The Voleon Group is a machine learning hedge fund founded in 2007 by Michael Kharitonov and Jon McAuliffe and headquartered in Berkeley, California, with offices in New York and London. It reported about $23bn in assets in early 2026 and employs roughly 430 to 530 people, many of them with doctorates in statistics, computer science and mathematics. Its research postings say that experience in finance is not a primary factor in hiring. In 2025 it added Voleon Securities, a market making business that now hires its own traders and researchers.
The academic culture runs through the interview. Candidates describe questions on eigenvalues, the pseudoinverse and the bias-variance decomposition written out in expectations, asked by researchers who want the right answer rather than a discussion.
No timed online assessment is reported for quant research or trading candidates. The first technical stage is a live 1-hour screen on statistics, linear algebra, probability and machine learning, followed by a 2-round virtual onsite on statistics and machine learning. If your invitation names a coding platform, expect Python with pandas, not mental maths.
- 1Recruiter call. About 30 minutes on Zoom to confirm your background.
- 2Technical screen. 1 hour with a researcher on statistics, linear algebra, probability and brainteasers.
- 3Virtual onsite. 2 back-to-back rounds, one on statistics and one on machine learning, sometimes the next day.
- 4Further rounds. Coding and research questions, sometimes a presentation for PhD candidates, then references.
Voleon sends candidates an interview prep guide. Despite what the prep material says, expect interviewers to care about reaching the right answer, not only about your process. Know the material rather than hoping to reason toward it.
Roles for students and graduates
| Role | Who | Notes |
|---|---|---|
| Trading Intern | Undergraduates. The summer 2026 posting asked for graduation in 2027 | Shadow trading desks across regions and asset classes, Bloomberg and automation projects. Wants Python or R, Linux and a prior financial internship |
| Associate Member of Trading Staff | Graduates, Berkeley and a UK version | Monitor execution algorithms and portfolio risk, work with brokers, contribute code. Python or R, Unix shell and SQL. $115,000 to $160,000 base posted |
| Research Intern | PhD students. The London posting targets PhDs finishing in December 2026 or 2027 | Predictive models and components of automated trading systems |
| Member of Research Staff | PhDs | $250,000 to $275,000 base posted. Reinforcement learning and causal inference teams at Voleon Securities |
| Member of Technical Staff | Researchers and engineers | No current posting found |
| Software Engineer Intern, summer 2027 | Undergraduates and graduates | Open from September 2026 |
| Quantitative Trader, Voleon Securities | Experienced hires only | Requires 2 or more years of institutional trading experience and licences |
The trading seats are execution and trading operations roles, not market making seats. The market making trader role at Voleon Securities is not open to new graduates. Postings are rolling, on Ashby.
Stage 1: Technical screen
A 60-minute call with a researcher, often a PhD. Earlier candidates describe it as a triage interview on maths, statistics and probability. The screen can open with "explain a sorting algorithm of your choice" and runs about 3 questions.
Linear algebra and matrices
- Properties of doubly stochastic matrices and their eigenvalues.
- A matrix whose rows and columns all sum to 1: must it be square, and can it be nilpotent?
- A pseudoinverse question. Fitting a plane through nearly coplanar 3D points with the SVD, with and without noise.
Regression and statistics
- What range R-squared can take when y is regressed on x1 and on x2 separately.
- The closed form of OLS and when it exists, and what to do when the design matrix does not fit in memory.
- The slope of Y on X against the slope of X on Y. Testing a median with no distributional assumptions. p-values and the OLS assumptions.
- A basic regression question that jumps to one needing a good understanding of matrices, variable rotation and the coefficients.
Probability and brainteasers
- Puzzles solved by scaling and symmetry arguments, and a geometry question about points on a 2D grid.
- Two heads in a row before the other player's first tail. Conditional probability with biased coins. The expected number of rolls to reach a multiple of 6.
- A conditional expectation of a normal given a sum, coin and dice problems, and Markov transition matrix properties.
How to prepare
Work the multivariate statistics section until eigenvalues, covariance matrices and the conditional Gaussian are automatic, then OLS and multiple regression with the matrix algebra written out. The Markov chain lesson covers the dice and coin processes, and the probability questions the puzzles.
Stage 2: Virtual onsite
Two interviews back to back, one on statistics and one on machine learning, each with a researcher. They can come as soon as the day after the screen. Earlier versions ran as two 2-hour sessions.
“Define bias and variance in expectation terms.”
What it testswhether you can write the decomposition rather than describe it.
Candidates report this question almost word for word. For an estimator of at a point, the expected squared error over training sets is : squared bias, variance, and irreducible noise. The follow-ups candidates report are what regularisation does to each term, and why training error far below test error points at the variance term. Say what happens to each term with more data and with a stronger penalty, and you have answered the next 2 questions before they are asked.
What else the machine learning round covers
- Training error far below test error, and what regularisation and more data do to both.
- How a regression tree is trained and what the best split costs to find. Choosing k in k-nearest neighbours.
- Ridge against lasso, perturbing a logistic regression, and k-means from scratch.
- Feature leakage in a time series pipeline, asked in a machine learning engineer screen.
- An applied case: a fake dataset, how you would clean it, which model you would fit and why. How to choose which horse to bet on from the history of races.
- A construction task for data scientists: generate a series with a given normalised error, correlation and rank correlation against a target.
How to prepare
The cross-validation lesson and choosing a model family on noisy data cover the model selection questions, decision trees and bagging the tree cost, and ridge, lasso and elastic net the regularisation. For leakage, why k-fold fails on time-ordered data and feature and hyperparameter leakage are the 2 lessons to read.
Stage 3: Further rounds
Earlier candidates describe one more round of coding and research questions after the 2 onsites. Coding is in Python. Reported problems include grouping anagrams, pandas aggregation with file input and output, an order matching engine, and a trailing-volume tracker with participation-rate execution. Software engineers sit a live HackerRank CodePair screen at around mid LeetCode hard, under an NDA. About a tenth of research candidates mention a presentation, so PhD candidates may give a research talk, though none describes what it involves.
New hires may be assigned to groups somewhat randomly, so do not count on a particular team. References are checked late, and candidates have been rejected at that stage without a reason.
The trading track
The Associate Member of Trading Staff process currently runs to 5 rounds. Round 1 covers probability, statistics, your resume and finance. Round 2 has a LeetCode medium problem and a machine learning case study. Then come onsite interviews. In London that means a screening, 3 technical interviews, a final interview and references.
Earlier candidates fill in the content. One trading associate's first round was with a current associate with 1 to 2 years of experience, on dividend payment specifics, Python data types and SQL concepts. Equity trader candidates can get brainteasers, maths and a pandas exercise under an hour to report basic statistics on a financial dataset. Junior traders may be asked what markets they trade and given a code project. No market making game appears in any report.
How to prepare
Equity mechanics down to how dividends are paid, Python data types, SQL, 1 LeetCode medium in Python, and 1 machine learning case you can structure out loud. The markets course covers the mechanics, and the coding practice catalog the Python.
Key tips for success
- Know it or do not: candidates say interviewers want the correct answer and give few hints. Learn the linear algebra and regression results to recall, not to derive on the spot.
- Write the bias-variance decomposition, the OLS solution and R-squared out by hand before the screen.
- Prepare for silence. Say your plan before you compute, so a wrong turn is visible to you without a prompt.
- Have 1 applied case ready: a messy dataset, the cleaning, the model, the validation.
- Keep your references warm. They are checked, and candidates have fallen at that stage.
Closing remarks
Voleon interviews like a statistics department with a trading floor attached. There is no speed test, no market making game and little finance. There are eigenvalues, regression theory, the bias-variance decomposition and tree and nearest-neighbour methods, asked precisely by people who know the answers. Candidates who prepare it as a course and can state results cleanly do well. Candidates who expect to be led through a puzzle do not.
Work through everything we have for Voleon, and read our quant researcher interview process guide for how a machine learning fund's process compares with a market maker's.