Fermi estimation questions — "how many piano tuners are in Chicago?", "how many golf balls fit in a school bus?", "what's the mass of all the ants on Earth?" — show up at nearly every trading firm, usually in the phone screen or superday. At Jane Street they are a named stage: you give an estimate, then a 90% confidence interval, and then the interviewer offers to bet against your interval. SIG and Optiver run close variants, often folded into a market-making game.
Why trading firms ask these
The question is never about pianos. Trading is the job of putting numbers on things you cannot look up — the fair value of an option thirty seconds after a headline, the size of a flow you're seeing. A Fermi question tests exactly that skill under observation:
- Decomposition: can you break an unknowable number into factors you can estimate?
- Calibration: does your stated uncertainty match your actual uncertainty? Overconfident intervals lose money; so do uselessly wide ones.
- Updating: when the interviewer gives you a new fact ("Chicago has about 9 million people in the metro area"), do you revise cleanly or anchor on your first guess?
Candidates who nail the arithmetic but quote a 90% interval that's really a 50% interval fail this round. The calibration is the point.
The core toolkit
Three tools cover almost every estimation question.
1. Anchor numbers. Memorize a small set of constants so every chain starts from solid ground:
| Anchor | Value |
|---|---|
| US population | ~330 million |
| World population | ~8 billion |
| Seconds in a year | ~$3.15 \times 10^7$ (π × 10&sup7; is a cute mnemonic) |
| US households | ~130 million (~2.5 people each) |
| Large metro area | 5–10 million people |
| Human working year | ~250 days, ~2,000 hours |
2. Geometric means for ranges. When a factor could plausibly be anywhere from $a$ to $b$, use $\sqrt{ab}$, not the midpoint. If between 1 in 10 and 1 in 100 households own a piano, estimate $\sqrt{\tfrac{1}{10} \cdot \tfrac{1}{100}} \approx \tfrac{1}{30}$. Fermi errors are multiplicative, so you should average in log space.
3. Error cancellation. This is the part interviewers love hearing. If each of $k$ independent factors carries a log-space error of $\sigma$, the product's log-error is $\sigma\sqrt{k}$ — not $\sigma k$ — because independent errors partially cancel. Five factors each off by up to a factor of 2 leave you off by roughly $2^{\sqrt{5}} \approx 4.5$, not $2^5 = 32$. More factors of moderate quality often beat fewer factors of guessed quality.
Worked example: piano tuners in Chicago
Decompose into a chain:
$$N_{\text{tuners}} = \frac{\text{population} \times \tfrac{1}{\text{household size}} \times P(\text{piano}) \times \text{tunings/yr}}{\text{tunings per tuner per yr}}$$
- Chicago metro: ~9 million people, ~2.5 per household → 3.6M households.
- Piano ownership: between 1/10 and 1/50 of households; geometric mean ≈ 1/22, call it 1/20 → ~180,000 pianos.
- Tuning frequency: about once per year → 180,000 tunings/year of demand.
- Tuner capacity: a tuning plus travel is ~2 hours, so ~4 per day × 250 working days = 1,000 per year.
$$N \approx \frac{180{,}000}{1{,}000} = 180 \text{ tuners}$$
Now the confidence interval. Five uncertain factors, each within roughly a factor of 2 (log-error $\sigma \approx \ln 2$). Combined: $\ln 2 \cdot \sqrt{5} \approx 1.55$, i.e. a one-sigma multiplicative factor of $e^{1.55} \approx 4.7$. So quote something like "180, and I'd put 90% probability between about 40 and 800." Directory listings put the true number around 100–200 — but stating the interval and its reasoning is worth more than landing near the truth.
The betting follow-up
At Jane Street especially, expect: "I'll pay you \$10 if the true number is inside your interval; you pay me \$X if it's outside. What X makes you indifferent?" If your interval is genuinely 90%, you should accept anything up to 9:1 against — expected value $0.9 \times 10 - 0.1 \times X = 0$ gives $X = 90$. Refusing a 3:1 bet on your own "90%" interval reveals it was never a 90% interval. This is the same calibration logic behind expected value questions and the sizing math in the Kelly criterion; some interviewers will push straight into "make me a market on the number of piano tuners," which turns it into a market-making exercise with a bid, an ask, and a width that reflects your uncertainty.
Traps that fail candidates
- Silent arithmetic. The interviewer is grading the chain, not the answer. Narrate every factor and flag which ones you'd bet on least.
- Arithmetic-midpoint ranges. Averaging 1/10 and 1/100 to get ~1/18 instead of the geometric ~1/30 biases every wide-range factor upward.
- Anchoring after new information. If told your household-size guess is off, recompute the whole chain; don't nudge the final answer by 10%.
- Symmetric intervals on a multiplicative quantity. "180 ± 100" is wrong-shaped; uncertainty here is a factor, so the interval should be 40–800, not 80–280.
- False precision. Saying "183.6" signals you don't understand your own error bars. Round to one significant figure and say why.
Practice this properly
Estimation rounds rarely appear alone — they sit next to probability puzzles and games in the same superday. Drill the adjacent material in our brain teaser bank and probability question bank, then pressure-test your calibration for real: play a few rounds of our betting game, where quoting overconfident intervals costs you immediately, or try the market-making game to practice turning an estimate into a two-sided quote.
More topic guides
- The Airplane Seat Problem: Why the Answer Is 1/2 (Three Proofs)
- Bayes' Theorem in Quant Interviews
- Behavioral Interview Questions at Trading Firms (With Answer Frameworks)
- Coin Flip Questions in Quant Interviews
- C++ Low-Latency Interview Questions at HFT Firms
- Dice Questions in Quant Interviews
- Gambler's Ruin in Quant Interviews
- The Kelly Criterion in Quant Interviews
- Linear Regression Interview Questions: OLS Assumptions, R² Traps & Regression to the Mean
- Machine Learning Quant Interview Questions: Overfitting, Cross-Validation & Feature Leakage
- The Market Making Game Interview: How to Answer 'Make Me a Market'
- Market Microstructure Interview Questions: Order Books, Spreads & Adverse Selection
- Markov Chains in Quant Interviews
- Martingales in Quant Interviews
- Mental Math for Trading Interviews: Training Plan, Zetamac Benchmarks & Firm Tests
- The Monty Hall Problem — and the Variants Interviews Actually Ask
- Number Sequence Tests in Trading Interviews: The 8 Pattern Types & How to Practice
- Optimal Stopping in Quant Interviews
- Options Pricing Interview Questions: Black-Scholes, Greeks & Put-Call Parity
- Quant Interview Cheat Sheet: Probability, Markov Chains, Options & Linear Algebra (Free PDF)
- Random Walks in Quant Interviews
- Statistics Interview Questions for Quant Roles: Hypothesis Testing, MLE & p-Value Traps
- Stochastic Calculus Interview Questions: Ito's Lemma, SDEs & Brownian Motion
- Time Series Interview Questions: Stationarity, ARMA & Autocorrelation Traps
- Top 50 Quant Interview Questions (With Full Solutions)
- All guides & explainers
Frequently asked questions
Which trading firms ask Fermi estimation questions?
Jane Street is the best-known example: candidates give an estimate plus a 90% confidence interval, and the interviewer often offers to bet against the interval. SIG, Optiver, and most options market makers run similar estimation or calibration exercises, frequently combined with a market-making game where you quote a bid and ask around your estimate.
How do I give a good confidence interval on a Fermi estimate?
Treat the uncertainty as multiplicative, not additive. If your chain has k independent factors each uncertain by a factor of roughly 2, the total log-space error is about ln(2) times the square root of k, so five such factors give a one-sigma factor of roughly 4-5. Quote the interval as your estimate divided and multiplied by that factor, and be prepared to accept a bet at odds consistent with the probability you claimed.
Does the final answer need to be close to the true number?
No. Interviewers grade the decomposition, the reasonableness of each factor, and whether your stated uncertainty matches your real uncertainty. A candidate who lands within a factor of 3 with a clearly narrated chain and a well-calibrated interval beats one who happens to guess the exact number with no visible reasoning.
What anchor numbers should I memorize for estimation questions?
A short list covers most questions: US population about 330 million, world population about 8 billion, roughly 130 million US households at 2.5 people each, about 3.15 x 10^7 seconds in a year, and a working year of about 250 days or 2,000 hours. From these you can derive most demographic, time-based, and capacity factors on the spot.
Practice the real thing
QuantVault has 2,800+ quant interview problems with full solutions, intuition, and hints, firm-by-firm interview funnels, and an auto-graded coding judge. Start free.