Company interview guide

Radix Trading Machine Learning Interview Guide (2026)

For ML, AI and research roles: hard algorithms, numerical analysis, geometric probability and classical statistics in a quant research loop.

Radix Trading's machine learning interview loop has 7 reported stages, from application & recruiter screen to final day & senior conversation; typical length: About 4-8 weeks; decision about 1-2 weeks after the final day.

Roles covered: Research Scientist, ML Engineer and Data Scientist. The question bank holds 12 questions reported from Radix Trading interviews, 4 of them free to solve in your browser.

Try a free question

Interview process length
About 4-8 weeks; decision about 1-2 weeks after the final day
Stages
7
Reports linked to bank questions
12
Bank report dates
Sep 2020–Jul 2026
Questions in bank
12 (4 free)
Last reviewed
October 2026

How the Radix Trading interview loop works

Application with transcripts, a short recruiter screen, an online test of two or three hard algorithm problems in about 90-120 minutes, then a 60-90 minute live C++ and algorithms round that often mixes in probability. Probability and numerical methods, linear algebra and optimization, and regression and research discussion each take 45-60 minutes. The process ends with a single in-person final day of five to seven interviews: the hardest variants of every topic, systems design for technologists, a research-style problem and a conversation with a senior member of the firm. Researcher-track candidates, the usual route for ML researchers, lean on Python, statistics and signal design; technologist-track candidates, the usual route for ML engineers, get deeper C++, memory and low-latency systems questions. There is no trading or market-making game, and mental-math speed drills and betting games are trader-track prep that does not apply here. The outline rests on few reports: we hold no first-hand ML-role account, so ask your recruiter which rounds your track has.

The Radix Trading interview process, stage by stage

01Application & recruiter screen

Online application with transcripts, then a short phone call on background and fit, sometimes with a quick technical warm-up

The application asks for an unofficial transcript, GPA for technologist roles, outstanding offers and their deadlines, and optional publication or Google Scholar links. You may apply to and interview for only one role at a time, and the firm says it will contact you if another role fits better. The call covers why the firm and why this role; compiled accounts say a short technical warm-up can follow.

What it tests

  • Why the firm and why a research or technologist role
  • How your ML or statistics work maps to finding real signal in noisy data
  • Quick technical basics answered aloud
  • Logistics: track, office (Chicago, New York or Amsterdam), deadlines

How to prepare

  • Choose the track before you apply: quantitative researcher for ML research, C++ technologist for ML engineering; parallel interviews are not allowed.
  • List competing offers and deadlines up front; the form asks for them.
  • Warm up basic arithmetic and probability so a quick warm-up question does not catch you cold.
  • Ask which of the later rounds apply to your track, including whether you take the online test.

Application rules (transcripts, offers, one role at a time, possible redirection) are on the firm's postings; the call itself comes from compiled accounts of the quant research loop only.

02Online assessment: hard algorithms

Timed online test, about 90-120 min, two or three hard algorithm problems; clearest for the technologist track

An algorithm-heavy test gates the live rounds. Problems tagged to the firm are classic hard data-structure tasks: range add with range min using lazy propagation, rank and select on a dynamic set, and union area of rectangles with a sweep line. Compiled accounts frame the test as engineering-first, so ML engineers on the technologist track should expect it; whether every researcher candidate takes it is unconfirmed.

What it tests

  • Segment trees with lazy updates
  • Augmented balanced trees for order statistics
  • Sweep line with coordinate compression
  • Dynamic programming and two-pointer patterns at hard difficulty
  • Correct, efficient code under a tight clock

How to prepare

  • Have a lazy segment tree with range add and range min you can write from memory.
  • Practise in the language you will use on the day; C++ is the house language.
  • State the complexity before coding and check it against the input bounds.
  • Budget time per problem; two solid solutions beat three partial ones.

Format and length from compiled accounts of the firm's loop, framed as engineering-first; no first-hand ML-role report. The three problems are tagged to engineering roles.

03C++ & algorithms round

60-90 min live coding by video with an engineer or researcher; usually C++, often with probability or algorithm questions mixed in

The first live round is C++ implementation plus algorithm reasoning: a reference-counted smart pointer, a fixed-buffer allocator with its own malloc and free, accurate summation of values with very different magnitudes, and the minimum comparisons to find the second largest element. Short warm-ups such as reversing a linked list also appear. Technologist candidates get a systems-design discussion on top; researcher candidates more often get probability alongside the coding.

What it tests

  • C++ internals: ownership, reference counting, memory layout and allocation
  • Numerically stable summation (compensated or pairwise)
  • Comparison-count lower bounds and tournament arguments
  • Clean live coding of standard data-structure operations
  • Explaining correctness and complexity as you go

How to prepare

  • Be able to write a reference-counted pointer with copy, move and destruction handled correctly.
  • Know why naive summation loses precision and code Kahan summation from memory.
  • Prove that max plus second max needs n + ceil(log2 n) - 2 comparisons, do not just quote it.
  • Expect follow-ups on memory, edge cases and thread safety even on easy warm-ups.

Free reported questions for this stage

Content and length from compiled accounts of the firm's quant research loop; no first-hand ML-role report. Both problems placed here are tagged to engineering roles.

04Probability & numerical

45-60 min by video; probability, combinatorics and statistics with a numerical bent

Probability is the largest topic among problems tagged to the firm, and most have a geometric or symmetry twist: the length of a random chord, whether three uniform lengths form a triangle, identifying a poisoned item in a limited number of tasting rounds, finding a heavy bean with a balance. The numerical bent shows in questions such as computing the variance of n plus a uniform draw without catastrophic cancellation, or sampling a returning random walk path uniformly.

What it tests

  • Geometric probability and choosing the right uniform model
  • Symmetry, parity and conditioning arguments
  • Information-counting bounds for weighing and group-testing puzzles
  • Stable variance computation and uniform sampling of combinatorial objects
  • Checking answers on small or extreme cases

How to prepare

  • For geometric questions, state the random model first; the answer depends on it.
  • Look for symmetry or an invariant before you integrate anything.
  • Count how many outcomes each test can separate to get lower bounds on rounds.
  • Know Welford's update and why subtracting two large nearby numbers destroys precision.

Free reported questions for this stage

Stage shape and themes from compiled accounts of the firm's quant research loop; no first-hand ML-role report. Three problems are tagged to the firm's research and trading roles, 2022-2026.

05Linear algebra & optimization

45-60 min research-style derivations, sometimes with a little coding

A distinctive block: inverting a matrix and what it costs, eigenvalues and conditioning of structured matrices such as a cyclic permutation or an equicorrelation matrix, the eigenvalue, determinant and trace identities, QR via Householder reflectors, and writing a sum of absolute values as a linear program. Stability is a theme here too, for example whether a recurrence for an integral is stable run forwards or backwards. Optimization extends to optimal stopping and the Kelly fraction for an uneven bet.

What it tests

  • Matrix inversion, factorizations and their cost
  • Eigenvalues, conditioning and structured matrices
  • L1 minimization as a linear program
  • Forward versus backward stability of recurrences
  • Optimal stopping and bet sizing

How to prepare

  • Know why you solve a system instead of forming an inverse, and what the condition number predicts.
  • Practise inverting an equicorrelation matrix and finding the eigenvalues of a permutation matrix by hand.
  • Turn absolute values into linear constraints with slack variables.
  • When a recurrence blows up, check whether running it in the other direction damps the error.

Free reported questions for this stage

Stage shape and themes from compiled accounts of the firm's quant research loop; no first-hand ML-role report. Three problems are tagged to the firm's research roles, 2023-2026.

06Statistics, regression & research discussion

45-60 min discussion and derivations, often as the research-style problem on the final day; may involve Python

The ML-heavy part of the loop. Problems tagged to the firm cover when the closed-form OLS solution exists and the Gauss-Markov conditions; compiled question sets add shrinkage estimators for a correlation matrix, scoring rules and calibration for classifiers that drive trades, testing whether two Sharpe ratios differ when returns are serially correlated, and detecting outliers in a stream. The firm's postings ask for quick discovery of real signal, classical statistics applied rigorously and fast abandonment of low-value ideas, so expect to defend every modelling and validation choice.

What it tests

  • OLS derivation, rank conditions, collinearity and Gauss-Markov
  • Shrinkage and regularization when estimates are noisy
  • Calibration and proper scoring rules for probabilistic classifiers
  • Hypothesis tests under serial correlation and non-stationary data
  • Overfitting, look-ahead leakage and time-ordered validation of a signal

How to prepare

  • Derive OLS from the normal equations and say exactly what breaks with collinear features.
  • Explain shrinkage as trading bias for variance and know how to pick the intensity.
  • Know why log loss and Brier score are proper and accuracy is not, and how to check calibration.
  • Bring one project where you found a real effect in noisy data and can explain how you ruled out a spurious one.

Themes from compiled question sets for the firm's quant research loop plus the skills its research postings list; no first-hand ML-role report. One regression problem is tagged here.

07Final day & senior conversation

One in-person day of five to seven interviews, ending with a senior member of the firm; decision about 1-2 weeks later

The final day draws the hardest variants of every topic above: hard coding, hard probability and stochastic processes such as a random walk escaping to infinity, hard linear algebra, a research-style statistics problem, and for technologists a systems-design round on low-latency code, memory and caches. It closes with a senior conversation on motivation and fit. The postings stress persistence, adaptability, self-awareness, a team mindset and dropping low-value work.

What it tests

  • Staying rigorous across five to seven back-to-back rounds
  • Hard variants and follow-ups on earlier topics
  • Low-latency systems design: memory model, false sharing, cache behaviour (technologist track)
  • Motivation for research at a trading firm and how you prioritize

How to prepare

  • Plan your energy for a full day; keep each answer structured even late in the day.
  • Expect the interviewer to push past your first answer with a harder follow-up.
  • Have a concrete example of abandoning a project that was not paying off.
  • Ask how researchers and technologists split work on the shared C++ stack.

Day shape, round count and decision timing from compiled accounts of the firm's loop; traits from its postings; no first-hand ML-role report.

Questions and themes reported most often

  • Geometric probability from uniform lengths: random chords and triangle formation (2 reports, probability & numerical)
  • Range-query trees: lazy segment tree and order-statistics tree (2 reports, online assessment: hard algorithms)
  • Fewest comparisons to find the largest and second largest (1 reports, c++ & algorithms round)
  • Numerical stability: summing mixed magnitudes, stable variance, unstable recurrences (1 reports, probability & numerical)
  • Union area of axis-aligned rectangles with a sweep line (1 reports, online assessment: hard algorithms)
  • Inverting a matrix: cost, conditioning and eigenvalue identities (1 reports, linear algebra & optimization)
  • Least absolute deviations as a linear program (1 reports, linear algebra & optimization)
  • Find the poisoned item with few tasting rounds (1 reports, probability & numerical)

What Radix Trading asks about

  • Algorithms & coding24%
  • Probability & combinatorics22%
  • Linear algebra & numerical methods20%
  • Statistics, regression & research discussion20%
  • C++ & low-latency systems (ML engineer track)9%
  • Fit & senior conversation5%

Radix Trading interview prep checklist

  • Coded a lazy segment tree with range add and range min in under 30 minutes
  • Wrote a reference-counted smart pointer and a fixed-buffer allocator in C++
  • Proved the n + ceil(log2 n) - 2 comparison bound for the two largest elements
  • Wrote Kahan summation and Welford's variance and explained catastrophic cancellation
  • Solved the random chord and uniform-triangle problems and justified the model choice
  • Inverted an equicorrelation matrix and wrote an L1 fit as a linear program
  • Derived OLS, explained the rank condition and when shrinkage helps
  • Prepared a research story about finding a real signal in noisy data and validating it out of sample

Tips

  • Pick the track deliberately: quantitative researcher for ML research, C++ technologist for ML engineering; you cannot interview for two roles at once.
  • Prepare a quant research loop with ML as a plus: algorithms, probability and linear algebra come before deep learning.
  • Write hard data structures from memory: lazy segment trees, order-statistics trees and sweep lines.
  • Treat numerical analysis as a core topic: compensated summation, stable variance, conditioning and recurrence stability.
  • For probability puzzles, state your model and look for symmetry or an invariant before computing.
  • Be ready to prove lower bounds and complexity, not only quote them.
  • For the research round, defend validation on noisy, serially correlated data without look-ahead, and explain shrinkage and calibration.
  • ML engineers: rehearse C++ ownership, allocators, false sharing and cache behaviour, plus a low-latency data pipeline design.
  • Expect a single long final day of five to seven interviews and plan your energy.
  • Skip mental-math speed drills and market-making games; there is no trading game in this loop.

See the rest of the Radix Trading loop

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

Unlock with Pro

Practice the Radix Trading loop

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

Radix Trading interview FAQ

How many interview rounds does Radix Trading have for ML roles?

Candidates report 7 stages: Application & recruiter screen, Online assessment: hard algorithms, C++ & algorithms round, Probability & numerical, Linear algebra & optimization, Statistics, regression & research discussion and Final day & senior conversation. Application with transcripts, a short recruiter screen, an online test of two or three hard algorithm problems in about 90-120 minutes, then a 60-90 minute live C++ and algorithms round that often mixes in probability. Probability and numerical methods, linear algebra and optimization, and regression and research discussion each take 45-60 minutes. The process ends with a single in-person final day of five to seven interviews: the hardest variants of every topic, systems design for technologists, a research-style problem and a conversation with a senior member of the firm. Researcher-track candidates, the usual route for ML researchers, lean on Python, statistics and signal design; technologist-track candidates, the usual route for ML engineers, get deeper C++, memory and low-latency systems questions. There is no trading or market-making game, and mental-math speed drills and betting games are trader-track prep that does not apply here. The outline rests on few reports: we hold no first-hand ML-role account, so ask your recruiter which rounds your track has.

How long does the Radix Trading ML interview process take?

Reported timelines run About 4-8 weeks; decision about 1-2 weeks after the final day.

What does Radix Trading ask in ML interviews?

By reported question mix: Algorithms & coding 24%; Probability & combinatorics 22%; Linear algebra & numerical methods 20%; Statistics, regression & research discussion 20%; C++ & low-latency systems (ML engineer track) 9%; Fit & senior conversation 5%. The most repeated themes: Geometric probability from uniform lengths: random chords and triangle formation; Range-query trees: lazy segment tree and order-statistics tree; Fewest comparisons to find the largest and second largest.

How should I prepare for the Radix Trading ML interview?

Pick the track deliberately: quantitative researcher for ML research, C++ technologist for ML engineering; you cannot interview for two roles at once. Prepare a quant research loop with ML as a plus: algorithms, probability and linear algebra come before deep learning. Write hard data structures from memory: lazy segment trees, order-statistics trees and sweep lines.

Are there free Radix Trading interview questions?

Yes. LeetML has 12 questions reported from Radix Trading interviews, and 4 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources