Company interview guide

Jane Street Machine Learning Interview Guide (2026)

For ML, AI and research roles: multi-part coding judged on clarity, a hands-on ML round, and probability and statistics on research-leaning loops.

Jane Street's machine learning interview loop has 8 reported stages, from application & recruiter contact to collaboration, fit & decision; typical length: 2-8 weeks.

Roles covered: ML Engineer, Research Scientist and Deep Learning Engineer. The question bank holds 36 questions reported from Jane Street interviews, 16 of them free to solve in your browser.

Try a free question

Interview process length
2-8 weeks
Stages
8
Reports linked to bank questions
38
Bank report dates
Aug 2022–Apr 2026
Questions in bank
36 (16 free)
Last reviewed
October 2026

How the Jane Street interview loop works

Two entry paths. ML engineer reports run: recruiter contact, a 60-minute coding interview over Zoom in a shared editor, usually a 30-60 minute ML screen, then an onsite of 2-4 technical rounds of which 1-2 test ML. The firm's quant research loop, which ML researchers may be routed into, opens instead with a timed probability set of about 4 questions in 30 minutes and up to three 45-minute probability video rounds, adds a live coding round of the same style, and ends with a final day whose afternoon block is coding plus data science and statistics. Market-making games and mental-math speed drills are trader-track only and are not reported for ML roles. The firm says the rounds are all technical, that it replies within about a week, and that you are considered for every open role. The outline rests on few reports: about 11 ML-role candidates, all describing the screens, none the onsite.

The Jane Street interview process, stage by stage

01Application & recruiter contact

Online application reviewed on a rolling basis; recruiter contact within a few days, often a short call; recruiters also reach out first or meet candidates at conferences

Applications are reviewed on a rolling basis and you are considered for all open roles, so the recruiter conversation mostly settles which track you are interviewing for: ML engineer, ML researcher, ML performance engineer or a quant role. The track decides what the second technical round looks like.

What it tests

  • Practical experience training models, building ML tooling or speeding up ML workflows
  • Which track fits you best: engineering, research or performance
  • Timing and any competing deadlines

How to prepare

  • Ask what the second round covers for your track; reports range from PyTorch modeling to GPU performance.
  • Mention competing offers early if timing matters; the firm asks candidates to do so.
  • No finance background is expected, so do not spend prep time on markets.

Three ML-role reports start with recruiter outreach, a conference contact or a recruiter screen; the firm's interviewing page describes rolling review, recruiter contact within a few days and consideration for all roles.

02Probability screens (research track)

Research track: a timed written or HackerRank set of about 4 questions in 30 minutes, then up to three 45-minute video rounds. ML engineer reports skip this stage

On the firm's quant research track the first screen is a short timed set of probability and expected-value questions with some no-calculator arithmetic: parity of a sum of drawn primes, a stop-or-reroll dice game, a Bayesian re-flip of a randomly chosen biased coin, a bet after drawing from one of two urns, plus counting and estimation. You write out your reasoning, not just the number. Video rounds follow: a warm-up of three to five easier puzzles, a batch of five to seven unrelated questions with some game theory and a Fermi estimate, and a single hard problem for the full 45 minutes, often a Markov chain, optional stopping or a generating-function trick. One ML engineer reported probability, statistics, brain teasers and an estimation question; no other ML report mentions this stage.

What it tests

  • Linearity of expectation, indicator variables and symmetry
  • Bayesian updates on coins and urns, then pricing the bet they imply
  • One-step optimal stopping: find the keep-or-reroll threshold
  • First-step recursions on absorbing Markov chains and when optional stopping applies
  • Bounds when only marginals are known, and valid correlation matrices

How to prepare

  • On the timed set you have about 7 minutes a question; write a two-line justification for each answer, since the reasoning is graded.
  • In the batch round, restate each new setup from scratch and aim for 6-8 minutes a question; long silences cost more than a recoverable slip.
  • In the hard round spend the first few minutes on the state space and target, solve n = 1, 2, 3, and offer a quick simulation to check a formula.
  • Reach for indicators before algebra; most expected-count puzzles collapse with linearity.

Free reported questions for this stage

Format and question families from compiled accounts of the firm's quant research screens; one ML-engineer report describes probability, brain teasers and estimation.

03Coding interview

About 60 min over Zoom in a shared editor such as CoderPad; real code in any language; one problem that grows over 2-4 parts

Almost every ML-role report opens with a general coding round that matches the software screen, often run by a software engineer with no ML content. Problems are small simulations or APIs that grow part by part: a board where pieces enter from the bottom and you check for full rows, columns or k in a row, an editor that folds indented blocks, a marketplace data store that fills orders and then reconciles failed payments, or an interpreter for a toy language. The quant research track's live coding round uses the same gravity-board problem, with follow-ups on efficiency and on shaping the code into a reusable library.

What it tests

  • Turning a loosely stated spec into a clean class with a sensible API
  • Handling the edge cases each new part introduces
  • Updating state incrementally instead of rescanning everything after each operation
  • Reshaping working code into a small reusable library when asked
  • Explaining your data structures and their complexity as you go

How to prepare

  • Finish the first part quickly and cleanly; the firm says most passing candidates also complete the second.
  • After each move or update, check only the row, column or item it touched; interviewers hint at this.
  • Ask about ordering, ties and failure cases before coding; several reported rejections came from unhandled edge cases or running out of time while optimizing.
  • Use your strongest language; the firm says there is no bonus for OCaml or functional style.

Nine ML-role reports from 2024 to 2026 describe a multi-part coding round first, several saying it matches the software screen; the firm's ML postings and blog describe it, and research-loop accounts report the same board problem.

04ML technical screen

30-60 min, live PyTorch coding or ML discussion; code is not always executed

Most candidates who pass the coding round report a second, ML-focused screen. Examples include forecasting a time series with a model and an end-to-end PyTorch training loop, implementing a network from a diagram, choosing a model structure for a given input and output, and a system-level discussion of ML optimization. A performance-engineer candidate was told to expect GPU kernel performance. The firm's internship posting goes straight from coding to the onsite, so ask whether your track has this round.

What it tests

  • Writing a PyTorch model, dataset windowing and training loop from memory
  • Picking an architecture that maps a stated input to a stated output, and saying why
  • Leakage-free validation for time series: walk-forward splits, baselines, backtesting
  • Designing reusable training code: config, checkpointing and resume, evaluation
  • Reasoning about where training time and memory go, for performance-focused roles

How to prepare

  • Rehearse writing a full PyTorch training loop without autocomplete; candidates reported feeling rusty.
  • Say why you chose the loss, horizon strategy and validation split, not only what you chose.
  • For performance roles, review profiling, mixed precision, kernel launch overhead and memory-bound versus compute-bound operations.

Free reported questions for this stage

Five ML-role reports from 2024 to 2026 describe an ML screen after coding, with content varying by team; the firm's own pages do not describe it, and its ML intern postings go straight from coding to the onsite.

05Onsite: coding & data structures

1-2 rounds of about 60 min within an onsite of 2-4 technical rounds; in person in New York, London or Hong Kong, or virtual

The firm says the onsite holds two to four technical rounds and otherwise mirrors the software process, and compiled accounts of its research loop place a coding block in the final-day afternoon: clean implementations such as an unbounded Connect-Four board, a streaming transform over a sparse time-indexed matrix, or a diff of two sequences. No ML-role report describes these rounds first-hand. Problems tagged to the firm lean towards data structures that support updates and queries on the fly, streaming algorithms with memory limits, and sliding-window array problems.

What it tests

  • Data structures with fast updates and queries: weighted sampling, order statistics
  • One-pass streaming algorithms and their error guarantees
  • Two-pointer and monotone-deque techniques on arrays and strings
  • Bug-free code with stated complexity

How to prepare

  • Know reservoir sampling, heavy hitters and approximate quantiles well enough to code and prove them.
  • Practise Fenwick or segment trees for sampling and rank queries under updates.
  • Talk through trade-offs out loud; how you get there counts more than the final snapshot.

Free reported questions for this stage

No first-hand ML-role report; the firm's ML intern postings describe a 2-4 round technical onsite, its software prep page says the final round is mostly collaborative coding, and research-loop accounts agree.

06Onsite: ML modeling case

1-2 rounds; collaborative work on a realistic modeling problem from a small dataset or description; research tracks may add a past-project discussion

The firm describes its ML interviews as working together on realistic modeling problems like the team's daily work, and its posting says one or two onsite rounds test ML skills. Compiled accounts of the research loop's data science block give the flavour: predict a quantity from a small, messy record, such as a finish time from a training log, route difficulty from an elevation profile, level difficulty from a layout, a week of pageviews from the first hours, or the sum of digits in an image without OCR. Expect the interviewer to change the data or constraints midway. A research-intern candidate expected a discussion of a past ML project.

What it tests

  • Framing a messy prediction task: target, features, validation and baselines
  • Feature engineering from raw sequences and logs, and what signal is plausible
  • Choosing model capacity for small, noisy samples
  • Spotting overfitting and leakage, and saying how you would check for each
  • Adapting the plan when the data or constraints change

How to prepare

  • Start with a simple baseline and a validation scheme before proposing anything deep.
  • Say what you would plot or inspect first; the round is a conversation, not a deliverable.
  • Expect the interviewer to change the data or constraints midway; adapt rather than defend your first plan.

Free reported questions for this stage

No first-hand report; the firm's interviewing page describes collaborative realistic modeling problems and its ML intern postings place 1-2 ML rounds onsite; example prompts come from compiled research-loop accounts.

07Statistics, regression & research judgement

Questions inside onsite or screen rounds, on paper or in a shared doc; more likely on research tracks

The research loop's final-day data science block mixes modeling with short statistics questions that reward clean derivations: why a noisy regressor biases the slope toward zero, how the y-on-x and x-on-y slopes relate, what R-squared can and cannot tell you, ridge and lasso in closed form, a Bayesian credible interval or model comparison that then sets a bet, and a constrained maximum-likelihood estimate. Problems tagged to the firm's research roles add autoregressive order selection and sample sizes for a target interval width. One ML engineer reported statistics questions in the loop.

What it tests

  • OLS identities, errors-in-variables bias and linear-model diagnostics
  • Ridge versus lasso shrinkage, in closed form and in the SVD basis
  • Likelihood, intervals and Beta posteriors, and turning a posterior into a decision
  • Overfitting and multiple testing on noisy, non-stationary data; walk-forward validation
  • Autoregressive models, the PACF and order selection

How to prepare

  • Be ready to derive ridge and lasso solutions and the slope relation between the two regressions on paper.
  • When asked how good a model is, ask how many variants were tried before it; the firm's research questions probe data snooping.
  • Finish Bayesian questions with the action: the price or bet the posterior implies.

Free reported questions for this stage

One ML-engineer report mentions statistics questions; content and examples come from compiled accounts of the firm's quant research final-day data science block and problems tagged to its research roles.

08Collaboration, fit & decision

No separate behavioral round is described; fit is judged through the technical rounds

The firm says its interviews are collaborative and that it values asking good questions over knowing every answer; one candidate noted there were no behavioral questions. A decision usually arrives within about a week of each round, and you may be considered for a different team than the one you applied to.

What it tests

  • Thinking out loud and taking hints well
  • Admitting and fixing mistakes during the interview
  • Curiosity about the problem rather than reciting known answers

How to prepare

  • Tell the interviewer if you have seen a question before; they will switch it.
  • Follow up if you have not heard back after about ten days.

Two ML-role reports heard back within a week of the ML screen; the firm's interviewing page and blog state the one-week reply, and its blog and software prep page describe all-technical rounds.

Questions and themes reported most often

  • Small class or data-store API that grows over parts, such as order filling then end-of-day reconciliation (3 reports, coding interview)
  • Board simulation with gravity: insert pieces, then check rows, columns or k in a row after each move (2 reports, coding interview)
  • Code editor that folds and unfolds indented blocks (2 reports, coding interview)
  • Pick or build a network for a given input and output, sometimes from a diagram (2 reports, ml technical screen)
  • End-to-end time-series forecasting in PyTorch with a reusable training framework (1 reports, ml technical screen)
  • Tokenizer, parser and interpreter for a toy language, with design follow-ups (1 reports, coding interview)
  • ML training performance: profiling, kernels and memory limits (1 reports, ml technical screen)
  • Probability, statistics and quick estimation questions in an ML loop (1 reports, probability screens (research track))

What Jane Street asks about

  • General coding & simulation30%
  • ML modeling & PyTorch24%
  • Statistics & regression18%
  • Probability & expected value16%
  • ML performance & systems6%
  • Communication & collaboration6%

Jane Street interview prep checklist

  • Solved a gravity board simulation with row, column and k-in-a-row checks in under 45 minutes, then refactored it into a reusable class
  • Built a marketplace or similar multi-part API and handled every follow-up edge case
  • Wrote a PyTorch forecasting model and training loop with walk-forward validation, without running it
  • Talked through a small messy prediction case: baseline, features, validation, overfitting checks
  • Derived ridge, lasso, errors-in-variables and Beta-posterior results on paper
  • Did a timed set of four probability questions in 30 minutes with written reasoning (research track)
  • Can code reservoir sampling and a weighted sampler with updates, and explain both

Tips

  • Treat the coding interview as the main gate: drill multi-part simulation problems against a 60-minute clock.
  • Clarify ordering, ties and failure cases before coding; several rejections came from unfinished edge cases or slow optimization, not the core idea.
  • Be able to write a PyTorch model, data windowing and training loop from memory, since code may not be run.
  • Know time-series validation cold: walk-forward splits, look-ahead traps and simple baselines.
  • Ask your recruiter whether your track has the probability screens, a separate ML screen, or both.
  • On research tracks, practise writing short justifications for probability answers under a 30-minute clock.
  • Prepare the statistical side of regularization and regression, not only the optimisation side.
  • For performance roles, prepare to explain where training time goes and how you would profile and fix it.
  • Narrate your reasoning; the firm says the path through a problem matters more than the final snapshot.
  • Use the language you are fastest in; OCaml is not expected.

See the rest of the Jane Street loop

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

Unlock with Pro

Practice the Jane Street loop

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

Jane Street interview FAQ

How many interview rounds does Jane Street have for ML roles?

Candidates report 8 stages: Application & recruiter contact, Probability screens (research track), Coding interview, ML technical screen, Onsite: coding & data structures, Onsite: ML modeling case, Statistics, regression & research judgement and Collaboration, fit & decision. Two entry paths. ML engineer reports run: recruiter contact, a 60-minute coding interview over Zoom in a shared editor, usually a 30-60 minute ML screen, then an onsite of 2-4 technical rounds of which 1-2 test ML. The firm's quant research loop, which ML researchers may be routed into, opens instead with a timed probability set of about 4 questions in 30 minutes and up to three 45-minute probability video rounds, adds a live coding round of the same style, and ends with a final day whose afternoon block is coding plus data science and statistics. Market-making games and mental-math speed drills are trader-track only and are not reported for ML roles. The firm says the rounds are all technical, that it replies within about a week, and that you are considered for every open role. The outline rests on few reports: about 11 ML-role candidates, all describing the screens, none the onsite.

How long does the Jane Street ML interview process take?

Reported timelines run 2-8 weeks.

What does Jane Street ask in ML interviews?

By reported question mix: General coding & simulation 30%; ML modeling & PyTorch 24%; Statistics & regression 18%; Probability & expected value 16%; ML performance & systems 6%; Communication & collaboration 6%. The most repeated themes: Small class or data-store API that grows over parts, such as order filling then end-of-day reconciliation; Board simulation with gravity: insert pieces, then check rows, columns or k in a row after each move; Code editor that folds and unfolds indented blocks.

How should I prepare for the Jane Street ML interview?

Treat the coding interview as the main gate: drill multi-part simulation problems against a 60-minute clock. Clarify ordering, ties and failure cases before coding; several rejections came from unfinished edge cases or slow optimization, not the core idea. Be able to write a PyTorch model, data windowing and training loop from memory, since code may not be run.

Are there free Jane Street interview questions?

Yes. LeetML has 36 questions reported from Jane Street interviews, and 16 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources