Optimal Stopping — The 37% Rule
Candidates arrive one at a time and every "no" is final: look first, then leap, and discover why the best moment to start choosing is about 37% of the way through
This concept builds on Probability Trees — Mapping Every Outcome. Consider exploring it first for the best experience.
One Pass, No Going Back
You must choose one of several candidates who arrive in random order. You can rank each one only against those already seen, and once you pass someone they are gone for good. The simplest plan is to take the first candidate. Set "Candidates" to 3 and "Look phase" to 0: you pick the best only 1 time in 3, a 33.3% chance.
Three Candidates, Six Orders
Now set "Look phase" to 1: let the first candidate go, but remember them as a benchmark, then take the first later one who beats it. List all six orders of the scores 1, 2 and 3. The rule finds the best in 3 of them, so the "Exact chance of the best" card reads 50.0% — better than taking the first.
Look, Then Leap
With 10 candidates the look phase is the shaded band. Cards there are never chosen; the tallest of them sets the dashed yellow benchmark. In the leap phase the first card taller than the benchmark is chosen, green if it is the best of all and orange if not. With "Look phase" at 0 you simply take the first card, which wins only 10% of the time.
Too Short or Too Long
Drag "Look phase" up to 8. Now the benchmark is strong, but the best candidate is often among the 8 you let go, and then nobody beats the benchmark, so you are stuck with the last card. The chance falls to 18.9%. Look too briefly and the benchmark is weak; look too long and the best has already left.
Worked Example: Ten Candidates
With n = 10 and a look phase of r = 3, you win when the best arrives at some position i after the look phase and the best of the first i − 1 sits inside the look phase. Adding these cases gives P = (3/10) × (1/3 + 1/4 + … + 1/9) = 0.3 × 1.329 = 0.399, a 39.9% chance. Press "Best look phase" to confirm that 3 is the best choice for 10 candidates.
The 37% Rule
Set "Candidates" to 50 and press "Best look phase": it picks 18, which is 36% of 50, and the exact chance is 37.4%. As n grows, the best share and the best chance both approach 1/e ≈ 0.368, where e ≈ 2.718. Press "Run 100 rounds" a few times and watch the pink cross settle onto the purple curve.
How It Works — Step by Step
-
1
One Pass, No Going Back
You must choose one of several candidates who arrive in random order. You can rank each one only against those already seen, and once you pass someone they are gone for good. The simplest plan is to take the first candidate. Set "Candidates" to 3 and "Look phase" to 0: you pick the best only 1 time in 3, a 33.3% chance.
-
2
Three Candidates, Six Orders
Now set "Look phase" to 1: let the first candidate go, but remember them as a benchmark, then take the first later one who beats it. List all six orders of the scores 1, 2 and 3. The rule finds the best in 3 of them, so the "Exact chance of the best" card reads 50.0% — better than taking the first.
-
3
Look, Then Leap
With 10 candidates the look phase is the shaded band. Cards there are never chosen; the tallest of them sets the dashed yellow benchmark. In the leap phase the first card taller than the benchmark is chosen, green if it is the best of all and orange if not. With "Look phase" at 0 you simply take the first card, which wins only 10% of the time.
-
4
Too Short or Too Long
Drag "Look phase" up to 8. Now the benchmark is strong, but the best candidate is often among the 8 you let go, and then nobody beats the benchmark, so you are stuck with the last card. The chance falls to 18.9%. Look too briefly and the benchmark is weak; look too long and the best has already left.
-
5
Worked Example: Ten Candidates
With n = 10 and a look phase of r = 3, you win when the best arrives at some position i after the look phase and the best of the first i − 1 sits inside the look phase. Adding these cases gives P = (3/10) × (1/3 + 1/4 + … + 1/9) = 0.3 × 1.329 = 0.399, a 39.9% chance. Press "Best look phase" to confirm that 3 is the best choice for 10 candidates.
-
6
The 37% Rule
Set "Candidates" to 50 and press "Best look phase": it picks 18, which is 36% of 50, and the exact chance is 37.4%. As n grows, the best share and the best chance both approach 1/e ≈ 0.368, where e ≈ 2.718. Press "Run 100 rounds" a few times and watch the pink cross settle onto the purple curve.
Key Takeaways
Look First, Then Leap
Let the first r candidates go, remember the best of them, then take the first later candidate who beats that benchmark.
The Chance Has an Exact Formula
P(r) = (r/n) × (1/r + 1/(r+1) + … + 1/(n − 1)); for 10 candidates the best look phase is 3, with a 39.9% chance of choosing the very best.
About 37% Is the Sweet Spot
For many candidates the best look phase is about n/e — roughly 37% of them — and the chance of ending with the best also approaches 1/e ≈ 36.8%.
A Rule Beats Guessing by a Lot
Picking at random finds the best of 50 only 2% of the time; the look-then-leap rule finds it about 37% of the time, no matter how many candidates there are.
Connected Concepts
Probability Trees
With three candidates you can list all six arrival orders as branches and count the ones where the look-then-leap rule lands on the best.
Harmonic Series
The success chance is (r/n) times a slice of the harmonic series, 1/r + 1/(r+1) + … + 1/(n − 1) — and that slice grows like a logarithm.
Compound Interest
The number e that appears when interest compounds continuously is the same e behind the rule: look at n/e candidates and win with chance 1/e.
Logarithms
For large n the harmonic slice becomes ln(n/r), so the chance turns into x · ln(1/x) for a look share x — a curve whose peak sits at x = 1/e.
Monte Carlo Method
Pressing "Run 100 rounds" plays random arrival orders and counts wins — a Monte Carlo check of the exact formula.
Law of Large Numbers
The simulated rate jumps around after 100 rounds but settles onto the exact curve as hundreds more rounds pile up.
Expected Value
Optimal stopping is about choosing a strategy before the data arrive — the same habit of averaging over every possible outcome that defines expected value.