10 quant researcher interview questions — and how to actually think about them
Most candidates prepare for quant interviews by collecting puzzle answers. That helps less than people think. The better preparation is learning how to choose a representation, state assumptions, test edge cases, and explain uncertainty while someone is watching your reasoning.
The questions below cover probability, statistics, coding, linear algebra, signal research, market reasoning, and technical behavioral discussion. They are not meant as a leaked question bank. They are a compact map of the habits interviewers are usually trying to observe.
Core signal
Probability states
Core signal
Statistical inference
Core signal
Research judgment
Before the questions: what strong candidates do differently
Narrate assumptions before equations. Interviewers cannot evaluate reasoning they cannot hear.
Use small state spaces and toy examples. Most hard problems become manageable after the right representation.
Separate estimation, inference, and decision-making. A number can be correct and still not justify a trade.
Check edge cases out loud. Zero variance, tiny samples, overlapping returns, and transaction costs are not details.
Practice recovery. A corrected wrong path is often more informative than a memorized perfect answer.
Interview pattern
1. Coin flips until two heads in a row
Prompt
You flip a fair coin until you see two heads consecutively. What is the expected number of flips?
What it tests: This tests whether you can build a state space instead of trying to force a memorized geometric-distribution formula onto a process that has memory.
How to think
- Define states by the current suffix, not by the full history: state 0 means the last flip was not a head streak, state 1 means the last flip was H, and done means HH has appeared.
- Let E0 be the remaining expected flips from state 0 and E1 from state 1. Then E0 = 1 + 0.5E1 + 0.5E0 and E1 = 1 + 0.5·0 + 0.5E0.
- Solving gives E0 = 6. The number matters less than the modeling habit: name the state, write the recurrences, then solve.
Watch out
Do not say E = 1 / P(HH) = 4. Overlapping patterns make that shortcut wrong here.
Practice drill
Repeat the same setup for HTH, HHH, and two tails in a row. The recurrences should change in a controlled way.
Interview pattern
2. Two players roll for the first six
Prompt
Players A and B alternate rolling a fair die. A goes first. The first player to roll a 6 wins. What is the probability A wins?
What it tests: Interviewers like this because it is simple enough to solve live, but it reveals whether you notice turn order, conditioning, and infinite series.
How to think
- A wins immediately with probability 1/6.
- If both players miss in a round, probability (5/6)², the game restarts with A to move.
- So p = 1/6 + (5/6)²p, which gives p = 6/11. You can also derive this as a geometric series over complete failed rounds.
Watch out
A common weak answer is 1/2 because the die is fair. The die is fair; the turn order is not.
Practice drill
Change the target event to rolling an even number, or let B win on a 5 or 6. Write the recurrence before doing arithmetic.
Interview pattern
3. Expected time to collect all six die faces
Prompt
A fair die is rolled repeatedly. What is the expected number of rolls until all six values have appeared?
What it tests: This is a standard coupon collector question, but a strong answer explains why the decomposition works rather than only naming the result.
How to think
- Break the process into stages: after seeing k distinct faces, the probability the next roll is new is (6 − k) / 6.
- The waiting time for the next new face has expectation 6 / (6 − k). Sum k = 0 through 5.
- The expected value is 6(1 + 1/2 + 1/3 + 1/4 + 1/5 + 1/6), about 14.7 rolls.
Watch out
Do not simulate mentally. The point is to identify independent waiting-time stages, even though the raw rolls are not grouped in advance.
Practice drill
Generalize to n coupons, then estimate the asymptotic behavior as n log n plus lower-order terms.
Interview pattern
4. A backtest has Sharpe 1.5 over five years
Prompt
A strategy has daily returns over five years and an annualized Sharpe ratio of 1.5. Is it good?
What it tests: This separates candidates who know a metric from candidates who know how research decisions are made under uncertainty.
How to think
- Ask what was selected. Was this one idea tested once, or the best result among thousands of variants? Multiple testing changes the interpretation completely.
- Inspect stability: subperiod Sharpe, drawdowns, turnover, capacity, costs, borrow, slippage, and regime dependence matter more than the headline number.
- Estimate uncertainty. Daily returns are noisy, often autocorrelated, and not normally distributed. Standard errors based on independent Gaussian returns can be misleading.
- Require out-of-sample evidence or a clean walk-forward design before treating the result as tradable.
Watch out
Do not answer with only yes or no. In a research interview, the defensible answer is a checklist of failure modes.
Practice drill
Take any published equity factor and write down five ways its backtest could overstate live performance.
Interview pattern
5. OLS residuals are autocorrelated
Prompt
You run a linear regression for a signal model. The residuals are autocorrelated. What breaks?
What it tests: This checks whether you understand the difference between coefficient estimation, uncertainty estimation, and prediction quality.
How to think
- Under the usual exogeneity assumptions, OLS coefficients can remain unbiased even when residuals are autocorrelated.
- The standard errors are generally wrong, so t-statistics and confidence intervals can be overconfident.
- For time-series signals, autocorrelation may also indicate a missing state variable, stale prices, seasonality, or overlapping returns.
- A practical answer mentions Newey-West-style corrections, blocked cross-validation, and checking whether the signal survives after fixing the dependence structure.
Watch out
Do not just say the model is invalid. Be precise: which part of inference fails, and which assumptions would rescue it?
Practice drill
Explain the same issue for heteroskedastic residuals, then for correlated features in a cross-sectional model.
Interview pattern
6. Implement rolling variance on a stream
Prompt
Write a function that maintains rolling mean and variance over the last N observations in a data stream.
What it tests: Quant coding rounds often use small problems to test numerical habits. The interviewer is not only looking for a working loop.
How to think
- Start with the simple data structure: a queue or ring buffer, plus running sum and sum of squares for O(1) updates.
- Then discuss numerical stability. Variance from E[X²] − E[X]² can lose precision when values are large and variance is small.
- For expanding windows, Welford updates are natural. For fixed rolling windows, removing an old observation makes the stable update more delicate, so you may choose periodic recomputation or a more careful online algorithm depending on latency constraints.
- State edge cases: fewer than N observations, N = 1, NaNs, timestamp gaps, and sample versus population variance.
Watch out
A solution that is asymptotically fast but silently produces negative variance due to floating-point cancellation is not production-quality research code.
Practice drill
Implement the simple version, then generate values around 1e9 with tiny noise and compare numerical error against a two-pass recomputation.
Interview pattern
7. First principal component of equity returns
Prompt
You run PCA on a matrix of standardized equity returns. What do you expect the first principal component to represent?
What it tests: This tests linear algebra plus market intuition. Good candidates connect eigenvectors to risk factors without overclaiming.
How to think
- The first component often resembles a broad market mode because many equities co-move with the market.
- The loading signs and magnitudes can be affected by universe construction, standardization, sector concentration, and the return horizon.
- A useful answer explains variance maximization, eigenvectors of the covariance or correlation matrix, and why PCA factors are statistical objects rather than guaranteed economic causes.
- Then ask how stable the component is over time and whether it improves risk control out of sample.
Watch out
Do not claim PCA discovers the true factors. It discovers directions of historical variance under your preprocessing choices.
Practice drill
Compare PCA on raw returns, demeaned returns, and volatility-standardized returns. Explain how each preprocessing choice changes the first component.
Interview pattern
8. Bayes rule and base rates
Prompt
A classifier flags 1% of trades as suspicious. It catches 90% of truly bad trades and falsely flags 2% of normal trades. If a trade is flagged, how likely is it truly bad?
What it tests: Base-rate questions appear in many forms: fraud, anomaly detection, adverse selection, rare event prediction, and signal evaluation.
How to think
- Make the base rate explicit. If 1% of trades are truly bad, then in 10,000 trades there are 100 bad and 9,900 normal.
- The classifier catches 90 bad trades and falsely flags 198 normal trades.
- So P(bad | flagged) = 90 / (90 + 198), about 31.25%.
- Then discuss whether that precision is useful: it depends on the cost of investigation, false positives, and missed bad trades.
Watch out
Many candidates answer 90% by confusing sensitivity with posterior probability.
Practice drill
Redo the calculation with base rates of 0.1%, 1%, and 10%. Notice how much the posterior changes.
Interview pattern
9. A signal worked yesterday and stopped today
Prompt
A signal that looked strong in backtests and recent paper trading stops working after launch. What do you check first?
What it tests: This is a research-debugging question. It rewards structured thinking more than cleverness.
How to think
- Start with data integrity: timestamp alignment, corporate actions, survivorship, stale inputs, missing values, and vendor changes.
- Then execution assumptions: costs, queue position, fill model, latency, borrow, market impact, and whether live orders changed the opportunity.
- Then research validity: overfit parameters, multiple testing, regime change, capacity, and whether the signal was a proxy for another exposure.
- Finally, compare live versus backtest feature distributions. If the inputs are different, the model is not being asked the same question.
Watch out
Do not start by tuning parameters. First determine whether the live system matches the researched system.
Practice drill
Build a two-column checklist: things that would make the backtest wrong versus things that would make the market opportunity decay.
Interview pattern
10. Explain a wrong answer you gave
Prompt
Tell me about a technical problem where your first solution was wrong. How did you find the mistake?
What it tests: Behavioral questions in quant interviews are often technical in disguise. They test calibration, honesty, and debugging process.
How to think
- Pick a real example with enough technical detail to be credible, but not so much that the story becomes hard to follow.
- Explain the wrong assumption, the evidence that contradicted it, and the specific test or calculation that changed your mind.
- End with what you changed in your process: independent checks, simpler baselines, dimensional analysis, code review, or adversarial test cases.
Watch out
Avoid a fake weakness story. Quant teams care whether you can update under evidence, not whether you can perform confidence.
Practice drill
Write three stories from coursework, research, trading projects, or programming bugs. For each, name the assumption that failed.
A compact prep plan
Spend less time memorizing finished answers and more time doing timed explanations. Take one probability problem, one statistics problem, and one coding problem each day. For each one, record yourself stating the assumptions, solving, checking edge cases, and summarizing what would change if the setup changed.
For research roles, add a weekly backtest critique. Pick a factor, strategy write-up, or small project and ask: where could the data be wrong, where could the inference be overconfident, what costs were omitted, and what evidence would make this believable out of sample?
The target is not to sound polished. The target is to be technically legible. A good interviewer can work with a candidate who thinks clearly, catches mistakes, and updates. They cannot work with an answer that jumps from prompt to final number with no model in between.
Next step
If you want a sharper prep loop
Ask a Quant has deeper write-ups and 1:1 sessions with a former Jump Trading quant researcher (École Polytechnique + Stanford ICME, 3× IMO medalist). The emphasis is practical: interview reasoning, research judgment, resume positioning, and what to prepare next.