Fermi Estimation: Getting the Order of Magnitude Right

2 min read

"How many windows are in this building?" isn't trivia — it's a test of whether you can decompose an unknown into knowable pieces, estimate each honestly, and multiply without losing track of magnitude. Market-making interviews then make you trade on your estimate, so the skill compounds.

The method

Decompose into factors you can anchor: population sizes, frequencies, rates, physical dimensions.
Estimate each factor to the nearest half order of magnitude — is it closer to 1, 3, or 10? Precision beyond that is fake at this stage.
Multiply, tracking powers of ten separately from the leading digits.
Sanity-check the result against an independent anchor.

Classic: piano tuners in Chicago. ~3M people → ~1M households → ~1-in-20 owns a piano → 50k pianos → tuned yearly → 50k tunings/year → a tuner does ~4/day × 250 days = 1,000/year → ~50 tuners. Every factor is defensible; errors partially cancel (that's not luck — independent over- and under-estimates offset, which is why decomposition beats one wild guess).

Error discipline

Geometric, not arithmetic, thinking. If your range for a factor is 10–40, your point estimate is $\sqrt{10 \times 40} = 20$ , not 25. Estimates multiply, so means should too.
Confidence interval on the product: with $k$ independent factors each good to a factor of 2, the product is good to roughly a factor of $2^{\sqrt{k}}$ , not $2^k$ — cancellation again. Five decent factors still land within ~3× of truth.
Know your anchors cold: world ~8B people, US ~340M, a day ~ $10^5$ seconds, a year ~ $3 \times 10^7$ seconds ("π × 10⁷" as the mnemonic), US GDP ~ $28T, S&P daily volume ~$ 500B notional traded across US equities. A dozen memorized anchors cover most questions.

The trading version

The interviewer will follow your estimate with "make me a market on it." Your spread must reflect your confidence interval — this is where Fermi meets market-making. If you estimated 600 windows with a 90% CI of 300–1,200, quoting 550/650 is malpractice; quote wide (say 350/1,000), and tighten only as the conversation adds information. State your decomposition aloud: it lets the interviewer trade with your reasoning rather than against a mystery number, and it shows the CI is derived, not vibes.

Worked micro-example

"How many trades print on a major US stock in a day?" Anchor: a liquid large-cap trades ~ $1–5B notional/day; average trade size on lit venues is small — order of$ 10^4 $dollars (~100 shares × ~$ 100). So $10^9 / 10^4 = 10^5$ — on the order of 100k trades, more for the very biggest names. Whether the truth is 60k or 300k, you're in the right decade, which is what "right" means here.

The interview version

Rules of engagement: never say "I don't know" and stall — start decomposing; never quote false precision ("about 47,500") — magnitude plus a CI reads as calibrated; and when the interviewer offers information mid-question ("actually there are 60 floors"), visibly update — the update is worth more than the original estimate. Estimation questions are calibration questions in costume, and calibration is the trait trading firms are actually hiring.