Company interview guide
Headlands Technologies Machine Learning Interview Guide (2026)
For ML, AI and research roles: a 70-minute statistics and C++ test, streaming-data coding, and regression and modeling questions on noisy market data.
Headlands Technologies's machine learning interview loop has 6 reported stages (Application & recruiter screen, Online assessment: statistics and coding, Live coding interview: C++ and streaming, Probability & expectation, Statistics, regression & ML modeling, Final rounds: hardest variants & fit); typical length: Often under two weeks from test to decision; results after each round usually arrive within one or two days.
Roles covered: Research Scientist, Data Scientist, ML Engineer and Applied Scientist. The question bank holds 7 questions reported from Headlands Technologies interviews, 2 of them free to solve in your browser.
- Interview process length
- Often under two weeks from test to decision; results after each round usually arrive within one or two days
- Stages
- 6
- Reports linked to bank questions
- 7
- Bank report dates
- Jul 2026–Jul 2026
- Questions in bank
- 7 (2 free)
- Last reviewed
- October 2026
How the Headlands Technologies interview loop works
Apply to a single posting, with transcripts and short written answers, then a brief recruiter call. Research candidates take a 70-minute online test of 17 questions in four timed blocks (statistics multiple choice, statistics free response, one C++ problem, two problems in any of Python, Java or C++); developer candidates take a coding test of about 90 minutes with two problems, the first in C++. Passing candidates get one or more live video interviews: a 60-minute coding round with an engineer, and 45-60 minute rounds on probability and on statistics, regression and modeling. Final rounds bring the hardest variants of each topic plus fit questions. Quant researchers, the usual route for ML researchers, are weighted to statistics and modeling; developers, the usual route for ML engineers, get deeper C++ and systems. Mental-math drills and betting-style games are trader-track only. The outline rests on few reports: candidate accounts confirm the online test well but say little about later rounds.
The Headlands Technologies interview process, stage by stage
01Application & recruiter screen
Online application with transcripts and short written answers, then a brief call on background and fit, sometimes with a quick technical warm-up
The firm asks you to apply to one posting only and reviews each candidate individually. Research postings ask for recent transcripts, a GPA field, competing offers and deadlines, notice period or non-compete, your top three programming languages with a self-rating, and short written answers on motivation, fit, strengths and weaknesses. Compiled accounts describe a short intro call that can include a quick technical warm-up.
What it tests
- Why the firm and why systematic research or engineering
- Evidence that you have applied statistics or ML to noisy real-world data
- Honest self-rating of your programming languages
- Logistics: office (Amsterdam, Chicago, London or New York), offers, notice period
How to prepare
- Pick the posting with care, since you may apply to only one: quantitative researcher for ML research, C++ or Java developer for ML engineering.
- Rate your languages honestly; the C++ block in the test will show the truth.
- Write the motivation answers about turning research into results in production, which is what the posting stresses.
- Warm up basic probability and arithmetic in case the call includes a quick question.
Application rules and questions come from the firm's careers page and research postings; the call itself is from compiled accounts only, with no first-hand report describing it.
02Online assessment: statistics and coding
Research track: one 70-minute timed sitting of 17 questions in four blocks (10 statistics multiple choice in 10 min, 4 statistics free response in 24 min, 1 C++ problem in 20 min, 2 problems in Python, Java or C++ in 16 min).
Developer track: about 90 min, two coding problems, the first in C++
For research candidates the statistics blocks decide most outcomes: reviewers describe the multiple-choice items as multi-select and harder than the coding, which runs from easy to a medium follow-up. Developer tests center on table and stream processing, such as an as-of join over timestamped rows exposed through a struct API; an older version was one long C++ problem with deliberately wordy instructions, reconstructing job chains from CSV rows and rejecting malformed lines. Compiled problem themes include top-K by notional, maximum concurrent orders and a streaming queue with mean and snapshots.
