Jane Street Estimathon Guide

A complete guide to Fermi estimation, confidence ranges, Jane Street Estimathon questions, and structured quantitative reasoning under uncertainty.

What is a Jane Street Estimathon?

An Estimathon is a competition built around a deceptively difficult skill: producing useful numerical estimates when you do not know the exact answer. Instead of asking you to calculate a value from perfectly specified inputs, an Estimathon gives you questions where much of the information is missing. You might be asked to estimate the number of stop signs in a city, the number of people doing something at this exact moment, the mass of a large structure, or a mathematical quantity that is inconvenient to compute exactly.

Jane Street's public Estimathon format is team-based. The current public rules give teams 30 minutes to work on 13 estimation problems and ask them to submit an interval for each answer rather than one number. Jane Street has run the event since 2013 across universities, mathematics conferences, student programs, and internal groups. The important idea is not trivia recall. It is whether you can build a model quickly, identify the variables that matter, and express how uncertain you are about the result.

This is why a question such as "How many cups of coffee are consumed in New York City each day?" is useful even if nobody in the room knows the official statistic. A weak approach is to guess a number that feels plausible. A much stronger approach is to reason from population, the fraction of people who drink coffee, average cups per coffee drinker, commuter and tourist effects, and the difference between coffee consumed inside the city and coffee consumed by residents elsewhere. The final estimate is simply the output of that model.

It helps to separate four different things that are often blurred together:

The final category is what makes the Estimathon especially interesting. It is possible to have a decent central estimate and still perform poorly by being overconfident. It is also possible to protect yourself with an enormous interval, but then lose much of the value of having estimated anything at all. Good performance therefore requires both estimation and calibration.

Preparing for Jane Street more broadly?
Estimation is only one style of quantitative reasoning. You can combine it with the Jane Street preparation hub, Jane Street interview questions, probability questions, brainteasers, and market-making games.

How Estimathon Questions Work

The public Estimathon rules are more specific than a normal "guess the number" game. Teams submit a closed interval [L, U], where L is the lower bound and U is the upper bound. The interval is successful if the true answer lies inside it. The answer must be positive, and the current public rules allow scientific notation while disallowing external reference material such as the internet or a calculator.

The event also creates a strategic constraint. Teams have 13 problems but only 18 total interval submissions. Because a later submission replaces an earlier submission for that question, you can spend some of those extra submissions correcting a miss or tightening a range after receiving scoreboard feedback. You cannot endlessly revise every problem. Time and submissions are both scarce resources.

The scoring intuition

In the current public rules, the team's total score combines two effects. Correct intervals contribute a penalty related to their multiplicative width, while every missing or incorrect problem creates an exponential penalty. In simplified notation, the published format can be thought of as:

Current public scoring formula

Score = (10 + Σ floor(Uᵢ/Lᵢ)) × 2^(13 - G)

where G is the number of good intervals, Lᵢ is the submitted lower bound, and Uᵢ is the submitted upper bound for each good interval. Lower scores are better.

The exact integer-width convention has varied across circulated Estimathon rule sheets. Current materials use floor(U/L); some older rule sheets use the closely related ceil(U/L) - 1 convention. Those formulas differ at some bucket boundaries, so if you are attending a real event, use the scoring sheet you are given. The strategic lesson is the same: narrow correct intervals are better, but a missed answer is far more damaging than making a correct interval somewhat wider.

For intuition, imagine every good interval has a width term of 1. Getting all 13 questions good gives (10 + 13) × 1 = 23. Getting 12 good gives (10 + 12) × 2 = 44. Getting only 10 good gives (10 + 10) × 8 = 160. The exact score depends on interval widths, but this shows why covering an uncertain answer can be worth more than shaving a small amount from the width of an already-good interval.

Notice that the width is naturally multiplicative rather than additive. The interval [1, 2] and the interval [1,000,000, 2,000,000] are equally wide in relative terms: the upper endpoint is twice the lower endpoint. This is exactly what you want in estimation problems that span many orders of magnitude.

This also explains why an interval such as [99, 101] is not automatically "better" reasoning than [70, 140]. If you genuinely know the answer is approximately 100, the first interval is excellent. If your model relies on three uncertain assumptions and could easily be off by 50%, the first interval is reckless. Calibration means making the width reflect the uncertainty in the model, not your desire to look precise.

A final terminology point: Estimathon participants often talk about "confidence intervals", but these are not necessarily statistical confidence intervals in the textbook sense. You are usually not sampling from a distribution and constructing a formal 95% interval. It is better to think of the submitted range as a subjective uncertainty interval: a defensible region in which you believe the true value is likely to lie.

A Repeatable Estimation Framework

The best way to improve at estimation is to stop treating every question as a new puzzle. Most strong solutions follow the same underlying process. The surface details change, but the structure does not.

1. Define exactly what is being estimated

Before calculating anything, make the target precise. "How many coffees are consumed in New York each day?" could mean New York State or New York City; cups purchased or cups consumed; residents only or everyone physically inside the city; brewed coffee only or espresso-based drinks too. A factor-of-two error often comes from solving the wrong version of the problem rather than from bad arithmetic.

2. Identify useful anchors

Write down facts you plausibly know. You may know that the United States has a few hundred million people, New York City has under ten million residents, a working year has about 250 weekdays, a day has 86,400 seconds, a litre of water weighs roughly one kilogram, and commercial aircraft carry hundreds rather than thousands of passengers. Anchors do not need to be perfect. They only need to be much better than starting from nothing.

3. Decompose the unknown

Convert the target into quantities that are easier to estimate. If N is hard to estimate directly but you can write N = A × B × C, the problem becomes three smaller estimation tasks. This is the core Fermi move.

For coffee consumption, for example:

Daily cups ≈ people present × fraction who drink coffee × cups per coffee drinker per day

For piano tuners:

Number of tuners ≈ annual piano tunings required ÷ annual tunings performed by one tuner

4. Choose explicit assumptions

Do not hide uncertainty inside a single unexplained number. If you assume 40% of people drink coffee, say so. If you assume a piano is tuned once per year, say so. Explicit assumptions are easier to challenge, replace, and bound. They also make your reasoning easier to communicate to teammates or an interviewer.

5. Calculate at the right precision

Estimation is usually harmed, not helped, by unnecessary decimal places. If your population input could be wrong by 10% and your behavioural assumption could be wrong by 50%, spending time computing 8.42 × 0.37 × 1.84 precisely is false precision. Round to numbers that preserve the order of magnitude and make mental arithmetic easy.

6. Sanity-check the order of magnitude

Ask whether the answer is physically, economically, or demographically possible. If your calculation implies that a city with eight million people consumes 80 million takeaway coffees every morning, you have either assumed ten coffees per person or made a unit error. A good sanity check often catches mistakes faster than redoing the arithmetic.

7. Find the assumptions that dominate uncertainty

Not all inputs deserve equal attention. If one input is known within ±5% and another could plausibly vary from 0.2 to 0.8, improving the first number will barely change your interval. Spend your time on the assumption with the largest multiplicative uncertainty.

8. Build lower and upper cases

Recompute the model using assumptions that are jointly conservative and jointly aggressive, while keeping them plausible. This gives you a reasoned interval rather than an arbitrary ±20% around the midpoint.

9. Adjust the interval to your actual confidence

Your first lower and upper cases are not sacred. Ask what you might have omitted: tourism, substitution effects, duplicated people, seasonality, capacity constraints, definitions, or an anchor that may be badly remembered. If there is a credible failure mode that would move the answer by 3×, your final range should reflect that unless you can rule it out.

Fermi Estimation: The Core Skill

Fermi estimation is named after physicist Enrico Fermi, who became famous for producing surprisingly useful estimates from very limited information. The method is powerful because many real-world quantities are difficult to know directly but can be reconstructed from simpler pieces.

Consider a deliberately simple question: How many heartbeats does a person have in a year? You do not need a medical database. Use an anchor of roughly 70 beats per minute:

70 beats/min × 60 min/hour × 24 hours/day × 365 days/year

≈ 70 × 60 × 9,000
≈ 38,000,000

A sensible answer is therefore on the order of 4 × 10^7 heartbeats per year. We do not care whether the exact arithmetic is 36.8 million or 38.2 million because the resting heart-rate assumption, sleep, exercise, age, and individual variation are larger sources of uncertainty than the multiplication error.

Decomposition

Decomposition turns an unfamiliar quantity into familiar building blocks. It is often helpful to look for one of a few recurring identities: population × behaviour, rate × time, area × density, customers × spend, inventory ÷ throughput, or volume ÷ object volume. These patterns show up repeatedly in estimation interviews because they force you to build a model rather than recall a fact.

Order-of-magnitude reasoning

When the exact answer is uncertain, getting the exponent right is often the first objective. If you estimate a quantity as 6 × 10^5 and the truth is 1.1 × 10^6, your reasoning is much healthier than an answer of 8 × 10^8. Powers of ten help you separate small arithmetic errors from structural errors.

Scientific notation also makes approximate multiplication easier:

(3 × 10^6) × (4 × 10^-2) = 12 × 10^4 ≈ 1.2 × 10^5

Base rates and anchors

A base rate is a reference frequency that constrains a guess. If you are estimating how many people use a service, knowing the city's population is useful. If you are estimating how many restaurants exist, knowing roughly how often people eat out and how many meals a restaurant can serve gives you an independent route. Strong estimators build a mental library of anchors: populations, common dimensions, time conversions, rough salaries, household sizes, transport capacity, densities, and consumption rates.

Dimensional reasoning

Units are a debugging tool. If the question asks for trips per day and your formula ends in people, you are missing a rate. If you multiply watts by hours, you obtain watt-hours, not watts. If you estimate a volume using square metres without multiplying by depth, your model is dimensionally incomplete. Tracking units prevents many silent errors.

Independent routes

A particularly strong technique is to estimate the same target from two different directions. Suppose you want the number of taxis in a city. A demand-side estimate might use trips per resident divided by trips per taxi. A supply-side estimate might use taxis per 1,000 residents based on a city you know. If both routes land near 20,000, your confidence should rise. If one gives 5,000 and the other 80,000, you have learned that at least one model contains a weak assumption.

Useful Mathematical Techniques

Estimathon mathematics is rarely difficult in the traditional sense. The challenge is choosing the right operation quickly and keeping the scale under control. The following tools are especially useful.

Percentages and ratios

Many estimates become easier if you convert percentages into simple fractions. Use 33% as roughly one-third, 20% as one-fifth, 12.5% as one-eighth, and 67% as roughly two-thirds. If a number rises from 80 to 100, that is a 25% increase, because the change of 20 is one-quarter of the original 80. The ability to move comfortably between fractions, percentages, and ratios saves time.

Logarithmic thinking

Logs are useful even when you never explicitly calculate one. Thinking multiplicatively helps you compare uncertainty. Being off by +100 on a quantity near 1,000 is minor; being off by +100 on a quantity near 10 is enormous. A range from 1 million to 3 million is a factor of three wide. A range from 100 million to 300 million has the same relative width. Estimathon scoring reflects this multiplicative view.

Sensitivity analysis

Suppose your estimate is N = A × B × C. If A might be 10% higher, the final answer is 10% higher. If B could plausibly be five times larger than your central assumption, B dominates the uncertainty. This simple observation is one of the most important habits in estimation: ask which assumption has the largest elasticity before spending time refining anything else.

For small independent percentage uncertainties, you can sometimes think of them as accumulating approximately in relative terms. But Estimathon questions rarely justify formal error propagation. The more practical method is to construct explicit scenarios and examine what happens when the uncertain inputs move together.

Building Good Assumptions and Defensible Bounds

Estimation becomes much easier when you recognise that assumptions come in different classes. Some are demographic: population, household size, age distribution. Some are geographic: land area, urban density, distance. Some are temporal: hours per day, working days per year, average duration. Some are physical: capacity, dimensions, speed, mass. Others are economic or behavioural: price, purchase frequency, adoption, utilisation, or willingness to pay.

The best assumption is usually not the number that "sounds right". It is the number that can be connected to something you already know. If you do not know how many cafés are in a city, perhaps you know the population and can estimate how many people one café can serve. If you do not know the total electricity consumption of Sydney, perhaps you know approximate household power usage and can separately add commercial and industrial demand.

Use ranges at the input level

Suppose you are estimating annual spending on a product using:

Annual spend = number of buyers × purchases per buyer per year × price per purchase

Instead of producing one guess for each input, write down a conservative, base, and aggressive case:

A range such as $90m–$500m now has a reason behind it. The width is telling you where the uncertainty comes from. You can then ask whether all three low assumptions could realistically occur together. If price tends to be higher when purchase frequency is lower, for example, multiplying every low input may create an unnecessarily extreme lower bound. Correlation matters.

10% uncertainty versus 5× uncertainty

Imagine the population anchor could be 8.0–8.8 million, while your estimate of the percentage using a service could be anywhere from 5% to 30%. The population is not where you should spend your effort. Even perfect knowledge of population barely tightens the final range. The adoption assumption can move the answer sixfold.

This is a general interview skill as well. A strong candidate does not merely calculate quickly; they identify which unknown is decision-relevant. In trading language, you care about the variable that moves fair value, not the variable you happen to be able to measure most precisely.

Avoid arbitrary ±20% bounds

Taking a central estimate and attaching ±20% because it feels reasonable is usually a sign that the uncertainty has not been modelled. If the estimate depends on three near-known physical constants, ±20% might be very wide. If it depends on an unknown behavioural adoption rate and an uncertain historic population, ±20% might be absurdly narrow. Build the interval from the assumptions, then widen it only for failure modes your model does not capture.

Worked Estimathon-Style Examples

The following examples are intentionally solved from first principles. The exact real-world answers are not the point. What matters is how the estimate is structured, how the largest uncertainties are identified, and how the final interval is justified.

Example 1 — How many cups of coffee are consumed in New York City each day?

Step 1: define the target. We will estimate cups of coffee physically consumed inside New York City on an average day, including residents, commuters, and visitors.

Step 2: choose a population anchor. Use roughly 8.5 million residents. Instead of trying to model tourists and commuters perfectly, we can treat their net contribution as an adjustment later.

Step 3: estimate coffee participation. Suppose roughly 60% of the effective adult-equivalent population drinks coffee on a given day. This is uncertain, but it is a reasonable behavioural starting point.

Step 4: estimate cups per drinker. Use about 1.7 cups per coffee drinker per day. Some have one; some have several.

8.5m people × 0.60 coffee drinkers × 1.7 cups
≈ 8.7m cups/day

That number should immediately trigger a definition check. We treated all 8.5 million residents as equally eligible to drink coffee, including children. A cleaner model would first reduce the population to an adult/older-teen share. Suppose 75% is in a coffee-relevant age group:

8.5m × 0.75 × 0.60 × 1.7 ≈ 6.5m cups/day

Now add a rough net inflow for commuters and visitors. If they add the equivalent of 5–15% to daily consumption, the central estimate lands around 7 million cups per day.

Sanity check: seven million cups in a city with roughly eight to nine million residents means somewhat less than one cup per resident per day. That is plausible once children and non-coffee drinkers are included.

Uncertainty: the dominant assumptions are daily coffee participation and cups per drinker. Population is comparatively well constrained. A reasonable Estimathon-style interval might be 3.5 million to 12 million cups per day. The lower end corresponds to fewer drinkers and closer to one cup each; the upper end allows heavy consumption and larger commuter/visitor effects.

