two boundaries: enough evidence for one hypothesis, enough for the other, or an intermediate region where testing continued. The procedure controlled specified error risks without always requiring a fixed batch of observations. Wald presented the framework in his 1947 book, Sequential Analysis.
Sequential testing was one major route into optimal-stopping theory. Later work developed general methods for deciding when to stop observing a changing process. Candidate-selection puzzles made the idea especially memorable outside statistics, while mathematical treatments connected it to broader sequential decision-making. The wartime problem and the hiring puzzle share a structure: another observation may help, but you need a rule for deciding when it no longer helps enough.
08How solid is this?
Optimal-stopping results are mathematically established for specified objectives, information, and search conditions. Their practical value depends on whether those assumptions fit. For everyday choices, the 37% rule applies only when its specific assumptions hold.
09Connections
- Often confused with Satisficing
- Helps counter Sunk Cost Fallacy, Information Bias
- Part of Option Value, Value of Information, Dynamic Programming, Bellman's Principle of Optimality
- IncludesSecretary Problem
- See also Explore-Exploit Trade-Off, Opportunity Cost, Expected Value
+ 1 more in the list
10Origin and sources
Developed through sequential analysis and optimal-stopping theory. Abraham Wald’s work in the 1940s established a major foundation through sequential statistical testing; later researchers broadened the theory.
- [1]Wald, A. (1947). Sequential Analysis. John Wiley & Sons.
- [2]Chow, Y. S., Robbins, H., & Siegmund, D. (1971). Great Expectations: The Theory of Optimal Stopping. Houghton Mifflin.
- [3]Ferguson, T. S. (1989). Who solved the secretary problem? Statistical Science, 4(3), 282–289.
Suggest an edit· Updated 2026-10-02