Company interview guide

The Voleon Group Machine Learning Interview Guide (2026)

For 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.

The Voleon Group's machine learning interview loop has 9 reported stages, from recruiter contact and reading list to references and background check; typical length: Not well reported; plan for several weeks.

Roles covered: Research Scientist, ML Engineer, Data Scientist and Applied Scientist. The question bank holds 31 questions reported from The Voleon Group interviews, 13 of them free to solve in your browser.

Try a free question

Interview process length
Not well reported; plan for several weeks
Stages
9
Reports linked to bank questions
31
Bank report dates
Jul 2019–Sep 2026
Questions in bank
31 (13 free)
Last reviewed
October 2026

How the The Voleon Group interview loop works

Compiled accounts of the firm's quant research loop describe a recruiter contact (often with a reading list), a 60-minute statistics and ML phone screen, a 30-45 minute culture discussion with a manager (added in 2025), then a virtual onsite in two roughly 2-hour sessions: a statistics round and an ML round with hands-on data first, a coding round and a research deep dive second. Three references and a background check close it. Engineering-track candidates, which can include ML engineers, start with a 1-1.5 hour coding test and face a harder coding onsite with system design; data-science candidates do pandas, regression and statistics work. Our two first-hand ML-role reports (2025, 2026) cover only the screen and show a newer applied format. There are no trading games or mental-math tests at the firm; its market-flavoured content is coding (a trailing-volume tracker, an order-matching engine).

The The Voleon Group interview process, stage by stage

01Recruiter contact and reading list

Short recruiter call; research candidates are often sent a reading list before the screen

Hiring is described as largely recruiter-sourced and small in headcount. Research candidates are often sent a reading list: The Elements of Statistical Learning, Bishop's Pattern Recognition and Machine Learning, All of Statistics and an algorithms text (CLRS). Treat it as the syllabus, because later rounds ask textbook results at derivation depth. Data-science candidates go from this call to a pandas, regression and statistics screen.

What it tests

  • Fit between your background and the research, ML engineering or data track
  • Whether your statistics and ML depth matches a PhD-level research bar
  • Logistics: track, level, timeline and whether an online test applies to you

How to prepare

  • Ask which track you are on; research, engineering and data candidates get different screens and onsites.
  • If a reading list arrives, work the linear-regression, classification and model-assessment chapters with pen and paper.
  • Ask whether the screen will be statistics-led or hands-on so you can weight your preparation.

No first-hand ML-role report covers this call; the format and the reading list come from compiled accounts of the firm's quant research loop.

02Online assessment (engineering and data tracks)

Engineering: auto-graded coding test or take-home, about 1-1.5 hours, 5 problems, stdin/stdout.

Data analyst: 1 hour, 5 questions (4 SQL, 1 pandas). Research candidates usually skip it

The engineering test mixes classics (group anagrams, parsing an input file and totalling quantities, a stack with a bulk-increment operation, the largest square sub-grid whose blocks stay under a budget) with a market-data simulation: track each ticker's trading volume over a trailing 60-second window and use it for participation-rate execution. The data-analyst version is four SQL items and one pandas aggregation and filter task. Compiled accounts call the bar high and rejections quick.

What it tests

  • Correct stdin/stdout programs that pass hidden test cases
  • Sliding time windows with per-key state over an ordered event stream, with exact boundary handling
  • Lazy updates and precomputation instead of repeated rescans
  • Pandas group-by, filter and aggregate chains, plus SQL on the data track

How to prepare

  • Clear the easy items quickly so the trailing-volume simulation gets most of the time.
  • Decide whether a trade exactly 60 seconds old is in or out before coding, then test that boundary.
  • Keep one pointer or deque per ticker; recomputing the window for every query is the usual timeout.
  • For data roles, drill pandas groupby, merge and boolean filtering until they are routine.

Free reported questions for this stage

No first-hand ML-role report; format, length and problems from compiled accounts of the firm's engineering and data tracks, which say research candidates usually skip it.

03Technical screen

Research phone about 60 min (about 10 intro, 40 technical, 10 for questions); 2025-2026 ML screens ran up to 90 min, concepts then live hands-on work; engineering phone 45-60 min of stdin/stdout coding

Three versions are on record. The research phone follows the reading list: derive the OLS closed form and say when the inverse of X^T X exists, fit least squares by gradient descent in plain numpy, combine R-squared values, compare the Y-on-X and X-on-Y slopes, find E[X | X+Y] for jointly normal variables, plus doubly stochastic matrices, a small-sample median test and perhaps one numpy task. A 2025 research-staff screen instead spent the first half of 90 minutes on LLM and RL concepts, then live-coded a tool-calling flow against an LLM API; a 2026 ML engineer screen was a hands-on audit of a leaky time-series pipeline. The engineering phone is one hard coding problem with an optimisation follow-up, most often an expression evaluator that grows from addition to parentheses and inter-dependent variables.

What it tests

  • OLS derivation, invertibility, cost and a gradient-descent fit in numpy
  • Regression-range facts: product of reverse slopes, R-squared bounds, normal conditional expectations
  • LLM fine-tuning on question-answer pairs, memory accounting, mixture-of-experts versus dense, MDPs and Bellman equations
  • Live implementation of an LLM tool-calling loop
  • Time-series leakage, walk-forward validation and train/inference parity

How to prepare

  • Derive OLS three ways (normal equations, projection, gradient descent) and code the gradient-descent version with numpy only.
  • Have a fine-tuning memory ledger ready: weights, gradients, optimizer state, activations, and what LoRA removes.
  • Practise wiring an LLM API call to a tool and back end to end in a plain editor; the 2025 candidate failed on exactly this.
  • In a pipeline audit, check every feature against the prediction timestamp before touching hyperparameters.

Free reported questions for this stage

Two first-hand ML-role screen reports (2025 research staff, 2026 ML engineer) with different content; the research-phone and engineering-phone formats come from compiled accounts.

04Culture discussion with a manager

About 30-45 min with a technical or hiring manager; added in 2025 on both research and engineering tracks

A conversation that compiled accounts describe as a real gate, usually between the screen and the onsite: your opinions on engineering and ML practice, a project you are proud of, a decision you drove with data, and your weaknesses. Candidates describe the firm's interviewers overall as reserved and sparing with feedback, so do not read silence as a bad sign.

What it tests

  • Defensible opinions on ML and engineering practice
  • Ownership of a project you are proud of
  • A data-driven decision told with numbers
  • Self-awareness about weaknesses

How to prepare

  • Prepare two or three reasoned positions, for example on validation discipline, code review or when a simple model beats a complex one.
  • Tell the data-driven decision story with the metric, the baseline and the outcome.
  • Name a real weakness and what you do about it; a vague answer reads as poor self-awareness.

No first-hand ML-role report; compiled accounts of the firm's research and engineering loops describe this conversation since 2025.

05Onsite part 1: statistics and regression round

First of two roughly 2-hour virtual sessions; a statistics PhD interviews, derivations on a shared doc, sometimes live data work in R or pandas

The centre of gravity of the research loop. Derive the OLS estimator, state when it exists and what it costs, go through the assumptions, what breaks when each fails and how you would diagnose it, and answer sharp regression puzzles: the range of the reverse-regression slope, bounds on joint R-squared from pairwise correlations, a regression after rotating two uniforms, least absolute deviations through the origin, a hyperplane fitted to noisy points. Matrix facts (unit row and column sums) and the expectation of one normal given a sum also recur. Interviewers push on edge cases rather than accept a recited result.

What it tests

  • OLS closed form, invertibility of X^T X and computational cost
  • Reverse-regression slopes and R-squared bounds from correlations
  • Assumption violations and diagnostics
  • Conditional expectation for jointly normal variables
  • Least-squares variants: absolute deviations, total least squares

How to prepare

  • Know why the product of the two regression slopes is the squared correlation and what that bounds.
  • State assumptions out loud before each derivation; the follow-ups target the case where one fails.
  • Practise small-sample tests without distributional assumptions, such as a sign test for a median.
  • Practise in pandas or R as well as on paper: fit, read coefficients and p-values, then explain them.

Free reported questions for this stage

No first-hand ML-role report; format and content from compiled accounts of the firm's quant research loop; the placed problems were reported in earlier research interviews there.

06Onsite part 1: ML round with hands-on data

Same session as the statistics round; a same-field PhD interviews, ML theory at derivation depth plus live 'run the data' work in R or pandas

Classical ML from first principles: how K in nearest neighbours and the number of cross-validation folds move bias and variance, the cost of finding the best split when growing a regression tree, how a perturbation shifts logistic-regression predictions, the logistic gradient and the optimal step size from the Hessian, why L1 and L2 penalties rather than an L1/2 penalty, k-means from scratch and least squares when the design matrix does not fit in memory. The hands-on part fits a regression, reads p-values and debugs a deliberately odd data setup live. A 2026 data-science account adds a construction task: produce predictions with a set normalised MSE, a Pearson correlation near 0.5 and a rank correlation near 0.99.

What it tests

  • Bias-variance reasoning for KNN and k-fold choices
  • Tree-building complexity and optimisation step sizes
  • Regularisation: why convex L1 and L2 penalties
  • Out-of-core least squares
  • Live regression fitting, p-values and data debugging

How to prepare

  • Work out complexities explicitly: sorting cost per feature, scan per split, depth.
  • Write k-means and gradient-descent least squares in plain numpy until you can do it without notes.
  • When the data look strange, say what you check first (duplicates, time order, units, leakage) and why.
  • Know how Pearson and rank correlation come apart under a strong monotone transform.

Free reported questions for this stage

No first-hand ML-role report; format and content from compiled accounts of the firm's quant research loop; one 2026 ML-engineer screen used a similar hands-on style.

07Onsite part 2: coding round

Second roughly 2-hour session; live coding with a CS-background interviewer, clean and refactored code graded.

Engineering track instead: 4 rounds (about 1 hour plus 2.5 hours of coding, then design)

Research candidates get one coding round in the second session, where structure counts as well as correctness. The engineering track, which ML engineers may be on, is much harder: an eight-direction word search over a grid for many words (trie plus depth-first search with pruning to avoid timeouts), a four-part order-matching engine (limit, then market, then stop, then cancel orders, with heaps), a sparse-matrix class with add and multiply or hash-map internals on a whiteboard, and system design such as point-to-country lookup on a huge map, dependency resolution or a distributed crawler. Cache designs and numeric methods have also been reported.

What it tests

  • Data-structure design with clean interfaces
  • Stateful simulations extended part by part
  • Numeric methods from scratch, such as Newton's method
  • Complexity analysis and an optimisation follow-up
  • Spatial and dependency design on the engineering track

How to prepare

  • Refactor as each part arrives; a tidy class beats a long function.
  • For the matching engine, keep bid and ask books as heaps and handle a stop order that triggers another stop.
  • Prune the word search by prefix and remove found words from the trie.
  • Practise O(1) cache designs and their eviction edge cases.

Free reported questions for this stage

No first-hand ML-role report; formats from compiled accounts of the firm's research and engineering loops; placed problems come from earlier engineering-track interviews.

08Onsite part 2: research deep dive

Same session as the coding round; a senior researcher probes your PhD or main project

A rigorous discussion of your thesis or main project: why that method, what the baselines were, how you validated it, what failed and what you would do next. Compiled accounts say fit weighs heavily here and in the loop overall.

What it tests

  • Depth and ownership of your own research
  • Experimental rigour and validation choices
  • How you reason when challenged on a weak point

How to prepare

  • Prepare a 10-minute walkthrough and defend every modelling choice.
  • Know the weakest point of your work before the interviewer finds it.
  • Link your methods to noisy, low-signal prediction where that is honest.

No first-hand ML-role report; from compiled accounts of the firm's quant research loop.

09References and background check

Three references plus a background check, on all tracks

Compiled accounts describe thorough reference calls, with detailed probing of research claims and awards, and a background check, and say candidates who passed every technical round have been rejected at this step.

What it tests

  • Consistency of your claims with what references say
  • A clean, verifiable record

How to prepare

  • Line up three strong references early who can speak to depth and rigour.
  • Brief references on the work you highlighted in interviews.

No first-hand report; compiled accounts of all the firm's tracks describe three references and a background check. Likely but unconfirmed for ML roles.

Questions and themes reported most often

  • Fine-tune an LLM on question-answer pairs and estimate the memory it needs (1 reports, technical screen)
  • Mixture-of-experts versus dense transformers (1 reports, technical screen)
  • Define an MDP and write the Bellman equations (1 reports, technical screen)
  • Live-code an LLM tool-calling loop that answers a current-weather question (1 reports, technical screen)
  • Audit a time-series pipeline for leakage, tune it with walk-forward validation, then add inference (1 reports, technical screen)
  • Derive OLS, say when it exists, and compute it out of core (onsite part 1: statistics and regression round)
  • Regression-range puzzles: Y on X versus X on Y, bounds on R-squared (onsite part 1: statistics and regression round)
  • KNN, cross-validation folds and tree-building cost through bias and variance (onsite part 1: ml round with hands-on data)

What The Voleon Group asks about

  • Statistics & regression28%
  • ML theory & LLM concepts22%
  • Hands-on data & ML pipelines18%
  • Coding16%
  • Probability & linear algebra8%
  • Research deep dive & behavioral8%

The Voleon Group interview prep checklist

  • Derived OLS by normal equations and gradient descent, with when X^T X is invertible
  • Solved the reverse-regression slope, R-squared bound and E[X | X+Y] puzzles
  • Explained KNN and k-fold bias-variance, regression-tree cost and the Hessian step size
  • Wrote a fine-tuning memory estimate for full fine-tuning and LoRA
  • Built a working LLM tool-calling loop against a real API
  • Audited a time-series pipeline for leakage and set up walk-forward validation
  • Implemented a trailing-window volume tracker and LRU and LFU caches cleanly
  • Prepared a defensible research deep dive and confirmed three references

Tips

  • Public detail is thin: only the screen has first-hand ML reports, so prepare for both the classic statistics phone and the newer applied screen.
  • Make regression your strongest topic: derivations, assumptions, reverse regression and R-squared bounds.
  • Work the reading list with pen and paper if you get one; questions are textbook results at derivation depth.
  • For applied ML roles, revise LLM fine-tuning, memory accounting, mixture-of-experts and RL basics, and practise an LLM tool-calling loop.
  • Rehearse a time-series audit: prediction timestamp, lagged features, walk-forward splits, train/inference parity.
  • Be hands-on in R or pandas: fit a regression, read p-values and debug odd data live.
  • Write clean, refactored code in every coding round; structure is graded, and engineering-track problems read stdin and write stdout.
  • Expect a reserved, low-feedback interview style; keep reasoning out loud anyway.
  • Line up three strong references early; compiled accounts say the final checks reject candidates who passed.

See the rest of the The Voleon Group loop

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

Unlock with Pro

Practice the The Voleon Group loop

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

The Voleon Group interview FAQ

How many interview rounds does The Voleon Group have for ML roles?

Candidates report 9 stages: Recruiter contact and reading list, Online assessment (engineering and data tracks), Technical screen, Culture discussion with a manager, Onsite part 1: statistics and regression round, Onsite part 1: ML round with hands-on data, Onsite part 2: coding round, Onsite part 2: research deep dive and References and background check. Compiled accounts of the firm's quant research loop describe a recruiter contact (often with a reading list), a 60-minute statistics and ML phone screen, a 30-45 minute culture discussion with a manager (added in 2025), then a virtual onsite in two roughly 2-hour sessions: a statistics round and an ML round with hands-on data first, a coding round and a research deep dive second. Three references and a background check close it. Engineering-track candidates, which can include ML engineers, start with a 1-1.5 hour coding test and face a harder coding onsite with system design; data-science candidates do pandas, regression and statistics work. Our two first-hand ML-role reports (2025, 2026) cover only the screen and show a newer applied format. There are no trading games or mental-math tests at the firm; its market-flavoured content is coding (a trailing-volume tracker, an order-matching engine).

How long does the The Voleon Group ML interview process take?

Reported timelines run Not well reported; plan for several weeks.

What does The Voleon Group ask in ML interviews?

By reported question mix: Statistics & regression 28%; ML theory & LLM concepts 22%; Hands-on data & ML pipelines 18%; Coding 16%; Probability & linear algebra 8%; Research deep dive & behavioral 8%. The most repeated themes: Fine-tune an LLM on question-answer pairs and estimate the memory it needs; Mixture-of-experts versus dense transformers; Define an MDP and write the Bellman equations.

How should I prepare for the The Voleon Group ML interview?

Public detail is thin: only the screen has first-hand ML reports, so prepare for both the classic statistics phone and the newer applied screen. Make regression your strongest topic: derivations, assumptions, reverse regression and R-squared bounds. Work the reading list with pen and paper if you get one; questions are textbook results at derivation depth.

Are there free The Voleon Group interview questions?

Yes. LeetML has 31 questions reported from The Voleon Group interviews, and 13 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources