Quant interview questions cover probability, expected value, statistics, brain teasers, mental math and coding, plus market-making games for trading roles and behavioral questions for every role. Interviewers grade how you reason out loud as much as the number you land on, so practice every day and talk through your work as you go.
I'm Fiona Taft, founder and CEO of QuantMinds, former Executive Director of UC Berkeley's MFE program and former Associate Director at Stanford University.
The practice questions below are classic, widely used problem types, each with a worked answer. Firm details come from each firm's own careers pages, checked October 2026.
How the quant interview process works
Most firms run some version of five stages, covered more fully in our guide on how to become a quant.
- Resume screen. A clean format that shows quickly who you are. See why your quant resume isn't getting responses.
- Online assessment. A timed technical test.
- HR interview. Sometimes before the assessment, sometimes as a recorded video.
- Technical interviews. Live problem solving and questions about your projects.
- Superday. A final round of several interviews.
IMC's US recruitment page, for example, lists an application, an assessment for some roles and a series of interviews. I'd plan on the interviews taking 4 to 8 weeks for an internship and 6 weeks to 3 months for a graduate role.
The assessment can cover anything from mental math to probability, statistics and coding, and the format and platform vary by firm. Optiver's Europe recruitment FAQ, for one, refers to its own Optiver Assessment Platform. Passing isn't enough. You need a high score to reach the HR interview, and if you don't get invited to the next round, you didn't pass.
You'll get a deadline, whether that's 3, 5, 7 or 30 days. Don't take it on the last day, because technical problems happen and their site might be down. Some firms won't let you retake one for months, so treat each as one shot. Apply to 5 to 15 roles at a time, not 50 in a day, since each can send back an assessment with its own clock. And practice a recorded video interview a few times first, because some don't let you try again. To find people at the firms, see how to use LinkedIn to get a quant job.
What changes by role
The topics overlap, but the weight shifts with the role. If you haven't picked one, read QR, QT or QD, which quant path is right for you.
- Quant researcher (QR). Heaviest on probability, statistics and modeling, and the best fit for math- and stats-heavy profiles. Jane Street's interviewing page says its research interviews are a mix of its trading and software engineering ones.
- Quant trader (QT). Speed under pressure and fast mental math. If you're applying to market-making firms, study poker and market making.
- Quant developer (QD). Coding first, and you'll need more languages than a researcher does. Optiver's interview tips for software engineers say its technical interviews skip standard LeetCode puzzles and brainteasers in favor of problem-solving discussions.
For the math bar by role, read the math you actually need to break into quant finance.
How to answer in the room
Technical questions are how interviewers judge your way of thinking. They're watching how you solve the problem, and whether you're open to feedback.
- Think out loud. An interviewer can only credit reasoning they hear.
- Ask clarifying questions, then state your assumptions.
- If the interviewer suggests another approach, try it instead of defending your path.
- If you're stuck, say where and ask for a hint.
- Learn methods, not answers. Interviewers change the numbers.
Probability questions
Two fair eight-sided dice total at least 13. What's the chance at least one shows an 8?
It's 7/10. Of the 64 equally likely ordered outcomes, totals of 13, 14, 15 and 16 come from 4, 3, 2 and 1 pairs, so 10 meet the condition. Seven contain an 8, namely (5,8), (8,5), (6,8), (8,6), (7,8), (8,7) and (8,8), and all 10 are still equally likely.
It tests whether you shrink the set of outcomes correctly and count ordered pairs.
Two percent of strategy ideas have a real edge. A backtest passes 90% of those and 10% of the rest. If a strategy passes, what's the chance it's real?
It's 9/58, about 15.5%. Picture 1,000 ideas. Twenty have an edge and 18 of those pass, while 98 of the 980 without one pass too. So 18 of the 116 passes are real, and 18/116 = 9/58. When real edges are rare, even an accurate-sounding filter mostly lets through false positives.
It tests Bayes' rule, and whether you respect the base rate.
Starting with $3, you play independent rounds, each winning $1 with probability 0.6 or losing $1. You stop at $6 or $0. What's the chance you reach $6?
It's 27/35, about 77.1%. Let P(k) be the chance of reaching $6 from $k. Conditioning on the next round gives P(k) = 0.6 P(k+1) + 0.4 P(k-1), with P(0) = 0 and P(6) = 1, which is solved by P(k) = (1 - rk) / (1 - r⁶) with r = 0.4/0.6 = 2/3. Since 1 - r⁶ = (1 - r³)(1 + r³), P(3) = 1 / (1 + r³) = 1 / (1 + 8/27) = 27/35.
It tests setting up a recursion on the first step and spotting a shortcut.
Expected value questions
You roll a fair die repeatedly, adding each result, and stop at the first 6, which isn't added. What's the expected total?
It's 15. Let E be the expected total. With probability 1/6 the first roll is a 6 and you finish with 0. Otherwise you roll 1 to 5, averaging 3, and the game starts over, so E = (5/6)(3 + E), which gives E/6 = 5/2 and E = 15. Or, more quickly, you expect 5 non-six rolls before the first 6, each averaging 3.
It tests a first-step equation, and noticing that a roll known not to be a 6 averages 3, not 3.5.
Five orders are each sent independently and uniformly at random to one of 10 venues. How many venues do you expect to get at least one?
It's 4.0951. Write the count as a sum of 10 indicators, one per venue, each equal to 1 if that venue gets an order. A venue gets none with probability 0.9⁵ = 0.59049, so each is used with probability 0.40951. Expectations add even when the indicators depend on each other, so the answer is 10 × 0.40951 = 4.0951.
It tests linearity of expectation, which saves you an ugly case-by-case count.
A fair eight-sided die pays its face value in dollars, and you can pay $1 to reroll, up to three rolls in total. What's the game worth?
It's 169/32, about $5.28. Work backwards. The third roll is worth its average, 4.5. On the second roll, rerolling is worth 4.5 - 1 = 3.5, so keep 4 or higher, which makes the second roll worth (5/8) × 6 + (3/8) × 3.5 = 5.0625. On the first roll, rerolling is worth 5.0625 - 1 = 4.0625, so keep 5 or higher, and the game is worth 0.5 × 6.5 + 0.5 × 4.0625 = 5.28125, or 169/32. Without the fee you'd hold out for 6 and then 5, so the fee makes you less picky.
It tests backward induction, comparing each roll with the value of continuing after costs.
Statistics questions
A strategy truly averages 0.05% a day with a 1% daily standard deviation. Roughly how many independent days until its expected t-statistic is 2?
About 1,600 days, a little over six years at 252 trading days a year. The standard error of the average after n days is 1% / √n, so you need 0.05% = 2 × 1% / √n, which gives √n = 40 and n = 1,600. The average you measure is itself random, so 1,600 days clears t = 2 only about half the time.
It tests how standard errors shrink, and how noisy daily returns are.
On one data set, least squares regressions with intercepts give slopes of 0.8 for y on x and 0.45 for x on y. What's the correlation, and which varies more?
The correlation is 0.6, and y varies more, with sd(y)/sd(x) = 4/3. The slope of y on x is the correlation times sd(y)/sd(x), and the slope of x on y is the correlation times sd(x)/sd(y). Their product is the correlation squared, 0.8 × 0.45 = 0.36, so the correlation is 0.6, positive because both slopes are. Then sd(y)/sd(x) = 0.8 / 0.6 = 4/3. The slopes are reciprocals only when the correlation is exactly 1 or -1.
It tests what a slope is made of, and that flipping a regression doesn't invert it.
Brain teasers
One of 25 identical-looking coins is slightly heavier. What's the fewest balance-scale weighings that guarantees you find it?
Three. Each weighing has three outcomes (left heavier, right heavier, balanced), so two can tell apart at most 3 × 3 = 9 possibilities, fewer than 25. To do it in three, weigh 9 against 9 with 7 off, which leaves 9 or 7 suspects. Next, weigh 3 against 3 from those. That leaves 3 suspects, unless 3 of the 7 balance against 3, in which case the coin left off is the heavy one. Finally, weigh 1 against 1.
It tests pairing an information-count lower bound with a method that meets it.
The numbers 1 to 10 are on a board. A move replaces any two with their absolute difference. Can the last number be 0, and how big can it be?
It can't be 0, and the most it can be is 9. For any a and b, a + b and |a - b| are both even or both odd, so a move never changes whether the board's total is odd. The total starts at 55, so the last number is odd and can't be 0. A move never creates a number bigger than the largest on the board, so the last number is at most 10, and since it's odd, at most 9.
To reach 9, set the 9 aside. Turn 10 and 8 into 2, then 3 and 2, 5 and 4, and 7 and 6 into three 1s. That leaves 9, a 2 and four 1s. Turn the 2 and a 1 into 1, pair the four 1s into two 0s, combine the 0s, then combine 9 with 0.
It tests finding an invariant (a property no move can change), then backing a bound with a construction.
Mental math questions
What's 68 × 72, in your head?
It's 4,896. Both numbers sit 2 either side of 70, so the product is 70² - 2² = 4,900 - 4 = 4,896, using (a - b)(a + b) = a² - b².
It tests number sense, spotting structure instead of doing long multiplication.
At 7% a year, compounded annually, roughly how long does money take to double?
About 10.2 years. The rule of 72 gives 72 / 7, about 10.3, as a first guess. More precisely, doubling means 1.07t = 2, so t = ln 2 / ln 1.07, which is about 0.693 / 0.0677, or 10.24. As a check, 1.07¹⁰ is about 1.967 and 1.07¹¹ is about 2.105.
It tests compounding, logarithms and a quick approximation you can check.
Market making and betting questions
A fair die is rolled and you're paid the square of the result in dollars. What's the fair value, and what market would you quote?
The fair value is 91/6, about $15.17, because the squares 1, 4, 9, 16, 25 and 36 add to 91. A sensible market brackets that with a small edge on each side, say 14.75 bid (the price you'd buy at) and 15.60 ask (the price you'd sell at). The payout's standard deviation is about $12.2 a roll, so a few tens of cents of edge is small next to the risk, which argues for modest size or a wider quote. The fair value isn't 3.5² = 12.25, because the average of the squares is bigger than the square of the average.
It tests a fast expectation, a well-centered market and a sense of risk.
You make a market on the total of three fair dice. Where do you center with no information, if the first die is a 6, and if at least one die is a 6?
At 10.5, 13 and 1143/91, about 12.56. Each die averages 3.5, so three average 10.5. If the first die is a 6, the other two still average 7, giving 13.
For the third case, outcomes with no 6 have each die uniform on 1 to 5, so they average 9 in total, with probability (5/6)³ = 125/216. The overall average is still 10.5, so 10.5 = (91/216) × E + (125/216) × 9, which gives E = 1143/91. The third statement moves you less than the second because 75 of the 91 outcomes with a 6 have exactly one, and there the other two dice average 3 each, not 3.5.
It tests updating a fair value on information that only sounds similar.
You can repeatedly bet any fraction of your bankroll at even money with a 55% chance to win. What fraction grows it fastest, and what happens at 25%?
Bet 10%. At 25% your bankroll shrinks in the long run, even though every bet has positive expected value. Betting a fraction f multiplies your wealth by 1 + f on a win and 1 - f on a loss, so long-run growth per bet is g(f) = 0.55 ln(1 + f) + 0.45 ln(1 - f). Setting the derivative to zero, 0.55/(1 + f) = 0.45/(1 - f), gives f = 0.55 - 0.45 = 0.10, the Kelly fraction. Then g(0.10) is about +0.50% per bet and g(0.25) is about -0.67%. Losses compound harder than gains. After a 25% loss you need a 33% gain to get back to even.
It tests whether you separate expected value from compounding.
Coding questions
Given daily profit and loss (P&L) figures and a target, how do you check whether two different days sum to the target faster than checking every pair?
Make one pass with a hash set (a set with fast lookups). For each value x, return true if target - x is already in the set, and otherwise add x. That's O(n) time and O(n) memory. Checking before inserting stops a day pairing with itself and still handles repeats, such as two days of 50 with a target of 100. If memory is tight, sort in place and walk two pointers inward from both ends, which takes O(n log n) time and no extra memory.
It tests trading memory for speed and catching edge cases unprompted.
Trades stream in and you don't know how many are coming. In one pass with constant memory, how do you pick one so every trade is equally likely?
Keep a count and the first trade, and when the k-th arrives, swap it in with probability 1/k. That's reservoir sampling. Trade i is picked on arrival with probability 1/i and survives each later arrival j with probability (j - 1)/j, so its chance of being the final pick is (1/i) × (i/(i+1)) × ... × ((n-1)/n) = 1/n, the same for every trade. To keep k trades, the m-th arrival (for m greater than k) enters with probability k/m and replaces a randomly chosen kept trade.
It tests designing under a memory limit, and proving a randomized method is fair.
Given positive prices in time order, how do you find the maximum drawdown, the largest percentage fall from a peak to a later low, in one pass?
Track the running peak and the worst drawdown so far. At each price p, update the peak, compute (peak - p) / peak and keep the largest value. That's O(n) time and constant memory, and it's correct because the fall to any price is largest from the highest earlier price, which is exactly the running peak. Don't compare the overall high with the overall low, because a low that comes before the high isn't a drawdown.
It tests replacing an all-pairs comparison with the right running state.
Questions for research and bank roles
Interviews for quant research and for bank quant teams often add derivatives pricing, linear algebra, C++ and random processes.
A $100 stock will be worth $120 or $80 in a year, and interest rates are zero. What's a one-year call option with a $100 strike worth?
It's $10. The call pays $20 if the stock goes up and nothing if it goes down. Copy that payoff with Δ shares and B dollars, so 120Δ + B = 20 and 80Δ + B = 0. That gives Δ = 0.5 and B = -40, meaning you borrow $40, and the portfolio costs 0.5 × 100 - 40 = $10 today. Any other price is a riskless profit for someone. Equivalently, the risk-neutral up probability solves 120q + 80(1 - q) = 100, so q = 0.5 and the option is worth 0.5 × 20 = $10. The real chance of an up move never appears, because the option can be hedged with the stock.
It tests pricing by replication, and why real-world probabilities drop out.
Three assets have correlations of 0.9 between A and B, 0.9 between B and C, and 0.2 between A and C. Is that possible?
No. A correlation matrix must be positive semidefinite, meaning no portfolio of the assets can have negative variance, or equivalently no eigenvalue is negative. For three assets with correlations a, b and c, the determinant is 1 + 2abc - a² - b² - c², which here is 1 + 2 × 0.9 × 0.9 × 0.2 - 0.81 - 0.81 - 0.04 = -0.336. The determinant is the product of the eigenvalues, so at least one is negative (about -0.18). If A tracks B and B tracks C that closely, A and C can't be nearly unrelated. Their correlation has to be at least 0.9 × 0.9 - √((1 - 0.81)(1 - 0.81)) = 0.81 - 0.19 = 0.62.
It tests tying an abstract matrix property to real risk, and checking it by hand.
In C++, why does a base class with virtual functions need a virtual destructor?
Because deleting a derived object through a base-class pointer is undefined behavior when the base destructor isn't virtual. Create a Derived with new, hold it in a Base pointer and delete it, and on a typical compiler only Base's destructor runs, so whatever Derived owns, such as memory or a file handle, leaks. Declare the destructor virtual in Base, for example virtual ~Base() = default, and the delete runs Derived's destructor and then Base's. A unique_ptr to Base doesn't save you, because it still deletes through the Base pointer.
It tests object lifetime, virtual dispatch and who owns a resource in C++.
What's the distribution of the integral of a standard Brownian motion W from 0 to T?
It's normal with mean 0 and variance T³/3. It's normal because it's a limit of sums of jointly normal values, and its mean is 0 because W has mean 0 at every time. Its variance is the integral of the covariance E[W(s)W(t)] = min(s, t) over s and t from 0 to T, which by symmetry is 2 × ∫ t²/2 dt from 0 to T = T³/3.
It tests whether you know Brownian motion's covariance and can use it.
Behavioral questions
Every role gets behavioral questions, and they're the gateway to the technical rounds. Firms want to see that you can communicate and work on a team, and very analytical people often find that the hardest part. Expect your resume, your projects, why this firm and how you keep up with markets.
Tell me about a time an analysis or model of yours turned out to be wrong.
Pick a real case with numbers, not a strength in disguise. Say what you believed and why, how the error came to light (ideally because you checked your own work), and what it cost or could have cost. Spend most of your time on what you changed afterward, such as a new test or check, and show you've used it since. Own it without blaming a teammate or the data.
It tests self-awareness, and whether you build safeguards after a mistake.
How to prepare
Practice every day, and start before your first assessment arrives.
- Mental math, 30 to 60 minutes a day. Prop trading firms like quick math and brain teasers because they show how your mind works.
- Probability, statistics and brain teasers, a few problems a day, solved out loud.
- Coding on LeetCode and HackerRank, which lets you sign up free. See our guide to Python for quant roles.
- Behavioral stories, said out loud to a friend, a family member or your roommate. Feeling confident about a story you've never said out loud isn't the same as being ready.
- For trading roles, poker, market-making games and trading competitions.
For books, the Green Book (A Practical Guide to Quantitative Finance Interviews, by Xinfeng Zhou) is the top resource in my view, and our guide to the Green Book covers how to use it. Heard on the Street, by Timothy Falcon Crack, is another well-known collection. Jane Street says its Probability and Markets guide covers many of the core concepts in its interviews.
Frequently asked questions
Do I need to know finance for a quant interview?
Often not for entry-level trading roles. Jane Street says it won't test quant trading candidates on finance or economics, and IMC asks interns and graduates for an interest in financial markets and a strong analytical or technical background rather than industry experience. Research and bank interviews can still ask pricing questions like the option example above.
Can I retake a quant online assessment?
Don't count on it. Optiver's Europe recruitment FAQ says its assessments can be completed only once every 8 months, with no resets or retakes in that window once you've started one. So don't open an assessment until you're ready to finish it.
Which programming language should I use in a quant interview?
Use the one you know best. Jane Street says you can interview in whichever language you know best, and Optiver looks for confident knowledge of C++, Java or Python in software engineering interviews. For research roles, Python with libraries like NumPy is the baseline.
Get a second opinion on your prep
If you'd like another set of eyes on your interview prep, request a free Discovery Session, a Zoom session with one of our coaches, and we'll work through it together.