What it tests
- Basic inference under time pressure: estimators, tests, p-values, terminology
- Explaining a statistical answer in a short free-response box
- Clean C++ for a small problem in 20 minutes
- Streaming and table processing: as-of joins, running aggregates, top-K
- Defensive parsing of messy input
How to prepare
- Budget about one minute per multiple-choice item and six minutes per free-response question.
- Warm up C++ before the sitting: the single C++ problem is where candidates lose the most time.
- Review statistics vocabulary and standard quantities; reviewers say terminology, not depth, caught them.
- On developer tests, read the whole prompt twice and handle malformed lines before you optimize.
Free reported questions for this stage
- Sliding Window Maximum with a Monotonic DequeCoding & AlgorithmsMediumSolve
- K-th Smallest Element in a Sorted MatrixCoding & AlgorithmsMediumSolve
Five quant-research reports from 2025-2026 describe the statistics plus coding test, one giving 17 questions in 70 minutes; two developer reports (2020, 2026) give the C++ coding test. Block timings from compiled accounts.
03Live coding interview: C++ and streaming
About 60 min by video with an engineer; mostly C++, with follow-ups on performance and edge cases
A short problem grows into a discussion of how fast and safe your code is. One candidate summed 1 to n without the formula, then had to compare recursive, iterative and tail-recursive versions, reason about overflow and stack depth, and suggest compile-time evaluation, before designing a fast map from short stock tickers to integers. Research-tagged items in compiled accounts include packing a ticker into a 64-bit integer, a configurable C++ filter over a stream of order structs that must compile and pass tests, and an O(1) moving average. An engineering report from 2026 paired a non-adjacent max-sum DP with a lock-free data structure question.
What it tests
- C++ fundamentals: integer overflow, recursion limits, compile-time evaluation
- Hashing and memory layout for small fixed-size keys
- Object-oriented design that stays configurable and testable
- O(1)-per-update streaming aggregates
- Explaining expected-time arguments and when they break
How to prepare
- Give the simplest correct answer first, then expect to be pushed on speed, memory and overflow.
- Know how to pack up to eight ASCII characters into a uint64 and why that beats string hashing.
- Practise a rolling mean and variance that stays numerically stable over long streams.
- Think aloud: interviewers judge the reasoning as much as the final code.
One developer report describes a 60-minute coding round with an engineer and one quant intern report confirms a coding phone screen; research-track content from compiled accounts.
04Probability & expectation
45-60 min by video, a chain of probability and expected-value questions with follow-ups that extend each result
Compiled accounts list classic puzzles with a quantitative bent: optimal stopping (keep or reroll a die up to three times, stop on the maximum of uniforms), waiting time for three heads in a row, random walks to the edge of a square, simulating a 10-sided die with a 6-sided one, card ordering, splitting fresh and stale items between two boxes, and a three-way duel. Order statistics and waiting times recur, and each answer is extended with a follow-up.
What it tests
- Expectation by conditioning and recursion
- Optimal stopping and dynamic programming over outcomes
- Order statistics of uniforms
- Markov chains and random walks with absorbing states
- Rejection sampling and constructing distributions from coins or dice
How to prepare
- Estimate the answer first, then set up the recursion out loud.
- Know the expected maximum and minimum of n uniforms and the standard waiting-time results by heart.
- Expect a follow-up that changes one parameter; solve generally so it is a substitution.
- Skip mental-math speed drills; they belong to the trader track.
Format and topics from compiled accounts of the quant research loop; four quant-role reports from 2025-2026 confirm live phone or video interviews after the test but not what was asked.
05Statistics, regression & ML modeling
45-60 min by video: rapid-fire regression and inference questions, then open questions on building and debugging predictive models on noisy, time-ordered data
The deepest round for ML researchers. Inference questions include two studies with the same p-value but very different sample sizes, a correlation that flips sign once the sample is restricted, low R-squared alongside a highly significant coefficient, ordinary versus total least squares, ridge versus lasso (closed form, sparsity, factor models), an AR(1) with unknown mean and a unit-root test, the sample size needed to detect a given Sharpe ratio, and normal-normal Bayesian updating. Modeling questions from research reports cover features and target both shifting between training and live trading, adding a feature with a short history, a model that underperforms live versus its backtest, combining per-second and per-day features, how lasso, ridge, trees and neural nets extrapolate, clustering all-categorical data, and research when data arrives 500 ms late.
What it tests
- Effect size versus significance, and selection or collider bias
- Linear-model diagnostics: R-squared, collinearity, errors in variables
- Regularization and shrinkage, and when each penalty helps
- Covariate and concept drift; backtest-to-live gaps and leakage
- Feature engineering across horizons and update frequencies
How to prepare
- Name the estimator, its bias and its variance before giving an answer.
- For any live-underperformance question, walk a checklist: leakage and timestamps, costs and latency, regime change, overfitting from many trials, then pipeline bugs.
- Know why lasso zeroes coefficients and ridge does not, and derive the ridge closed form on paper.
- When features update at different rates, say what value is known at each decision time before proposing a model.
Topics from compiled accounts of the quant research loop, including several research-role modeling questions; two candidate reports mention ML questions alongside statistics, without detail.
06Final rounds: hardest variants & fit
Several back-to-back interviews covering the hardest versions of each earlier topic, plus motivation and fit; developer candidates get deeper C++ and systems design
Compiled accounts describe the final rounds as the hardest version of everything: a sliding-window median in O(log k), the shortest subarray with sum at least S when values can be negative, an order matching engine, merging K sorted streams with bounded concurrency, and hard probability such as meeting on a grid. Developer candidates are also probed on virtual dispatch, shared_ptr internals and thread safety, shared-memory inter-process communication, recent C++ standards and a large-scale code-analysis design. Expect questions on motivation, strengths and weaknesses like those in the application. Few candidate reports reach this stage, so confirm the format with your recruiter.
What it tests
- Hard streaming data structures with tight complexity bounds
- Monotonic deques, heaps and balanced structures under pressure
- C++ object model, memory ownership and concurrency (developer track)
- Hard probability and game-theory follow-ups
- Motivation and fit for a research-driven team
How to prepare
- Have the two-heap sliding median with lazy deletion and the prefix-sum monotonic deque ready from memory.
- Stay composed when a follow-up breaks your first approach; restate the invariant and adapt.
- Prepare one research story told the firm's way: hypothesis, test, what failed, what reached production.
- Ask your recruiter how many interviews the final round has and whether it is in person.
Shape and content from compiled accounts of the quant research loop; one 2025 report of an other-role final stage focused on behavioral questions.
Questions and themes reported most often
- Timed statistics block: multi-select definitions and short free-response answers (4 reports, online assessment: statistics and coding)
- As-of join of timestamped rows through a struct API, first problem in C++ (1 reports, online assessment: statistics and coding)
- Reconstruct job chains from CSV rows, summarize them and reject malformed lines (1 reports, online assessment: statistics and coding)
- Sum 1 to n without the formula, then overflow, stack depth and compile-time options (1 reports, live coding interview: c++ and streaming)
- Fast mapping from short stock tickers to integers (1 reports, live coding interview: c++ and streaming)
- Streaming statistics at O(1) per update: moving average, running mean and covariance (live coding interview: c++ and streaming)
- Top-K by notional and maximum concurrent orders over a stream (online assessment: statistics and coding)
- Same p-value, very different sample sizes: what the result really says (statistics, regression & ml modeling)
What Headlands Technologies asks about
- Statistics & regression24%
- ML modeling on noisy, time-ordered data14%
- Algorithms & streaming data structures26%
- Probability & expectation16%
- C++ & systems (developer track)12%
- Research discussion & fit8%
Headlands Technologies interview prep checklist
- Completed a timed mock of 10 statistics multiple-choice items in 10 minutes and 4 free-response items in 24
- Solved a small problem in C++ in under 20 minutes without autocomplete
- Wrote an as-of join and a sliding-window median from memory
- Can explain same p-value at different sample sizes and a correlation flip under selection
- Can derive ridge, explain lasso sparsity and say how each model class extrapolates
- Have a checklist for a model that underperforms live: leakage, costs, regime, overfitting, bugs
- Prepared one research story from hypothesis to production, including what failed
Tips
- Treat the 70-minute test as the main gate: drill basic statistics definitions under a clock until a multiple-choice item takes under a minute.
- Get C++ warm enough to solve a small problem in 20 minutes, even if you plan to use Python elsewhere.
- Practise online estimators (rolling mean, variance, covariance, median) until they are routine; they recur in the test, the live coding and the finals.
- For every modeling answer, state what is known at decision time; timestamp leakage is the trap behind several reported questions.
- Prefer simple, regularized, well-diagnosed models in discussion; the firm's research postings stress noisy data and results in production.
- Read the core statistical learning chapters on linear methods and on model assessment and selection, which the firm's blog recommends to new researchers.
- Expect quick decisions after each round, and plan your other processes around a short timeline.
- Skip market-making games and mental-math speed drills; there is no trading game in this loop.
See the rest of the Headlands Technologies loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Headlands Technologies loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Headlands Technologies interview FAQ
How many interview rounds does Headlands Technologies have for ML roles?
Candidates report 6 stages: Application & recruiter screen, Online assessment: statistics and coding, Live coding interview: C++ and streaming, Probability & expectation, Statistics, regression & ML modeling and Final rounds: hardest variants & fit. Apply to a single posting, with transcripts and short written answers, then a brief recruiter call. Research candidates take a 70-minute online test of 17 questions in four timed blocks (statistics multiple choice, statistics free response, one C++ problem, two problems in any of Python, Java or C++); developer candidates take a coding test of about 90 minutes with two problems, the first in C++. Passing candidates get one or more live video interviews: a 60-minute coding round with an engineer, and 45-60 minute rounds on probability and on statistics, regression and modeling. Final rounds bring the hardest variants of each topic plus fit questions. Quant researchers, the usual route for ML researchers, are weighted to statistics and modeling; developers, the usual route for ML engineers, get deeper C++ and systems. Mental-math drills and betting-style games are trader-track only. The outline rests on few reports: candidate accounts confirm the online test well but say little about later rounds.
How long does the Headlands Technologies ML interview process take?
Reported timelines run Often under two weeks from test to decision; results after each round usually arrive within one or two days.
What does Headlands Technologies ask in ML interviews?
By reported question mix: Statistics & regression 24%; ML modeling on noisy, time-ordered data 14%; Algorithms & streaming data structures 26%; Probability & expectation 16%; C++ & systems (developer track) 12%; Research discussion & fit 8%. The most repeated themes: Timed statistics block: multi-select definitions and short free-response answers; As-of join of timestamped rows through a struct API, first problem in C++; Reconstruct job chains from CSV rows, summarize them and reject malformed lines.
How should I prepare for the Headlands Technologies ML interview?
Treat the 70-minute test as the main gate: drill basic statistics definitions under a clock until a multiple-choice item takes under a minute. Get C++ warm enough to solve a small problem in 20 minutes, even if you plan to use Python elsewhere. Practise online estimators (rolling mean, variance, covariance, median) until they are routine; they recur in the test, the live coding and the finals.
Are there free Headlands Technologies interview questions?
Yes. LeetML has 7 questions reported from Headlands Technologies interviews, and 2 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- Hudson River TradingFor 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.
- Jump TradingFor ML, AI and research roles: a timed coding test, a probability and linear algebra screen, then a final day built around live time-series modelling, with no trading games.
- Tower Research CapitalFor ML, AI and research roles: a team-run loop of probability, regression, research statistics and fast coding, with ML depth set by your team.
- Radix TradingFor ML, AI and research roles: hard algorithms, numerical analysis, geometric probability and classical statistics in a quant research loop.
- Virtu FinancialFor ML, AI and research roles: a coding test gate, quick brain teasers, probability and regression phone rounds, then a logic-test onsite and a modeling final.
- Two SigmaInterview guide for ML, AI and research roles: a numeric HackerRank assessment, then data science, coding and regression-heavy statistics rounds, three manager rounds and a senior closer.