Company interview guide
DRW Machine Learning Interview Guide (2026)
For ML, AI and research roles: a timed Python and ML test, then probability, regression and an ML design or research deep dive over a two-day final.
DRW's machine learning interview loop has 7 reported stages, from online technical challenge to final round: research deep dive & behavioral; typical length: Varies; applications are reviewed on a rolling basis.
Roles covered: ML Engineer, Research Scientist, Data Scientist and Applied Scientist. The question bank holds 31 questions reported from DRW interviews, 18 of them free to solve in your browser.
- Interview process length
- Varies; applications are reviewed on a rolling basis
- Stages
- 7
- Reports linked to bank questions
- 34
- Bank report dates
- Jun 2022–Jul 2026
- Questions in bank
- 31 (18 free)
- Last reviewed
- October 2026
How the DRW interview loop works
Apply with a resume (rolling review, so apply early), then take a timed technical challenge at home. The test depends on the track: ML engineering candidates report a 140-minute proctored Python and ML test, quant research candidates get a 45-minute math form of about 6 fill-in questions, and software candidates get a 150-minute, 3-problem coding test. A pass leads to a Zoom or phone interview of about 45 minutes, mixing resume questions with short technical ones, and on research tracks a 45-60 minute probability and expectation round. The final is on site over two days: introductions and another technical challenge on the first afternoon, then a full day of back-to-back interviews covering probability, regression and linear algebra, coding, ML design and your past research. Market-making estimation drills, options and game-theory rounds are trader-track only. ML-specific first-hand reports cover only the online test, so later stages follow the firm's own pages and its quant research loop.
The DRW interview process, stage by stage
01Online technical challenge
Timed and auto-graded at home.
ML track: about 140 min on Codility with camera and screen sharing, three parts. QR track: 45 min, about 6 fill-in math questions, webcam-proctored, calculator allowed. SDE track: 150 min, 3 coding problems
For ML engineering roles the test has three parts: about ten single-choice ML and deep learning questions, a pandas and numpy task on a returns table (simulate random portfolios, compute annualized return, volatility and Sharpe ratio, return the best weights), and a modelling task. One candidate built a scikit-learn plus imbalanced-learn pipeline for a heavily imbalanced target; another had to beat a hidden baseline with a fixed LinearSVC, writing only train and test functions, and found that only data-side changes helped. Formulas are given and no finance is needed. Quant research titles get the 45-minute fill-in form instead, where numbers rotate each sitting and a slip in arithmetic scores as wrong.
What it tests
- Short ML and deep learning facts: PCA eigenvectors, Gini impurity, the Bellman equation, dropout, gradient clipping and ResNets, non-convexity, attention scaling
- Vectorized pandas and numpy on a returns table: random long-only weights, annualized return and volatility, Sharpe ratio, missing histories
- Leak-free scikit-learn pipelines with a ColumnTransformer, resampling or class weights, and precision, recall and F1
- Improving a fixed model through scaling, features, cleaning and rebalancing rather than a new estimator
- QR form: matrix unknowns, a dice expected-value game, covariance, a CLT tail, a Markov stationary share, a Gamma-function area, and since 2025 an invert-then-differentiate matrix item
How to prepare
- Rehearse the ML test end to end in one sitting; it runs well over two hours under proctoring.
- Fit every preprocessing step on training data only, and know the main imbalanced-learn samplers and when class weights are enough.
- On the hidden-baseline task, fix the data first: once you pass the baseline you stop seeing test feedback, so validate locally.
- If your title says quant researcher, drill the fill-in families: about 7 minutes per question, exact answers, check the stated rounding.
Free reported questions for this stage
- Simulated Max-Sharpe Weights with Missing HistoriesData WranglingMediumSolve
- Imbalanced Classification Pipeline with SMOTENCClassical MLMediumSolve
- Transformer Attention and the 1/sqrt(d_k) ScalingLLMs & GenAIMediumSolve
- Information Gain of a Decision-Tree SplitStatistics & InferenceEasySolve
Two ML-engineer reports from mid-2025 describe the same three-part test; a third listing repeats one. Firm pages describe a timed take-home challenge by role; QR and SDE specifics from compiled accounts.
02Zoom or phone interview
About 45 min with a developer, quantitative trader or researcher; weekdays, business hours
One interview that walks through your resume, coursework and research, with short technical questions mixed in. The firm suggests brushing up on mental math and computer science fundamentals and having a computer ready if coding is likely. Compiled accounts describe it as behavioral plus light technical, sometimes with a quick probability or mental-math warm-up, and they reward thinking aloud.
What it tests
- Depth behind every project and number on your resume
- Why this firm and this kind of work, and which teams interest you
- Quick, accurate reasoning on a short probability or estimation question
- Computer science fundamentals and clear communication
How to prepare
- Prepare a two-minute summary of each project: the question, the key decision, the result and what you would change.
- Do some mental arithmetic and basic probability practice the week before; the firm mentions it explicitly.
- Offer several weekday slots quickly when asked, and bring your own questions about how the role grows.
The firm's phone-interview article describes this 45-minute technical and resume call; compiled accounts of the quant loop agree. No ML-role report covers it.
03Probability & expectation
45-60 min video round on research tracks, and again in the final; verbal or whiteboard, with follow-ups that extend each answer
The core quant screen: expected value of a game and whether it is worth a given price, conditional updates, waiting times and streaks, gambler's ruin, order statistics and optimal stopping. Interviewers reportedly mark how fast you see the key insight and how cleanly you reason under follow-ups, so estimate first and then formalize. How much of this an ML engineer sees depends on the team; ML researchers should assume it.
What it tests
- Expected payoff of a game, conditioning on the first outcome
- Waiting times to a pattern, streaks and gambler's ruin
- Order statistics and optimal stopping thresholds
- Conditional probability and Bayes updates with careful counting
- Linearity of expectation and symmetry under follow-up variations
How to prepare
- Say a rough answer first, then derive it; the follow-up usually changes one parameter.
- Set up a recursion for stopping and game problems, and check small cases and limits.
- Know expected flips to a pattern, the secretary problem and uniform order statistics well enough to adapt them.
Free reported questions for this stage
- Third Ball Red, Given a Red in the First TwoProbabilityEasySolve
- The Secretary ProblemProbabilityHardSolve
- Expected Number of Chord Intersections Inside a CircleProbabilityMediumSolve
- Expected Crossing Handshakes in a CircleProbabilityMediumSolve
- Birthday ProblemProbabilityEasySolve
Format and content from compiled accounts of the firm's quant research loop; problems here were tagged to quant research and trading roles. No ML-role report.
04Statistics, regression & linear algebra
Questions in the final round, sometimes in the Zoom interview, for research-leaning roles; derivations on a whiteboard or paper
A small but pointed regression core: closed-form OLS and what breaks with collinear features or a duplicated dataset, why ridge is always well defined, OLS versus ridge versus lasso for a factor model with 200 candidate factors and 500 daily observations, the bias from regressing the other way round, and what doubling the data does to t-statistics. Linear algebra comes up fast: eigenvalues from trace and determinant, diagonalizability, minimum-norm least squares. Compiled accounts also list PCA derivations and Bayesian updating with known variance.
What it tests
- OLS derivation, rank deficiency and the pseudoinverse minimum-norm answer
- Ridge as a Gaussian-prior MAP estimate, lasso for sparse factor selection, and when p is close to n
- Regression of y on x versus x on y, prediction bias and t-tests under duplicated data
- PCA from the covariance eigendecomposition, and correlation and covariance calculations
- Quick eigenvalue, trace and determinant arguments
How to prepare
- Derive the normal equations and the ridge solution aloud without notes, then explain the prior that ridge implies.
- Know why duplicating rows leaves the coefficients unchanged but shrinks standard errors, and why that is a trap.
- Practise one-line linear algebra: trace is the sum of eigenvalues, determinant is their product, exponentiate eigenvalues not entries.
Free reported questions for this stage
- Why Ridge Regression Is Well-DefinedStatistics & InferenceMediumSolve
- Inverse Regression Prediction BiasStatistics & InferenceMediumSolve
- Correlation of Two Indicators Over Three Coin TossesStatistics & InferenceEasySolve
- Proving the Cauchy-Schwarz InequalityLinear Algebra & OptimizationMediumSolve
Format and content from compiled accounts of the firm's quant research loop; problems here were tagged to quant research roles. No ML-role report.
05ML coding & ML system design
Final-round interviews for ML-track roles: live coding in PyTorch from skeleton scripts, and an open design discussion
Compiled accounts list two ML interview problems past the screen. One is live coding a small CNN binary image classifier in PyTorch: fill in the model (convolution, ReLU, pooling, normalization, a fully connected head), the data loaders and the training loop with BCE loss and Adam for a few epochs, then report metrics. The other is designing a system to predict stock returns: choose the target, features from market, fundamental and alternative data, the model, a validation scheme that avoids leakage, and how it is served.
What it tests
- Writing a working PyTorch model, data loader and training loop quickly and correctly
- Choosing loss, optimizer and metrics for binary classification and reporting them honestly
- Framing a prediction target and horizon for noisy, non-stationary financial data
- Leakage-safe, time-ordered validation and how signals decay out of sample
- Training and inference pipeline choices: retraining cadence, data freshness, latency
How to prepare
- Practise writing a CNN training loop from a blank file in under 20 minutes, including eval mode and metric logging.
- In the design, define the label and the as-of time of every feature before talking about models.
- Use walk-forward or purged validation and say why random K-fold overstates performance on time series.
From compiled accounts of the firm's ML interview questions only; no first-hand ML-role report describes this round, so treat its place in the loop as unconfirmed.
06Algorithmic coding
Engineering-leaning roles: 60 min of algorithms in the final, often with C++; problems below come from the software track's 150-minute, 3-problem coding test
The software track's timed coding test is judged on correctness and performance with no hidden-test feedback, and final-round coding goes deep on data structures and complexity. Compiled accounts mention streaming medians and order-matching engines alongside short greedy, string and simulation tasks like those placed here. ML engineers on engineering-heavy teams may meet similar coding; research and data roles report it less.
What it tests
- Greedy choices on arrays and strings, and careful simulation
- Data-structure design: streaming statistics, a price-time priority order book
- Stating complexity first, then optimizing
- Writing your own edge-case tests when hidden results are not shown
How to prepare
- Read the stated assumptions closely; the firm says tasks reward efficient code under them.
- Test edge cases yourself before submitting: silence from the grader does not mean coverage.
- If the role lists C++, review memory layout and the standard containers you would use.
Free reported questions for this stage
- Minimum Water Tanks to Serve All HousesCoding & AlgorithmsEasySolve
- Minimum Cleaning-Robot Runs to Clean a BuildingCoding & AlgorithmsEasySolve
The firm's prep article describes a Codility coding challenge judged on correctness and performance; length and round from compiled accounts. Problems tagged to software roles.
07Final round: research deep dive & behavioral
On site over two days: introductions and a technical challenge on the first afternoon, then a full day of back-to-back interviews; schedule given on arrival
The firm describes the final round as mostly technical, with several back-to-back interviews, networking time on both days and a chance to meet potential teammates. Researchers are told to be ready to discuss their research and publications in detail; developers to expect in-depth computer science. Compiled accounts say the day spans the full technical bar, from probability to coding, with the same reward for seeing the key insight fast. Behavioral questions run through the second day rather than in a separate round.
What it tests
- Detailed defence of your research or main project: design, results, what failed
- The probability, regression and coding themes from earlier stages, in more depth
- How you work with traders, researchers and developers
- Stamina across a full day of interviews
How to prepare
- Prepare one project in depth, including the experiments that did not work and why.
- Treat every conversation, including informal ones, as part of the assessment.
- Ask what the first-day technical challenge covers for your role; it is not described for ML candidates.
The firm's graduate, intern and final-round prep pages describe this two-day onsite and the research discussion; compiled accounts agree. No ML-role report covers it.
Questions and themes reported most often
- Simulate random portfolios from a returns table and return the max-Sharpe weights (2 reports, online technical challenge)
- About ten single-choice ML and deep learning concept questions (2 reports, online technical challenge)
- Imbalanced binary classification pipeline with resampling and precision, recall, F1 (1 reports, online technical challenge)
- Beat a hidden baseline with a fixed linear SVM using only data-side changes (1 reports, online technical challenge)
- Six fill-in math questions: matrix unknowns, dice expected value, covariance, Markov share (online technical challenge)
- Least squares, ridge and lasso: closed forms, collinearity, many factors on few days (statistics, regression & linear algebra)
- Waiting-time, stopping and order-statistic puzzles with follow-ups (probability & expectation)
- Live-code a small CNN binary image classifier in PyTorch (ml coding & ml system design)
What DRW asks about
- Python data & ML coding25%
- Probability & expectation20%
- Statistics, regression & linear algebra20%
- ML & deep learning concepts15%
- ML system design & research deep dive12%
- Algorithmic coding8%
DRW interview prep checklist
- Wrote a max-Sharpe random-portfolio simulator in pandas and numpy without loops over assets
- Built an imbalanced classification pipeline with no preprocessing leakage and tuned the threshold
- Improved a fixed linear model by changing only preprocessing and features
- Answered ten quick ML and deep learning concept questions in under twenty minutes
- Derived OLS, ridge and PCA on a whiteboard and explained collinearity and duplicated data
- Worked waiting-time, stopping and Bayes problems aloud with follow-ups
- Coded a CNN training loop in PyTorch and sketched a leakage-safe return-prediction design
- Prepared a detailed deep dive on one research project, including what failed
Tips
- Ask the recruiter which online test you will sit: the ML test, the 45-minute quant math form or the software coding test.
- Rehearse the ML test in one sitting of over two hours: concept questions, a pandas returns task and a scikit-learn modelling task.
- Get fluent with pandas and numpy on a returns table: random weights that sum to one, annualized return and volatility, Sharpe ratio.
- Build leak-free pipelines and, when the model is fixed, look for gains in the data: scaling, features, cleaning and rebalancing.
- For research roles, drill OLS, ridge, lasso and PCA derivations until you can give them aloud, and tie each to low signal-to-noise data.
- In every probability answer, estimate first, then formalize, then expect a follow-up that changes one parameter.
- Prepare a PyTorch training loop from scratch and a return-prediction design that starts from the label and point-in-time features.
- Expect to defend everything on your resume from the Zoom interview onwards, with the numbers behind each result.
- Skip market-making games, options and game-theory drills unless the recruiter says your team uses them; they are trader-track rounds.
See the rest of the DRW loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the DRW loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
DRW interview FAQ
How many interview rounds does DRW have for ML roles?
Candidates report 7 stages: Online technical challenge, Zoom or phone interview, Probability & expectation, Statistics, regression & linear algebra, ML coding & ML system design, Algorithmic coding and Final round: research deep dive & behavioral. Apply with a resume (rolling review, so apply early), then take a timed technical challenge at home. The test depends on the track: ML engineering candidates report a 140-minute proctored Python and ML test, quant research candidates get a 45-minute math form of about 6 fill-in questions, and software candidates get a 150-minute, 3-problem coding test. A pass leads to a Zoom or phone interview of about 45 minutes, mixing resume questions with short technical ones, and on research tracks a 45-60 minute probability and expectation round. The final is on site over two days: introductions and another technical challenge on the first afternoon, then a full day of back-to-back interviews covering probability, regression and linear algebra, coding, ML design and your past research. Market-making estimation drills, options and game-theory rounds are trader-track only. ML-specific first-hand reports cover only the online test, so later stages follow the firm's own pages and its quant research loop.
How long does the DRW ML interview process take?
Reported timelines run Varies; applications are reviewed on a rolling basis.
What does DRW ask in ML interviews?
By reported question mix: Python data & ML coding 25%; Probability & expectation 20%; Statistics, regression & linear algebra 20%; ML & deep learning concepts 15%; ML system design & research deep dive 12%; Algorithmic coding 8%. The most repeated themes: Simulate random portfolios from a returns table and return the max-Sharpe weights; About ten single-choice ML and deep learning concept questions; Imbalanced binary classification pipeline with resampling and precision, recall, F1.
How should I prepare for the DRW ML interview?
Ask the recruiter which online test you will sit: the ML test, the 45-minute quant math form or the software coding test. Rehearse the ML test in one sitting of over two hours: concept questions, a pandas returns task and a scikit-learn modelling task. Get fluent with pandas and numpy on a returns table: random weights that sum to one, annualized return and volatility, Sharpe ratio.
Are there free DRW interview questions?
Yes. LeetML has 31 questions reported from DRW interviews, and 18 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- CitadelFor ML, AI and research roles: the firm's quant research loop, heavy on regression, ML theory on noisy data, probability derivations and classic coding.
- Jane StreetFor ML, AI and research roles: multi-part coding judged on clarity, a hands-on ML round, and probability and statistics on research-leaning loops.
- Two SigmaInterview guide for ML, AI and research roles: a numeric HackerRank assessment, then data science, coding and regression-heavy statistics rounds, three manager rounds and a senior closer.
- Hudson River TradingFor ML, AI and research roles: a recycled timed coding test, a probability or programming screen, then a final day of code, systems and statistical depth on very noisy data.
- OptiverFor ML, AI and research roles: a long timed assessment battery, a live betting game with a trader, a probability whiteboard, a P&L-graded take-home and a blunt fit final.
- IMC TradingFor ML, AI and research roles at a global options market maker: a track-split timed test, probability and ML-fundamentals rounds, and a fast superday.