Company interview guide
Optiver Machine Learning Interview Guide (2026)
For ML, AI and research roles: a long timed assessment battery, a live betting game with a trader, a probability whiteboard, a P&L-graded take-home and a blunt fit final.
Optiver's machine learning interview loop has 7 reported stages, from online assessment battery (track-dependent) to final culture-fit round; typical length: Several weeks; the online assessments are due within 7 days of the invite.
Roles covered: Research Scientist, ML Engineer and Deep Learning Engineer. The question bank holds 3 questions reported from Optiver interviews, 2 of them free to solve in your browser.
- Interview process length
- Several weeks; the online assessments are due within 7 days of the invite
- Stages
- 7
- Reports linked to bank questions
- 3
- Bank report dates
- Oct 2025–Nov 2025
- Questions in bank
- 3 (2 free)
- Last reviewed
- October 2026
How the Optiver interview loop works
Research candidates share the early funnel with traders: a multi-module online battery (with an extra three-problem coding section for research), a live betting game with a trader, and a short recruiter call. Research candidates then get a 60-75 minute probability and estimation whiteboard, a take-home trading-data project with a roughly 45-minute deep dive about two days after submission, and a final fit round with a senior partner. Software and ML engineering candidates take a separate coding test, technical and design interviews and a behavioural round. The interval-quoting game day, horse-race odds games, group trading and the trading-programme cognitive interview are trader-track only. After an assessment or interview for a quant graduate or intern role the firm applies an 8-month cool-off before you can reapply. Every stage rests on the firm's pages and compiled accounts of its quant research loop re-cut for ML roles; we hold no first-hand ML-role reports.
The Optiver interview process, stage by stage
01Online assessment battery (track-dependent)
Quant tracks: about 3 hours of timed modules on CodeSignal, often split across days, no calculator; research adds a 90-min, 3-problem HackerRank coding section.
Engineering tracks: 1-2 long simulation problems (75-90 min each) plus a ~20-question CS quiz
The heaviest filter in the loop, deliberately long and frequently auto-rejected on submission. Compiled accounts describe Number Logic (find the next term in sequences in about 25 minutes), Beat the Odds (20-30 probability and expected-value items at about 90 seconds each, pick the closest of five options), Zap-N reaction and memory mini-games, and a separate 80-question, 8-minute mental-arithmetic screen with no skips. The probability and sequence modules score +1 right, -1 wrong and 0 for a skip. A 2026 variant adds a Likelihood Test (rank three events by probability from a chart or table); a September 2026 sitting ran six separately clocked modules, including a 25-minute Likelihood module, Intervals at roughly 40-45 seconds a question and Order Books with 20 boards in 8 minutes. The research coding section uses trading-flavoured problems: counting valid buy and sell sequences, a proportional-allocation backtest returning the mean and standard deviation of daily log returns, order matching with two heaps, currency-arbitrage detection with fees, and an optimal-execution DP with a broker backstop.
What it tests
- Recognising recurring probability families in seconds: gambler's ruin, return times on a cycle, coupon collector, remove-pairs symmetry
- Sequence pattern triage: alternating sub-sequences, a(n) = 2a(n-1) + a(n-2), product and difference ladders
- Deciding when to answer or skip under +1/-1/0 scoring
- Clean algorithmic Python under time pressure: DP counts, heaps, Bellman-Ford on negative log rates
- Engineering track: modelling a small class method by method with corner cases
How to prepare
- With +1/-1/0 scoring, answer only when you are better than even; a skip beats a low-confidence guess.
- Bank the first ~15 sequences and skip the near-impossible last few rather than burning the clock on one.
- Drill the Beat the Odds families until each is a 30-second recall, and estimate to the closest option instead of solving exactly.
- For the research coding section, rehearse Bellman-Ford on -log(rate) with per-leg fees and a two-heap order matcher until both are routine.
Free reported questions for this stage
- Sex Ratio with a Girl-Stopping RuleProbabilityEasySolve
No first-hand ML-role report. Modules, timings and scoring from compiled accounts (2019-2026 sittings); the firm's quant postings say online assessments follow the application and must be done within 7 days.
02Betting game with a trader (Tech 1 and 2)
25-45 min, interactive, one-on-one with a trader; usually two rounds
The firm's signature filter, reported for research as well as trading candidates. You start with a bankroll of about $1,000 and bet across simultaneous boards, such as three coin flips, two dice and three-card draws, whose odds reshuffle every round, plus a market-making sub-game where you quote a two-sided market on the sum of three cards. The trader trades against you and keeps asking why you made each bet. Scoring is on sizing relative to edge, spotting mispriced odds that allow a riskless split, and risk discipline, not on final P&L; candidates who play too safe are reported to be cut.
What it tests
- Computing the expected value of each offered bet quickly and out loud
- Sizing by edge, roughly fractional Kelly, without risking ruin
- Spotting arbitrage when implied probabilities across complementary outcomes sum below 100%
- Quoting a fair value and a spread that widens with uncertainty and adverse selection
- Staying composed while being challenged
How to prepare
- Say the EV of every bet before you size it; the reasoning is graded more than the result.
- Bet big on large edges, small on thin ones; being too conservative is a reported failure mode.
- Take an arbitrage the moment you see it; it is often disallowed after the first use.
- On the market-making sub-game, adjust your quote as fills reveal what the trader knows.
No first-hand ML-role report. Format and scoring from compiled accounts of the firm's research and trader loops, which describe this game for both tracks.
03Recruiter / behavioural call
About 20 min with a recruiter
A short conversation that checks motivation and logistics and lays out the rest of the process: why the firm, why trading or quant research, other offers and deadlines, visa and sponsorship, and a walk-through of one project. Accounts say sponsorship comes up early and that trading and research roles are more often sponsored than software roles.
What it tests
- A specific reason for the firm and for markets
- A clear, structured story about one project
- Logistics: offers, timing, location, sponsorship
How to prepare
- Tie your motivation to something specific: fast feedback from live results, owning a model end to end.
- Prepare one project in situation, task, action, result form that you can later defend in depth.
- Know your sponsorship position before the call.
No first-hand ML-role report. Length and content from compiled accounts of the firm's quant recruiter call.
04Probability & estimation whiteboard (research track)
60-75 min on a Zoom whiteboard with a researcher: 3-4 problems plus a Fermi estimate
The research technical round mixes probability, conditional expectation, number theory and brainteasers, and wants a clean argument rather than a bare number. Recurring themes in compiled accounts: a two-piles-of-stones game solved with a doubling invariant, the bus-waiting (inspection) paradox, a proof about diagonal sums in a Latin square, and how many years until you are 97.5% sure a drifting price exceeds 100. Fermi questions ask for a point estimate and an interval, then add a fact, such as a subscriber count, and ask you to tighten it. Interviewers like to ask you to generalise a worked case to all n.
What it tests
- First-step analysis and conditional expectation for waiting-time and random-walk problems
- Size-biased sampling: why the wait for a bus exceeds half the mean gap
- Drift versus volatility: when mean minus two standard deviations clears a threshold
- Fermi estimation with a calibrated interval that updates on new information
- Turning an example into an induction or invariant proof
How to prepare
- Write E[X] = E[E[X | first step]] explicitly; most of these collapse to a small recursion.
- For the price question, note the mean grows like t and the standard deviation like sqrt(t), then solve for t.
- Give an interval, not just a number, and show how a new fact narrows it.
- Talk through the set-up before computing; framing is scored.
Free reported questions for this stage
- Optimal Pick Time for a Ripening FruitStatistics & InferenceMediumSolve
No first-hand ML-role report. Format and recurring problems from compiled accounts of the firm's quant research technical round.
05Take-home trading-data project & deep dive (research track)
Take-home strategy build and backtest on an algorithmic-trading dataset, then a ~45-min presentation and deep dive about two days after you submit
Research candidates build and backtest a strategy on a supplied trading dataset. Accounts stress that it is graded on realized profit, not R-squared: a clean, well-validated linear model is enough to pass, and costs should be accounted for. The follow-up is a research review rather than a demo: interviewers probe each feature, horizon and validation choice, with heavy live questions on pandas and time-series handling such as resampling, rolling windows, look-ahead and time-ordered splits. Saying you would search for the syntax is reported as a red flag.
What it tests
- Building a signal that survives costs, not one that maximises in-sample fit
- Leakage control: features computed only from information available at decision time
- Time-ordered validation, with a final untouched period for the reported number
- Fluent pandas: resample, rolling, shift, groupby on timestamps
- Defending feature, horizon and model choices under pressure
How to prepare
- Start from a regularised linear baseline and only add complexity that improves out-of-sample P&L after costs.
- Use walk-forward or purged splits and report how the edge varies across periods; markets drift.
- Check for signal decay across horizons and say how many variants you tried before choosing one.
- Rehearse rolling and resampling operations from memory, including off-by-one shifts that leak the future.
No first-hand ML-role report. Task, grading and deep-dive format from compiled accounts of the firm's quant research take-home; the firm's research pages stress live, out-of-sample results.
06Coding, design & behavioural (software / ML engineering track)
Software funnel: a long simulation-style coding test and CS quiz, then technical interviews in one object-oriented language, a system design discussion and a behavioural assessment
ML engineers most likely interview through the software funnel. The firm's own interview-tips page says its technical rounds are collaborative problem-solving rather than stock puzzles: know one object-oriented language (C++, Java or Python) and its standard library, the time and space cost of common data structures and how they sit in memory, plus computer architecture, networking, concurrency and memory management. In design, ask about load, latency and edge cases, justify choices and revise when given a hint. Compiled accounts add trading-system design themes such as co-location, TCP versus UDP and a lock-free ring buffer. The behavioural part looks for collaboration, transparency and curiosity, answered in STAR form.
What it tests
- Data-structure costs and memory layout, not just big-O
- Concurrency and memory management in your main language
- Latency-aware system design: networking choices, queues between threads
- Modelling a long simulation problem cleanly within the time
- Behavioural: collaboration and learning from a hard project
How to prepare
- Pick one language and know its collections and algorithms in depth.
- In design, state the latency budget and throughput first, then choose structures.
- Engage with hints; the firm says revising a design is part of the evaluation.
- Prepare STAR stories from projects, internships and teams, and practise them out loud.
No first-hand ML-engineer report. The firm's interview-tips page for software engineers describes the technical and behavioural assessments; test and design themes from compiled accounts.
07Final culture-fit round
About 45 min with a hiring manager or senior partner
Nominally behavioural, functionally a fit decision, and the round most often described as rejecting technically strong candidates. It is free-flowing and the partner interrupts out of curiosity. Expect why quant rather than academia, your personal story, what your edge is over other candidates, an honest ranking of where you are stronger and weaker than your cohort, and career goals. Accounts say the firm values ownership, someone who finds the next problem and asks to expand it, and that the floor culture is loud and direct, so timid answers hurt. Decisions come fast.
What it tests
- A personal, specific reason for industry over academia
- Self-assessment against peers without hedging
- Ownership and initiative in past work
- Holding a conversation through interruptions
How to prepare
- Replace generic answers like faster pace or more impact with the specific moment you chose this path.
- Treat interruptions as interest, not as a sign the answer is wrong.
- Frame ambitions around owning and extending a model or research area.
- Speak with conviction and defend your view when challenged.
No first-hand ML-role report. Format and themes from compiled accounts of the firm's research and trader finals; the cool-off rule is on the firm's quant postings.
Questions and themes reported most often
- Gambler's ruin, including biased and dice-disguised versions (online assessment battery (track-dependent))
- Remove items until one type remains: symmetry and stopping rules (online assessment battery (track-dependent))
- Coupon collector and expected waits for a pattern (online assessment battery (track-dependent))
- Order matching with two heaps (online assessment battery (track-dependent))
- Kelly-style sizing and arbitrage across mispriced bets (betting game with a trader (tech 1 and 2))
- Bus-waiting and inspection paradox (probability & estimation whiteboard (research track))
- How long until a drifting price clears a level with 97.5% confidence (probability & estimation whiteboard (research track))
- Optimal stopping and when to act on a changing value (probability & estimation whiteboard (research track))
What Optiver asks about
- Probability & expected value28%
- Statistics, regression & time series20%
- Research methodology & backtesting16%
- Coding: Python, pandas & algorithms14%
- Betting & position sizing8%
- ML depth & your research8%
- Fit & motivation6%
Optiver interview prep checklist
- Ran timed probability sets at about 90 seconds an item with +1/-1/0 scoring and a skip rule
- Can do gambler's ruin, cycle return times, coupon collector and remove-pairs problems from memory
- Played a mock betting game out loud with Kelly-style sizing and one arbitrage
- Can explain the bus-waiting paradox and solve the drifting-price confidence question
- Built a small strategy on trading data with walk-forward splits, costs and a reported out-of-sample P&L
- Can write resample, rolling and shift code without looking it up
- Have one ML or ML-systems project ready to defend in detail
- Confirmed with the recruiter which track, tests and final rounds apply
Tips
- Make probability the first priority: the battery, the betting game and the whiteboard all run on expected value, gambler's ruin, first-step analysis and symmetry.
- Practise the battery modules on a clock with +1/-1/0 scoring until you skip without hesitation.
- Rehearse the betting game out loud: EV, size by edge, take arbitrage, never freeze.
- Learn the whiteboard classics, including the inspection paradox and drift-versus-volatility timing, and give intervals for every estimate.
- Treat the take-home as a P&L problem: simple model, costs, walk-forward validation, no look-ahead.
- Get pandas time-series operations to recall speed; the deep dive tests them live.
- Prepare one ML project to defend end to end, including what failed live or out of sample.
- ML engineers: prepare the software funnel, latency-aware design and a measured GPU or inference project.
- Have a personal answer to why industry over academia for the final.
- Ask early which assessment, game and take-home your track gets; a rejection after an assessment means an 8-month wait.
See the rest of the Optiver loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Optiver loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Optiver interview FAQ
How many interview rounds does Optiver have for ML roles?
Candidates report 7 stages: Online assessment battery (track-dependent), Betting game with a trader (Tech 1 and 2), Recruiter / behavioural call, Probability & estimation whiteboard (research track), Take-home trading-data project & deep dive (research track), Coding, design & behavioural (software / ML engineering track) and Final culture-fit round. Research candidates share the early funnel with traders: a multi-module online battery (with an extra three-problem coding section for research), a live betting game with a trader, and a short recruiter call. Research candidates then get a 60-75 minute probability and estimation whiteboard, a take-home trading-data project with a roughly 45-minute deep dive about two days after submission, and a final fit round with a senior partner. Software and ML engineering candidates take a separate coding test, technical and design interviews and a behavioural round. The interval-quoting game day, horse-race odds games, group trading and the trading-programme cognitive interview are trader-track only. After an assessment or interview for a quant graduate or intern role the firm applies an 8-month cool-off before you can reapply. Every stage rests on the firm's pages and compiled accounts of its quant research loop re-cut for ML roles; we hold no first-hand ML-role reports.
How long does the Optiver ML interview process take?
Reported timelines run Several weeks; the online assessments are due within 7 days of the invite.
What does Optiver ask in ML interviews?
By reported question mix: Probability & expected value 28%; Statistics, regression & time series 20%; Research methodology & backtesting 16%; Coding: Python, pandas & algorithms 14%; Betting & position sizing 8%; ML depth & your research 8%. The most repeated themes: Gambler's ruin, including biased and dice-disguised versions; Remove items until one type remains: symmetry and stopping rules; Coupon collector and expected waits for a pattern.
How should I prepare for the Optiver ML interview?
Make probability the first priority: the battery, the betting game and the whiteboard all run on expected value, gambler's ruin, first-step analysis and symmetry. Practise the battery modules on a clock with +1/-1/0 scoring until you skip without hesitation. Rehearse the betting game out loud: EV, size by edge, take arbitrage, never freeze.
Are there free Optiver interview questions?
Yes. LeetML has 3 questions reported from Optiver interviews, and 2 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- Jane StreetFor ML, AI and research roles: multi-part coding judged on clarity, a hands-on ML round, and probability and statistics on research-leaning loops.
- IMC TradingFor ML, AI and research roles at a global options market maker: a track-split timed test, probability and ML-fundamentals rounds, and a fast superday.
- Susquehanna (SIG)For ML, AI and research roles: a timed probability test, regression-heavy technical calls, a trade-data take-home and a long onsite.
- Flow TradersFor ML, AI and research roles: a stats-and-ML or language-and-coding online test, a recruiter call, probability and regression interviews, a codepair for engineers, then a research deep dive with a team lead.
- CitadelFor ML, AI and research roles: the firm's quant research loop, heavy on regression, ML theory on noisy data, probability derivations and classic coding.
- 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.