Company interview guide

Hudson River Trading Machine Learning Interview Guide (2026)

For ML, AI and research roles: a recycled timed coding test, a probability or programming screen, then a final day of code, systems and statistical depth on very noisy data.

Hudson River Trading's machine learning interview loop has 7 reported stages, from application & first calls to project deep-dive & team fit; typical length: Not well reported; students should allow up to two weeks for application review.

Roles covered: Research Scientist, ML Engineer, Deep Learning Engineer, Applied Scientist and Data Scientist. The question bank holds 37 questions reported from Hudson River Trading interviews, 20 of them free to solve in your browser.

Try a free question

Interview process length
Not well reported; students should allow up to two weeks for application review
Stages
7
Reports linked to bank questions
37
Bank report dates
Jun 2011–Aug 2026
Questions in bank
37 (20 free)
Last reviewed
October 2026

How the Hudson River Trading interview loop works

An application review, a timed online coding test, one or two phone screens, then a full final day of back-to-back rounds that the firm prefers to hold at least partly in person. Research-leaning candidates usually get a 45-minute probability and statistics screen and engineering candidates a 45-60 minute programming screen; some get both. The final day mixes harder programming, a systems or design discussion, a project deep-dive and, for research and modelling roles, statistics and ML depth. Market-making games, microstructure models and mental-arithmetic drills are trader-track items and get no stage here. Only one first-hand ML-role report backs this outline; the rest follows compiled accounts of the firm's research loop and its own hiring pages.

The Hudson River Trading interview process, stage by stage

01Application & first calls

Application review (the firm asks students to allow up to two weeks), then Zoom calls that are either exploratory or already a technical screen

Apply to the single posting that fits best and list other interests in the application questions; the firm says most roles share one pipeline and recruiters can adjust placement later. ML candidates come in through one of three doors: the AI lab's researcher posting, its research-engineer postings for training and inference, or the algorithm developer track, whose interns are matched to teams that include the AI lab. Recruiters set the format of each later round, so ask early.

What it tests

  • Fit between your work and a team: modelling research, training or inference infrastructure, or data
  • Interest in the problems themselves rather than in finance
  • Logistics: posting, office, start date and coding language (C++ or Python)

How to prepare

  • Ask the recruiter which online test and which phone screens your posting uses.
  • Prior finance experience is not expected; lead with your ML, systems or research work.
  • Choose the researcher or algorithm developer door if your strength is modelling, the research-engineer door if it is kernels, data loading or distributed training.

No ML-role report describes this step; the firm's student FAQ and its 2025 engineering post describe the review, email contact and first Zoom calls.

02Online coding test

Timed and auto-graded on HackerRank or CodeSignal: about 150 min for 3 problems on the algorithm developer track, about 70-120 min for 2-4 problems on software tracks; C++ or Python

Compiled accounts call this the biggest cut in the loop. The algorithm developer set draws 3 problems from a small recycled pool of implementation-heavy tasks: counting numbers whose base-4 digits are only 0 and 1, board simulations such as 2048 or Reversi, a trie-based word game, rectangle counting on grid points and tree problems. An August 2026 sitting reportedly kept two pool problems with small changes and added a new tree construction and merge task solved with DFS, the hardest of the three. Software-track sets are shorter classic algorithm problems, such as replaying operations on an in-memory filesystem. Hidden edge cases and clean code are graded, not just the visible tests.

What it tests

  • Turning a precise spec into exact code under a long clock
  • Board and grid state simulation with direction and merge rules
  • Digit and representation tricks in place of an O(n) loop
  • Building trees from serialized input, merging them and DFS
  • Edge cases the sample tests miss

How to prepare

  • Budget about 45 minutes a problem and bank partial passes before sinking time into one.
  • Write your own edge-case tests; the firm says the sample cases are not comprehensive.
  • Drill board simulations and tree merges in a plain editor until they are routine.
  • Do not use AI tools; the firm says cheaters are caught in later rounds and not welcome back.

Free reported questions for this stage

Format and content from compiled accounts of the firm's research and software tests; the firm's student FAQ and 2021 post confirm a timed coding test. No ML-role report yet.

03Probability & statistics screen (research tracks)

About 45 min over video; derivations out loud or on a shared doc, sometimes followed by a quick Python simulation

Compiled accounts describe a conversational screen of expected-value and probability questions derived cleanly without a calculator: gambler's ruin, coupon collecting, waiting times for coin patterns, chord and handshake crossings, Bayesian coin updates. They call several such questions back to back the firm's signature filter, and the same style can return on the final day. A recurring research-track pattern is a statistics question followed by a request to simulate it in Python and check the answer. Software-track candidates usually get the programming screen instead.

What it tests

  • Linearity of expectation and symmetry arguments
  • First-step conditioning and hitting times for simple Markov chains
  • Bayesian updating with a clean sample space
  • Order statistics, the uniform MLE and quick hypothesis tests
  • Checking a derivation with a short Monte Carlo

How to prepare

  • Derive out loud, state assumptions and check a small or limiting case.
  • Pair every practice derivation with a ten-line numpy simulation that confirms it.
  • Be ready to generalise a puzzle from a fixed number to n.

Free reported questions for this stage

Content from compiled accounts of the firm's quant research screen; no ML-role report, and the firm's own pages do not mention a probability round.

04Programming phone screen

45-60 min on CoderPad in C++ or Python; the firm's 2021 guide pairs it with a 45-min technical discussion

A live coding round tuned to the team. Reach a correct first solution, then expect a complexity and optimisation conversation; for core and low-latency teams, C++ specifics such as pointers versus references, ownership and move semantics come up. The firm says it avoids flash-of-insight puzzles and wants real understanding of your tools, and a second screen may be a discussion of systems, data structures or problem solving. Compiled question lists favour one-pass and streaming solutions: selection with few comparisons, the k-th element of two sorted arrays, range-minimum queries, online mean and variance with decay.

What it tests

  • Clean, correct code on a standard algorithmic problem
  • Complexity analysis and an optimisation conversation
  • One-pass and streaming solutions with tight memory
  • Knowing what your language does under the hood

How to prepare

  • Restate the problem and confirm inputs before writing code; the firm lists poor listening as a common failure.
  • State complexity as you go and propose the optimisation before being asked.
  • Say when you need quiet time, change approach or have seen the problem before; the firm stresses honesty.

Free reported questions for this stage

Format from compiled accounts; the firm's student FAQ gives one or two phone screens and its 2021 post a programming round plus a technical discussion. No ML-role report.

05Final day: programming, systems & design

Full day of back-to-back rounds, virtual or in person (the firm prefers some in person): several programming rounds plus a systems or technical design discussion

Programming rounds are harder than the screen and judged on idiomatic, readable, well-encapsulated code and on optimisation. The firm describes coding and debugging rounds and design discussions, sometimes with code in a language you do not know; its 2021 guide adds systems knowledge such as memory, I/O and process management. One 2026 ML-engineer report got a verbally stated simulation (people walking toward a boundary while a turning watcher freezes anyone it sees), solved by brute force, then asked how to handle continuous time by splitting it at the turn points.

What it tests

  • Exact simulation from a spoken spec, then generalising it
  • Debugging and reading unfamiliar code
  • Memory, I/O and process fundamentals behind a design
  • Breaking down an unfamiliar problem and iterating with hints

How to prepare

  • Take notes when a problem is described aloud; one candidate got no written prompt.
  • After a brute force, expect a follow-up that removes the discrete-step assumption or scales the input.
  • Treat each round as a collaboration: take hints and apply earlier feedback to later questions.

One 2026 ML-engineer onsite report (spoken simulation, continuous-time follow-up); compiled accounts and the firm's 2025 and 2021 posts describe the programming, debugging and design rounds.

06Final day: ML & statistics research depth

Research and modelling roles: derivation and discussion of statistics and ML foundations; compiled accounts call the AI research track heavy on ML fundamentals and statistical depth

The firm's AI lab says market data carries very little signal, distributions shift, other traders adapt and events arrive at irregular times, so this round asks whether a result is real. Compiled question lists for the firm cover multiple testing across many backtested strategies (false-discovery control, data-snooping checks), whether one Sharpe ratio truly beats another, forecast-comparison tests, block-bootstrap intervals for dependent data, change-point and CUSUM detection, shrinkage of noisy correlation matrices, closed-form lasso, online ridge updates and the effect of doubling the sample. A smaller ML strand asks how a perturbation moves logistic-regression predictions and why to clip gradients.

What it tests

  • Separating signal from noise when many hypotheses are tested
  • Validation on time-ordered, non-stationary data and backtest pitfalls
  • Regression mechanics: regularisation, shrinkage and online updates
  • Detecting a change in a stream of returns
  • Training stability and regularisation in deep models

How to prepare

  • Rehearse explaining why a strong backtest can be noise and how you would correct for the number of trials.
  • Derive the lasso soft-threshold and a Sherman-Morrison ridge update on paper.
  • Prepare a story about a model that failed out of sample and what you changed.

Free reported questions for this stage

No ML-role report; content from compiled accounts of the firm's research loop and question lists, plus the modelling challenges its AI lab page names.

07Project deep-dive & team fit

A project or behavioural deep-dive and a team-fit component on the final day

The firm rates communication as heavily as technical skill and lists shallow fundamentals, poor listening, over-complex solutions, arrogance, indifference and cheating as common reasons candidates fail. Compiled accounts say the project round carries real weight and probes collaboration, teachability and communication; the firm's 2021 guide asks you to explain past work at the right level of detail.

What it tests

  • Explaining past work clearly to a non-specialist
  • Taking hints and applying earlier feedback
  • Humility, curiosity and genuine interest in the work
  • Honesty about what you know and have seen

How to prepare

  • Pick one or two projects and practise explaining them to someone outside your field.
  • Bring a story with real technical depth, including what failed.
  • Do not use AI tools in any round; the firm says cheaters are not welcome back.

No ML-role report; compiled accounts and the firm's 2025 and 2021 posts describe the project discussion, team fit and failure reasons.

Questions and themes reported most often

  • Simulate people walking past a watcher that turns at set times, then handle continuous time (1 reports, final day: programming, systems & design)
  • Exact grid or board simulation from a precise spec (1 reports, online coding test)
  • Replay a list of operations against an in-memory filesystem (online coding test)
  • Derive a probability, then confirm it with a quick simulation (probability & statistics screen (research tracks))
  • Selection with tight bounds: k smallest, k-th of two sorted arrays, second largest (programming phone screen)
  • Many range queries or updates on a large array (programming phone screen)
  • Bit tricks and memory-bounded search over huge inputs (final day: programming, systems & design)
  • Expected crossings of random chords in a circle (probability & statistics screen (research tracks))

What Hudson River Trading asks about

  • Coding & simulation32%
  • Probability & expectation18%
  • Statistics & inference on noisy data18%
  • Regression & ML fundamentals16%
  • Systems & language internals8%
  • Communication & past work8%

Hudson River Trading interview prep checklist

  • Solved a 2048 or Reversi-style board simulation and a tree merge in under 45 minutes each, with your own edge-case tests
  • Answered five expectation questions by linearity or first-step conditioning and confirmed one by simulation
  • Explained how you would correct a backtest for the number of strategies tried
  • Derived a lasso soft-threshold and implemented an online ridge or decayed-variance update
  • Can explain what your main language costs in memory and time for common operations
  • Practised a project story for a non-specialist, including its validation and what failed
  • Asked the recruiter about language, round types and whether a probability screen applies

Tips

  • Find out early which online test and which phone screen your posting uses; research and engineering doors differ.
  • Treat clean, edge-case-correct code as the main gate: the online test and most final-day rounds are programming.
  • Drill the recycled test style: board simulations, digit-representation counting and tree build-and-merge tasks under a 150-minute clock.
  • For research tracks, pair each probability derivation with a short numpy simulation that checks it.
  • Prepare to argue whether a result is real on noisy, shifting data: multiple testing, time-ordered validation, shrinkage.
  • Review regression and regularisation at derivation level, including online and closed-form updates.
  • Know the internals of C++ or Python well enough to explain memory and performance costs.
  • Think aloud, confirm the problem before solving and say when you need a hint; communication is rated as heavily as code.
  • Never use AI tools in any round; the firm says cheaters are caught later and not welcome back.

See the rest of the Hudson River Trading loop

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

Unlock with Pro

Practice the Hudson River Trading loop

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

Hudson River Trading interview FAQ

How many interview rounds does Hudson River Trading have for ML roles?

Candidates report 7 stages: Application & first calls, Online coding test, Probability & statistics screen (research tracks), Programming phone screen, Final day: programming, systems & design, Final day: ML & statistics research depth and Project deep-dive & team fit. An application review, a timed online coding test, one or two phone screens, then a full final day of back-to-back rounds that the firm prefers to hold at least partly in person. Research-leaning candidates usually get a 45-minute probability and statistics screen and engineering candidates a 45-60 minute programming screen; some get both. The final day mixes harder programming, a systems or design discussion, a project deep-dive and, for research and modelling roles, statistics and ML depth. Market-making games, microstructure models and mental-arithmetic drills are trader-track items and get no stage here. Only one first-hand ML-role report backs this outline; the rest follows compiled accounts of the firm's research loop and its own hiring pages.

How long does the Hudson River Trading ML interview process take?

Reported timelines run Not well reported; students should allow up to two weeks for application review.

What does Hudson River Trading ask in ML interviews?

By reported question mix: Coding & simulation 32%; Probability & expectation 18%; Statistics & inference on noisy data 18%; Regression & ML fundamentals 16%; Systems & language internals 8%; Communication & past work 8%. The most repeated themes: Simulate people walking past a watcher that turns at set times, then handle continuous time; Exact grid or board simulation from a precise spec; Replay a list of operations against an in-memory filesystem.

How should I prepare for the Hudson River Trading ML interview?

Find out early which online test and which phone screen your posting uses; research and engineering doors differ. Treat clean, edge-case-correct code as the main gate: the online test and most final-day rounds are programming. Drill the recycled test style: board simulations, digit-representation counting and tree build-and-merge tasks under a 150-minute clock.

Are there free Hudson River Trading interview questions?

Yes. LeetML has 37 questions reported from Hudson River Trading interviews, and 20 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources