Company interview guide

D. E. Shaw Machine Learning Interview Guide (2026)

For ML, AI and research roles: a slow, research-first loop of deep dives, algorithm coding, probability, regression and open-ended modelling, with no trading games.

D. E. Shaw's machine learning interview loop has 9 reported stages, from application review & recruiter contact to closing gates: samples, references, background check; typical length: Often 2-3 months; the firm says several weeks.

Roles covered: ML Engineer, Research Scientist, AI Engineer and Data Scientist. The question bank holds 18 questions reported from D. E. Shaw interviews, 7 of them free to solve in your browser.

Try a free question

Interview process length
Often 2-3 months; the firm says several weeks
Stages
9
Reports linked to bank questions
18
Bank report dates
Feb 2022–Feb 2026
Questions in bank
18 (7 free)
Last reviewed
October 2026

How the D. E. Shaw interview loop works

A person reviews the application, then engineering-track candidates usually take a 90-minute, four-problem online coding test while research-track candidates go straight to a 45-60 minute technical phone round. The first virtual onsite is three rounds of about an hour (probability and statistics, coding, and an open-ended modelling round with a VP, the last often held only if the first two go well); the second is two rounds with an SVP or hiring manager covering behavioral questions, deeper statistics and more modelling. A research deep dive of about 20 minutes opens most rounds. Offers then wait on a code or writing sample, references and a background check. Market-making and betting games, mental-math tests and the macro take-home case are trader-track only and are left out. Only one first-hand ML-role report (2025, first two rounds) backs the outline.

The D. E. Shaw interview process, stage by stage

01Application review & recruiter contact

A person reviews every application, then a recruiter emails to set up the phone round; applying to several roles can mean several interviewer groups

No screening algorithm sits in front of the first conversation: a person reads the application and decides whether to call. Recruiters, including outside search firms for newer AI roles, then schedule the phone round. If you apply to more than one opening you may be interviewed by more than one group, each running its own loop.

What it tests

  • Whether your degree, research and code fit the role
  • Which group or groups should interview you
  • Logistics: timing, location and deadlines from other processes

How to prepare

  • Share competing deadlines early; the firm asks candidates to do so.
  • Ask whether your loop runs on the research track or the engineering track, and whether a code or writing sample will be needed.
  • Check that recruiter emails come from the firm's own domain or the search firm's own domain; the firm warns about recruiting scams.

The firm's careers pages describe a human application review before the phone round. One 2025 ML-role report (GenAI engineering role) began with an outside recruiter.

02Online coding test (engineering track)

90 min, 4 medium-hard algorithm problems on CodeSignal; referred candidates and research-track candidates often skip it

The software gate, often described as harder than at peer firms and drawn from a stable bank: count pairs with a given difference while values are inserted and deleted, count pairs that become equal after one digit swap, count free gaps of a minimum length as trees are planted, find the largest plus shape in a 1000 x 1000 grid, pick a subset of coins closest to a target with a tie rule, and count string pairs where one is a suffix of the other. Brute force times out on the grid and gap problems. A separate data-analyst track has used a pandas and chi-squared take-home case instead.

What it tests

  • Keeping a count correct while the data changes (insert, delete, query)
  • Grid precomputation: run lengths in four directions instead of expanding from every center
  • Interval and gap bookkeeping under repeated updates
  • Subset sum with exact tie handling
  • Pacing four problems in 90 minutes

How to prepare

  • Read all four first and bank the two cheapest (suffix pairs, pair-difference counts) before the hard ones.
  • Name the maintained quantity before you code: a counts map, gap lengths, arm lengths or reachable sums.
  • Budget about 22 minutes per problem and leave a buffer to fix time-outs.

Format and recurring problems from compiled accounts of the firm's software loop; no ML-role report mentions a test, and the firm's pages do not describe one.

03First technical phone

45-60 min by video: about 20 min on your research or resume, then 2-3 problems

The round opens with a long look at your background, including meta questions such as what makes a good research project. Research-track problems are probability, statistics and short puzzles: when to stop and keep a draw to get the maximum, how many uniform draws it takes for the sum to pass one, how to combine readings from two inaccurate scales into one best estimate, and how the slopes of Y on X and X on Y relate. Engineering-track problems are algorithms in any language, where you describe the approach and its complexity, then optimise; a library-free integer square root and a sliding-window maximum recur. The one ML-role report got a square root to a stated precision after a background chat.

What it tests

  • Explaining your research and its key decisions concisely
  • Numerical coding from first principles: binary search, Newton steps, stopping rules
  • Streaming problems such as a sliding-window maximum with a monotonic deque
  • Optimal stopping and expected-count puzzles
  • Combining noisy estimates and simple regression slopes

How to prepare

  • Have a tight research story; about 20 minutes go to it and the questions probe how you think, not just what you did.
  • Say the approach and its complexity before writing code, then optimise when asked.
  • For the square root, state the error bound and the stopping condition, and handle 0 and values below 1.

Free reported questions for this stage

One 2025 ML-role report had background plus a square-root coding task; the round's length and problem mix come from compiled accounts of the research loop. The firm's pages list a phone interview first.

04Research deep dive (opens most rounds)

About 20 min at the start of most rounds, sometimes most of a round; any line of your resume is fair game

Interviewers mine what you have built or researched in depth, line by line: why you chose an approach, what failed, what you would change and how you knew a result was real. Expect meta questions too, such as what makes a research project good and how your approach has changed over time. The one ML-role report describes a second round made up mostly of project deep dives.

What it tests

  • Motivation and key decisions behind your main projects
  • Whether a reported improvement was real or noise
  • What you learned and how your approach evolved
  • Your own share of team work

How to prepare

  • Prepare two projects in depth, including the experiments that failed and why.
  • Know your numbers: dataset sizes, baselines, effect sizes and how stable they were across runs.
  • Never share confidential details from past employers; the firm's guide warns against it.

One 2025 ML-role report had a round of project deep dives; compiled accounts put a 20-minute dive in most rounds. The firm's guide asks candidates to discuss projects in depth.

05Virtual onsite 1: coding & algorithms

About 60 min, any mainstream language, medium-hard problems with a probability or counting follow-up

One of the three first-onsite rounds is classic data structures and algorithms rather than ML code. The ML-role report had to rearrange an array in place so every value at an even position is smaller than every value at an odd one. Problems tagged to the firm lean on medians and percentiles over streams, randomised data structures and comparison-count proofs, and interviewers often attach a combinatorics or probability question to the algorithm.

What it tests

  • In-place array work with O(1) extra memory (partition and selection)
  • Order statistics: medians and percentiles over arrays or windows
  • Data structures with O(1) operations, including random removal
  • Proving a bound, such as the minimum comparisons for the second largest

How to prepare

  • Know quickselect and three-way partitioning well; they unlock many in-place rearrangements.
  • Explain correctness and complexity aloud, then optimise; a clear brute force beats silence.
  • Expect one constraint to change and be asked how your solution changes.

Free reported questions for this stage

One 2025 ML-role report has this coding task in its second round; the other problems are tagged to software roles. The firm's guide lists algorithms and programming.

06Virtual onsite 1: probability & statistics

About 60 min, often with a math or physics PhD; a few puzzles solved aloud, each extended with follow-ups

The research-track core: clean probability puzzles followed by an extension, such as letting n grow, adding a cost or changing the target. Firm-tagged questions include expected loops when random rope or noodle ends are tied, geometric probability with random cuts and points on a circle, expected non-dominated items, optimal stopping with uniform draws and conditional-probability traps. Many expectation problems reduce to conditioning on the first step and solving for a fixed point.

What it tests

  • Expected values via indicators, symmetry and linearity
  • Geometric probability on lines, sticks and circles
  • Recursions from first-step conditioning
  • Conditional probability and Bayes updates
  • Extending a result once the base answer is found

How to prepare

  • Write the setup and random variables before computing anything.
  • Check answers on a small case and a limiting case before stating them.
  • Practise saying derivations aloud; the method is graded, not just the number.

Free reported questions for this stage

No first-hand ML-role report. Format and themes from compiled accounts of the firm's research loop; the firm's guide names probability and mathematical statistics.

07Virtual onsite 1: open-ended modelling (VP round)

About 60 min with a VP, discussion style; often the third round of the first onsite, held only if the first two go well

Build a model for a vague question from scratch: predict a baseball team's season wins, decide whether ride-hailing helps or hurts city traffic, or size an investment. Interviewers want the simplest sensible baseline, stated assumptions, then complexity added one step at a time with a reason, and a view on combining several simple models. The firm's own guide describes the same case structure: clarify the question, state assumptions, compare options, support the answer, then name risks and next steps.

What it tests

  • Choosing a baseline and the data you would need
  • Stating and testing assumptions
  • Validating honestly on few, noisy observations
  • Combining simple models rather than jumping to a complex one
  • Naming risks and next steps

How to prepare

  • Treat it as a conversation: propose, check with the interviewer, then refine.
  • Say how you would know the model works before you say how you would build it.
  • Close with risks, failure modes and what you would try next.

No first-hand ML-role report. Format and examples from compiled accounts of the firm's research loop; the firm's guide describes case questions with this answer structure.

08Virtual onsite 2: senior rounds (statistics depth, experiment design, behavioral)

2 rounds with an SVP or the hiring manager; later rounds can be cancelled mid-loop if earlier ones go poorly

More senior interviewers mix behavioral and research-taste questions with deeper statistics and another open-ended modelling prompt. Regression goes deep: the slopes of Y on X and X on Y and why their product is R-squared, whether R-squared can be negative and the no-intercept case, what demeaning changes, how the t-statistic moves when the dataset is duplicated, and correlation after duplication. Experiment-design prompts recur, such as testing whether pressing an elevator button twice helps.

What it tests

  • OLS algebra: slopes both ways, R-squared bounds, demeaning, constrained fits
  • What duplicated data does to t-statistics and correlation
  • Designing a test with clear hypotheses and controls
  • Behavioral: why the firm, high standards, collaboration
  • Research taste: which problems are worth working on

How to prepare

  • Derive OLS from the normal equations rather than quoting formulas.
  • For experiment design, name the hypothesis, the control, the noise sources and the sample size logic.
  • Prepare real questions for senior staff; the firm treats the interview as a two-way fit check.

Free reported questions for this stage

No first-hand ML-role report. Round shape and questions from compiled accounts of the research loop; the firm's guide lists behavioral, situational and statistics questions.

09Closing gates: samples, references, background check

Code sample and/or writing sample (your own unedited expository piece, up to 15 pages), reference conversations and a background check; weeks between steps

Even after the loop, an offer depends on a code sample or a writing sample, conversations with your references and a background check. Candidates describe the reference check as a real screen rather than a formality, and multi-week silences between steps as normal.

What it tests

  • Quality of code you wrote on your own
  • Clear, concise technical writing
  • What references say about your rigour and collaboration

How to prepare

  • Prepare a clean code sample and an unedited essay of your own in advance.
  • Pick references who can speak to the depth and rigour of your work.
  • Be patient; silence for a few weeks is not a rejection.

No first-hand ML-role report reaches this step. The firm's pages describe references and writing or code samples; the background check comes from compiled accounts.

Questions and themes reported most often

  • Square root without a math library, to a stated precision (1 reports, first technical phone)
  • Deep dive on a past project: decisions, failures and whether the gain was real (1 reports, research deep dive (opens most rounds))
  • Rearrange an array in place so every even position is below every odd position (1 reports, virtual onsite 1: coding & algorithms)
  • Streaming statistics: sliding-window maximum, medians and rolling percentiles (virtual onsite 1: coding & algorithms)
  • Optimal stopping and how many uniform draws push the sum past one (virtual onsite 1: probability & statistics)
  • Expected loops from tying random ends, and geometric probability with random cuts (virtual onsite 1: probability & statistics)
  • Regress Y on X versus X on Y, R-squared bounds and demeaning (virtual onsite 2: senior rounds (statistics depth, experiment design, behavioral))
  • What duplicating a dataset does to the t-statistic and correlation (virtual onsite 2: senior rounds (statistics depth, experiment design, behavioral))

What D. E. Shaw asks about

  • Coding & algorithms24%
  • Probability & expectation22%
  • Statistics & regression22%
  • Research deep dive & behavioral16%
  • Open-ended modelling & ML judgment16%

D. E. Shaw interview prep checklist

  • Can walk through two projects in depth, including failures and how you validated the results
  • Wrote a library-free square root with an error bound, plus an integer version by binary search
  • Solved a sliding-window maximum with a deque and the in-place even/odd rearrangement with quickselect
  • Can derive the two-scales estimate, the pick-the-maximum stopping rule and both regression slopes aloud
  • Can explain negative R-squared, the no-intercept case and the effect of duplicated data on t-statistics
  • Have a structured modelling answer: baseline, assumptions, time-ordered validation, combining models, risks
  • Engineering track: completed one timed 90-minute, four-problem mock
  • Prepared a code sample and an unedited essay of your own

Tips

  • Ask in the first call whether your loop runs on the research or the engineering track; it decides the online test and the problem mix.
  • Prepare two projects for a 20-minute probe in most rounds: baselines, failed ideas, and how you knew a gain was real.
  • Know the research-track trio cold: pick-the-maximum stopping, the best combined estimate from two noisy scales, and the two regression slopes.
  • Make regression routine: R-squared bounds including the negative and no-intercept cases, demeaning, and what duplicated data does to t-statistics.
  • Keep classic coding sharp: library-free square root, monotonic deque, quickselect and medians, with complexity stated first.
  • On the engineering track, rehearse a timed 90-minute, four-problem session and bank the easy problems first.
  • In modelling questions, start from the simplest baseline, state assumptions, validate in time order and combine simple models before complex ones.
  • Solve probability puzzles out loud and add one extension to each.
  • Skip trader-track prep such as market-making games, mental-math drills and macro case write-ups.
  • Prepare your code and writing samples and references early, and plan for a process of two to three months.

See the rest of the D. E. Shaw loop

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

Unlock with Pro

Practice the D. E. Shaw loop

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

D. E. Shaw interview FAQ

How many interview rounds does D. E. Shaw have for ML roles?

Candidates report 9 stages: Application review & recruiter contact, Online coding test (engineering track), First technical phone, Research deep dive (opens most rounds), Virtual onsite 1: coding & algorithms, Virtual onsite 1: probability & statistics, Virtual onsite 1: open-ended modelling (VP round), Virtual onsite 2: senior rounds (statistics depth, experiment design, behavioral) and Closing gates: samples, references, background check. A person reviews the application, then engineering-track candidates usually take a 90-minute, four-problem online coding test while research-track candidates go straight to a 45-60 minute technical phone round. The first virtual onsite is three rounds of about an hour (probability and statistics, coding, and an open-ended modelling round with a VP, the last often held only if the first two go well); the second is two rounds with an SVP or hiring manager covering behavioral questions, deeper statistics and more modelling. A research deep dive of about 20 minutes opens most rounds. Offers then wait on a code or writing sample, references and a background check. Market-making and betting games, mental-math tests and the macro take-home case are trader-track only and are left out. Only one first-hand ML-role report (2025, first two rounds) backs the outline.

How long does the D. E. Shaw ML interview process take?

Reported timelines run Often 2-3 months; the firm says several weeks.

What does D. E. Shaw ask in ML interviews?

By reported question mix: Coding & algorithms 24%; Probability & expectation 22%; Statistics & regression 22%; Research deep dive & behavioral 16%; Open-ended modelling & ML judgment 16%. The most repeated themes: Square root without a math library, to a stated precision; Deep dive on a past project: decisions, failures and whether the gain was real; Rearrange an array in place so every even position is below every odd position.

How should I prepare for the D. E. Shaw ML interview?

Ask in the first call whether your loop runs on the research or the engineering track; it decides the online test and the problem mix. Prepare two projects for a 20-minute probe in most rounds: baselines, failed ideas, and how you knew a gain was real. Know the research-track trio cold: pick-the-maximum stopping, the best combined estimate from two noisy scales, and the two regression slopes.

Are there free D. E. Shaw interview questions?

Yes. LeetML has 18 questions reported from D. E. Shaw interviews, and 7 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources