Citadel Data Analyst (Engineering) Case Interview — Market Data Signal Vetting

This case simulates an on-desk analytics problem at Citadel: quickly assessing whether a noisy market signal is real, scalable, and worth capital. Modeled on common candidate experiences, it blends data hygiene, statistical rigor, coding fluency, and business judgment under time pressure. What you’ll work on: - Data quality triage and microstructure awareness: identify bad prints, outliers, time-zone mismatches, stale quotes, and corporate actions; articulate survivorship-bias and look-ahead risks; build a reproducible cleaning pipeline. - SQL proficiency: window functions, event-time joins, deduplication, sessionization, and computing event-study returns across large equity universes. - Python/statistics: vectorized pandas workflows, time-series resampling, feature engineering, cross-validation that respects temporal order, hypothesis testing and false-discovery control, back-of-the-envelope PnL with realistic frictions (slippage, fees) and capacity limits. - Risk/portfolio thinking: translate signal metrics into capital allocation language (hit rate, expectancy, Sharpe, drawdown), scenario/sensitivity analysis, and clear kill-criteria when evidence is weak. - Communication under scrutiny: crisp reasoning, defend assumptions, surface trade-offs; expect PM-style pushback aimed at stress-testing your conclusions. Format (typical flow): - 5 min problem brief and clarifying questions. - 35–40 min hands-on analysis in a shared environment (notebook/CoderPad/SQL console). Internet browsing is disabled; basic docs/help may be available. - 15–20 min readout: 2–3 plots/tables, a concise narrative, and a go/no-go recommendation with risk and implementation notes. - 5 min targeted Q&A and “what would you do next” discussion. Representative prompt: “You’re given two CSVs: trades_quotes_intraday.csv (top-of-book aggregates for a liquid large-cap basket, Q1–Q2 2024) and econ_calendar.csv (macro releases with timestamps and surprise magnitudes). Investigate whether payroll-surprise or another feature you engineer predicts next-day open-to-close returns after accounting for transaction costs. Show your data checks, method, and a recommendation suitable for a quick PM decision.” What Citadel screens for: - Speed with precision: bias to executable analysis over exhaustive perfection; time-boxing and prioritization. - Evidence over assertion: quantify uncertainty, avoid p-hacking, and separate in-sample patterns from deployable edges. - Engineering mindset: clean, readable, vectorized code; correct joins and indexing; reproducible outputs. - Commercial impact: clear link from metrics to capital deployment, risk controls, and operational feasibility. Deliverables: a cleaned dataset or reproducible steps, a few compact visuals/tables, and a one-paragraph decision with risks, assumptions, and next steps.