Company interview guide

Two Sigma Machine Learning Interview Guide (2026)

Interview 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.

Two Sigma's machine learning interview loop has 7 reported stages, from online assessment to final: senior manager; typical length: 3-8 weeks.

Roles covered: Research Scientist, Data Scientist, Applied Scientist and ML Engineer. The question bank holds 100 questions reported from Two Sigma interviews, 52 of them free to solve in your browser.

Try a free question

Interview process length
3-8 weeks
Stages
7
Reports linked to bank questions
100
Bank report dates
Jun 2014–Jul 2026
Questions in bank
100 (52 free)
Last reviewed
October 2026

How the Two Sigma interview loop works

An online assessment usually arrives first (often soon after applying or a referral) and passing it leads to a recruiter call. Three live technical rounds of about 45-60 minutes follow, in recent reports in this order: an open-ended data science round, a coding round and a statistics and math round; some candidates meet the data round alone as a screen. Three hiring-manager rounds with different teams and a conversational senior-manager round close the loop. Older reports describe one or two phone screens and then a virtual onsite of about four interviews. The firm interviews over video and may end the day early if the first rounds show no fit. Market-making games and mental-math speed tests are trader-track only and do not appear in research or ML loops. ML-engineer loops are thinly documented beyond the assessment.

The Two Sigma interview process, stage by stage

01Online assessment

HackerRank, auto-graded.

Research and data track: 3 problems (interpolator, temperature regression, lasso extension), reported windows of 1 to about 3 hours. Engineering track: 2 problems in about 75 minutes from a small pool, refreshed late 2025

A hard gate before the recruiter call, and the stage candidates describe most consistently. The research and data set has barely changed since about 2019: a piecewise-linear interpolator with extrapolation, a regression of one city's daily temperature on nearby towns with a search for the most predictive towns, and an extension that picks five towns by lasso; a 2026 report saw a batch-then-streaming no-intercept regression instead of the lasso part. The engineering set draws two problems from a pool (allocating IPO shares by price then timestamp round robin, cutting a tree edge to balance subtree sums, distance triples in a tree, peak concurrent intervals, budgeted pipeline throughput); a late-2025 refresh added counting arrays from fixed differences and splitting a wildcard bracket string. Older research sets used connected components and longest chains.

What it tests

  • Piecewise-linear interpolation over sorted knots, with extrapolation, sort plus binary search
  • Regression on tabular data in pandas: fit, compare MSE, greedy or exhaustive input selection, lasso by coefficient size
  • Streaming or online updates of a regression from running sums
  • Engineering pool: ordered allocation simulations, subtree sums on parent arrays, interval sweeps, binary search on a budget
  • Hidden-test edge cases and efficiency

How to prepare

  • Ask which set you will get; research and data titles should drill interpolation and regression code, engineering titles the algorithm pool.
  • Reduce each wordy prompt to one invariant before coding (allocation order, subtree complement, one valid range for the first value).
  • Handle points outside the range, ties and empty inputs explicitly; candidates say hidden cases decide the score.
  • Sort once and reuse the order; repeated sorting is a common cause of timeouts.

Free reported questions for this stage

Assessment in 13 candidate reports from 2023-2026 (12 research or data, 1 ML engineer), research versions consistently interpolation and regression. Pool and timing details from compiled accounts of the research loop.

02Recruiter call

Short phone or video call, usually after the assessment; some 2026 candidates also had an HR chat after the technical rounds

A recruiter confirms background, motivation and which title fits (research, data science, modelling or engineering), then schedules the technical rounds. Compiled accounts place this call after a passed assessment, which matches most recent reports; one 2026 candidate met a recruiter before two researcher interviews, and another had a separate HR round after the technical rounds.

What it tests

  • Fit between your background and a research, data science or engineering title
  • Motivation for systematic investing and for the firm
  • Logistics: timeline, location, level

How to prepare

  • Ask which title and team you are being considered for; later rounds follow the title.
  • Have a two-minute summary of one data-heavy project ready, including what did not work.

A recruiter or HR conversation in 3 research-role reports from 2023-2026; position varies. Compiled accounts put the call after the assessment.

03Technical round 1: open-ended data science

45-60 min video round with a researcher, shared screen; often the first live round and sometimes a standalone screen

The exploratory round. You get a fuzzy problem, such as forecasting bike counts, judging whether a house price is reasonable, finding a signal in return data or validating someone else's research finding, and then a long run of follow-ups. Interviewers probe inputs, outputs, features and how each assumption changes the model more than model sophistication. Expect discussion across regression, ML basics and time-series intuition; a standard workflow (check stationarity, fit a baseline regression, inspect residuals, then look for non-linearity) is a good spine.

What it tests

  • Framing a vague prediction problem: target, inputs, baseline, metric
  • Validation for time-ordered data: leakage, look-ahead bias, walk-forward splits, regime change
  • Overfitting on noisy data: many features tried, multiple testing, why a strong backtest fails live
  • Regularization, ensembles and the bias-variance trade-off
  • Feature engineering and how a signal decays over its horizon

How to prepare

  • State inputs, outputs and assumptions before naming a model; candidates report this is what gets probed.
  • Start from a simple, explainable regression baseline and say how you would know it is wrong.
  • Split by time and raise leakage, look-ahead bias and regime change yourself.
  • Think aloud; structured exploration is scored more than the final answer.

Free reported questions for this stage

An open-ended data or modelling round in 7 research and data reports from 2022-2026; recent ones put it first among live rounds. Order and scope agree with compiled accounts.

04Technical round 2: coding

About 45 min live coding in a shared editor (Python preferred); engineering titles may get more than one coding round

A medium-hard algorithm problem rather than a puzzle: interviewers want clean, correct code, explicit edge cases and one optimization conversation after the first working version. Research candidates describe it as more straightforward than the statistics hour. The firm's own engineering advice stresses memory and performance over brute force; streaming and range-query problems recur in reported questions.

What it tests

  • Classic data structures: caches, heaps, order statistics
  • Streaming algorithms: reservoir sampling, approximate quantiles, heavy hitters
  • Range queries and sliding windows
  • Dynamic programming and graph basics
  • Stating complexity and memory trade-offs

How to prepare

  • Write a brute-force version first, then optimize while talking through complexity.
  • Write code a teammate could review tomorrow: clear names, edge cases and a quick test.
  • Be ready to explain why a streaming algorithm is correct, not only to implement it.

Free reported questions for this stage

2 more free questions for this stage are in the app.

A live coding hour in 3 research-role reports from 2025-2026; the firm's engineering interview page gives coding advice. Length and editor from compiled accounts; ML-engineer coding rounds unreported.

05Technical round 3: statistics & math

About 45-60 min on a whiteboard or shared doc; follow-ups build on the first answer

The signature round. Expect linear regression from first principles and then its twists: the matrix form and variance of the estimator, regressing y on x versus x on y, duplicated rows, collinear or omitted variables and the direction of the bias, residualization, OLS versus weighted least squares under heteroskedasticity, and ridge. Maximum likelihood setups (Bernoulli, normal mean, uniform maximum), time-series basics and classic probability at textbook depth also appear, with the follow-up 'and if the noise is not Gaussian?'. Some loops split this across a phone screen and a later round.

What it tests

  • OLS derivation, variance, assumptions and what breaks when each fails
  • Omitted-variable bias, collinearity, measurement error, ridge and lasso closed forms
  • Maximum likelihood for simple families; WLS and when it equals OLS
  • Time series: autocorrelated errors, AR order, change detection
  • Classic probability and linear algebra: order statistics, stopping, correlation matrices, PCA

How to prepare

  • Re-derive the OLS estimator and its variance on paper until it is reflex.
  • Know that the two slopes (y on x, x on y) multiply to R-squared and when they are equal.
  • Derive ridge both ways: the closed form and choosing the penalty by cross-validation or bias-variance.
  • Ask clarifying questions early on probability puzzles and check answers on a small or extreme case.

Free reported questions for this stage

23 more free questions for this stage are in the app.

Statistics, regression or probability in 10 research-role reports from 2023-2026; the firm's research interview page lists statistics, theoretical math and ML. Round shape from compiled accounts.

06Hiring manager rounds

Three sequential 45-60 min rounds, each with the manager of a different team

Each manager is from a team that might take you, and the topic depends on that team: research teams dig into your past projects, signal-combination teams ask how correlated forecasts should be blended, execution teams ask market-microstructure questions. One round usually leans behavioral. Because teams share data and infrastructure, managers also ask how you would hand work to another team and what you would instrument so others can debug it.

What it tests

  • Depth behind your past research or projects
  • Reasoning on an unfamiliar, team-specific problem
  • Combining correlated signals and judging what a new one adds
  • Intellectual honesty about failed or inconclusive work
  • Hand-off and collaboration across research and engineering

How to prepare

  • Prepare one project in depth, including a result that did not hold up.
  • Rehearse a reasoning framework for unfamiliar setups instead of studying one domain.
  • Have a clear answer for why systematic investing and why this team.

Hiring-manager rounds with several teams in 3 research-role reports from 2023-2026; the firm's internship page describes technical and behavioral fit interviews. Three-round shape from compiled accounts.

07Final: senior manager

About 60 min, conversational

A senior manager closes the loop, mostly on behavior and fit: would they want to work with you for years? Technical content tends to frame the conversation, such as walking through the first hour with a new data set, a research call you got wrong, or why this team. The expected answer to a data walkthrough is a structured workflow (clean, explore, form a hypothesis, bound the answer, list what you would need next), not the signal itself.

What it tests

  • Structured approach to a new data set
  • Reflection on a wrong research call
  • Long-term fit and motivation
  • Questions back that show you know the team

How to prepare

  • Prepare one short story per behavioral theme and a clear three-to-five-year narrative.
  • Bring two specific questions about the team's work.
  • Do not speculate about the firm's performance or leadership; reason, do not repeat rumors.

A senior-manager or senior-staff final in 2 research-role reports (2023, 2026); format and content from compiled accounts of the research loop.

Questions and themes reported most often

  • Uniform order statistics, records and nested uniforms (7 reports, technical round 3: statistics & math)
  • OLS from first principles: derivation, assumptions, demeaning, rescaling, duplicated data (6 reports, technical round 3: statistics & math)
  • Time-series econometrics: AR order, GLS, change detection, instruments (6 reports, technical round 3: statistics & math)
  • Ridge and lasso: closed forms, fitting with a plain solver, tuning (5 reports, technical round 3: statistics & math)
  • Broken sticks and geometric probability (5 reports, technical round 3: statistics & math)
  • Open-ended prediction case: bike counts, demand from event counts, house prices (4 reports, technical round 1: open-ended data science)
  • Streaming algorithms: sampling, quantiles, heavy hitters (4 reports, technical round 2: coding)
  • Range-query data structures (4 reports, technical round 2: coding)

What Two Sigma asks about

  • Statistics & regression28%
  • Coding & algorithms24%
  • ML & open-ended data modelling20%
  • Probability & math18%
  • Research discussion & behavioral10%

Two Sigma interview prep checklist

  • Coded a linear interpolator with extrapolation and a streaming no-intercept regression within an hour, with edge-case tests
  • Selected the most predictive inputs for a regression by held-out error and by lasso in pandas
  • Solved a parent-array tree split and an IPO-style allocation simulation (engineering assessment)
  • Can derive OLS, its variance, ridge and the y-on-x vs x-on-y slope relationship from memory
  • Explained omitted-variable bias with its direction, collinearity, duplicated data and WLS
  • Outlined an open-ended prediction case with features, assumptions and walk-forward validation
  • Explained why a strong backtest fails live, in four causes
  • Have a short story about a research call that was wrong and what you changed

Tips

  • Find out your exact title early: research and data titles get the numeric, regression-flavoured assessment; engineering titles get the algorithm pool.
  • Code a linear interpolator, a town-selection regression with a lasso extension and a streaming regression from scratch before the assessment, with edge-case tests.
  • Make regression your first priority for the live rounds: derive OLS, then drill every twist (swap x and y, duplicate rows, collinear or omitted variables, WLS, ridge).
  • For the data science round, define inputs, outputs and assumptions first, start from a baseline regression and expect many follow-ups.
  • Split time-ordered data by time and raise leakage, multiple testing and regime change before anyone asks.
  • Do not neglect coding: it is its own round even for research titles, and clean code with one optimization pass is the bar.
  • Work the classic probability problems on order statistics, records and stopping times until the tricks are automatic.
  • Follow the firm's own advice: clarify, give a first pass, name its weaknesses and iterate, and take hints when offered.
  • Prepare one ML or research project for three different managers, including what failed and how another team could reuse it.

See the rest of the Two Sigma loop

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

Unlock with Pro

Practice the Two Sigma loop

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

Two Sigma interview FAQ

How many interview rounds does Two Sigma have for ML roles?

Candidates report 7 stages: Online assessment, Recruiter call, Technical round 1: open-ended data science, Technical round 2: coding, Technical round 3: statistics & math, Hiring manager rounds and Final: senior manager. An online assessment usually arrives first (often soon after applying or a referral) and passing it leads to a recruiter call. Three live technical rounds of about 45-60 minutes follow, in recent reports in this order: an open-ended data science round, a coding round and a statistics and math round; some candidates meet the data round alone as a screen. Three hiring-manager rounds with different teams and a conversational senior-manager round close the loop. Older reports describe one or two phone screens and then a virtual onsite of about four interviews. The firm interviews over video and may end the day early if the first rounds show no fit. Market-making games and mental-math speed tests are trader-track only and do not appear in research or ML loops. ML-engineer loops are thinly documented beyond the assessment.

How long does the Two Sigma ML interview process take?

Reported timelines run 3-8 weeks.

What does Two Sigma ask in ML interviews?

By reported question mix: Statistics & regression 28%; Coding & algorithms 24%; ML & open-ended data modelling 20%; Probability & math 18%; Research discussion & behavioral 10%. The most repeated themes: Uniform order statistics, records and nested uniforms; OLS from first principles: derivation, assumptions, demeaning, rescaling, duplicated data; Time-series econometrics: AR order, GLS, change detection, instruments.

How should I prepare for the Two Sigma ML interview?

Find out your exact title early: research and data titles get the numeric, regression-flavoured assessment; engineering titles get the algorithm pool. Code a linear interpolator, a town-selection regression with a lasso extension and a streaming regression from scratch before the assessment, with edge-case tests. Make regression your first priority for the live rounds: derive OLS, then drill every twist (swap x and y, duplicate rows, collinear or omitted variables, WLS, ridge).

Are there free Two Sigma interview questions?

Yes. LeetML has 100 questions reported from Two Sigma interviews, and 52 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources