Statistics Interview Questions
Statistics questions in quant interviews cluster into six recurring families: hypothesis testing, maximum-likelihood estimation, estimator bias and variance, regression diagnostics, Bayesian updating, and time-series or risk applications of all of the above.
The common thread: every answer is an estimate plus an honest measure of how wrong it could be.
Last updated 2 July 2026 · sub-areas, difficulty mix and firm attributions on this page are compiled directly from the 185 statistics problems in the QuantVault bank; firm tags are candidate-reported, not employer-verified, and no individual author is named in our source data.
- Core sub-areas
- Hypothesis testing · maximum likelihood · estimator bias & variance · regression diagnostics · Bayesian updating · time-series & risk
- Typical difficulty
- Medium-heavy — 31 easy, 96 medium and 58 hard across the 185-problem set
- Who leans on it
- Research-flavored loops above all — problems here carry candidate tags from Two Sigma, Jane Street, Citadel, Renaissance, Barclays, HRT and others
- Practice pool
- 185 problems in the bank · 15 free to open with the full worked solution
Where statistics shows up: mostly in quant-research phone screens and data-analysis onsites, where “is this effect real?” doubles as a test of whether you can tell an edge from noise — the same reasoning that decides whether a backtest goes to production. For how it sits inside specific funnels, see the Two Sigma interview questions guide and the Citadel interview questions guide.
The surfaceThe statistics sub-areas quant interviews test
Six families cover essentially every statistics question in the bank; hypothesis testing alone touches roughly half the set, with maximum likelihood and estimator-quality questions close behind. Each row pairs a sub-area with the recurring question shape and a representative type — a flavor drawn from real problems, not a leaked wording.
| Sub-area | Recurring shape | Representative type (flavor, not a real question) |
|---|---|---|
| Hypothesis testing & p-values | Standardize an effect by its own standard error, then judge the size | Decide whether a strategy's year of daily returns shows a real edge — and what happens to that verdict after testing 100 candidate signals. |
| Estimation & maximum likelihood | Write the likelihood, take the log, differentiate, defend the answer | Estimate a coin bias or an arrival rate from a sample, then state the estimator's asymptotic variance via Fisher information. |
| Estimator bias & variance | Take the expectation of an estimator and hunt for the bias term | Explain a degrees-of-freedom correction, or argue when a slightly biased low-variance estimator beats an unbiased noisy one. |
| Regression & diagnostics | Identify which OLS assumption breaks and which test or fix applies | Detect heteroskedasticity in residuals, or reason through what clipping outliers does to a fitted slope. |
| Bayesian updating | Prior times likelihood, normalize, read the posterior | Update a belief about a biased die after a few rolls — or about a market after an informative non-event. |
| Time-series & risk applications | Adapt iid tools to autocorrelated, fat-tailed market data | Test returns for serial correlation, or estimate tail risk when the normal approximation quietly fails. |
What's confirmed vs. what varies: the sub-areas and the difficulty mix above come straight from the problem set, so they are stable. Which firm asks which flavor is candidate-reported through problem tags — treat the firm attributions as directional, and expect research tracks to push deeper into estimation theory while trading tracks favor the quick signal-versus-noise judgments.
The patternsSignature statistics question patterns
Three reasoning templates generate most correct answers in this topic. Each worked box below is a 60–90 second micro-example in the interview's actual cadence — applied to a generic setup rather than any firm's wording.
Signal versus noise — standardize before you judge
Takeaway: never judge an effect by its raw size — divide it by its own standard error first. A profitable year can be pure noise.
Shape. A strategy averaged +4 basis points per day over 252 trading days with a daily standard deviation of 80 bp. Is the edge real?
1. Price the noise. The standard error of the mean is \(80/\sqrt{252} \approx 5.0\) bp — the wobble of the average itself, not of a single day.
2. Standardize. The t-statistic is \(4/5.0 \approx 0.8\), nowhere near the conventional bar of about 2. A full profitable year is statistically indistinguishable from luck.
3. Say what would change it. Since \(t\) grows like \(\sqrt{n}\), you would need roughly six times the history — or a much larger edge — before the data forces the conclusion. The move: state the standard error before the verdict, assuming roughly independent days.
Maximum likelihood — a three-line routine, then a defense
Takeaway: MLE is a fixed cadence — likelihood, log, derivative — and the follow-up is always about the estimator's quality, not the algebra.
Shape. You observe \(k\) successes in \(n\) independent trials with unknown success probability \(p\). Estimate \(p\).
1. Write and log the likelihood. \(L(p)=p^k(1-p)^{n-k}\), so \(\ell(p)=k\log p+(n-k)\log(1-p)\).
2. Differentiate and solve. \(\ell'(p)=k/p-(n-k)/(1-p)=0\) gives \(\hat p = k/n\) — the sample frequency, as intuition demanded.
3. Defend it. \(\hat p\) is unbiased with variance \(p(1-p)/n\), which matches the Cramér–Rao bound from the Fisher information — no unbiased estimator does better. The move: finish with the quality argument unprompted; that is the part that scores.
Bayesian updating — the posterior is a compromise
Takeaway: with a conjugate prior the update is bookkeeping — add the counts — and the posterior mean visibly shrinks the raw frequency toward the prior.
Shape. Same coin as above: 7 heads in 10 flips, but now you start from a uniform prior on \(p\). What do you believe now?
1. Update by adding counts. A uniform prior is \(\text{Beta}(1,1)\); after 7 heads and 3 tails the posterior is \(\text{Beta}(8,4)\).
2. Read off the estimate. The posterior mean is \(8/12 = 2/3\) — pulled from the raw \(0.7\) toward the prior's \(0.5\), with the pull shrinking as data accumulates.
3. Contrast with MLE. Same data, different answer: the prior acts like two phantom flips, one of each kind — regularization in its simplest form. The move: name the shrinkage and when it matters (small samples), not just the number.
Free practiceStatistics practice questions by difficulty
All 15 free problems from the 185-problem statistics set — 4 easy, 6 medium, 5 hard — every link opens the full worked solution, so you can check your reasoning line by line. The free pool leans toward the applied end of the set; the remaining 170 problems behind the paywall fill out the same six sub-areas.
Easy — the core machinery
- MLE for a Bernoulli ParameterMaximum likelihoodeasyfree
- Correct Interpretation of a Frequentist Confidence IntervalEstimationeasyfree
- Fisher Information for Exponential SamplesMaximum likelihoodeasyfree
- ETF Tracking Error: Tail Probability and VaRNormal & riskeasyfree
Medium — the interview mainline
- Why We Divide by n-1 in Sample VarianceBias & variancemediumfree
- False Discovery in Alpha MiningHypothesis testingmediumfree
- Bayesian Posterior for a Biased DieBayesian updatingmediumfree
- Bayesian Update from a No-Trade EventBayesian updatingmediumfree
- Bayesian ML, Overfitting, and Error MetricsRegularizationmediumfree
- Online Mean and Variance for Streaming ReturnsStreaming estimationmediumfree
Hard — the separators
- Breusch-Pagan Heteroskedasticity TestRegression diagnosticshardfree
- Winsorization Impact on OLSRegression diagnosticshardfree
- Ljung-Box Test for Serial Correlation in ReturnsTime-series testinghardfree
- Estimation with Equicorrelated Normal SamplesCorrelated sampleshardfree
- VaR and ES via Peaks-Over-ThresholdTail riskhardfree
The planHow to prepare for statistics questions
Five techniques to make reflexive, in the order they pay off. Each maps onto a sub-area above, so you can drill it immediately after reading.
- Drill the MLE cadence until it takes two minutes. Likelihood, log, derivative, solve — then the defense: bias, variance, Fisher information, Cramér–Rao. Interviewers grade the defense at least as heavily as the algebra.
- Standardize everything by its standard error. A t-statistic is just signal over noise; knowing that the error of a mean shrinks like \(1/\sqrt{n}\) lets you sanity-check any “is this real?” question in your head before formal machinery.
- Hunt bias by taking expectations. When a question smells like a trick, compute the expectation of the estimator and watch for the term where the sample mean chases the data — degrees-of-freedom corrections and shrinkage arguments both fall out of this habit.
- Make conjugate Bayesian updates mental arithmetic. Beta–Bernoulli updates are just adding counts; practice reading the posterior mean as “raw frequency pulled toward the prior” and saying when the pull matters.
- Learn how market data breaks the iid textbook. Autocorrelation, heteroskedasticity and fat tails are the three standard complications the hard problems add; know one named test or fix for each. Then run the free practice set and take the interactive statistics playlist on a clock.
FAQStatistics interview questions — frequently asked
How important is statistics for quant interviews?
Core, not optional: alongside probability it is the backbone of quant-research and data-heavy trading loops, because estimating an edge and deciding whether it is real is the day job. Research-track interviews in particular can spend an entire round inside estimation and testing.
What statistics topics should I focus on?
Hypothesis testing and maximum likelihood dominate — together they touch most of the problems in this set. After those, drill estimator bias and variance, regression diagnostics, and quick Bayesian updates; time-series wrinkles like autocorrelation matter most for markets-facing roles.
How hard are statistics interview questions?
Medium-heavy: of the 185 problems in this set, 31 are easy, 96 are medium and 58 are hard. The hard tail usually adds a market complication — correlated samples, heteroskedastic noise, fat tails — on top of a standard technique rather than deeper pure math.
Are these real quant interview questions?
They are representative, not verbatim. The problems are curated from our bank to match the statistics shapes candidates report, rewritten for clarity with worked solutions we author ourselves — we never claim any wording is a leaked question.