Mock Screens · SIG Problem-Solving Assessment
SIG Problem-Solving Assessment Guide
A single 20-minute sitting of 17 probability and logic problems drawn from a recycled archetype bank - the pre-resume screen SIG is reported to send every candidate before anyone reads their application.
What is a SIG problem-solving assessment?
Susquehanna International Group is reported to send a Problem-Solving Assessment to essentially every applicant before a resume is even reviewed - a probability and logic battery that stands entirely apart from the candidate's background, GPA, or coursework. An assessment that runs before any human judgment enters the pipeline treats reasoning ability as the first filter, ahead of pedigree.
The test is widely reported to be drawn from a recycled bank of archetypes - the same handful of problem shapes (a gambler's-ruin question, a Bayes update, a combinatorics count, a deductive logic puzzle) reappear across candidates and across years, with only the specific numbers changed. A fixed, leaked question bank would normally be a security flaw, but if the reasoning has to be redone for new numbers each time, memorizing the final answer to one instance doesn't transfer to the next. It also means the test can be graded and compared across an enormous applicant pool without any variation in difficulty between administrations.
The archetypes are the toolbox a trading or quantitative research role uses: martingale arguments for gambler's-ruin and random-walk questions, Bayes' rule for updating a belief on new information, combinatorics for counting arrangements, and single-variable optimization for questions like maximizing a product under a fixed sum. None of these require more than an undergraduate probability course, but recognizing which tool a five-line word problem is asking for, under a hard clock, is a different skill from having once solved the specific problem in a textbook.
Firms favor this style of screen over a resume-only cut because it is difficult to fake - a candidate who has crammed the specific numbers of a leaked bank still has to reapply the same reasoning to new ones in the room, and the deductive-logic and knights-and-knaves items in the mix cannot be gamed by rote memorization at all. It is a cheap, scalable way to find candidates who reason well under uncertainty before spending any interviewer's time confirming it in person.
How it works
The sitting is a single section called Problem-Solving Assessment: 17 items in 20 minutes, which is the internship-length version of the format - full-time candidates reportedly get the same bank at 60 minutes. There is no penalty for a wrong answer and no going back once an item is left behind.
The item bank covers 18 archetypes, and a 17-item sitting draws one of each with a single archetype held back and shuffled into a fresh order each attempt: gambler's ruin, expected steps of a random walk to a boundary, lattice-path combinatorics, a two-coin Bayes update, single-variable product optimization, an alternating add-then-double sequence, a knights-and-knaves logic puzzle, optimal-stopping die re-roll expected value, a poker-pair probability, a double-or-nothing betting chain, the geometric distribution's expected trials to a first success, a binomial "not enough muffins" probability, expected fixed points of a random permutation, the expected maximum of two uniform arrival times, converting betting odds to expected value, counting binary strings that avoid three consecutive matches, choosing a committee from a group, and a conditional dice probability.
Every item is either a multiple-choice question with four options or a direct numeric-entry question, presented one at a time. You answer or skip, and the next item loads immediately - there is no palette to jump around within the section, since going back is not permitted.
How scoring works
The section's penalty is 0, so raw score is simply the count of correct answers - a wrong answer costs nothing beyond not being counted as correct, identical to leaving the item blank in terms of score.
Since there is only one section, the sitting's overall scaled score equals that section's raw score, out of a maxScaled of 17. The review screen reports this as a percentage: round(scaled / 17 x 100).
The section's allowBack is false, so once you leave an item you cannot return to it within the run. Flagging an item during the run (the F key or the flag button) has no effect on navigation - it only marks that item to appear on the final review screen, with your answer, the correct answer, and the worked explanation, once the sitting ends.
Problem-Solving Assessment
Because the whole test is one section drawn from a fixed archetype list, the highest-leverage preparation is pattern recognition on the setup. A "stop when you reach X or 0" wording is gambler's ruin, solvable by the martingale shortcut (probability of success equals starting stake over the target, independent of the bet size). A "draw a stone, replace it, draw again" or "flip a possibly biased coin" setup is Bayes - write the numerator and denominator of the update explicitly. Any "choose k of n" phrasing is combinatorics, and the answer almost never needs a formula more advanced than n-choose-k.
The logic puzzles (knights-and-knaves, syllogism-style deduction) reward writing out the two or three possible worlds explicitly and checking each statement against each world - a wrong guess here costs the same as anywhere else, so there's no reason to rush past a puzzle that brute enumeration solves in under a minute.
Since there is no penalty for a wrong answer, an educated partial guess - eliminating one or two clearly wrong choices on a multiple-choice item, or estimating the order of magnitude on a numeric item - is always worth submitting over a blank. The only real cost in this section is time spent on one item that could have answered two others.
Pacing across the whole sitting
Twenty minutes for 17 items is a little over a minute a question, but the archetypes vary sharply in how long they take to set up correctly - a fixed-points or coupon-collector question is one line of linearity of expectation, while a lattice-path or extended-Bayes question takes longer just to write down correctly. Budget time by how quickly you can identify which archetype you're looking at.
Because no item can be revisited, resist re-deriving an answer you've already submitted from scratch if you have time left over - use any spare minutes on items you skipped or flagged as uncertain instead, since those are the only ones still worth marginal time.
A worked example
Setup: you start with $15, betting $5 at a time on a fair coin flip, and stop when you reach $35 or $0. The bet size is a distractor - it will not appear in the final answer.
Because the coin is fair, your wealth over time is a martingale: its expected value never changes from one flip to the next, so it must still equal the starting wealth once the game stops.
The game can only stop at $35 or $0, so the expectation on the left is just the probability of reaching $35, p, times 35.
Solve for p. Enter 15/35 - the choice format is start over target, not reduced.
Sanity check: the bet size of $5 never entered the calculation, as the martingale argument predicts - only the starting stake and the target matter. The wrong choices in this section are built to catch anyone who puts the bet size back in.
Common mistakes
• Treating the coin-flip bet size as part of the answer. The stopped-wealth martingale argument depends only on the starting amount and the target, not on how much is wagered per round - a choice that reintroduces the bet size is a distractor.
• Defaulting to p = 1/2 on a gambler's-ruin question regardless of the starting amount. That is only correct when the starting stake is exactly half the target; in general the probability of success is start / target.
• On Bayes items, forgetting to weight by the prior probability of each hypothesis before combining with the observed evidence - a bag with two equally likely coins needs both coins' likelihoods in the denominator, not just the biased one's.
• Using permutations where the problem asks for combinations, or vice versa - a committee or group-selection question that doesn't care about order needs n-choose-k, not a factorial count of orderings.
• Spending equal time on every item regardless of how quickly its archetype is recognizable. A fixed-points or geometric-distribution question is a single formula; a lattice-path or extended-Bayes question takes longer to set up correctly, and the score treats them identically.
Why interviews test this
SIG is widely reported to run essentially this exact screen - a recycled probability and logic bank, sent before a resume is reviewed, where only the numbers change between candidates. The design is meant to be scalable and hard to game: a candidate who has seen the same archetype before still has to correctly redo the reasoning with new numbers in real time, so rote memorization of a specific past answer does not transfer.
The format filters for fluency with a small toolkit - martingale arguments, Bayes updating, combinatorial counting and deductive logic - applied quickly and correctly to an unfamiliar instance of a familiar shape of problem. That is a closer proxy for how a trader or quant researcher reasons under uncertainty on the job than a GPA or a resume line, which is why it runs ahead of any human review.
Play SIG Problem-Solving Assessment on Outcry
SIG Problem-Solving Assessment is free to play on Outcry, in the Mock Screens section, and every run is timed and scored. Read the guide once, then play SIG Problem-Solving Assessment and check your score against the strategy above. The guide book has a written guide for every other game, and the practice hub and mock assessments put the same skills under a real screen's clock.