Company interview guide

GTS Machine Learning Interview Guide (2026)

For ML, AI and research roles: a timed DP and sliding-window coding test, probability and regression rounds, then a final day of the hardest variants.

GTS's machine learning interview loop has 5 reported stages (Application & recruiter screen, Online algorithms test, Probability & expected value, Statistics & regression, Final rounds: hardest variants & research fit); typical length: Not reported.

Roles covered: Research Scientist, Applied Scientist, Data Scientist and ML Engineer. The question bank holds 5 questions reported from GTS interviews, 4 of them free to solve in your browser.

Try a free question

Interview process length
Not reported
Stages
5
Reports linked to bank questions
5
Bank report dates
Nov 2025–Jul 2026
Questions in bank
5 (4 free)
Last reviewed
October 2026

How the GTS interview loop works

An application through the firm's open roles or a general resume submission, a recruiter phone screen, a timed online algorithms test, then a probability and expected-value interview and a statistics and regression interview, each about 45-60 minutes over video, and final rounds made of the hardest problems across all of these topics. ML work at the firm sits inside the quant researcher role, so ML researchers and data scientists usually go through the quant research loop; ML engineers would enter through the core technology (software) track, whose interviews are coding and C++ heavy and less reported. Compiled accounts say there is no market-making game; mental-arithmetic speed is a trader-track emphasis and gets no stage here. Every stage rests on compiled accounts and the firm's own pages rather than first-hand ML-role reports.

The GTS interview process, stage by stage

01Application & recruiter screen

Apply to a posted role or submit a general resume; then a phone call about background and fit, sometimes with a quick technical warm-up

The firm hires quant traders, quant researchers and core technology engineers, plus graduate co-ops, 3-6 month internships and 6-week summer internships; students are pointed to the contact page for program questions. The first call covers why the firm and why this role, and compiled accounts say it can include a short probability or arithmetic warm-up, so do not treat it as purely logistics.

What it tests

  • Why a market maker, and why research rather than trading
  • How your modelling or ML work maps to pricing, forecasting or risk models
  • Comfort with a quick spoken probability or arithmetic question
  • Logistics: track, office, start date and coding language

How to prepare

  • Ask which track you are being considered for (quant research or core technology); it decides the later rounds.
  • Warm up basic expectation and arithmetic before the call and think aloud if a question comes.
  • Have a one-minute summary of a model you built, validated and saw used.

No first-hand report; the firm's careers page describes applying to open roles or by general resume and its student programs; the call itself comes from compiled accounts.

02Online algorithms test

Timed and auto-graded, one track; platform, length and question count are not reported

Compiled accounts describe a coding test built from sliding-window, two-pointer and dynamic-programming problems. Named examples include the cheapest roof that covers k of the parked cars (reported in 2024), maximum-sum circular subarray, a rolling rank inside a moving window, maximum profit with at most k trades and splitting an array into exactly m segments to maximise a score. The bar is code that is correct on edge cases and linear time where possible.

What it tests

  • Sliding window and two pointers over sorted positions
  • DP with a clearly defined state: at most k trades, exactly m segments
  • Kadane-style subarray maximisation, including wrap-around
  • Order statistics over a moving window

How to prepare

  • Name what each DP state means before writing code; it is the fastest way to avoid off-by-one errors.
  • For window problems, decide what the window invariant is and when it shrinks.
  • Write your own edge cases: all negatives, k larger than n, empty windows.

Free reported questions for this stage

Format and problem families from compiled accounts of the firm's research and software tests; one tagged question; no first-hand ML-role report.

03Probability & expected value

About 45-60 min over video; derivations spoken or on a shared document

The largest block in compiled question lists for the firm. Expect to price simple games by expected value and to reason about beliefs: the guess-the-next-card game whose answer is a harmonic number (reported in 2025), coin pattern races such as HTH against HHT, whether five heads in a row should make you doubt a fair coin, a capped coin-flip game, the St Petersburg paradox under log utility, the gap between two random cut points and when two top seeds meet in a knockout bracket.

What it tests

  • Linearity of expectation with indicator variables
  • First-step conditioning for waiting times and pattern races
  • Bayesian updating from a short run of data
  • Clean counting and symmetry arguments
  • Expected utility versus expected value

How to prepare

  • Estimate first, then formalise; say the rough number before the exact one.
  • Reach for indicators and symmetry before brute-force enumeration.
  • State the prior explicitly in any is-the-coin-fair question.

Content from compiled accounts and question lists for the firm's quant research loop; no first-hand ML-role report and no firm page mentions this round.

04Statistics & regression

About 45-60 min over video; short regression and estimation brainteasers answered with reasons

The stage closest to ML. Signature questions ask what happens to an OLS fit when you duplicate every data point (slope and R-squared stay, standard errors shrink by about a factor of root two and t-statistics grow), when you shift or rescale the data, and whether pooling two datasets that each have a positive slope must give a positive slope. Others recover a correlation from the variance of a sum, size a sample with the central limit theorem, and estimate the informed share of order flow with a two-component mixture fitted by EM.

What it tests

  • OLS mechanics: which outputs change under duplication, shifting and scaling
  • Why duplicated or dependent rows fake statistical significance
  • Simpson-style reversals when datasets are pooled
  • Variance and correlation algebra
  • Mixture models and EM for a latent class

How to prepare

  • For each transformation, walk through slope, intercept, R-squared, standard error and t-statistic in that order.
  • Explain the why: duplication adds no information, so any gain in significance is fake.
  • Be ready to write the E and M steps for a two-class mixture on paper.

Free reported questions for this stage

Content from compiled accounts of the firm's quant research loop and four tagged questions; no first-hand ML-role report.

05Final rounds: hardest variants & research fit

Final interviews that bring back the hardest versions of every earlier topic; number and length of rounds not reported

Compiled accounts describe the final rounds as the toughest problems across probability, statistics, coding and combinatorics rather than new topics: conditional probability for uniforms on a simplex, maximum profit with at most k trades, the largest number with a digit property below n, completing a sequence with no three-term arithmetic progression and the EM estimate of informed flow. Engineering-leaning candidates also see C++ questions such as virtual dispatch and object slicing, and classic data-structure tasks. Research roles at the firm own their models end to end, so expect to discuss how you would build, ship and measure one.

What it tests

  • Hard probability with geometric or conditional arguments
  • Harder DP and combinatorics under time pressure
  • C++ object model basics on the engineering side
  • Explaining a past model: from first principles to live impact
  • Composure under layered follow-ups

How to prepare

  • Expect the hard version of something you already practised; review the earlier stages rather than new topics.
  • Keep a clear chain of reasoning when follow-ups stack; restate what has been established.
  • Connect the math back to pricing, forecasting or risk on the desk.

Content from compiled accounts of the firm's final rounds; research-ownership framing from the firm's research postings; no first-hand ML-role report.

Questions and themes reported most often

  • Duplicate every row of a regression: what changes and what does not (statistics & regression)
  • Shift or rescale the data, then track slope, R-squared and t-statistics (statistics & regression)
  • Two datasets each have a positive slope: must the pooled slope be positive? (statistics & regression)
  • Recover a correlation from the spread of a sum (statistics & regression)
  • Sample size from the central limit theorem, and the effect of more data (statistics & regression)
  • Estimate a hidden informed share of flow with a two-class mixture and EM (statistics & regression)
  • Maximum-sum subarray variants, including wrap-around (online algorithms test)
  • Order statistics over a moving window: rolling rank or median (online algorithms test)

What GTS asks about

  • Statistics & regression intuition26%
  • Probability & expected value24%
  • Coding: DP & sliding window24%
  • Modelling & research discussion14%
  • C++ & CS fundamentals6%
  • Fit & communication6%

GTS interview prep checklist

  • Explained what duplicating, shifting and scaling data do to slope, R-squared, standard errors and t-statistics
  • Showed with an example why two positive slopes can pool into a negative one
  • Solved a sliding-window cover problem and an at-most-k-trades DP in linear or near-linear time
  • Priced the guess-the-next-card game and a coin pattern race without enumeration
  • Wrote the E and M steps for a two-class mixture on paper
  • Practised a project story: model, time-ordered validation, deployment, measured impact
  • Asked the recruiter which track and which round types apply

Tips

  • Ask early whether you are on the quant research or core technology track; ML roles sit inside research.
  • Treat the statistics round as your strongest card: know OLS outputs under duplication, shifting, scaling and pooling cold.
  • Drill sliding-window and DP problems until you can state the state or invariant before coding.
  • Answer probability questions with indicators and conditioning, and estimate before you formalise.
  • Tie every statistics answer back to validation: leakage, dependent samples and overstated significance.
  • Prepare one model story from first principles to measured live impact, including what decayed or failed.
  • Expect final rounds to be harder versions of earlier topics, not new ones.
  • Engineering-leaning candidates should review C++ object model basics and latency costs.

See the rest of the GTS loop

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

Unlock with Pro

Practice the GTS loop

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

GTS interview FAQ

How many interview rounds does GTS have for ML roles?

Candidates report 5 stages: Application & recruiter screen, Online algorithms test, Probability & expected value, Statistics & regression and Final rounds: hardest variants & research fit. An application through the firm's open roles or a general resume submission, a recruiter phone screen, a timed online algorithms test, then a probability and expected-value interview and a statistics and regression interview, each about 45-60 minutes over video, and final rounds made of the hardest problems across all of these topics. ML work at the firm sits inside the quant researcher role, so ML researchers and data scientists usually go through the quant research loop; ML engineers would enter through the core technology (software) track, whose interviews are coding and C++ heavy and less reported. Compiled accounts say there is no market-making game; mental-arithmetic speed is a trader-track emphasis and gets no stage here. Every stage rests on compiled accounts and the firm's own pages rather than first-hand ML-role reports.

How long does the GTS ML interview process take?

Reported timelines run Not reported.

What does GTS ask in ML interviews?

By reported question mix: Statistics & regression intuition 26%; Probability & expected value 24%; Coding: DP & sliding window 24%; Modelling & research discussion 14%; C++ & CS fundamentals 6%; Fit & communication 6%. The most repeated themes: Duplicate every row of a regression: what changes and what does not; Shift or rescale the data, then track slope, R-squared and t-statistics; Two datasets each have a positive slope: must the pooled slope be positive?.

How should I prepare for the GTS ML interview?

Ask early whether you are on the quant research or core technology track; ML roles sit inside research. Treat the statistics round as your strongest card: know OLS outputs under duplication, shifting, scaling and pooling cold. Drill sliding-window and DP problems until you can state the state or invariant before coding.

Are there free GTS interview questions?

Yes. LeetML has 5 questions reported from GTS interviews, and 4 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources