Company interview guide

Jump Trading Machine Learning Interview Guide (2026)

For ML, AI and research roles: a timed coding test, a probability and linear algebra screen, then a final day built around live time-series modelling, with no trading games.

Jump Trading's machine learning interview loop has 8 reported stages, from recruiter call to behavioral & senior conversations; typical length: Usually a few weeks; about a week per round, though one AI-role candidate waited three weeks between steps.

Roles covered: Research Scientist, ML Engineer and Data Scientist. The question bank holds 28 questions reported from Jump Trading interviews, 14 of them free to solve in your browser.

Try a free question

Interview process length
Usually a few weeks; about a week per round, though one AI-role candidate waited three weeks between steps
Stages
8
Reports linked to bank questions
29
Bank report dates
Aug 2019–Jul 2026
Questions in bank
28 (14 free)
Last reviewed
October 2026

How the Jump Trading interview loop works

Recruiter call, then a timed coding test (both ML engineer reports had two problems in 125 minutes, and an AI research candidate also sat one), then a 45-60 minute technical screen. Research-style screens mix probability, linear algebra and one coding task; engineering-titled screens add a resume deep dive, language internals and sometimes a rapid-fire CS quiz. The final day for research tracks has a math round, a roughly 60-minute live modelling round on a time-series dataset, a roughly 75-minute session finishing or fixing a long ML script, and behavioral talks with increasingly senior people; engineering tracks swap in coding and low-latency systems design. There is no market-making or betting game on any track; the brain-teaser coding test is trader-track only. This outline rests mainly on compiled accounts, since first-hand ML reports stop at the first screen.

The Jump Trading interview process, stage by stage

01Recruiter call

30-45 min phone call

Background, resume, motivation, communication and logistics such as location and visa. Compiled accounts describe it as mostly a fit check that clear communicators pass, with recruiters following up quickly; some roles are tied to one office.

What it tests

  • Why the firm, and why ML research or ML engineering in markets
  • How your resume maps to the AI/ML group or to a trading team
  • Location, visa and start timing

How to prepare

  • Ask which track you are on and whether your loop has the online test, the live modelling round and a systems round.
  • Know every line of your resume; later rounds can deep-dive it.
  • Genuine recruiting email comes only from the firm's own domain; the careers site warns about scams.

No first-hand ML-role report describes this call; format and length come from compiled accounts of the firm's research and engineering loops.

02Online coding assessment

Timed, auto-graded test on Codility; two problems in 125 min in both ML engineer reports, 2-4 problems in about 80 min to 4 hours for engineering tracks

Both 2025 ML engineer reports describe the same two-problem test: return the shares held each day when buy and sell rules fire on runs of up and down price moves, and repeatedly delete fixed letter pairs from a string. Compiled accounts describe long, wordy statements where parsing and edge cases cause more failures than the algorithm, with a final problem that needs the efficient solution to pass. Pure quant research candidates often skip this test, but an AI research candidate reported one, and some research or quant-developer candidates get a multi-hour single-dataset analysis task instead.

What it tests

  • Turning a long rule-based statement into an exact simulation
  • Stack machines with error handling, such as an RPN evaluator
  • Dynamic programming over choices and game positions
  • Combinatorics such as building a permutation directly from its rank

How to prepare

  • Restate the rules and edge cases in comments before coding; misreading the statement is the usual failure.
  • Choose the optimal approach before typing on the last problem; brute force times out.
  • Practise both a 2-hour two-problem sitting and a longer multi-problem one, since lengths vary by track.

Free reported questions for this stage

Two 2025 ML engineer reports describe a two-problem, 125-minute test; one AI research report confirms a test before the first screen. Wider format from compiled accounts.

03Technical screen

45-60 min video call with a shared editor; research versions have two or three short problems

Research-style screens go straight into problems, mixing probability, linear algebra and one coding task, with little time on the resume. Engineering-titled screens are a resume deep dive, language internals and one coding problem, and compiled accounts add a rapid-fire CS-knowledge quiz of about 20 questions where only one or two misses are tolerated. Reported linear algebra asks you to build or characterize a matrix from its properties.

What it tests

  • Valid ranges for correlation matrices, such as equal pairwise correlation
  • Mean and variance of a limiting random process
  • Eigenvalues of symmetric, rotation and reflection matrices
  • One short coding problem explained as you go

How to prepare

  • Know why equal pairwise correlation among n variables lies between -1/(n-1) and 1, and what the extreme case looks like.
  • Use eigen-decomposition arguments for matrix polynomial equations on symmetric matrices.
  • Narrate your reasoning; follow-ups and process are reported to matter more than speed.

Free reported questions for this stage

Two AI-role candidate reports confirm a technical screen right after the online test but give no content; content comes from compiled accounts of the research loop.

04Final round: probability & linear algebra

Final-day interview on a whiteboard or shared document; research tracks

Probability, Bayesian reasoning, combinatorics, optimal stopping, game theory and linear algebra. A reported favourite is splitting an array into contiguous groups to minimise total variance, first in linear time with running sums and then as a dynamic program over k groups. Other reported themes are drawing balls until a condition holds, guess-a-fraction-of-the-average games and repairing a covariance matrix that is not quite positive semidefinite.

What it tests

  • Bayes updates with an extreme prior
  • Expected stopping times in urn and drawing problems
  • DP over contiguous partitions with O(1) segment variance
  • Making a noisy covariance estimate positive semidefinite

How to prepare

  • Set up states and condition on the first step instead of recalling formulas.
  • For the variance split, derive the prefix-sum segment cost before writing the DP.
  • Write full derivations; interviewers are reported to give few hints.

Free reported questions for this stage

No first-hand ML-role report; one AI-role report confirms an onsite without detail. Content from compiled accounts of the firm's quant research loop.

05Final round: live time-series modelling

About 60 min building a model live on a time-series dataset in pandas and scikit-learn, with intuition follow-ups

The firm's most distinctive research round: you get a time-series dataset and build a model end to end within the hour, from a time-aware train/test split and rolling-window features to a regularized linear fit such as Lasso and an honest evaluation. Interviewers then ask about the intuition behind each feature and are reported to give few hints. Some candidates were asked to bring a favourite public dataset to summarise, plot and model.

What it tests

  • Fluent pandas and scikit-learn under time pressure
  • No look-ahead: walk-forward splits and causal rolling features
  • Regularized regression, what Lasso selects and why
  • Linear-model diagnostics such as residual behaviour

How to prepare

  • Rehearse load, split, feature, fit and evaluate until you can do it without checking the API.
  • Say out loud how you know no feature uses future data, and validate walk-forward rather than shuffled.
  • Prepare a public dataset you can summarise, plot and model quickly.

No first-hand ML-role report. Format and content from compiled accounts of the firm's quant research loop; the one problem here was tagged to an engineering role.

06Final round: ML script completion & debugging

About 75 min completing or fixing a long, pre-written ML script

A separate final-day session hands you a long ML script to finish or repair against the clock. Compiled accounts describe it alongside the live modelling round and give little more detail, so treat it as reading unfamiliar code quickly, finding the broken or missing steps and making the pipeline run correctly.

What it tests

  • Reading someone else's ML code quickly
  • Spotting leakage, shape and indexing bugs in a pipeline
  • Completing training and evaluation steps correctly

How to prepare

  • Read the whole script first and write down what each block is meant to do before changing anything.
  • Check the split, feature construction and evaluation for look-ahead before chasing small bugs.
  • Run small checks as you go rather than one big run at the end.

No first-hand ML-role report. Described only briefly in compiled accounts of the firm's quant research loop; no practice problems are tagged to the firm.

07Final round: coding & systems

About 60 min per round of live coding; engineering tracks get two coding rounds plus a systems design round

Classic algorithm problems solved live, often with a complexity target. Reported problems lean on dynamic programming, graphs, sorting and geometry. Engineering-titled roles add language internals and systems design around latency and networking, such as pushing updates to thousands of clients over TCP or UDP, which research candidates usually do not get.

What it tests

  • Dynamic programming with an O(n log n) target
  • Graphs: dependency deadlines, functional graphs, spanning trees
  • Sweep-line and rectangle geometry
  • Latency, memory and network trade-offs on engineering tracks

How to prepare

  • State the data structure, complexity and edge cases before typing.
  • If your role is engineering-titled, ask whether C++ internals or systems design are in scope.
  • Expect reserved interviewers; drive the discussion and state assumptions.

Free reported questions for this stage

No first-hand ML-role report. Matches compiled accounts of the firm's research and engineering loops; problems here were tagged to software engineering roles.

08Behavioral & senior conversations

Short conversations woven through the final day, with increasingly senior people

Standard prompts such as a failure and how you recovered, explaining a project to a non-specialist and how colleagues would describe you. The firm is secretive and teams are siloed, so this is also your chance to learn what the specific group does.

What it tests

  • Clear, concise stories about your own contribution
  • Explaining technical work simply
  • Genuine interest in research on markets

How to prepare

  • Prepare three stories: a technical win, a failure with a lesson and a teamwork example.
  • Keep answers tight and do not oversell; the tone is reserved.
  • Ask about the group you would join, since day-to-day work varies a lot.

No first-hand ML-role report. Matches compiled accounts of the firm's research and engineering loops.

Questions and themes reported most often

  • Shares held each day when buy and sell rules fire on up/down price patterns (2 reports, online coding assessment)
  • Repeatedly delete fixed adjacent letter pairs from a string (1 reports, online coding assessment)
  • RPN stack machine with duplicate, pop and error handling (online coding assessment)
  • k-th permutation built directly from the factorial number system (online coding assessment)
  • 0/1 knapsack and a win/lose turn game by dynamic programming (online coding assessment)
  • Range of equal pairwise correlation and the extreme case (technical screen)
  • Nested random intervals: mean and spread of the limit point (technical screen)
  • Build or characterize a matrix from its properties (technical screen)

What Jump Trading asks about

  • Time-series modelling & ML coding30%
  • Coding & algorithms25%
  • Probability15%
  • Linear algebra & regression math15%
  • Systems & engineering depth5%
  • Behavioral & research fit10%

Jump Trading interview prep checklist

  • Solved the up/down pattern share-count problem and a string pair-deletion problem inside a timed two-problem sitting
  • Can derive the range of equal pairwise correlation and solve a matrix polynomial equation on a symmetric matrix
  • Can make a nearly-PSD covariance matrix valid and explain shrinkage
  • Wrote the variance-minimizing contiguous split with O(1) segment costs, then as a k-group DP
  • Built a leak-free time-series model with rolling features, walk-forward validation and Lasso in under an hour
  • Found and fixed planted leakage bugs in someone else's forecasting script against the clock
  • Coded LIS in O(n log n) and a union-find MST from scratch
  • Have three concise behavioral stories and questions about the specific group

Tips

  • Ask the recruiter first which track you are on; research and engineering-titled ML roles get noticeably different loops.
  • For the online test, practise long, rule-heavy simulations and stack machines where reading the statement is the hard part.
  • Drill probability by first-step analysis and linearity of expectation rather than memorised formulas.
  • Get fluent with correlation and covariance matrices: valid ranges, eigenvalues, positive semidefiniteness and shrinkage.
  • Rehearse an end-to-end time-series model in pandas and scikit-learn in under an hour, with walk-forward validation and no look-ahead.
  • Know regularized regression well enough to explain what Lasso keeps, what ridge shrinks and how to tell a fit is overfit.
  • Practise reading and fixing a long ML script you did not write, starting with the split and feature steps.
  • Keep classic dynamic programming and graph coding sharp, and state complexity before you code.
  • Skip market-making games and speed arithmetic; no track at the firm is reported to use them.

See the rest of the Jump Trading loop

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

Unlock with Pro

Practice the Jump Trading loop

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

Jump Trading interview FAQ

How many interview rounds does Jump Trading have for ML roles?

Candidates report 8 stages: Recruiter call, Online coding assessment, Technical screen, Final round: probability & linear algebra, Final round: live time-series modelling, Final round: ML script completion & debugging, Final round: coding & systems and Behavioral & senior conversations. Recruiter call, then a timed coding test (both ML engineer reports had two problems in 125 minutes, and an AI research candidate also sat one), then a 45-60 minute technical screen. Research-style screens mix probability, linear algebra and one coding task; engineering-titled screens add a resume deep dive, language internals and sometimes a rapid-fire CS quiz. The final day for research tracks has a math round, a roughly 60-minute live modelling round on a time-series dataset, a roughly 75-minute session finishing or fixing a long ML script, and behavioral talks with increasingly senior people; engineering tracks swap in coding and low-latency systems design. There is no market-making or betting game on any track; the brain-teaser coding test is trader-track only. This outline rests mainly on compiled accounts, since first-hand ML reports stop at the first screen.

How long does the Jump Trading ML interview process take?

Reported timelines run Usually a few weeks; about a week per round, though one AI-role candidate waited three weeks between steps.

What does Jump Trading ask in ML interviews?

By reported question mix: Time-series modelling & ML coding 30%; Coding & algorithms 25%; Probability 15%; Linear algebra & regression math 15%; Systems & engineering depth 5%; Behavioral & research fit 10%. The most repeated themes: Shares held each day when buy and sell rules fire on up/down price patterns; Repeatedly delete fixed adjacent letter pairs from a string; RPN stack machine with duplicate, pop and error handling.

How should I prepare for the Jump Trading ML interview?

Ask the recruiter first which track you are on; research and engineering-titled ML roles get noticeably different loops. For the online test, practise long, rule-heavy simulations and stack machines where reading the statement is the hard part. Drill probability by first-step analysis and linearity of expectation rather than memorised formulas.

Are there free Jump Trading interview questions?

Yes. LeetML has 28 questions reported from Jump Trading interviews, and 14 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources