Finance7 min read·

G-Research Interview: Process & Questions 2026

The G-Research interview process explained - online assessment, technical phone screens, the onsite, real probability and ML question types, and how to prepare.

What to Expect at a G-Research Interview

G-Research runs a three to four stage interview process over three to six weeks, and it is one of the more mathematically demanding loops in UK finance. Expect a timed online assessment of probability, statistics and logic, one or two technical phone screens, and a half or full day onsite that mixes harder quantitative problems with a discussion of your own academic or research background. For software engineering roles the shape is similar but the online assessment and coding rounds lean algorithmic rather than statistical.

This guide focuses purely on the interview itself - the stages, the real question types candidates report, and how to prepare for the researcher, machine learning and engineering tracks separately. For background on the firm, its structure and culture, see our G-Research guide, and for what the roles pay once you are through the process, see our G-Research salary guide.


G-Research Interview at a Glance

  • Typical rounds: 3 to 4, over 3 to 6 weeks
  • Online assessment: Timed probability, statistics and logic test (researchers); algorithmic coding test (engineers)
  • Phone screens: 1 to 2, each 45 to 60 minutes
  • Onsite: Half or full day in London, 3 to 6 sessions
  • Question style: Puzzle-heavy probability, applied statistics, ML fundamentals, and coding
  • Roles covered: Quantitative Researcher, Machine Learning Researcher, Software Engineer, Data Engineer

The Process, Stage by Stage

Stage 1: Application and CV Screen

Applications go through an initial CV review, with strong maths, physics, computer science and engineering backgrounds prioritised. PhD candidates and students with strong olympiad or competitive programming results are fast-tracked more often than at most banks, though undergraduate and master's candidates are hired every year.

Stage 2: Online Assessment

Researcher and ML candidates face a timed test built around probability puzzles, combinatorics and quick quantitative reasoning, pitched noticeably harder than a typical bank test. Engineering candidates instead see an algorithmic coding assessment with two or three problems under time pressure. Candidates consistently describe the researcher assessment as one of the harder online tests in the industry, closer to a top prop firm than a bank.

Stage 3: Technical Phone Screen

A 45 to 60 minute call with a working researcher or engineer. Researchers solve probability and statistics problems live, with interviewers pushing into follow-up variants once the first answer lands. Engineers get live coding plus discussion of complexity, data structures and language depth. A second phone screen is common for candidates whose first call was strong but left specific gaps to check.

Stage 4: Onsite

A half or full day, either in London or over video, with three to six sessions. Researchers typically face harder probability and statistics questions, a machine learning fundamentals discussion (bias-variance trade-offs, regularisation, why deep models overfit financial data), a walk-through of your own research or dissertation work, and further coding. Puzzle-style questions in the spirit of the firm's public monthly puzzle appear throughout rather than in one dedicated round.

Stage 5: Offer

Decisions typically arrive within one to two weeks of the onsite. G-Research is reported to move decisively once it has made up its mind, and strong PhD candidates often have very little back-and-forth on the numbers.


How the Tracks Differ

DimensionQuantitative ResearcherMachine Learning ResearcherSoftware Engineer
Online assessmentProbability, statistics, logicProbability, statistics, logicAlgorithmic coding
Phone screen focusProbability and statistics, liveML fundamentals plus probabilityCoding, data structures, complexity
Onsite centrepieceResearch walk-through, harder probabilityModel architecture and training decisionsSystems and infrastructure discussion
Typical backgroundMaths, physics or statistics, often PhDML-focused PhD or strong master'sStrong CS or engineering fundamentals
Puzzle frequencyHighModerate to highLow to moderate

Engineering candidates without a finance background should not be put off - platform and infrastructure roles are hired on engineering strength, and G-Research's public job adverts read like a technology company's rather than a trading firm's.


Real Question Types

Probability

The last passenger. 100 passengers board a plane with assigned seats. The first sits in a random seat; every subsequent passenger takes their own seat if free, otherwise a random free seat. What is the probability the last passenger gets their own seat?

Approach: The answer is 1/2. At every point where a displaced passenger chooses randomly, the only two seats that matter by symmetry are seat 1 and seat 100, and each is equally likely to be taken first. Interviewers want the symmetry argument stated cleanly, not a simulation. Our probability interview questions collection covers more problems in this family.

Biased coin estimation. You flip a coin of unknown bias 20 times and see 14 heads. Give an estimate of the bias, a measure of uncertainty, and describe how you would test fairness.

Approach: The maximum likelihood estimate is 0.7, the standard error is roughly the square root of p(1-p)/n, giving about 0.10, so a rough 95% interval runs from 0.5 to 0.9. A two-sided binomial test of p = 0.5 gives a p-value around 0.115, so fairness cannot be rejected at the 5% level. The question checks whether you produce an estimator, an uncertainty measure and a test, in that order.

Statistics and Machine Learning

Overfitting in finance. Why do machine learning models that work well in vision or language often fail on financial prediction?

Approach: Signal-to-noise ratios are orders of magnitude lower, the data-generating process is non-stationary because the market adapts to exploited patterns, effective sample sizes shrink once cross-sectional and serial correlation are accounted for, and iterative researcher choices introduce selection bias. Strong answers mention multiple-testing corrections and honest out-of-sample discipline.

Puzzles

The 100 lockers. 100 lockers start closed. Person k toggles every kth locker for k from 1 to 100. Which lockers end open?

Approach: A locker is toggled once for every divisor of its number, so it ends open when it has an odd number of divisors, which happens exactly for perfect squares - lockers 1, 4, 9, and so on up to 100. Fast recognition of divisor-counting arguments is very much in the house style here.

Coding

Streaming median. Design a structure that supports inserting a number and querying the median, both efficiently.

Approach: A max-heap for the lower half and a min-heap for the upper half, rebalanced so the sizes differ by at most one. Insertion is O(log n) and querying the median is O(1). Interviewers expect this produced cleanly under time pressure; our coding interview questions collection covers the wider repertoire.


A Five-Week Preparation Plan

Week 1: Probability fundamentals. Conditional probability, Bayes' theorem, expectation and combinatorics, drilled daily until the mechanics are automatic rather than effortful.

Week 2: Applied statistics and estimation. Hypothesis testing, confidence intervals, and the specific failure modes of applying standard statistics to financial and noisy data - autocorrelation, overfitting and multiple testing.

Week 3: Machine learning fundamentals (researcher and ML tracks). Bias-variance trade-offs, regularisation, cross-validation with dependent data, and being ready to defend design choices in your own past work under scrutiny.

Week 4: Coding and puzzles. Standard algorithmic practice for engineers; for researchers, a lighter but real coding refresher plus deliberate practice on divisor, parity and symmetry-style puzzles.

Week 5: Mock interviews. Full mock loops under time pressure with a study partner, including the research walk-through format specifically - explaining your own work out loud to a sceptical interviewer is a distinct skill from having done the work.


Common Mistakes Candidates Make

The most frequent failure is treating the online assessment as a formality. It is genuinely hard, and candidates who do not practise timed probability puzzles specifically tend to run out of time on the harder half of the test. The second most common mistake at the onsite is answering the research walk-through defensively rather than analytically - interviewers are testing intellectual honesty about what went wrong in your own work, not looking for a polished highlight reel.


Compensation & recruiting notes

G-Research runs an unusually visible public puzzle programme (the monthly puzzle competition) but is otherwise quite closed about its actual interview loop. Candidates typically report a research-heavy machine learning process for research roles and a systems-heavy loop for engineers, but the exact stage count and question style vary by team and between cycles. This guide is drawn from candidate reports and public sources; nothing here guarantees any specific interview format, question or outcome.


Frequently Asked Questions

How many interview rounds does G-Research have?

Typically three to four stages across three to six weeks: an online assessment, one or two technical phone screens, and an onsite of three to six sessions. Engineering and research tracks follow the same broad shape with different content.

How hard is the G-Research online assessment?

Candidates consistently describe it as one of the harder tests in UK quant recruiting, comparable to top prop firms rather than banks. It covers probability, combinatorics and quick quantitative reasoning under time pressure, with little room for hesitation on the easier questions.

Do I need a PhD to pass a G-Research interview?

Not strictly. A PhD is the norm for the core researcher track, but exceptionally strong master's and undergraduate candidates are hired every cycle, particularly for engineering roles where a doctorate is not expected at all.

What programming languages does G-Research test?

Python is standard across research and much of the platform. C# and C++ have both featured in engineering interviews depending on the team, alongside substantial questions on Kubernetes and large-scale compute for infrastructure roles.

How long does the G-Research hiring process take?

Most candidates report three to six weeks from application to offer, with decisions moving quickly once the onsite is complete. This is faster than the multi-month cycles typical of graduate schemes at large banks.

Should I read the G-Research guide before interviewing?

Yes - our G-Research guide covers the firm's structure, culture and confidentiality expectations in more depth, which is useful context for the onsite and for deciding whether the research-lab environment actually suits you before you accept an offer.

Practise the questions G-Research Interview: Process & Questions 2026 actually asks

Reading about the interview is one thing - sitting one is another. Open your free Quantt prep workspace for a real course lesson plus interview-style coding tests modelled on firms like Jane Street, Citadel, Hudson River and Optiver.

Free lesson + interview practice · No credit card required