What is the best strategy for finding a job?

When should you stop looking and take a job? A famous rule says: spend the first part of your search only looking, then accept the first offer that beats everything you have seen so far. Behind it is a real theorem. We will derive it, and then question it.

1. 1The classical answerDerive 37% step by step from seven premises.
2. 2Rethink the premisesWhich of them are true of job hunting?
3. 3A one-sided modelLoosen the premises and simulate 10,000 job searches.
4. 4Companies can say noA two-sided model: rejections, self-knowledge, waiting time.
5. 5So, what is best?What a model can and cannot answer.

Part 1

The classical answer: 37%

The question, stripped down

Imagine you will get a known number of job offers, N, one at a time and in random order. You can see how each offer compares with the ones before, but you have no absolute scale. You must answer each offer before the next one comes, and an offer you decline is gone. You take only one job, and you "win" only if you take the single best of all N. Which strategy gives you the best chance of winning? This is the classic secretary problem, usually told from the employer's side: interview applicants one by one and hire on the spot (Ferguson, 1989). Here we look at it from the job seeker's side. Its premises:

1. One choice. You take exactly one job.
2. Known number. You know in advance how many offers you will get.
3. Random order. Offers arrive in random order.
4. Ranks only. You can only compare an offer with the ones before; you have no idea what is out there.
5. No going back. An offer you decline is gone.
6. Only the best counts. The second-best job is as bad as no job.
7. The job is always yours. The premise nobody states. We return to it in Part 4.

Try it

You will get 20 job offers during a 34-week search, one at a time. Each offer has a score, and higher is better. Nobody tells you what a typical score is or how high scores go: you learn that only from the offers you see. For each offer, accept it or decline and keep looking. If you decline all 20, you end the search with no job.

Where 37% comes from, step by step

1. Choose a family of strategies

   Consider the rule: look at the first k offers without accepting any, then leap at the first offer that beats every offer so far. Under the premises above, the best possible strategy has exactly this form (Ferguson, 1989). So our only job is to find the best k.

2. When does the rule win?

   Suppose the very best offer comes at position i.

   • If i ≤ k, we skipped it, and we lose.
   • If i > k, we reach it only if no offer between k and i tempted us first. That happens exactly when the best of the first i − 1 offers came in the looking zone. Because the order is random, that offer is equally likely to be anywhere among the i − 1, so the chance is k/(i − 1).
   Looking zone k

   3

   Best offer at position i

   7

   Shuffle a random order

3. Add up over every position

   The best offer comes at each position with probability 1/N, so the chance of winning is

   P(win) = (k/N) · (1/k + 1/(k+1) + … + 1/(N−1))

   Looking zone k (with N = 10)

   3

   Best at position i	Chance	Term	Running total

4. Let N grow

   Write x = k/N for the share you look at. The sum in brackets is close to ln(N/k) = ln(1/x), so

   P(win) ≈ x · ln(1/x)

5. Find the peak

   Differentiate and set to zero: ln(1/x) − 1 = 0, so x = 1/e ≈ 0.368. The chance of winning there is also 1/e ≈ 0.368.

6. The price

   The same logic shows that you take no job at all exactly when the best offer was in the looking zone, which happens with probability k/N = x. At the optimum, 37% of people who follow the rule end up with no job.

   Number of offers N

   40

What the answer does and does not say

The 37% rule is the right answer to one question: how to maximise the chance of the single best job, under all seven premises. Change the question and the answer changes:

• If what matters is how good your job is, not only whether it is the very best, the best look phase shrinks to about √N, about 10 out of 100 (Bearden, 2006).
• If you can see how good offers are and know what is out there, there is no look phase at all: start with a high bar and lower it as time runs out (Gilbert & Mosteller, 1966; Guttman, 1960).

Part 2

Rethink the premises

A theorem is only as true as its premises. For each one, decide whether it is true of job hunting, then see what our model will do with it.

Show what the model does with every premise

Loosened in our modelsKept as in the theorem

Loosening P2 already breaks a tempting shortcut

Suppose offers come mostly early in a search: the rate is steady for the first 4 weeks, then falls to about a sixth of that by week 34. This shape is illustrative, not measured. It is in the spirit of the finding that employers call back applicants less often the longer they have been looking for work (Kroft, Lange & Notowidigdo, 2013). With such a search, "37% of your search time" (week 12.6) comes only after about 61% of your offers have passed. That version leaves 61% of people with no job. "37% of your offers" switches in week 6.6 and leaves 37% with no job. The probability literature points the same way: when arrivals are spread out in time, set the cutoff by the expected share of arrivals, not by the calendar (Bruss, 1984).

Once premises are loosened there is no tidy formula left. So we do what computational social scientists do: write down the rules for one simulated job seeker, an agent, and let a computer run thousands of searches.

Part 5

So, what is the best strategy?

Inside the models, "best" depends on what you want, how many offers you will get, what you know about jobs and about yourself, and whether companies choose you back. Each loosened premise changed the answer:

• The classical answer (look 37%, then leap) is best only for "the single best job or nothing".
• One-sided: a bar that starts high and falls with time.
• Two-sided: know yourself, aim a little above it, and learn from rejections.

Even inside a toy world there is no single answer.

Outside the models, every result is conditional on assumptions that research on real job seekers contradicts:

Three ideas to take away

1. A model can explain, predict or illustrate, and these are different jobs (Hofman et al., 2021). These models illustrate. Even large teams with rich data struggle to predict individual life outcomes (Salganik et al., 2020).
2. Matching a pattern is not proof. The longer people have been out of work, the less often they find a job. That could mean employers avoid people with a long gap, or simply that the strongest candidates leave the pool first. The pattern alone cannot tell these apart. One way is an experiment: send résumés that differ in how long the applicant has been out of work, and see whether callbacks fall (Kroft, Lange & Notowidigdo, 2013). Many mechanisms can grow the same pattern (Epstein, 1999).
3. The value is in the disciplined argument. Writing the famous rule down exposed seven premises and a goal. Simulating it showed which ones matter: the goal, how many offers you get, how you learn, and whether companies choose you back.

The best strategy in these models is a falling bar set by what you have learned, about jobs and about yourself. The best strategy for your own job search is not something a model like this can tell you. Seeing exactly why is the point of the exercise.

Discussion

1. Goals. The 37% rule maximises the chance of "the best". What objective would you write down for a job search instead? Try the "value of ending with no job" slider.
2. Premises. Which premise did you vote "false" that our models still keep? How would you loosen it, and what would you predict?
3. Measurement. Our models need the number of offers or openings a person sees. How would you measure it for real, and what would that measurement miss?
4. Self-knowledge. In Part 4, self-image mattered more than strategy. Where do job seekers learn how strong a candidate they are, and how could that learning go wrong?
5. Validation. In Japan, 98.0% of university graduates who wanted a job had one by 1 April 2026 (MEXT & MHLW, 2026). Should a toy model try to match this? What would it prove if it did?
6. Ethics. Job platforms and hiring software rank applicants and suggest jobs. What changes when a model of human choice is used to steer choices rather than to understand them?

Further reading

Every reference was checked against its DOI registry or source page.

The puzzle and its mathematics

• Ferguson, T. S. (1989). Who solved the secretary problem? Statistical Science, 4(3), 282–289. doi:10.1214/ss/1177012493
• Gilbert, J. P., & Mosteller, F. (1966). Recognizing the maximum of a sequence. Journal of the American Statistical Association, 61(313), 35–73. doi:10.1080/01621459.1966.10502008
• Guttman, I. (1960). On a problem of L. Moser. Canadian Mathematical Bulletin, 3(1), 35–39. doi:10.4153/cmb-1960-008-8
• Bearden, J. N. (2006). A new secretary problem with rank-based selection and cardinal payoffs. Journal of Mathematical Psychology, 50(1), 58–59. doi:10.1016/j.jmp.2005.11.003
• Bruss, F. T. (1984). A unified approach to a class of best choice problems with an unknown number of options. The Annals of Probability, 12(3), 882–889. doi:10.1214/aop/1176993237

Job search as a model of behaviour

• McCall, J. J. (1970). Economics of information and job search. The Quarterly Journal of Economics, 84(1), 113–126. doi:10.2307/1879403
• Jovanovic, B. (1979). Job matching and the theory of turnover. Journal of Political Economy, 87(5), 972–990. doi:10.1086/260808
• Shimer, R., & Smith, L. (2000). Assortative matching and search. Econometrica, 68(2), 343–369. doi:10.1111/1468-0262.00112
• Baumann, C., Singmann, H., Gershman, S. J., & von Helversen, B. (2020). A linear threshold model for optimal stopping behavior. PNAS, 117(23), 12750–12755. doi:10.1073/pnas.2002312117
• Goldstein, D. G., McAfee, R. P., Suri, S., & Wright, J. R. (2020). Learning when to stop searching. Management Science, 66(3), 1375–1394. doi:10.1287/mnsc.2018.3245

Evidence about real job seekers

• Bertrand, M., & Mullainathan, S. (2004). Are Emily and Greg more employable than Lakisha and Jamal? A field experiment on labor market discrimination. American Economic Review, 94(4), 991–1013. doi:10.1257/0002828042002561
• Kroft, K., Lange, F., & Notowidigdo, M. J. (2013). Duration dependence and labor market conditions: Evidence from a field experiment. The Quarterly Journal of Economics, 128(3), 1123–1167. doi:10.1093/qje/qjt015
• Mueller, A. I., Spinnewijn, J., & Topa, G. (2021). Job seekers' perceptions and employment prospects: Heterogeneity, duration dependence, and bias. American Economic Review, 111(1), 324–363. doi:10.1257/aer.20190808
• Fujita, S., & Moscarini, G. (2017). Recall and unemployment. American Economic Review, 107(12), 3875–3916. doi:10.1257/aer.20131496
• Mas, A., & Pallais, A. (2017). Valuing alternative work arrangements. American Economic Review, 107(12), 3722–3759. doi:10.1257/aer.20161500
• Ministry of Education, Culture, Sports, Science and Technology & Ministry of Health, Labour and Welfare (2026, May 22). 令和7年度大学等卒業者の就職状況調査（4月1日現在） [Survey of the employment of 2026 university graduates, as of 1 April]. mext.go.jp

Why model? Computational social science

• Epstein, J. M. (2008). Why model? Journal of Artificial Societies and Social Simulation, 11(4), 12. jasss.org
• Epstein, J. M. (1999). Agent-based computational models and generative social science. Complexity, 4(5), 41–60. doi:10.1002/(SICI)1099-0526(199905/06)4:5<41::AID-CPLX9>3.0.CO;2-F
• Hofman, J. M., et al. (2021). Integrating explanation and prediction in computational social science. Nature, 595, 181–188. doi:10.1038/s41586-021-03659-0
• Salganik, M. J., et al. (2020). Measuring the predictability of life outcomes with a scientific mass collaboration. PNAS, 117(15), 8398–8403. doi:10.1073/pnas.1915006117