Skip to main content
This is a mock interview for Quantitative Research roles. Each scenario has a main question, the follow-ups interviewers commonly ask, and answer points. The questions are open-ended and meant for practicing structured thinking. For basic formula questions, see the pages under Interview questions. Mock questions closer to trading desk practice (ETF market making, Delta-1, warrants) are on the Equities & Derivatives Mock Interview.
Status of this page. Last checked on 15 November 2026. These are general practice questions and answer frameworks, not any firm’s interview content. The numerical questions have been recomputed.

Research design

Answer structure: hypothesis, data, signal, validation, costs.
  • Hypothesis: If the deviation comes from a liquidity shock or a temporary supply-demand imbalance, it reverts within the arbitrage cost band. If it comes from a stale NAV, it is not an opportunity. First clarify the source of the deviation.
  • Data: ETF quotes and trades, synchronized prices for constituents or futures, FX, and cost data. Timestamps must be aligned and only information known at the time may be used (point-in-time).
  • Signal: The size of the deviation (the gap divided by its historical volatility or by estimated cost), how long it has persisted, and a volume condition.
  • Validation: Use event-time markouts after a deviation, check the speed of reversion (half-life), and check stability across time of day, names and market regimes.
  • Costs: Spread, fees, borrow and funding, and hedging slippage. The expected value after costs is the answer.
Follow-up: How do you confirm it is not a false signal from a stale NAV?→ Recompute the deviation using a more live fair value estimate (related futures, ADRs, FX). If the deviation disappears, the original signal was an information lag, not a mispricing. See ETF and Delta-1 Q1 and Q3.
First ask how it was obtained, then ask whether it will persist.
  • How many versions and parameters were tried? (multiple testing and data snooping)
  • Is there out-of-sample or walk-forward validation, and was the out-of-sample period ever used to tune parameters?
  • Does the data have survivorship bias, look-ahead bias or corporate action errors?
  • What transaction cost, impact and tradable volume assumptions were used? What is the capacity?
  • What does the return distribution look like: skew, maximum drawdown, concentration in a few days? Is there a short-tail-risk profile?
Numerical check: Annualized Sharpe 2.0 over 3 years gives t ≈ 2.0 × √3 ≈ 3.5, which looks very significant. But if it was the best of 100 candidates, the Bonferroni threshold is also about t = 3.5, so it sits right at the edge. You would need a deflated Sharpe ratio and truly unused out-of-sample data. See Regression and backtesting Q9.Follow-up: Would you recommend going live?→ Not directly. Recommend a small-size forward test with explicit stop losses and monitoring, and define in advance when you will conclude the signal has failed.
Treat each standardized signal as having correlation 0.03 with future returns. The equal-weight sum has covariance 0.03 plus 0.03 with returns and a standard deviation of the square root of (2 plus 2 times 0.6), which is about 1.79. So the combined IC is 0.06 divided by 1.79, or about 0.034: only a small gain, because the signals overlap heavily.Fundamental law link: the information ratio is roughly the IC times the square root of breadth (the number of independent bets per year). With an IC of 0.034 and 1,000 independent bets a year, that is about 0.034 times 31.6, or roughly 1.1, before costs.Follow-up 1: When would you drop one of them?→ Regress one signal on the other. If the residual has no predictive power, the second signal adds nothing but cost and complexity. Also drop the weaker one if it has higher turnover, so its net contribution after costs is negative.Follow-up 2: The IC of a signal halved over the last year. Is it decaying or just in a bad regime?→ Look at the IC over a longer history to see if similar dips recovered, split it by sector, size and volatility regime, check whether similar public signals weakened at the same time (crowding), and check capacity: if your own trading grew, you may be eating the edge. Decide in advance what evidence would make you retire it.
Approach: use related assets that are still trading to estimate where the underlying market would price at the open.
  • Find leading indicators: related index futures (if they trade close to 24 hours), ADRs, FX, and similar ETFs.
  • Estimate the relationship: regress the ETF or index return on these variables, check the residuals and estimate an error band.
  • Output not just a point estimate but also the uncertainty, and use it to set the quoted spread.
Follow-up: When does this relationship fail?→ Opening gaps, stock-specific news or events, changes in the futures-spot basis, ex-dividend dates, sharp FX moves, and differences between constituent weights and the futures’ underlying. Monitor with rolling windows and stability tests rather than estimating once.

Probability and market-making games

Step 1: Fair value. The expected sum of two dice is 7. The variance is 2 × 35/12 = 35/6 ≈ 5.83, so the standard deviation is about 2.42.Step 2: Quote. Center on 7 and leave a spread on each side, for example bid 6.5 and ask 7.5. The spread compensates for risk and possible adverse selection.Follow-up 1: If you could see one of the dice, what would your fair value become?→ If one die shows d, fair value is d + 3.5. That is an information advantage: a counterparty who sees one die more than you will trade against your quote whenever it is off.Follow-up 2: The counterparty always buys your ask or sells your bid. What do you conclude?→ The direction of the trade is itself information. If they trade only in specific situations, they may know something you do not, so widen the spread, reduce size or move your midpoint. This is adverse selection.Follow-up 3: How do you handle inventory once it builds up?→ Shift the quote midpoint against the inventory so it is easier to flatten. See ETF and Delta-1 Q6.
The most you would pay equals the expected number of flips times the payment per flip. The expected number of flips to see HH is 6 (see Probability Q2), so the fair price is 6.Follow-up: What if the game ends on HT instead?→ The expected number of flips is 4, so the fair price is 4. They differ because after a failed HH attempt you start over, while after a failed HT attempt you still hold a head.
The Kelly fraction for an even-money bet is the win probability minus the loss probability: 0.6 minus 0.4, or 20% of capital per bet. It maximizes the long-run growth rate of capital.Follow-up 1: Why do traders often bet half-Kelly?→ The edge is estimated, not known. Betting more than Kelly lowers growth and raises the chance of deep drawdowns, while betting half-Kelly keeps about three quarters of the growth rate with much lower volatility. Errors in the estimated edge are therefore much cheaper on the low side.Follow-up 2: How does this relate to position sizing on a desk?→ Size scales with edge divided by variance, and is capped by risk limits and liquidity. A strategy with a high Sharpe ratio but fat tails gets less size than its Kelly fraction suggests.

Data and systems

Design principles: Immutable raw data, re-runnable transforms, output partitioned by date, and every step verifiable.
  • Ingest: Keep the raw files and record the source and receive time.
  • Clean: Deduplicate, handle out-of-order records, filter invalid quotes, and flag outliers rather than silently fixing them.
  • Align: Use UTC, and use each exchange’s trading calendar to define sessions and closures.
  • Aggregate: Define bar boundaries and timestamps explicitly (marked at interval start or end). Decide whether minutes with no trades are left empty or carry the previous value, and flag them.
  • Validate: Reconcile against another source or daily data, and compare summed volume with total volume.
What can go wrong: Clock desynchronization, handling of auction and closing prices, late or replayed data, corporate actions, and ticker changes after exchange mergers or splits.Follow-up: How do you make it safe to backfill?→ Partition by date, overwrite whole partitions or use upserts, and make tasks depend only on explicit inputs. See Data engineering Q8.
Split P&L into explainable sources, so that only what does not reconcile is “unexplained”:
  • Delta: Position delta times the price move.
  • Gamma: (1/2) × Γ × (price move)².
  • Theta: Time value decay.
  • Vega: The change in implied volatility times vega.
  • Other: Dividends, interest rates, borrow cost, trading cost and slippage, basis.
  • Unexplained: Total P&L minus the above. A persistently large residual points to a model or data problem.
Follow-up: The unexplained part is persistently large. How do you investigate?→ Check the data first (price timing consistency, corporate actions, FX), then the Greeks calculation and the higher-order terms for large moves (a second-order approximation is not enough when the price moves a lot). Only then suspect the strategy itself.

Sources