Fermi Estimation: Producing a Number When You Have No Data

You can produce a useful number for almost anything

Someone asks: how many piano tuners are there in your city? How much does your phone habit cost you a year? Is this "£2 billion market" claim even plausible? The instinct is "I have no idea, I'd need to look it up." But you can produce a genuinely useful answer — right now, no data, no internet — and it will usually land within a factor of a few of the truth. This is Fermi estimation, named after the physicist who reputedly estimated the yield of the first atomic bomb by dropping bits of paper as the blast wave passed.

The trick isn't knowing the answer. It's breaking an impossible question into small questions you can guess, estimating each roughly, and combining them — and then, magically, the errors in your guesses tend to cancel out, leaving a total that's far better than any single wild guess. It's one of the highest-leverage thinking skills there is: it lets you sanity-check claims (that "£2bn market"), scope decisions, and answer "roughly how big?" without waiting for data you don't have.

What you'll have at the end

• The decompose → estimate → multiply method
• Why order-of-magnitude ("within 10×") is the right target, and enough
• Three real estimates you worked out from nothing

The method: decompose, estimate, combine

The move is always the same:

1. Break the big unknown into factors you can estimate. Don't guess the answer directly — build it from pieces small enough to have a feel for.
2. Estimate each factor roughly. You know more than you think: rough populations, how often things happen, typical sizes and prices.
3. Combine (usually multiply) to get the answer.
4. Sanity-check the result against anything you do know.

The classic: how many piano tuners in a city of a million?

People in the city ............................ ~1,000,000
People per household .......................... ~2.5  → ~400,000 households
Fraction with a piano ......................... ~1 in 20 → ~20,000 pianos
Tunings per piano per year .................... ~1
→ tunings needed per year ..................... ~20,000
Tunings one tuner does per year
   (~4/day × ~250 working days) ............... ~1,000
→ tuners needed = 20,000 / 1,000 ............. ~20 piano tuners

No data was consulted, and the answer is in the right ballpark. Every step was a question you could guess, and multiplying them built an answer to a question you couldn't.

Why it works: errors cancel

The part that feels like magic has a real reason. Each of your estimates is off — some too high, some too low. But if your guesses are honest (not all biased the same way), the over- and under-estimates tend to cancel when you combine them, so the total is far more accurate than its parts. Guess five factors each maybe 2× off in random directions, and the product is usually within a small factor of the truth, not 2⁵ off.

This is why decomposition beats a direct guess: a single "how many piano tuners?" guess has nothing to cancel against and can be wildly wrong; a five-step estimate lets the errors police each other.

Aim for order of magnitude

The target is not the exact number — it's the right order of magnitude: is it about 20, or about 200, or about 2,000? Getting within a factor of ~10 (ideally ~3) is the goal, and it's usually enough to make the decision or catch the lie. "Is this closer to thousands or millions?" answers most real questions.

Think in powers of ten and round hard: use 1, 2, or 5 × a power of ten, and don't sweat whether it's 2.5 or 3 — at this resolution it doesn't matter, and the rounding keeps the arithmetic doable in your head. Precision you can't justify is fake precision; order-of-magnitude honesty is the whole game.

Anchor on what you know

You have more usable knowledge than you feel you do — build from anchors:

• Populations and scales: your city, your country (hundreds of millions? tens of millions?), the world (~8 billion). Round numbers you can peg to.
• Ratios and rates: how often you do a thing, roughly what fraction of people own/do X, typical prices, how long tasks take.
• Yourself as a sample: how much you spend on coffee / data / snacks, scaled up (carefully — DT-04 warns you're not always representative, but you're a starting anchor).
• Bracketing: if unsure, estimate a "surely more than" low bound and a "surely less than" high bound, and take something in between. The true answer is caged.

Sanity-check at the end

Always finish by asking "does this feel right?" against reality. ~20 piano tuners for a million people — plausible. If your method had said 20,000, you'd know a step was wrong (probably an order of magnitude slipped somewhere) and go hunt it. This end-check is also how you use Fermi defensively: when someone quotes a big impressive number, do the back-of-envelope and see if it survives. A surprising share of confident public numbers collapse under one honest Fermi estimate.

What this means for you

• You can produce a useful number for almost anything — no data needed, usually within a factor of a few.
• Decompose → estimate each factor → multiply → sanity-check. Never guess the answer directly.
• Errors cancel when guesses are honest, so a multi-step estimate beats a single wild guess.
• Aim for order of magnitude (within ~10×) — round to 1/2/5 × powers of ten; fake precision is worthless.
• Anchor on what you know (populations, rates, yourself, brackets), and use it to sanity-check big claims — many don't survive.

Exercise (30 min, verifiable output)

1. Pick three real questions you don't know the answer to — e.g. how much you spend a year on [habit], how many [something] are in your country, whether some quoted "market size" is plausible.
2. For each, decompose into factors, estimate each (listing your assumptions), and combine.
3. State the order of magnitude and sanity-check it against anything you know or can quickly verify.
4. Where you can, look up one afterwards — how close were you? (Within 10× is a win.)
5. Note which assumption mattered most — the one your answer was most sensitive to.

✅ Finish check: three estimates worked start to finish with assumptions listed, each given as an order of magnitude and sanity-checked — and at least one checked against reality.

Summary card

• Fermi estimation produces a useful number from nothing by decomposing an impossible question into guessable factors, then multiplying.
• Errors cancel across honest guesses, so the multi-step estimate beats a direct one.
• Target order of magnitude (within ~10×), rounding to 1/2/5 × powers of ten — fake precision is worthless.
• Anchor on populations, rates, yourself, and brackets ("surely more than / less than").
• Always sanity-check the result — and use Fermi to test big public numbers, which often collapse.

Sources

1. Weinstein, L. & Adam, J. — Guesstimation, 2008
2. von Baeyer, H. — The Fermi Solution, 1988
3. Weinstein, L. — Guesstimation 2.0, 2012

Next lesson: DT-06 — p-values, Statistical Significance and Effect Size (L3) Related: DT-10 Decision Frameworks · DT-01 Reading Data · MN-04 Compound Interest · DT-02 How Charts Lie Path: related — the estimation toolkit