Company interview guide

Citadel Machine Learning Interview Guide (2026)

For ML, AI and research roles: the firm's quant research loop, heavy on regression, ML theory on noisy data, probability derivations and classic coding.

Citadel's machine learning interview loop has 9 reported stages, from recruiter call to team-fit and final review; typical length: 4-8 weeks.

Roles covered: Research Scientist, ML Engineer, Data Scientist and Applied Scientist. The question bank holds 100 questions reported from Citadel interviews, 60 of them free to solve in your browser.

Try a free question

Interview process length
4-8 weeks
Stages
9
Reports linked to bank questions
100
Bank report dates
Jun 2015–Jul 2026
Questions in bank
100 (60 free)
Last reviewed
October 2026

How the Citadel interview loop works

A recruiter call, a track-dependent online test (engineering and data-science tracks; research candidates usually go straight to live rounds), one 45-60 minute remote CoderPad round (some candidates had two or three), then a Super Day of three to five interviews of 45-60 minutes on probability, coding, regression and ML, with behavioral questions in each. Hiring managers then review your feedback and an interested team holds a final fit interview; if several teams want you, you choose with your recruiter. The firm states about four to five weeks for research roles and about eight for engineering. The hedge fund and its market-making affiliate run separate but similar processes. Market-making simulations, the short probability-speed test and mental-math drills are trader-track only and are left out here.

The Citadel interview process, stage by stage

01Recruiter call

Short phone call, reported at 15-30 min

A mostly logistical conversation about the role, location, level and start date. The firm usually considers you for one role at a time and says it will share your resume with other teams only when there is a strong match, so be clear about which track you want.

What it tests

  • Why the firm and why this track (ML research, ML engineering or data science)
  • Location, level and start-date fit
  • Competing deadlines and other processes

How to prepare

  • Mention competing offers early; the firm says it will try to work around your deadlines.
  • Ask whether your track includes an online test, how many technical rounds to expect and which business area the role sits in.

One 2025 AI-researcher report starts with a recruiter screen; a public guide relays a second ML engineer whose call was logistics only. The firm's experienced-engineering hiring page lists a recruiter conversation first.

02Online assessment (track-dependent)

Auto-graded HackerRank: engineering about 70 min for 2 problems; data science about 60 min, ~18 multiple-choice; research usually none

Engineering candidates (software, quant-dev and many ML engineers) get two problems drawn from a rotating pool of roughly 40 story-style algorithm questions, mostly medium; every hidden test must pass and complexity matters. Some accounts describe up to 90 minutes with a short probability and statistics multiple-choice section added. The data-science datathon is about 18 multiple-choice questions in 60 minutes mixing pandas and R (recent versions test both), statistics, probability and ML theory. Research candidates are usually taken straight to the live round. The problems listed were tagged to engineering roles.

What it tests

  • Classifying a story problem into its pattern fast: binary search on the answer, sliding window, two pointers, BFS
  • Counting with hash maps, prefix sums and modular buckets
  • Dynamic programming: coin change, house-robber style and two-state DP
  • Graph basics: BFS on a grid, tree diameter, union-find
  • Data-science variant: pandas filters, CLT and normal approximation, Bayes, regularization and model-choice questions

How to prepare

  • With two of about 40 problems drawn, cover pattern families rather than memorising rumoured questions.
  • Read both problems first, bank the easier one, and check constraints for an O(n log n) target before coding.
  • For the datathon, practise both pandas and R syntax plus quick CLT and Bayes arithmetic; questions on explainable model choice and correlation versus causation have appeared.

Free reported questions for this stage

5 more free questions for this stage are in the app.

Format and pools from compiled accounts of the firm's engineering, data-science and quant tests; no first-hand ML-role report in our set, and the firm's pages do not mention a test.

03First round (remote technical)

45-60 min video, CoderPad link sent beforehand; one round per the firm, two or three for some candidates

One interviewer opens with your background and why the firm, then gives a medium coding problem in Python or C++ and, on research tracks, a probability or statistics question; conditional expectation is a favourite. ML candidates also get questions tied to the resume: one 2025 ML engineer spent about 20 minutes on the resume and was then asked to make a hard routing decision differentiable, and a research candidate with an NLP background was probed on large models versus fast classical alternatives for financial text.

What it tests

  • Classic data structures: caches, heaps, union-find
  • Dynamic programming on strings and sequences
  • A quick probability question solved by conditioning
  • Depth on your own projects and the models you used
  • Reasoning about gradients through discrete choices

How to prepare

  • The firm's own guidance: clarify the problem before coding, think out loud and ask for hints when stuck; Python and C++ are core but any language is accepted.
  • Drill conditional expectation and first-step analysis; it is the most reported style for the probability question.
  • For a routing or argmax question, start with the simplest relaxation (a steep sigmoid or softmax, straight-through) before heavier methods.
  • Know a fast classical baseline and the inference cost for every deep model on your resume.

Free reported questions for this stage

The firm's hiring pages and 5 candidate reports (one ML engineer 2025, four research-track 2016-2025); compiled accounts add the probability question. Pages say one round; some had two or three.

04Behavioral and culture fit

Questions in every round; some accounts describe a separate ~30 min talk with a senior researcher before the Super Day

The firm's pages put behavioral questions in each round: technical interests, internships, school projects and why the firm. Compiled accounts also describe an early culture-fit conversation with a senior researcher covering why the firm, a project walk-through and one ethics or judgement scenario. Our first-hand reports do not mention it as a separate round, so treat it as possible rather than standard.

What it tests

  • A specific answer to why the firm and why systematic research
  • A project told as situation, task, action and result, with numbers
  • Judgement in an ethics or dilemma scenario
  • Genuine interest in markets without pretending to be a trader

How to prepare

  • Prepare three crisp reasons for systematic research at the firm, tied to the team's data and horizon.
  • Pick one project you can walk through in two minutes and defend for twenty.
  • Answer a dilemma with the principle first, then the action and who you would tell.

The firm's hiring pages describe behavioral questions in every round; the separate early culture-fit talk comes only from compiled accounts, and no first-hand report mentions it.

05Super Day: probability and quant math

About 60 min, whiteboard; two or three problems back-to-back with full derivations

Part of a Super Day of three to five back-to-back interviews. Puzzle-style probability and statistics where the derivation is graded, not the number: set up the probability space or recursion explicitly, then solve and sanity-check. Conditional expectation and first-step analysis carry most problems. Many problems listed were tagged to trading roles, where this round is heavier.

What it tests

  • Geometric probability on sticks and circles
  • Linearity of expectation with indicator variables
  • Bayes updates with urns and dice
  • Order statistics and random walks
  • Sharp tail bounds from the first two moments

How to prepare

  • Derive rather than recall: write the recursion or E[X] = E[E[X|Y]] step and talk through it.
  • Reach for indicators and symmetry first; most expectation questions collapse once you do.
  • Sanity-check with limits, symmetry and small cases (n = 2 or 3); catching your own error scores well.

Free reported questions for this stage

Two research-track candidate reports (2017, 2018) mention probability or random-walk questions onsite; format and content from compiled accounts of the firm's quant research loop.

06Super Day: coding

About 60 min, CoderPad, Python or C++

One medium-hard algorithmic question and, in compiled accounts, often a practical part where you clean up or refactor a messy piece of code. Correctness, edge cases and complexity under time pressure are graded. The reported problems lean towards order statistics, streaming data and range queries, plus a few string classics.

What it tests

  • Medians and k-th elements across sorted or unsorted inputs
  • Online data structures: sliding-window median, streaming quantiles
  • Range-query structures: sparse table, Fenwick tree, lazy segment tree
  • String algorithms: KMP, minimum window, longest repeat
  • Refactoring messy code: structure, naming, tests and edge cases

How to prepare

  • Be able to code a two-heap sliding median with deletions; it ties together several reported problems.
  • Know the trade-off between a sparse table (static, O(1) query) and a segment tree (updates, O(log n)).
  • In a refactor, ask what the code is for and how it is called before changing it; the questions you ask are graded as much as the code.

Free reported questions for this stage

Three research-track reports (2016-2021) name an algorithms or software round onsite; the firm's pages say the onsite repeats the first round's coding mix. Refactor part from compiled accounts.

07Super Day: statistics and regression

About 60 min, whiteboard; the core of the ML and statistics round for research and ML tracks

The most consistent thread in research-track reports. Expect to derive the OLS estimator, state the Gauss-Markov conditions, then defend the fit under heteroskedasticity, collinearity, endogeneity and omitted variables, serially correlated errors and heavy tails. Regularization needs a reason: ridge versus lasso, the bias-variance trade-off and how you would choose the penalty. Diagnose before you model: residuals, leverage and out-of-sample checks.

What it tests

  • Deriving OLS in matrix form, its variance and standard errors, R-squared and n-1
  • Gauss-Markov conditions and what breaks under heteroskedasticity, collinearity or endogeneity
  • Omitted variable bias, suppressor variables and slopes of y on x versus x on y
  • Ridge and lasso: closed forms, geometry and choosing the penalty
  • Time-series errors: AR(1) forecasts, Yule-Walker, GLS

How to prepare

  • Re-derive beta-hat = (X'X)^-1 X'y and its standard error on paper, then show how to build X'X in chunks when data does not fit in memory.
  • For every assumption, say what fails first in financial data: non-stationarity, autocorrelation, spurious fit in levels.
  • When asked how to pick lambda, give cross-validation on time-ordered folds (and the one-standard-error rule), not in-sample MSE.
  • Know why a correlation matrix must be positive semidefinite and use it to bound an unknown correlation.

Free reported questions for this stage

Four research-track candidate reports (2016-2025), most on linear regression and its assumptions; depth and diagnostics follow compiled accounts of the firm's ML and statistics round.

08Super Day: ML theory and modelling on noisy data

About 60 min, whiteboard plus occasional short coding; research and ML tracks

Research-track candidates describe one or more mostly-ML rounds: comparing classical models, open-ended modelling questions and, in a 2025 AI-researcher loop, designing a model to find a signal in very noisy price data. Teams that lean hardest on ML probe theory, feature engineering and model diagnostics: overfitting and cross-validation, multiple testing in feature selection, L1 sparsity with p much larger than n, PCA versus autoencoders, boosting versus random forests. The math under classical ML matters more than deep-learning tooling.

What it tests

  • Diagnosing overfitting and designing time-ordered cross-validation without leakage
  • Multiple testing and selection bias when screening many features or backtests
  • L1 sparsity, L0/L1/L2 geometry and the p much larger than n regime
  • PCA, eigenvalues and quadratic forms; PCA versus autoencoders
  • A model and validation plan for a low signal-to-noise target

How to prepare

  • For an open-ended prompt, go data first (cleaning, leakage, non-stationarity), then a simple baseline, then time-aware validation, then the model.
  • Explain why the best of many backtests or features is biased upward and how you would correct for it.
  • Be able to code k-means, gradient descent and a stable log-sum-exp in a few minutes without libraries.
  • Tie every answer to a concrete failure mode; interviewers probe assumptions.

Free reported questions for this stage

Five research-track candidate reports (one AI researcher 2025, four 2016-2021); the open-ended design question appears in two. Topic list from compiled accounts. No ML-engineer onsite report.

09Team-fit and final review

About 45 min with an interested team, after hiring managers review your feedback

Hiring managers review your feedback across teams; a team that wants you runs a final fit interview, on the engineering side a leadership interview with a senior engineer. For research candidates this is often one open-ended prompt: design an alpha, or pick a paper you would actually trade, and defend it. If several teams want you, you choose with your recruiter. One 2025 research candidate met all the team leads during the onsite.

What it tests

  • One idea defended end to end: signal, horizon, capacity, costs and decay
  • Why it would stop working: crowding, costs, regime change
  • Depth on your own research decisions
  • Good questions about how research reaches live trading

How to prepare

  • Pick an idea you understand fully rather than the cleverest one; senior interviewers push on robustness.
  • Treat it as a peer discussion and bring questions about the team's data, horizon and tooling.
  • If several teams are interested, compare their mandates and growth paths before choosing; placement varies a lot.

The firm's hiring pages describe hiring-manager review, a team interview and team choice; one 2025 research report mentions meeting team leads. The open-ended prompt comes from compiled accounts.

Questions and themes reported most often

  • Derive OLS and reason about its properties (slopes both ways, R-squared, n-1) (6 reports, super day: statistics and regression)
  • Geometric probability with broken sticks and points on a circle (5 reports, super day: probability and quant math)
  • Lasso and ridge: closed forms, sparsity and geometry (4 reports, super day: statistics and regression)
  • Expected counts via indicators and symmetry (4 reports, super day: probability and quant math)
  • Medians and k-th order statistics across arrays or streams (4 reports, super day: coding)
  • Eigenvalues, quadratic forms and PCA variance (4 reports, super day: ml theory and modelling on noisy data)
  • AR(1) forecasting, Yule-Walker and GLS for correlated errors (3 reports, super day: statistics and regression)
  • OLS when the data does not fit in memory or arrives online (3 reports, super day: statistics and regression)

What Citadel asks about

  • Statistics & regression25%
  • ML theory & modelling25%
  • Coding & algorithms25%
  • Probability15%
  • Research discussion & behavioral10%

Citadel interview prep checklist

  • Can derive OLS and its standard error in matrix form and say what fails when each Gauss-Markov condition breaks
  • Can explain omitted variable bias, multicollinearity and lasso versus ridge, and how to choose the penalty, without notes
  • Can design a leak-free, time-ordered validation scheme and explain multiple-testing bias in feature or backtest selection
  • Coded a sliding-window median and a segment tree from scratch, and practised refactoring a messy function
  • Solved a set of probability puzzles with written derivations using conditioning and indicators
  • Have one research idea or recent ML paper you can defend on signal, horizon, costs and decay
  • Have a crisp answer to why the firm and one project story with numbers

Tips

  • Make regression the first priority: derive OLS, state the Gauss-Markov conditions, then drill omitted variables, collinearity, heteroskedasticity, regularization and time-series errors.
  • Prepare validation for noisy, non-stationary data: walk-forward or purged splits, leakage checks, and why the best of many backtests is biased upward.
  • Know classical ML at the level of derivations: generative versus discriminative models, L1 sparsity, PCA, p much larger than n.
  • Keep classic data-structure coding sharp in Python or C++; the harder rounds favour medians, streams and range queries, and a refactor task may appear.
  • Drill conditional expectation and first-step analysis and write full derivations; the probability rounds grade the steps.
  • Expect deep questions on every model on your resume, including latency and simpler alternatives.
  • Have one research idea or recent paper you can defend end to end, including why it would stop working.
  • Ask your recruiter early whether your track has an online test, which business unit the team sits in, and how many Super Day rounds to plan for.
  • Skip trader-track prep such as market-making games and speed arithmetic unless your recruiter says otherwise.

See the rest of the Citadel loop

The later stages, the questions reported for them and how to prepare are in Pro.

Unlock with Pro

Practice the Citadel loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

Citadel interview FAQ

How many interview rounds does Citadel have for ML roles?

Candidates report 9 stages: Recruiter call, Online assessment (track-dependent), First round (remote technical), Behavioral and culture fit, Super Day: probability and quant math, Super Day: coding, Super Day: statistics and regression, Super Day: ML theory and modelling on noisy data and Team-fit and final review. A recruiter call, a track-dependent online test (engineering and data-science tracks; research candidates usually go straight to live rounds), one 45-60 minute remote CoderPad round (some candidates had two or three), then a Super Day of three to five interviews of 45-60 minutes on probability, coding, regression and ML, with behavioral questions in each. Hiring managers then review your feedback and an interested team holds a final fit interview; if several teams want you, you choose with your recruiter. The firm states about four to five weeks for research roles and about eight for engineering. The hedge fund and its market-making affiliate run separate but similar processes. Market-making simulations, the short probability-speed test and mental-math drills are trader-track only and are left out here.

How long does the Citadel ML interview process take?

Reported timelines run 4-8 weeks.

What does Citadel ask in ML interviews?

By reported question mix: Statistics & regression 25%; ML theory & modelling 25%; Coding & algorithms 25%; Probability 15%; Research discussion & behavioral 10%. The most repeated themes: Derive OLS and reason about its properties (slopes both ways, R-squared, n-1); Geometric probability with broken sticks and points on a circle; Lasso and ridge: closed forms, sparsity and geometry.

How should I prepare for the Citadel ML interview?

Make regression the first priority: derive OLS, state the Gauss-Markov conditions, then drill omitted variables, collinearity, heteroskedasticity, regularization and time-series errors. Prepare validation for noisy, non-stationary data: walk-forward or purged splits, leakage checks, and why the best of many backtests is biased upward. Know classical ML at the level of derivations: generative versus discriminative models, L1 sparsity, PCA, p much larger than n.

Are there free Citadel interview questions?

Yes. LeetML has 100 questions reported from Citadel interviews, and 60 of them are free to solve in the browser; they are listed by stage above.

Related interview guides