What could make us badly wrong? If "cups" means purchased servings rather than consumed cups, home-brewed coffee changes the answer. If the question means coffee consumed by NYC residents anywhere in the world rather than physically in the city, commuter effects reverse. Definition risk matters as much as arithmetic.

Example 2 — How many piano tuners operate in London?

This is a classic Fermi structure because the number of tuners is hard to estimate directly, but the demand for tuning and the capacity of one tuner are easier to reason about.

Demand side. Use a London population of about 9 million. Assume one acoustic piano per 50 people across homes, schools, studios, venues, and institutions. That gives roughly 180,000 pianos.

Assume the average piano is tuned 0.8 times per year. Home pianos may be tuned less; performance pianos much more.

180,000 pianos × 0.8 tunings/year ≈ 144,000 tunings/year

Supply side. Suppose a full-time tuner completes four tunings per working day across 240 working days:

4 × 240 ≈ 960 tunings per tuner per year

Therefore:

144,000 ÷ 960 ≈ 150 tuners

Central estimate: about 150 piano tuners. The arithmetic is easy; the uncertainty is not. The biggest question is piano prevalence. If there is one piano per 100 people rather than one per 50, the answer halves. If institutional and professional instruments are tuned much more frequently than assumed, it rises again.

A defensible interval might be 60–400 tuners. Notice that this is much wider than ±20%, because the key behavioural and inventory assumptions are genuinely uncertain.

Example 3 — How many tennis balls fit inside a commercial aircraft?

This problem looks like geometry, but the hardest part is defining usable volume. We will imagine a narrow-body commercial aircraft with seats and fittings removed and ask how many tennis balls could physically occupy the interior fuselage and cargo volume.

A tennis ball has radius of roughly 3.3 cm, or 0.033 m. The volume of a sphere is:

V = (4/3)πr³

≈ 4.2 × (0.033)³
≈ 1.5 × 10^-4 m³ per ball

Spheres do not fill space perfectly. A reasonable random packing fraction is around 0.6–0.65, so only about 64% of the available volume becomes tennis-ball material.

Now estimate the interior volume. A rough cylindrical fuselage model might use an interior cross-sectional area of around 9–10 m² over 25–30 metres of useful length, but much of that cylinder is inaccessible structure. Instead of pretending the full geometric cylinder is usable, take roughly 120 m³ as a central usable-volume estimate for the cleared cabin and relevant cargo space.

balls ≈ 120 m³ × 0.64 ÷ 1.5×10^-4 m³
≈ 5.1 × 10^5

Central estimate: about 500,000 tennis balls.

Sanity check: one cubic metre would hold roughly 0.64 / 0.00015 ≈ 4,000 balls. A hundred-plus cubic metres therefore naturally gives a few hundred thousand, not tens of millions.

Range: the aircraft-volume assumption dominates. A reasonable interval might be 250,000–1,000,000. A larger wide-body aircraft would be well above our centre; a smaller aircraft or stricter definition of usable volume would be below it.

Example 4 — How many rideshare trips happen in Sydney each day?

This is a behavioural market-sizing problem. Start with roughly 5.5 million people in Greater Sydney. Rather than assume everyone uses rideshare equally, estimate monthly active riders and then trips per active rider.

Suppose 30% of the population uses a rideshare service in a typical month, giving 1.65 million monthly active riders. Assume the average active rider takes 2.5 rides per month:

1.65m riders × 2.5 rides/month ≈ 4.1m rides/month
4.1m ÷ 30 ≈ 137,000 rides/day

Tourists, airport trips, business travel, and people outside the resident-population frame push the number upward; lower weekday demand pushes some days downward. A central estimate of 150,000 rides per day is reasonable from this model.

Alternative route: estimate the driver fleet. If roughly 25,000 drivers are active on a typical day and each completes six rides, that also produces 150,000 rides. The fact that two very different approaches can agree is more informative than adding another decimal place to either model.

Range: perhaps 70,000–300,000 rides/day. The largest uncertainty is not Sydney's population. It is service adoption, active-driver definition, and average rides per user.

Example 5 — How many people are flying in the world right now?

This example is useful because the first obvious approach is poor. You might start with world population and guess the fraction currently flying. But almost nobody has useful intuition for that fraction. Saying "maybe one person in 5,000" is just moving the guess into a hidden parameter.

A better model uses flow × duration. If a certain number of passenger journeys begin each day, and each passenger spends an average amount of time in the air, then the average number simultaneously airborne is approximately:

people airborne ≈ passenger journeys per day × average airborne hours ÷ 24

Suppose global commercial aviation carries about 12 million passenger journeys per day on a representative day. Suppose average airborne time is 2.5 hours:

12m × 2.5 ÷ 24 ≈ 1.25m

Central estimate: about 1.3 million people airborne at a random moment. This is an application of the same intuition behind Little's Law: the number in a system is approximately the arrival rate multiplied by average time spent in the system.

Range: perhaps 500,000–2.5 million, depending on the day, season, what counts as a passenger journey, and whether private, military, and general aviation are included.

Why the second approach is better: daily passenger volume and flight duration are quantities for which people have at least some anchors. The original "fraction of humanity in the air" parameter has almost no intuitive base rate. Good decomposition is partly the art of replacing unanswerable assumptions with answerable ones.

Worked-example lesson: the strongest estimates are not always the ones with the cleverest arithmetic. They are the ones where the unknown has been decomposed into quantities you can actually reason about, then checked from a second direction.

Estimation vs Market Making

Estimathon questions and market-making questions are not the same exercise, but they are close relatives. An Estimathon asks you to place a range around an uncertain real-world quantity. A market-making problem asks you to quote a bid and an ask around an uncertain fair value, then potentially update those prices as you learn from trades or new information.

In both cases you need two things: a central estimate and an uncertainty model. If your estimate is 100 and you believe the value is tightly distributed around 100, you can justify a narrow interval or tight market. If the value could plausibly be anywhere from 40 to 250, quoting 99–101 is not confidence; it is a failure to represent uncertainty.

The cost of being too conservative is also similar. In an Estimathon, a range of [1, 1,000,000] may contain the true value but provides little information and receives a large width penalty. In a market-making game, quoting an unnecessarily wide bid-ask spread protects you from adverse selection but may mean you never trade. Good quantitative judgement lives between those extremes.

There is also an updating connection. Suppose you estimate an uncertain quantity at 100 with a broad interval. You then receive evidence that somebody with better information is willing to buy aggressively at 120. A trader asks whether that order flow should move fair value. An estimator asks whether the new information should move the centre or narrow the range. Both are Bayesian in spirit even when no formal Bayes formula is written down.

EverythingQuant's interactive market-making games are a natural way to practise this adjacent skill. They force you to turn a belief about value into a quoted range, experience the cost of being too confident, and update when the market gives you information.

Why This Matters for Jane Street and Quant Interviews

It is important not to overstate the direct connection. Jane Street publicly advertises Estimathons as events and has run them widely, but its public interview guidance does not describe the Estimathon as a standard interview round. You should therefore avoid preparing as if every Jane Street candidate will be handed 13 Fermi questions and asked to reproduce the competition format.

The underlying skills, however, are highly relevant. Jane Street describes quantitative trading interviews as collaborative problem solving rather than tests of finance or economics knowledge. Estimation exercises train exactly the kind of habits that are useful when a problem is underspecified and the interviewer cares about how you think.

Decision-making under uncertainty

Trading decisions are made before the true value is known. A trader must act on incomplete information, understand how sensitive the decision is to assumptions, and know when uncertainty is large enough to reduce risk. Estimation questions are a clean toy version of that problem.

Rapid quantitative reasoning

A trading environment rewards the ability to simplify. You often do not need a perfect model; you need a sufficiently good model quickly. Fermi problems force you to distinguish calculations that materially affect the answer from details that can safely be approximated.

Pricing uncertain quantities

A central Estimathon estimate resembles a rough fair value. The lower and upper bounds resemble a statement about valuation uncertainty. Market making goes one step further by attaching economic consequences to that uncertainty: if somebody trades against your quote, your model may be wrong or they may know something you do not.

Confidence calibration and risk management

One of the most dangerous quantitative mistakes is being precisely wrong. If your model is uncertain, position size and quoted width should reflect that. Estimathon scoring makes the trade-off visible: ranges that are too narrow miss; ranges that are too wide lose informativeness. Risk management is often the same tension expressed in financial terms.

Updating assumptions

In a strong interview conversation, your first model is not your final model. If the interviewer gives a new fact, you should be able to say which assumption it changes, how the estimate moves, and whether your confidence should increase or decrease. This is more valuable than defending your initial number at all costs.

Communicating the model

An estimate that only exists in your head is difficult to evaluate. State the decomposition, label assumptions, explain which ones matter, and narrate the sanity check. Clear communication lets another person audit your reasoning and collaborate with you. That is useful in an Estimathon team and in an interview.

How to Practise Estimation Effectively

Reading solved examples helps, but estimation only becomes intuitive when you repeatedly make predictions, commit to ranges, and compare them with reality. The feedback loop matters more than the number of clever examples you have seen.

1. Build an anchor sheet

Maintain a short page of quantities that recur across problems: world and major-city populations, seconds in a day and year, working days per year, rough household size, common transport capacities, dimensions of familiar objects, rough energy units, and common economic quantities. The goal is not to memorise trivia. It is to avoid rebuilding basic scale from zero every time.

2. Do timed Fermi questions

Give yourself three to five minutes per problem. Write the target, decomposition, assumptions, central estimate, and interval. Speed matters because a 30-minute team event rewards prioritisation. If you spend fifteen minutes perfecting one uncertain quantity, you may sacrifice easier problems whose missed-answer penalty is more costly.

3. Always submit a range in practice

Many people practise only the midpoint, which means they never train calibration. Before checking the answer, write down a lower and upper bound. After checking, record whether the truth was inside. Over 20 or 50 problems you will learn whether you are systematically overconfident or too conservative.

4. Keep a calibration log

For every miss, classify the failure. Was the central model wrong? Did you forget a unit conversion? Was one anchor off by an order of magnitude? Did you define the problem incorrectly? Was the midpoint fine but the interval unjustifiably narrow? Different errors require different fixes.

5. Practise second-route checks

After reaching an answer, spend another minute trying to attack it from the opposite direction. Demand versus supply, population versus capacity, per-person versus per-location, flow versus inventory. This habit catches structural mistakes and gives you evidence about whether a narrow interval is deserved.

6. Learn when not to refine

If your estimate is already dominated by a 4× uncertainty in one behavioural assumption, calculating the Earth's circumference more accurately will not help. Good Estimathon strategy is partly an optimisation problem: allocate time to the questions and assumptions where additional reasoning can actually improve the score.

7. Combine estimation with adjacent interview skills

Use mental maths to make the arithmetic automatic, probability questions to improve uncertainty reasoning, brainteasers to practise decomposition, and market-making games to turn uncertain fair values into decisions. The EverythingQuant Top 100 interview questions and general quant interview preparation are useful for integrating these skills into a broader interview routine.

Jane Street Estimathon FAQ

What is the Jane Street Estimathon?

The Estimathon is a team-based estimation contest associated with Jane Street. In the current public format, teams attempt 13 Fermi-style questions in 30 minutes and submit intervals that are intended to contain the true answer. Performance depends on covering the true value while keeping the interval reasonably tight. Jane Street's Estimathon site says the event debuted in 2013 and has since been run hundreds of times.

Does Jane Street ask Estimathon questions in interviews?

Jane Street does not publicly describe the Estimathon as a standard quantitative trading interview round. It publicly runs Estimathons as recruiting, educational, conference, university, and community events. That distinction matters. You should prepare estimation because the reasoning skills overlap with trading interviews, not because a 13-question Estimathon is guaranteed to appear in your interview.

How do you prepare for an Estimathon?

Practise Fermi questions under time pressure, build a library of useful anchor quantities, decompose each problem into smaller factors, do the arithmetic approximately, sanity-check from another direction, and always submit a lower and upper bound before seeing the answer. Then track whether your ranges are systematically too narrow or too wide.

What is Fermi estimation?

Fermi estimation is the process of approximating a difficult quantity by expressing it in terms of easier quantities. For example, instead of guessing the number of piano tuners directly, estimate the number of pianos, tunings per piano, and tunings one tuner can complete. It is especially useful when exact data is unavailable but the order of magnitude can be inferred from first principles.

How accurate does an Estimathon answer need to be?

You do not need to guess the exact value. Your submitted interval needs to contain the true value to be counted as good under the public format. Narrower successful intervals are rewarded more than unnecessarily wide ones, which is why calibration matters. The right range depends on how uncertain your model is; there is no universal "always use ±25%" rule.

How should I choose upper and lower bounds?

Start with the model inputs. If the estimate is A × B × C, ask for plausible low and high values of A, B, and C, especially for the assumptions that could move the answer by multiples rather than percentages. Recompute conservative and aggressive cases, then consider missing factors and correlations. This produces a much more defensible interval than adding an arbitrary percentage around the midpoint.

Are Estimathon questions the same as market-making questions?

No. Estimathon questions ask for ranges around uncertain quantities; market-making questions ask you to quote buy and sell prices around an uncertain value and often respond to trades. But the underlying thinking overlaps: estimate fair value, understand uncertainty, avoid overconfidence, and update when new information arrives.

What skills does Jane Street test with estimation questions?

Estimation exercises reveal decomposition, numerical intuition, use of base rates, order-of-magnitude reasoning, confidence calibration, prioritisation, and communication. These skills are consistent with the collaborative problem-solving approach Jane Street publicly describes for quantitative trading interviews, even though the Estimathon itself should not be assumed to be a standard interview stage.

Can I use a calculator in an Estimathon?

The current public rules say participants may not use the internet, a calculator or smartphone, or other external reference material. Scientific notation is allowed for submissions. Individual events can provide their own instructions, so follow the rules given by the organiser.

Should I always make my interval as wide as possible to avoid missing?

No. A very wide interval protects against a miss but increases the width component of the score and throws away information. The scoring system is designed to reward useful uncertainty estimates rather than vacuous ranges. The strategic objective is to make the interval as narrow as your model and confidence justify, not as narrow as you hope you can get away with and not as wide as the rules permit.

Conclusion

Jane Street Estimathon-style questions are useful because they expose a part of quantitative reasoning that conventional mathematics problems often hide. In a normal problem set, the relevant inputs are usually supplied and the task is to manipulate them correctly. In estimation, deciding which inputs matter is itself part of the problem.

The strongest approach is systematic: define the target, anchor what you know, decompose what you do not know, state assumptions, calculate approximately, sanity-check the scale, identify the dominant uncertainty, and build bounds from the inputs. Then be willing to update the model when new information arrives.

That is why Estimathon preparation transfers beyond the competition. The same habits appear in market sizing, expected-value reasoning, trading intuition, fair-value estimation, risk management, and interviews where the problem is deliberately unfamiliar. The objective is not to know every fact. It is to remain quantitative when the facts are incomplete.

Continue your Jane Street preparation:
Review the Jane Street preparation hub, work through Jane Street interview questions, read the Jane Street Trader Interview Guide, and practise uncertainty directly in the EverythingQuant market-making games.

Accuracy note: Jane Street publicly describes the Estimathon as a team event and publishes its rules at estimathon.com. Jane Street's interview guidance describes collaborative quantitative problem solving but does not state that the Estimathon is a standard interview round. The official Estimathon overview says the event debuted in 2013. Event formats and scoring conventions can change, so the organiser's current rules should always take precedence.