Company interview guide

Susquehanna (SIG) Machine Learning Interview Guide (2026)

For ML, AI and research roles: a timed probability test, regression-heavy technical calls, a trade-data take-home and a long onsite.

Susquehanna (SIG)'s machine learning interview loop has 6 reported stages (Online assessment (track-dependent), Technical recruiter screen, Technical calls: probability, regression & algorithms, Take-home data exercise & presentation (research track), Final round onsite, Coding & systems (ML engineering track)); typical length: About 2 months.

Roles covered: Research Scientist, ML Engineer and Deep Learning Engineer. The question bank holds 18 questions reported from Susquehanna (SIG) interviews, 6 of them free to solve in your browser.

Try a free question

Interview process length
About 2 months
Stages
6
Reports linked to bank questions
18
Bank report dates
Feb 2022–Jan 2026
Questions in bank
18 (6 free)
Last reviewed
October 2026

How the Susquehanna (SIG) interview loop works

Research and systematic-trading candidates share one funnel: a 17-question timed math test scored before anyone reads the resume, a technical recruiter call, one or two technical calls on probability, regression and algorithms, a 3-day take-home data exercise for research candidates, and a long onsite. The process often runs 8-11 rounds over about two months and is known for reschedules and late headcount changes. Engineering roles, likely including ML engineering, follow a separate coding test, C++ phone screen and onsite. Market-making spread games, the AKQ betting game and live poker are trader-track only. Every stage here rests on the firm's postings and compiled accounts of its quant research loop re-cut for ML roles; we hold no first-hand ML-role reports, so confirm your track's steps with your recruiter.

The Susquehanna (SIG) interview process, stage by stage

01Online assessment (track-dependent)

Research track: 17 questions in 60 min on Mettl with a calculator, sent before resume review. Engineering tracks: a proctored coding test of about 4 problems (about 70 min)

The first gate, scored before a recruiter reads your resume. The research-track test draws on a near-fixed bank that has recurred since about 2019 with only the numbers changed: roughly nine quick probability and expected-value items, then about eight multi-step ones such as Bayes posteriors, knights-and-knaves or seating logic, a binomial tail, a bold-play gambler's ruin to five tokens and a lattice-path count with no three equal steps in a row that wants a short DP. Answers go into typed blanks (reduced fractions, dollar and percent fields, a clock time); only one item is multiple choice. A 2027-cycle sitting added a two-dice payoff-matrix EV, a stars-and-bars count and expected consecutive pairs. Engineering candidates get a coding test that rewards clean modelling, with one update-and-query problem as the real optimisation gate. A perfect score is necessary, not sufficient.

What it tests

  • Naming a problem's family in seconds: Bayes table, absorbing random walk, linearity, DP state
  • Bayes posteriors from two-source, two-flip and table set-ups
  • Expected value of small dice and card games, including expected pairs by indicators
  • Fair-coin and bold-play gambler's ruin
  • Engineering track: clean modelling, edge cases and count-map updates for pair-sum queries

How to prepare

  • Average pace is about 3.5 minutes a question, but finish the first nine in 20-25 minutes so each of the last eight gets 4-5 minutes.
  • Enter fractions in simplest form, read each blank's rounding rule and use the revisit bookmark instead of stalling.
  • Know k/N for fair gambler's ruin, the biased-coin formula and the no-three-in-a-row lattice DP before test day.
  • Ask which test your track gets: research-style ML roles likely sit the math test, ML engineering roles may get the coding test.

Free reported questions for this stage

No first-hand ML-role report. Platform, length and question families from compiled accounts of the firm's research-track test (2019-2026 sittings); the firm's programme postings mention an online maths assessment.

02Technical recruiter screen

30-45 min audio call: a self-introduction and interests chat, then 2-3 probability puzzles read aloud and worked on paper

Run by the quant recruiting team and technical despite the label: the firm's recruiter postings say these recruiters conduct technical interviews and write their own questions in probability, statistics, calculus and linear algebra. Compiled accounts list recurring puzzles: catching a bus in time for a film with uniform arrival times, which of two coin patterns such as HHT and HTH appears first, the chance that the median of three uniforms on [0, 3] falls in [1, 2], whether a triangle from random points on a circle is obtuse, and the two-envelope switch.

What it tests

  • Geometric probability with uniform arrivals and overlap regions
  • Coin-pattern races solved with a small Markov chain
  • Order statistics of uniforms
  • Symmetry arguments for random points on a circle
  • Explaining a set-up clearly out loud

How to prepare

  • State the sample space and the event before computing; the clarity of the explanation is scored.
  • Make the classics reflexive: meeting-time squares, pattern races, median of uniforms, points on a circle, two envelopes.
  • In the interests chat, real experience with strategy games such as poker or bridge is reported to land well.

Free reported questions for this stage

No first-hand ML-role report. The firm's recruiter postings say quant recruiters run technical interviews with probability and logic questions; compiled accounts give the format and puzzles.

03Technical calls: probability, regression & algorithms

1-2 rounds of 45-60 min by phone or video; in the 2027 cycle the first was a 60-min call with 2-4 questions, followed on a pass by a short recruiter call

Harder probability, a regression or estimation derivation and an algorithm discussed for approach and complexity, with little live coding. One question set has been reused across candidates with changed follow-ups: choose design points to estimate a regression slope, then redo it when the noise variance depends on x; judge whether a two-stage vote is fair; keep a running median of a stream of digits, then discuss which C++ data types to use; price a bet on ten dice summing to a multiple of seven. Research-track calls also include a five-dice reroll game judged on expected-value intuition rather than exact counting. The firm's recruiter posting lists probability, statistics, calculus and linear algebra as interview subjects for teams including machine learning.

What it tests

  • Deriving the OLS estimator and its variance in matrix form
  • Design of experiments: where to place x values to minimise slope variance, with and without x-dependent noise
  • Estimation from samples, such as maximum likelihood for an unknown pool size
  • Waiting-time and conditional probability problems
  • Algorithm reasoning: a streaming median over bounded inputs, approach and complexity

How to prepare

  • Derive beta-hat and Var(beta-hat) = sigma^2 (X'X)^-1 from scratch, say when OLS is best, then show how weighted least squares changes the design when noise depends on x.
  • Learn why each answer holds; the follow-ups change between candidates.
  • For the streaming median, notice what a bounded input range (digits 0-9) buys you: counts instead of two heaps.
  • Give a reasoned EV estimate before refining it; a sound direction matters more than the last decimal.

Free reported questions for this stage

No first-hand ML-role report. Format and questions from compiled accounts of the firm's quant research calls (2026-2027 cycles); the firm's recruiter posting lists the subjects for research and ML teams.

04Take-home data exercise & presentation (research track)

3-day take-home on one stock's trades: a 1-2 page report plus code or notebook (8-10 hours suggested), then a presentation round onsite

Research-track candidates in the 2027 cycle describe about a year of trades in a single stock, one share per print, with buyer and seller trader ids on each. The goal is to maximise expected profit under rules: at most five buys and five sells a day, flat overnight, no new trades in the last 30 minutes and forced closing from 15:30. In live trading you would see ids only on your own fills, the rest at day end. The dataset recurs across cycles. You then present the work and field detail, technical, trading and market-making questions, including how sure you are that an edge will persist. Earlier cycles used a two-hour virtual review with two quants that pushed toward a game-theory view of how other traders respond.

What it tests

  • Exploratory analysis of noisy trade-level data
  • Testing which counterparties are informed, using markouts over several horizons
  • Avoiding look-ahead: using only information available at decision time
  • Multiple testing across many trader ids before calling an effect real
  • Turning a signal into a rule that obeys position and timing limits

How to prepare

  • Respect the information set: other traders' ids arrive only at day end, so a rule that conditions on them intraday leaks the future.
  • Split by time, not at random, and keep a final untouched period for the number you report.
  • Adjust markouts for the stock's drift and correct for how many traders you screened.
  • Prefer a simple, legible rule and write down every assumption; the presentation probes them one by one.

No first-hand ML-role report. Task, rules and presentation from compiled accounts of the firm's 2027 research-track cycle; earlier cycles reportedly used a live review instead.

05Final round onsite

A day at the firm's Philadelphia-area office: several 1-hour quant rounds and a roughly 75-min data round, with a dinner and tours; part may run by video

Compiled accounts describe a long final day: a dinner beforehand with a junior trader or researcher, tours of the building, data centre and trading floor, several one-hour quantitative rounds with open-ended estimation and problems such as estimating X from Z = X + Y for independent normals, and a data round of about 75 minutes on expected value and variance trade-offs. Trader-track candidates also play live poker and at times a team game. Across the whole process, 8-11 rounds are reported.

What it tests

  • Gaussian conditioning: E[X | X + Y] and its variance
  • Open-ended estimation with stated assumptions and a range
  • Expectation versus variance in decisions
  • Moment bounds and covariance-matrix facts
  • Staying clear and composed over a long day

How to prepare

  • For estimation, state your model and give an interval, not a single number.
  • Know the Gaussian conditioning formula cold; it is the same algebra as a shrinkage estimator or a Kalman update.
  • Treat the dinner and informal chats as part of the day.
  • Pace yourself and keep narrating; the day is long.

No first-hand ML-role report; shape of the day and question themes from compiled accounts of the firm's research and trader finals. The firm's research postings note automatic consideration for systematic trading.

06Coding & systems (ML engineering track)

Software funnel: proctored coding test of about 4 problems, a C++ phone screen, then onsite C++ internals, a trade-book design and live debugging

Engineering candidates go through a separate funnel. Compiled accounts describe a coding test that rewards clean modelling, a phone screen built on implementing a growable array like std::vector from scratch (2x growth, allocation and release, exception safety, type traits, ownership), and an onsite on pointers, polymorphism and virtual functions, a trade-book design that matches orders by price level (sorted map versus heap) and live C++ debugging. Strategy-developer candidates get a four-problem coding test of matrix manipulation, cyclic time arithmetic and pair counting. The firm's ML engineering posting asks for data structures, algorithms, concurrency, a feel for how software behaves on real hardware and a project you can walk through in detail.

What it tests

  • Implementing a growable container with correct ownership and exception safety
  • Count-map and hashing tricks for update-and-query problems
  • Order-book style design and data-structure trade-offs
  • C++ internals: virtual dispatch, object slicing, RAII, smart pointers
  • Walking through a system you built: what you designed, decided and did yourself

How to prepare

  • Be able to write a vector end to end, including move semantics and what happens when a copy throws during growth.
  • Prepare the trade-book design with latency in mind: which operations must be O(1) or O(log n).
  • Bring one ML-systems project with measured results, such as a training job you made faster or an inference path whose latency you cut.

Free reported questions for this stage

No first-hand ML-engineer report. Format from compiled accounts of the firm's software and strategy-developer funnels; the firm's ML engineering and ML systems postings describe the skills sought.

Questions and themes reported most often

  • Bayes posteriors from tables, two sources and repeated flips (online assessment (track-dependent))
  • Expected counts of adjacent pairs via indicator variables (online assessment (track-dependent))
  • Expected trials until a second distinct outcome (online assessment (track-dependent))
  • Birthday-style collision probabilities (online assessment (track-dependent))
  • Uniform arrivals and meeting-time overlaps (technical recruiter screen)
  • Coin-pattern races solved with a Markov chain (technical recruiter screen)
  • Random points on a circle and seating symmetry (technical recruiter screen)
  • Derive OLS, choose design points for a slope, then handle x-dependent noise (technical calls: probability, regression & algorithms)

What Susquehanna (SIG) asks about

  • Probability & expected value32%
  • Statistics & regression22%
  • Research methodology & backtesting14%
  • Coding, C++ & systems14%
  • ML depth & your research10%
  • Linear algebra4%
  • Fit & motivation4%

Susquehanna (SIG) interview prep checklist

  • Ran two timed 17-question, 60-minute sets with answers as reduced fractions and the first nine done in about 20-25 minutes
  • Can work uniform-arrival, coin-pattern, median-of-uniforms and points-on-a-circle puzzles out loud
  • Can derive OLS and its variance, place design points for a slope and redo it with x-dependent noise
  • Built a small rule on trade data with a time-ordered split, markouts at several horizons and a multiple-testing correction
  • Know E[X | X + Y] for independent normals and can frame it as shrinkage
  • Have a deep-learning or ML-systems project ready to defend in detail, with measured results
  • ML engineering: implemented a growable array in C++ and sketched a price-level order book
  • Confirmed with the recruiter which test, take-home and final-round format your track gets

Tips

  • Make probability the first priority: Bayes posteriors, linearity of expectation, uniform arrivals, pattern races and gambler's ruin run from the test to the onsite.
  • Drill the test's question families on a timer until recognition is automatic, and protect time for the harder second half.
  • Derive OLS, its variance and optimal design points from scratch, then redo them with x-dependent noise; regression is the main statistics thread.
  • Narrate every set-up; interviewers score how you frame a problem as well as the number.
  • Treat the take-home as a leakage and multiple-testing exercise: time-ordered splits, markouts at several horizons and a simple rule you can defend.
  • Prepare one deep-learning project on noisy time series to explain end to end, including how you validated out of sample and what failed.
  • For ML engineering roles, add a C++ vector implementation, an order-book design, C++ internals and a measured ML-systems project.
  • Ask early which track, test, take-home and final rounds apply; research applicants are also considered for systematic trading.
  • Budget about two months with reschedules, and do not read a quiet week as a rejection.

See the rest of the Susquehanna (SIG) loop

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

Unlock with Pro

Practice the Susquehanna (SIG) loop

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

Susquehanna (SIG) interview FAQ

How many interview rounds does Susquehanna (SIG) have for ML roles?

Candidates report 6 stages: Online assessment (track-dependent), Technical recruiter screen, Technical calls: probability, regression & algorithms, Take-home data exercise & presentation (research track), Final round onsite and Coding & systems (ML engineering track). Research and systematic-trading candidates share one funnel: a 17-question timed math test scored before anyone reads the resume, a technical recruiter call, one or two technical calls on probability, regression and algorithms, a 3-day take-home data exercise for research candidates, and a long onsite. The process often runs 8-11 rounds over about two months and is known for reschedules and late headcount changes. Engineering roles, likely including ML engineering, follow a separate coding test, C++ phone screen and onsite. Market-making spread games, the AKQ betting game and live poker are trader-track only. Every stage here rests on the firm's postings and compiled accounts of its quant research loop re-cut for ML roles; we hold no first-hand ML-role reports, so confirm your track's steps with your recruiter.

How long does the Susquehanna (SIG) ML interview process take?

Reported timelines run About 2 months.

What does Susquehanna (SIG) ask in ML interviews?

By reported question mix: Probability & expected value 32%; Statistics & regression 22%; Research methodology & backtesting 14%; Coding, C++ & systems 14%; ML depth & your research 10%; Linear algebra 4%. The most repeated themes: Bayes posteriors from tables, two sources and repeated flips; Expected counts of adjacent pairs via indicator variables; Expected trials until a second distinct outcome.

How should I prepare for the Susquehanna (SIG) ML interview?

Make probability the first priority: Bayes posteriors, linearity of expectation, uniform arrivals, pattern races and gambler's ruin run from the test to the onsite. Drill the test's question families on a timer until recognition is automatic, and protect time for the harder second half. Derive OLS, its variance and optimal design points from scratch, then redo them with x-dependent noise; regression is the main statistics thread.

Are there free Susquehanna (SIG) interview questions?

Yes. LeetML has 18 questions reported from Susquehanna (SIG) interviews, and 6 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources