SIG Quant Careers & Interview Prep: 4 Solved Exercises

SIG is Susquehanna International Group, often branded simply as Susquehanna. Its official history dates the firm's founding to 1987. Its quantitative research team applies mathematics, statistics and programming to market data, working with traders and developers to develop strategies. Company history, Quantitative Research + Machine Learning careers.

Start with the official website or careers site to find the employer's own information. If you are choosing a track, compare the roles; if you are checking eligibility, review the dated official posting examples. If you already have an invitation, read the assessment guidance and use your own instructions to choose the final rehearsal format.

A research candidate should be ready to explain why a model is credible. A trader should be ready to turn uncertain information into an actionable decision. An engineer should be ready to make a system correct, measurable and reliable. These skills overlap, but a probability worksheet cannot replace a research discussion or an engineering interview.

Use the employer's role names rather than treating every job containing “quant” as the same career.

Sources: quant careers, trading careers, technology careers, and the 2027 PhD researcher posting.

Choose from the work you enjoy. If you like investigating a noisy result and checking alternative explanations, research may fit. If you like making decisions and updating them as information arrives, consider trading. If you prefer building and debugging the infrastructure that makes either possible, examine technology roles.

This is a way to organise your search, not an eligibility decision. Teams vary, and the actual job description remains decisive.

The following are dated examples checked on 3 October 2026, not a promise that an application remains open. Direct automated access to some careers pages was restricted; the employer-domain search index exposed their descriptions. Reopen the original posting before relying on it.

The New York PhD QR description lists $300,000 annual base pay. That is one advertised base salary for one position, not SIG's average pay, a total compensation promise, or a graduate-trader range. It also says a QR application is automatically considered for Quantitative Systematic Trader; do not generalise this policy to other postings.

The London research example is a specific route for research master's candidates; the New York PhD row is not a firmwide degree rule. For broader exploration, the official quant page links experienced, graduate, internship and discovery opportunities. Read graduation dates, location and work-authorisation language before applying.

A recent preparation guide synthesising candidate reports describes a 17-question, approximately 60-minute mathematical assessment. Current candidate-report comparison. A different preparation page describes a 20-minute Quantitative Evaluation. Alternative reported format. These descriptions may refer to different assessments, roles or cycles; neither establishes your invitation's format. Reports also disagree about platforms and later stages.

Treat a report as a rehearsal suggestion. Your invitation establishes your actual test. Check the duration, number of sections, deadline and timezone, answer format, calculator policy, permitted materials, navigation and penalties. Ask the recruiting contact about a missing instruction or accommodation before starting.

The Dublin software graduate posting gives an official sequence: application, technical screen, recruiter interview, technical interview, final interview and offer. This is evidence for that software programme; it does not establish a universal trading or research “Superday”.

For live practice, explain assumptions before calculating. For research preparation, discuss experiments and validation. For developer preparation, practise implementation and complexity. No universal passing score or guaranteed number of rounds is established by the sources reviewed here.

Susquehanna's game theory and decision science page explains how strategic games teach decisions under uncertainty. Learn that reasoning without assuming you need poker expertise or that every interview includes a poker game.

Problem. A randomly selected device is equally likely to be type A or B. It produces a positive reading with probability 1/3 if it is A and 2/3 if it is B. Readings are independent conditional on the device's type. Two readings from that same device are positive. What is the probability the device is B?

Solution. Independence is conditional on the device type. The two-positive likelihoods are 1/9 for A and 4/9 for B. Weight them by their equal prior probabilities:

P(B | ++ ) = [(1/2)(4/9)] / [(1/2)(1/9) + (1/2)(4/9)] = 4/5.

A useful second route is posterior odds. Prior B-to-A odds are 1. Each positive reading multiplies the odds by 2, so two readings give odds 4, corresponding to probability 4/(1 + 4).

Follow-up. If the readings are one positive and one negative in either order, both types have likelihood 2 × (1/3)(2/3) = 4/9, so the posterior remains 1/2. A specified order, such as positive then negative, has likelihood 2/9 for either type. If readings are correlated even after conditioning on type, the likelihood multiplication fails.

What to explain. Distinguish the prior, observation likelihood and posterior. Check the direction of the update: positives favour B. A numerical answer alone hides whether you understood the dependence structure.

Problem. Roll a fair six-sided die. You receive its face value in dollars. After seeing the first roll, you may pay $1 to discard it and take one final independent roll. When should you reroll, and what is the expected payout after any reroll fee?

Solution. The expected value of the final roll is $3.50, so its net value after the fee is $2.50. Keep a first result of 3, 4, 5 or 6; reroll a 1 or 2. Average the values of the six first-roll branches:

Expected net payout = (2.5 + 2.5 + 3 + 4 + 5 + 6)/6 = 23/6 dollars.

You pay the fee only in the two reroll branches. Subtracting a dollar from every outcome would model a different game. The optional reroll improves on the original $3.50 mean because 23/6 is approximately $3.83.

Follow-up. With a reroll cost c, compare each observed roll with 3.5 − c. A tie gives the same expectation either way. A non-negative cost of at least $2.50 removes any strictly beneficial reroll.

What to explain. Define the decision after information arrives. Optimise each branch before taking the average; averaging first and then choosing a single action loses the value of the option.

Problem. A ticket pays the sum of two independent fair dice in dollars. What is its expected payout? If someone knows the first die and offers to buy at $8 only when it shows 5 or 6, what is the expected payout conditional on that trade?

Solution. Before receiving information, expected payout is 3.5 + 3.5 = $7. Conditional on the buyer's stated rule, the first die averages 5.5 and the second still averages 3.5. The expected payout when you sell is therefore $9. Selling at $8 has expected profit 8 − 9 = −$1 per executed trade.

The unconditional fair value did not make the offered transaction attractive. The trade itself disclosed information about which outcomes you were exposed to. This simple model illustrates adverse selection.

Follow-up. If you instead observe the first die to be 2, expected payout becomes $5.50. For a real quote, inventory exposure, costs, uncertainty about the buyer's behaviour and risk limits also matter. Expected payout is a starting calculation rather than a complete bid-and-ask policy.

What to explain. State whose information you condition on. Separate the value before a trade from the value in the subset of circumstances where a trade occurs.

Problem. You predict tomorrow's return from today's features. A colleague randomly divides rows into training and test sets, standardises using all rows, tries 50 models, and reports the best test result. What should change?

Solution. First identify when every input was available. A feature calculated with a future close cannot support a decision made earlier. Then place the training period before validation and reserve a later, untouched test period. Fit standardisation and other learned transformations on training data, carrying those fitted parameters forward.

Use validation to select among models. Reporting the best of 50 test scores turns the test set into part of model selection. Evaluate the final choice once on the reserved test, and state how many alternatives you investigated.

The split must also respect the target. If labels span multiple future days, observations near a boundary may share outcome information with another fold. Purging means excluding training observations whose target windows overlap the evaluation period. An appropriate gap or purging rule follows from that overlap, rather than from an arbitrary universal number of days.

What to explain. A chronological split helps, but timing correctness, preprocessing and repeated selection remain separate issues. A convincing research discussion explains the problem each safeguard addresses and admits what evidence still cannot establish.

The official PhD QR posting emphasises model development, validation and robustness. A useful personal preparation project should make those judgements visible. The following is our suggested protocol, not a claim about a SIG take-home assignment.

Keep an experiment log before looking at final results. Record each hypothesis, configuration, split and reason for the next experiment. A negative finding can demonstrate judgement when the test was well designed.

For a short project defence, explain the question, baseline, split, strongest result and biggest limitation. Be ready to describe one failed approach and the evidence that changed your mind. Prepare to reproduce a calculation or change a parameter without rebuilding the story around a preferred result.

The free returns lab lets you practise explaining a reproducible research output. It is general research preparation and has no affiliation with SIG.

Start with the available time and the exact role. This schedule is a suggested preparation plan; extend foundations if the diagnostic exposes gaps.

Use four error categories: concept, modelling assumption, calculation and communication. A dependence mistake needs different practice from a transcription error. Re-solve a missed problem several days later with changed inputs.

For “Why Susquehanna?”, connect a real experience to a concrete aspect of the work. A project where you tested an initially attractive hypothesis is stronger evidence than repeating that you like probability. Describe your own contribution and the evidence behind your conclusion.

Before applying, save the job URL, title, location, cycle and application date. Check whether that specific posting gives a rule about multiple applications. Use the official quant career page to find the right route, then use the general assessment to identify skill gaps. The assessment is a QuantSteps diagnostic, not a prediction of hiring outcomes.