DRW-Style Quantitative Challenge

How it works

The sitting is a single section, called Quantitative challenge: 45 minutes for 6 questions, none of them quick - about seven and a half minutes each if the time is split evenly, though a real sitting rarely divides that cleanly. The item pool mixes a binomial mean/variance question, a Bayes' rule signal-detection question, a two-step Markov chain, a regression-slope calculation from correlation and standard deviations, a game-theory optimal-stopping puzzle, and a coin-weighing logic puzzle - one of each is drawn and shuffled into a fresh six-question set on every attempt.

Every question is answered by selecting one of four multiple-choice options. There is no back button within the section: once you commit to an answer, that question is closed and the next one loads. That single rule reshapes the whole sitting - unlike an assessment with review passes, there is no coming back at the end to mop up an item you skipped under pressure.

How scoring works

No negative marking: a wrong answer costs nothing beyond the mark you would have earned for a right one, so the section's penalty is 0. Guessing on a question you cannot finish in time is always better than leaving it - there is no downside.

None of the six generators set a marks value on their items, so every question is worth the same one raw mark; the section result reports how many of the six you answered correctly, wrongly, or left unanswered.

allowBack is false for the section, so once an answer is locked in it is final - there is no return pass to revisit an earlier question with time freed up later in the sitting.

Quantitative challenge

Read all six prompts' topics before committing real time to any one of them if the interface lets you see the question number and skip forward is not available - since there is no going back here, decide your order of attack up front rather than discovering three questions in that an earlier one would have been faster.

For the binomial mean/variance question, write E = np and Var = np(1-p) before touching numbers - the standard trap is answering with the standard deviation (the square root of variance) in the Var slot, which is one of the offered distractors.

For the Bayes signal question, do not answer with the raw hit rate. The signal's accuracy on genuine days is only half the picture - the false-alarm rate on the many more numerous quiet days pulls the true answer down, sometimes a long way. Write out numerator and denominator explicitly rather than trying to eyeball it.

For the Markov two-step question, enumerate both paths back to the starting state (direct, and via the other state) before summing - forgetting the indirect path is the most common way to land on a distractor.

For the regression-slope question, remember correlation is unitless and the slope re-attaches the scale: beta = r times (sigma_y / sigma_x), not r alone.

For the card-counting game-theory question and the coin-weighing logic question, look for the closed-form shortcut (a random walk's optionality value, or log base 3 of the coin count) rather than trying to simulate every case by hand under the clock.

Pacing across the whole sitting

Forty-five minutes for six questions gives you about seven and a half minutes each, but treat that as a budget, not a target - some of these (the Markov chain, the regression slope) are one or two lines of arithmetic once you recall the right identity, while others (the card game, the coin-weighing puzzle) take longer to reason through than to compute.

Because there is no negative marking and no way to return to a locked question, the losing strategy is spending twelve minutes perfecting one answer and leaving two questions unanswered at the end. An imperfect answer on every question beats a perfect answer on four and blanks on two.

If you feel a question sliding past its budget, commit to your best current answer and move on - you can only lose the marks for that one question by staying, but you can lose marks on a question you never reach by not moving on.

A worked example

1. Read off the three given probabilities from the prompt: the base rate of a genuine move, and the two conditional firing rates.

   $$P(\text{genuine}) = 0.20, \qquad P(\text{fire}\mid\text{genuine}) = 0.80, \qquad P(\text{fire}\mid\text{quiet}) = 0.20$$

2. Bayes' rule gives the probability of a genuine move given the signal fired as the true-positive term divided by the total probability of firing, summed over both underlying states.

   $$P(\text{genuine}\mid\text{fire}) = \dfrac{P(\text{genuine})\,P(\text{fire}\mid\text{genuine})}{P(\text{genuine})\,P(\text{fire}\mid\text{genuine}) + P(\text{quiet})\,P(\text{fire}\mid\text{quiet})}$$

3. Compute the numerator: the prior times the hit rate.

   $$0.20 \times 0.80 = 0.16$$

4. Compute the denominator: add the false-alarm contribution from the quiet days, using P(quiet) = 1 - 0.20 = 0.80.

   $$0.20 \times 0.80 + 0.80 \times 0.20 = 0.16 + 0.16 = 0.32$$

5. Divide numerator by denominator for the final answer.

   $$P(\text{genuine}\mid\text{fire}) = \dfrac{0.16}{0.32} = 0.50$$

6. Sanity check the result: even with an 80% hit rate, the answer lands at only 50%, because quiet days vastly outnumber genuine-move days and each quiet day still fires 20% of the time - the false alarms from that much larger pool swamp the true positives.

Common mistakes

• Reporting the standard deviation in place of the variance on the binomial question - Var = np(1-p), not its square root.

• Answering a Bayes question with the raw hit rate instead of running the full calculation - the false-alarm rate on the far more numerous negative days pulls the true answer down.

• On the two-step Markov question, only tracking the direct path back to the starting state and forgetting the indirect path through the other state.

• Spending too long perfecting one answer - since there is no going back and no negative marking, an unanswered question is strictly worse than a fast guess.

• Redoing a lengthy calculation from scratch after a small slip instead of checking the last two lines of work - with roughly seven minutes per question there usually is not time for a full restart.

Why interviews test this

This format screens for breadth and setup speed rather than depth in any one topic, because live trading questions do not announce which branch of math they belong to before you start working them. A candidate who can only solve probability questions, or who needs to be told a problem is a Markov chain before recognizing the structure, will struggle in a seat where the next P&L-relevant question could be about any of these topics with no label attached.

The no-negative-marking, no-going-back combination also tests something interviewers care about beyond raw correctness: whether a candidate manages their own time and commits to a decision under uncertainty rather than second-guessing indefinitely. That is a direct proxy for a trading desk skill - a position has to be sized and entered with incomplete information, not perfected in isolation with unlimited time.