
Company interview guide
Qube Research Technologies Machine Learning Interview Guide (2026)
For ML, AI and research roles: classical ML and regression theory, probability, a project defence and, on some teams, a two-week ML take-home.
Qube Research Technologies's machine learning interview loop has 8 reported stages, from application & intro call to onsite & fit with the hiring manager; typical length: About 3-6 weeks is typical; reports range from one week to three months, and intern processes run slower.
Roles covered: Research Scientist, Data Scientist, ML Engineer and Applied Scientist. The question bank holds 6 questions reported from Qube Research Technologies interviews, 1 of them free to solve in your browser.
- Interview process length
- About 3-6 weeks is typical; reports range from one week to three months, and intern processes run slower
- Stages
- 8
- Reports linked to bank questions
- 6
- Bank report dates
- Jul 2026–Jul 2026
- Questions in bank
- 6 (1 free)
- Last reviewed
- October 2026
How the Qube Research Technologies interview loop works
Most research candidates apply online or through a recruiter, have a short intro or phone call (several interns skipped it), and on some teams take a timed online test on HackerRank or Codility. Then come two to five interviews, by video first and often onsite at the end: an ML and statistics round on classical models and regression, a deep dive on your CV and projects, a probability and brainteaser round, and a coding round in Python. Some teams replace or add a practical step: a reported two-week ML project defended in a remote round, or a data challenge. A final fit conversation with the hiring manager or senior people closes the loop. Order varies by team; one intern was told outright that there was no HR or behavioral round. Trader-style items such as market-making games and speed arithmetic are not reported for this research loop and would be trader-track only.
The Qube Research Technologies interview process, stage by stage
01Application & intro call
Apply online, through a recruiter or on campus; a short phone or video call with a recruiter or team member, often skipped for interns
A short opening conversation: introduce yourself, explain why quant research and why the firm, and walk through your most interesting projects. One Paris intern had this as a fit interview where the interviewer and candidate simply introduced themselves; a Singapore intern covered motivation and projects in a phone screen. Two intern reports say there was no HR call at all and the process went straight to technical rounds.
What it tests
- Why systematic research, and why this firm
- A clear two-minute tour of your strongest projects
- Logistics: office, internship dates, visa
How to prepare
- Have one project story ready that ends in a measurable result; it often gets reused in the next round.
- If you won or entered a competition the firm ran, mention it; one successful intern built the interview around a hackathon win.
- Ask which online test, if any, your team uses and which topics each round will cover.
Intro or phone call before technical rounds in 3 research-role reports since 2025 plus 1 risk-analyst report; 2 research-intern reports had no HR call.
02Online assessment (some teams)
Timed online test on HackerRank or Codility, used by some teams only.
Research track: easy-to-medium probability and statistics plus applied ML concepts. Technologist track: 1-2 easy-to-medium coding problems, or a long 2.5-3 hour test of hard problems plus multiple-choice concept questions. Data-adjacent roles: a pandas test plus one algorithm problem.
Whether you get a test depends on the team: one research intern reports that some teams start with an online assessment and others go straight to interviews, and another had no technical assessment at all. When it appears it gates the process and is auto-graded on correctness and speed. Problems reported for this firm are trading-flavored algorithm tasks: a min-cost array reduction that is Huffman-style heap greedy, a longest substring with at most k restricted characters by sliding window, and an event log that reports profit across tickers at each query. A technologist reported two competitive-programming problems plus easy probability multiple choice.
What it tests
- Heap greedy and sliding-window solutions under time pressure
- Careful spec reading on value classes and edge cases up to 1e5 elements
- Easy-to-medium probability and statistics multiple choice
- Pandas group-by and aggregation for data-adjacent roles
How to prepare
- Get medium problems down to 20-25 minutes with clean code; you cannot predict which test version you get.
- For 'repeatedly combine the two smallest', reach for a min-heap; for 'longest window with at most k bad items', a two-pointer window.
- Spend an hour on pandas group-by, merge and pivot in case your team uses the data-flavored test.
Format and content from compiled accounts of the firm's research and technologist tests; 1 research-intern and 1 technologist report confirm a test on some teams.
03ML & statistics interview
Usually a 1-on-1 video interview with a researcher, often the first technical round; questions on classical ML, regression and linear algebra
The most consistently reported round for research roles and the core of this page. Expect mechanics, not trivia: compare random forests with XGBoost, say whether boosting reduces bias or variance and what bagging does, explain KNN, derive the OLS estimator and ask how to improve it, and state whether ridge and lasso have closed forms. Linear algebra appears as PCA, eigenvalues and what makes a matrix positive semidefinite. Interviewers then move to financial data: how to detect and reduce overfitting, what to do with non-stationary series, and correlation versus causation. One 2026 report also asked about transformer architecture and the Adam optimizer.
What it tests
- Random forest versus gradient boosting: bias, variance and when each fails
- OLS derivation, ridge and lasso closed forms and when they exist
- PCA, eigen-decomposition and positive semidefinite covariance matrices
- Overfitting detection and control on noisy, non-stationary data
- Correlation versus causation and what a regression can claim
How to prepare
- Explain boosting as sequential bias reduction and bagging as variance reduction, then say what that means for a noisy target.
- Derive OLS from the normal equations and show what ridge adds; know why lasso has no general closed form but does under orthonormal design.
- Have a concrete answer to 'what do you do with non-stationary data': differencing or returns, rolling normalisation, time-ordered validation.
- Practise answering what matters beyond the architecture: data quality, features, validation and leakage.
Classical ML, regression or linear-algebra questions in 7 research-role reports from late 2024 to 2026; tree models and OLS recur in 4. Matches compiled accounts.
04CV & project deep dive
Part of one or more technical rounds; a researcher drills into your CV, a project you present, or a factor or signal you built
Several candidates say the interview mostly focused on past experience and that every CV line was questioned in depth. Typical prompts: present one project, describe a factor you built and how you improved it, explain how you dealt with overfitting while researching alphas, and discuss risk factors in a factor risk model. A Paris intern opened the technical interview by presenting a project; a Singapore intern spent a round on a hackathon the firm had organised. Interviewers also ask about the firm and your motivation here.
What it tests
- Depth on a project: data, features, validation and what failed
- How you guarded against overfitting and look-ahead in your own work
- Factor or signal construction and how you would improve it
- Understanding of factor risk models and risk exposures
How to prepare
- Remove any CV line you cannot defend from first principles; reports stress in-depth CV questions.
- Prepare a 5-minute project presentation with one slide's worth of numbers and a clear validation story.
- For a signal or factor project, know its turnover, decay and how it behaves out of sample.
In-depth CV, project or past-factor questions in 7 research-role reports from 2025-2026; consistent across London, Paris and Asia offices.
05Probability & maths round
A video or onsite round, sometimes merged with ML; easy-to-medium probability, statistics and brainteasers
Research candidates describe probability, statistics, brainteasers and logical reasoning at an easy-to-medium level, with some finding it hard. Reported examples include a classic random walk exercise and the value that results from random moves in a 2048-style game. Compiled accounts add Gaussian and dice classics tagged to the firm: correlated standard normals with correlation 0.5 asking for P(X > 3Y) and the conditional on X > 0, the chance that 100 dice sum above 400 by the central limit theorem, and the XOR of a fair and a two-headed coin.
What it tests
- Random walks, gambler's ruin and stopping problems
- Linear combinations of correlated normals and quadrant probabilities
- CLT approximations with exact mean and variance
- Conditional probability and expectation brainteasers
How to prepare
- Write correlated normals as Y = rho X + sqrt(1 - rho^2) Z and reduce half-plane events to one normal.
- Know mean 350 and variance 100 x 35/12 for 100 dice without stopping to compute.
- State the method before the arithmetic; interviewers care about clean reasoning more than speed.
Probability, statistics or brainteaser content in 7 research-role reports from 2025-2026; example problems from compiled accounts. Level described as easy to medium.
06Coding interview
Live coding by video, usually in Python for research roles; C++ and data structures for technologist roles
Research interns report Python coding questions and, in one accepted case, an algorithm problem rated medium to hard. Problems tagged to the firm are clean data-structure tasks: find two missing numbers in 1..k+2 in O(1) space, check whether two strings are one insert, delete or swap apart, and find the fewest elements to move to the back so every prefix sum stays nonnegative. Technologist candidates describe two technical rounds with stacks built from queues, C++ container complexities and graph traversal, and compiled accounts add designing and justifying a matrix class. Interviewers push a working answer toward optimal complexity.
What it tests
- Invariant tricks for O(1)-space answers (sums, sums of squares, XOR)
- Greedy with a heap and an exchange argument
- Careful case splits on string edits
- Complexity of standard containers and why you chose one
How to prepare
- Drill medium problems in plain Python until you can explain time and space without prompting.
- For the prefix-sum problem, evict the most negative element kept so far, not the one that broke the constraint.
- Be ready to justify a data-structure choice by memory layout and access pattern.
Free reported questions for this stage
- Minimum Elements Moved to the Back for Nonnegative PrefixesCoding & AlgorithmsHardSolve
Coding rounds in 3 research-intern reports since 2025; 2 technologist reports describe harder data-structure rounds. Tagged problems come from compiled accounts.
07ML project or data challenge (some teams)
Some teams only: a practical ML project on a real-world problem over about two weeks, then a remote technical interview on your approach and results; others report a data challenge
A Zurich research intern who accepted an offer describes a theory interview, then a two-week practical ML project, then a remote round dissecting the approach and results, then a longer onsite. A 2026 Paris researcher lists a data challenge as part of a three-to-four interview process. The defence matters as much as the score: expect questions on validation, leakage, model choice and error analysis.
What it tests
- An end-to-end pipeline on real, messy data
- Validation design that respects time ordering
- Defending model choice, for example boosted trees versus a linear model
- Honest error analysis and what you would try next
How to prepare
- Start with a regularised linear baseline, then add tree models and show the gain out of sample.
- Write down every variant you tried; you will be asked how you chose the final one.
- Keep a reusable project template so a two-week clock does not start from zero.
Two-week ML project with a remote defence in 1 research-intern report from 2024 and a data challenge in 1 report from 2026; compiled accounts agree.
08Onsite & fit with the hiring manager
Often onsite: one or more longer technical interviews with team leads, then a fit conversation with the hiring manager or a senior person
The closing stage goes deeper on technical questions and checks fit. Candidates describe longer, more in-depth onsite discussions, team leads who challenge answers and expect quick replies, and a final fit interview with the hiring manager. Behavioral questions are standard: strengths and weaknesses, where you see yourself in five years, and why the firm. Some teams have no separate behavioral round.
What it tests
- Speed and depth on follow-ups to earlier questions
- Motivation for systematic research at the firm
- How you work with others and handle pushback
- Fit with a specific team
How to prepare
- Revisit every earlier answer and prepare the next follow-up on it.
- Have short, honest answers to strengths, weaknesses and five-year plans.
- Bring specific questions about how the team uses data and models.
Onsite or fit stage in 2 research-role reports, plus 3 technologist and data-science reports with HR or behavioral rounds; fit is shared across tracks.
Questions and themes reported most often
- Random forest versus XGBoost; does boosting reduce bias or variance, and what does bagging do (4 reports, ml & statistics interview)
- OLS estimator, then ridge and lasso closed forms (4 reports, ml & statistics interview)
- Random walks and probability brainteasers (4 reports, probability & maths round)
- How to detect and reduce overfitting in signal research (2 reports, ml & statistics interview)
- PCA, eigenvalues and positive semidefinite matrices (2 reports, ml & statistics interview)
- Describe a factor or signal you built and how you improved it (2 reports, cv & project deep dive)
- What to do with non-stationary data (1 reports, ml & statistics interview)
- Transformer architecture and the Adam optimizer (1 reports, ml & statistics interview)
What Qube Research Technologies asks about
- Classical ML: trees, boosting, overfitting25%
- Regression, linear algebra & statistics20%
- Probability & brainteasers18%
- Coding: Python, algorithms, pandas17%
- Project deep dive & take-home defence15%
- Fit & motivation5%
Qube Research Technologies interview prep checklist
- Can compare random forests and gradient boosting, including which one reduces bias and which reduces variance
- Derived OLS, ridge and lasso closed forms and can say when lasso has one
- Can explain PCA through eigen-decomposition and test whether a matrix is positive semidefinite
- Have a 5-minute project presentation with numbers, a time-ordered validation story and a failure
- Solved correlated-normal, random-walk and CLT problems without notes
- Coded heap greedy, sliding-window and prefix-sum problems in Python under a timer
- Built one end-to-end model on time-ordered data with leakage checks, ready to defend
- Have short answers on why the firm, strengths, weaknesses and five-year plans
Tips
- Make classical ML your first priority: random forests, bagging, boosting and their bias-variance story, explained from mechanics.
- Derive OLS, ridge and lasso by hand and know PCA, eigenvalues and positive semidefinite matrices cold.
- Prepare to defend every CV line in depth; reports repeatedly say the interview centred on past experience.
- Have a concrete playbook for noisy, non-stationary data: returns or differencing, time-ordered validation, shrinkage and leakage checks.
- Work easy-to-medium probability: random walks, correlated normals, CLT approximations and conditional expectation.
- Practise medium algorithm problems in Python, plus pandas group-by and merge in case your team's test is data-flavored.
- If you get the two-week ML project, treat the remote defence as the real exam and keep a log of every variant tried.
- Ask your recruiter whether your team uses an online test and what each round covers; formats vary by team and office.
- Skip trader-track prep such as market-making games and speed arithmetic; none is reported for this research loop.
See the rest of the Qube Research Technologies loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Qube Research Technologies loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Qube Research Technologies interview FAQ
How many interview rounds does Qube Research Technologies have for ML roles?
Candidates report 8 stages: Application & intro call, Online assessment (some teams), ML & statistics interview, CV & project deep dive, Probability & maths round, Coding interview, ML project or data challenge (some teams) and Onsite & fit with the hiring manager. Most research candidates apply online or through a recruiter, have a short intro or phone call (several interns skipped it), and on some teams take a timed online test on HackerRank or Codility. Then come two to five interviews, by video first and often onsite at the end: an ML and statistics round on classical models and regression, a deep dive on your CV and projects, a probability and brainteaser round, and a coding round in Python. Some teams replace or add a practical step: a reported two-week ML project defended in a remote round, or a data challenge. A final fit conversation with the hiring manager or senior people closes the loop. Order varies by team; one intern was told outright that there was no HR or behavioral round. Trader-style items such as market-making games and speed arithmetic are not reported for this research loop and would be trader-track only.
How long does the Qube Research Technologies ML interview process take?
Reported timelines run About 3-6 weeks is typical; reports range from one week to three months, and intern processes run slower.
What does Qube Research Technologies ask in ML interviews?
By reported question mix: Classical ML: trees, boosting, overfitting 25%; Regression, linear algebra & statistics 20%; Probability & brainteasers 18%; Coding: Python, algorithms, pandas 17%; Project deep dive & take-home defence 15%; Fit & motivation 5%. The most repeated themes: Random forest versus XGBoost; does boosting reduce bias or variance, and what does bagging do; OLS estimator, then ridge and lasso closed forms; Random walks and probability brainteasers.
How should I prepare for the Qube Research Technologies ML interview?
Make classical ML your first priority: random forests, bagging, boosting and their bias-variance story, explained from mechanics. Derive OLS, ridge and lasso by hand and know PCA, eigenvalues and positive semidefinite matrices cold. Prepare to defend every CV line in depth; reports repeatedly say the interview centred on past experience.
Are there free Qube Research Technologies interview questions?
Yes. LeetML has 6 questions reported from Qube Research Technologies interviews, and 1 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- 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.
- Squarepoint CapitalFor ML, AI and research roles: a gating HackerRank test, two junior rounds on probability, regression and coding, then a 5-hour pandas exam.
- D. E. ShawFor 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.
- 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.
- The Voleon GroupFor ML, AI and research roles: a research-style loop of regression derivations, classical ML at proof level and hands-on data work, with newer applied screens on LLMs and leak-free pipelines.
- Point72For ML, AI and research roles: a track-dependent HackerRank test, then probability, regression and research talk with the team you would join.