Company interview guide

Tower Research Capital Machine Learning Interview Guide (2026)

For ML, AI and research roles: a team-run loop of probability, regression, research statistics and fast coding, with ML depth set by your team.

Tower Research Capital's machine learning interview loop has 7 reported stages, from online assessment (track-dependent) to team-matched rounds: ml depth and past work; typical length: Not well reported; varies by team.

Roles covered: ML Engineer, Research Scientist, AI Engineer and Deep Learning Engineer. The question bank holds 17 questions reported from Tower Research Capital interviews, 9 of them free to solve in your browser.

Try a free question

Interview process length
Not well reported; varies by team
Stages
7
Reports linked to bank questions
17
Bank report dates
Jun 2017–Jul 2026
Questions in bank
17 (9 free)
Last reviewed
October 2026

How the Tower Research Capital interview loop works

As widely compiled: a track-dependent online test (a 75-minute, four-problem coding test for developer tracks; research candidates more often go straight to a live probability screen), a recruiter call that matches you with a team, a 45-60 minute probability and expectation screen that also checks C++ basics, then rounds run by that team: algorithms and C++ internals, statistics and regression, and depth on your past work. ML candidates should expect the research-statistics questions (overfitting, multiple testing, leakage, shrinkage) to surface inside the statistics and team rounds rather than as a separate interview. There is no market-making game or trading simulation; fast mental-math drills are trader-track only. This outline rests on compiled accounts and the firm's pages; we hold no first-hand ML-role report.

The Tower Research Capital interview process, stage by stage

01Online assessment (track-dependent)

Developer track: HackerRank, 75 min, 4 problems with self-written input and output; some candidates get a dynamic-programming set instead.

Research track: usually a live probability screen in place of a test

The developer and quant-developer test is four medium problems in 75 minutes, reported to open with a short multiple-choice item on global-variable behaviour and then mix a difference-array string problem, a usage tracker ranked with a hash map plus heap and a deterministic tie-break, a stack-based tag parser, and a streaming failure detector. The alternative set leans on dynamic programming, including counting arrangements on a circle. The score leaks come from state, not tricks: replaying every update instead of keeping net counts, stale heap entries, missing tie-breaks. The problems listed here were tagged to software engineering roles at the firm.

What it tests

  • Reducing repeated updates to a difference array or net counts
  • Hash map plus heap ranking with lazy deletion and a fixed tie-break
  • Stack-based parsing of nested structured text
  • Streaming state over time-ordered events without peeking ahead
  • Circle-versus-line counting in a dynamic-programming problem

How to prepare

  • Read all four prompts first, bank the easiest medium inside about 25 minutes, and stop new logic a few minutes before time to test edge cases.
  • Name the invariant before coding: net delta per index, the open-tag stack, the stale-entry rule for the heap.
  • Practise reading standard input yourself so parsing does not eat minutes.

Free reported questions for this stage

Format and content from compiled accounts of the firm's quant research and developer loop; no first-hand ML-role report in our set. Problems here were tagged to software engineering roles.

02Recruiter call and team matching

Phone call on background, interests and fit; sometimes a quick probability or arithmetic warm-up

An introduction to your background and why the firm, which also routes you toward a specific team. Because teams run much of their own process, this is the call that decides what the rest of your loop looks like, and which team you land on matters more to the job than the firm as a whole. A short probability or arithmetic warm-up is sometimes added.

What it tests

  • Why the firm and why this track: ML platform, ML performance, ML research or quant research
  • A crisp two-minute story of your most relevant work
  • Which team fits your background
  • Location, level and timing

How to prepare

  • Ask which team you are being considered for and what its rounds cover; the answer predicts the rest of the loop.
  • Ask whether your later rounds lean research statistics or C++ and systems, and prepare accordingly.
  • Have one ML project ready that touches noisy time series, large data or low latency.

Format and content from compiled accounts of the firm's quant research and developer loop; no first-hand ML-role report. Team routing at this call is consistent across those accounts.

03Phone screen: probability and expectation

Live phone or video round, 45-60 min (about 45 min reported for research screens): several probability problems with follow-ups, plus a short C++ fundamentals check

The core live screen for research tracks and also reported for some engineers. Compiled accounts list coupon collecting, covariance of counts in multinomial draws such as dice faces, order statistics, random-walk distances and hitting times, a branching-process extinction question, the duplicate-the-dataset regression trap and combinatorial counting, each extended with a follow-up. A short C++ check on const, static, virtual, mutable and function pointers, and sometimes a streaming data-structure design question, rides along. The firm's careers page says it avoids gotcha brainteasers, so expect structured probability rather than riddles.

What it tests

  • Expectation of transformed or combined random variables, such as a ratio of uniforms or a clipped normal
  • Covariance of counts and zero correlation versus independence
  • Order statistics, random walks and hitting times
  • Generating a target distribution or fair bits from given random sources
  • Extending a result when the interviewer adds a constraint or sends n to infinity

How to prepare

  • Define the random variable and the conditioning before computing anything.
  • Give a quick estimate first, then derive it exactly and check a small case.
  • Know why E[X/Y] is not E[X]/E[Y] and what Jensen's inequality says about it.
  • Refresh const, static, virtual and mutable in C++ even for a research screen.

Free reported questions for this stage

Format and content from compiled accounts of the firm's quant research and developer loop; no first-hand ML-role report. The careers page says it avoids gotcha questions, which these accounts partly contradict.

04Algorithms and C++ internals

About 60 min per round: hard algorithm problems with a stated complexity target plus C++ internals; Python or C++ depending on the team

Engineering-heavy teams run coding rounds with tight optimal bounds and a streaming flavour: sliding-window median, online top-k sums with arbitrary deletes, shortest paths with unusual edge rules (weights in 0-2, a bounded edge count, state-dependent tolls, alternating moves on a grid), search in a sorted matrix, reverse Polish notation to infix, and the largest all-zero submatrix. C++ internals are a recurring second half: virtual destructors, virtual dispatch and object slicing, the rule of five, smart pointers with custom deleters, breaking shared_ptr cycles, placement new and alignment. Interviewers want the brute force, then the structure that buys the bound, then clean code, and may ask about memory and cache cost.

What it tests

  • Online data structures with deletions: sliding median, top-k under updates
  • Constant-time operations through clever indexing
  • Shortest paths under constrained edge rules
  • C++ object lifetime, virtual dispatch and ownership
  • Stating and meeting a complexity target, then discussing cache cost

How to prepare

  • Code a two-heap sliding median with lazy deletion from memory.
  • State the brute force, then the invariant that buys the bound, then code.
  • For C++ roles, be able to explain what goes wrong without a virtual destructor and how a vector grows and moves elements.

Free reported questions for this stage

Format and content from compiled accounts of the firm's quant research and developer loop; no first-hand ML-role report. Language split taken from the firm's ML postings.

05Statistics and regression round

45-60 min live round, whiteboard or shared doc; research and trading tracks

Research-track rounds return to linear regression under awkward data: what duplicating every row does to estimates and standard errors, fitted values with a redundant feature, equicorrelated predictors, the range of R-squared when combining regressors, whether R-squared can go negative or what changes without an intercept, multicollinearity, omitted-variable bias, ridge versus OLS, and the effect of winsorizing on OLS. Stochastic-process questions on Brownian motion, Markov-chain stationary distributions and hitting times sit in the same round. Interviewers want the math tied back to a signal or a trade.

What it tests

  • OLS in matrix form: duplication, redundant and correlated features
  • R-squared bounds, negative R-squared and the no-intercept case
  • Multicollinearity, omitted-variable bias and ridge shrinkage
  • Least squares under a known constraint
  • Brownian motion and Markov-chain properties, linked to a signal

How to prepare

  • Write OLS in matrix form from memory and reason from the projection view.
  • Know when two weak regressors can give a high joint R-squared, and why.
  • Practise explaining each result in one sentence a trader could act on.

Free reported questions for this stage

Format and content from compiled accounts of the firm's quant research and developer loop; no first-hand ML-role report. Problems here were tagged to research roles at the firm.

06Research judgment on noisy market data

Not a separate interview: questions asked inside the statistics round or team rounds for research-leaning roles

The firm's compiled question pool holds a research-hygiene strand beyond plain regression: how large the maximum of hundreds of null t-statistics gets, how long a backtest must be to reject an implausible Sharpe ratio, cross-validation leakage in financial time series, covariance shrinkage toward a structured target, testing for mean reversion and heavy tails, sequential tests with stopping boundaries, decision rules under a hidden regime, and telling the direction of time in a price series. These test whether you can tell a real signal from an artefact of search, overlap or non-stationarity.

What it tests

  • Multiple testing: the size of the best of many null strategies
  • Backtest length and overfitting to a Sharpe ratio
  • Leakage and time-ordered cross-validation with purging
  • Covariance shrinkage and conditioning
  • Stationarity, mean reversion and heavy tails in returns

How to prepare

  • Know the expected maximum of n independent standard normals grows like sqrt(2 log n), and what that means for strategy search.
  • Explain why random k-fold leaks on overlapping-label time series and how purging and an embargo fix it.
  • Be able to say why a shrunk covariance gives better portfolio weights than the sample one.

Topics from the compiled question pool for the firm's research tracks, not a reported separate round; no first-hand ML-role report.

07Team-matched rounds: ML depth and past work

Several rounds run by your matched team; format and number vary by team

There is no single onsite: the team you matched with runs its own rounds, some leaning probability and statistics, others C++ and systems. The firm's careers page says interviews are a conversation about your background, your past roles and what you could achieve there. For ML roles that conversation goes deep on the systems and models on your resume, measured against the postings: training and serving PyTorch or JAX models at scale, distributed training strategies and communication, GPU kernels and profiling, and, for inference roles, quantization and microsecond-level latency. Research-desk roles add market data, order-book basics and exchange simulation.

What it tests

  • Depth on a model or pipeline you built: data, training, evaluation, deployment
  • Scaling training across GPUs and nodes and finding the bottleneck
  • Inference latency: profiling, quantization, pruning, distillation
  • Working with large, noisy time-series data and market microstructure basics
  • Clear communication and calm when things go wrong

How to prepare

  • Prepare one project with numbers: throughput, latency or model gains, plus what did not work.
  • Rehearse a concrete story for each skill your team's posting lists.
  • Finance knowledge is not required, but know limit versus market orders and what an order book shows.

The firm's careers page describes interviews as a conversation about your background and past roles; team-run rounds from compiled accounts; no first-hand ML-role report.

Questions and themes reported most often

  • Regression traps: duplicated data, redundant features and R-squared bounds (4 reports, statistics and regression round)
  • Data structures with tight bounds: sliding median, random-pop multiset, bounded buffer (3 reports, algorithms and c++ internals)
  • Streaming and log-processing code: failure detection and login statistics (2 reports, online assessment (track-dependent))
  • Zero correlation versus independence, with a counterexample (2 reports, phone screen: probability and expectation)
  • Expectation of a transformed variable: ratio of uniforms, clipped normal (2 reports, phone screen: probability and expectation)
  • Sampling and geometric probability: uniform point in a disk, arc length on a sphere (2 reports, phone screen: probability and expectation)
  • Least squares with a known constraint (1 reports, statistics and regression round)
  • Birthday-style collision probability (1 reports, phone screen: probability and expectation)

What Tower Research Capital asks about

  • Probability & expectation25%
  • Statistics, regression & research validation25%
  • Coding & C++20%
  • ML depth & systems (team rounds)20%
  • Background, markets & fit10%

Tower Research Capital interview prep checklist

  • Solved a set of expectation, order-statistics and random-walk problems aloud with follow-ups
  • Can explain what duplicating a dataset does to OLS and when R-squared can jump or go negative
  • Can explain why shuffled cross-validation leaks on time series and how to purge it
  • Can size the best of many null strategies and say what it means for a backtest
  • Coded a sliding-window median with deletions from scratch, and finished a timed 75-minute, four-problem set with your own input handling
  • Have a deep-dive story on an ML system with latency or throughput numbers
  • Asked the recruiter which team you are matched with and what its rounds test

Tips

  • Ask early which team you are matched with and whether its rounds lean research statistics or C++ and systems; the loop is set largely by the team.
  • Make probability and expectation your first priority: conditioning, covariance of counts, order statistics, random walks and branching processes.
  • Drill regression theory in matrix form: duplicated rows, collinearity, R-squared bounds, the no-intercept case and ridge.
  • Prepare the research-hygiene strand: multiple testing across many strategies, backtest length, leakage in time-series validation and covariance shrinkage.
  • Keep data-structure coding fast and clean, with a complexity target stated before you write code; for C++ roles, know object lifetime and virtual dispatch cold.
  • For ML platform and performance roles, prepare concrete stories on distributed training, GPU profiling and inference latency, with numbers.
  • Think aloud and name the invariant or random variable first; the firm says it wants a conversation, not trick questions.
  • Skip trader-track prep such as market-making games and speed-arithmetic drills unless your recruiter says otherwise.

See the rest of the Tower Research Capital loop

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

Unlock with Pro

Practice the Tower Research Capital loop

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

Tower Research Capital interview FAQ

How many interview rounds does Tower Research Capital have for ML roles?

Candidates report 7 stages: Online assessment (track-dependent), Recruiter call and team matching, Phone screen: probability and expectation, Algorithms and C++ internals, Statistics and regression round, Research judgment on noisy market data and Team-matched rounds: ML depth and past work. As widely compiled: a track-dependent online test (a 75-minute, four-problem coding test for developer tracks; research candidates more often go straight to a live probability screen), a recruiter call that matches you with a team, a 45-60 minute probability and expectation screen that also checks C++ basics, then rounds run by that team: algorithms and C++ internals, statistics and regression, and depth on your past work. ML candidates should expect the research-statistics questions (overfitting, multiple testing, leakage, shrinkage) to surface inside the statistics and team rounds rather than as a separate interview. There is no market-making game or trading simulation; fast mental-math drills are trader-track only. This outline rests on compiled accounts and the firm's pages; we hold no first-hand ML-role report.

How long does the Tower Research Capital ML interview process take?

Reported timelines run Not well reported; varies by team.

What does Tower Research Capital ask in ML interviews?

By reported question mix: Probability & expectation 25%; Statistics, regression & research validation 25%; Coding & C++ 20%; ML depth & systems (team rounds) 20%; Background, markets & fit 10%. The most repeated themes: Regression traps: duplicated data, redundant features and R-squared bounds; Data structures with tight bounds: sliding median, random-pop multiset, bounded buffer; Streaming and log-processing code: failure detection and login statistics.

How should I prepare for the Tower Research Capital ML interview?

Ask early which team you are matched with and whether its rounds lean research statistics or C++ and systems; the loop is set largely by the team. Make probability and expectation your first priority: conditioning, covariance of counts, order statistics, random walks and branching processes. Drill regression theory in matrix form: duplicated rows, collinearity, R-squared bounds, the no-intercept case and ridge.

Are there free Tower Research Capital interview questions?

Yes. LeetML has 17 questions reported from Tower Research Capital interviews, and 9 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources