What a quant interview is actually testing

I want to describe a specific moment, because everything else in this piece hangs off it.

You are forty minutes into a final round. The interviewer asks something you have not seen. Not a variant of something you have seen — genuinely new. There is a silence of perhaps four seconds while you look for a handle and do not find one.

What you do in the next ninety seconds is the interview. Everything before it was qualification.

I have watched very strong mathematicians lose offers in that ninety seconds, and I have watched weaker ones win offers in it. The candidates who lose go quiet, or they start talking to fill the silence without actually thinking, or — most commonly — they reach for the nearest technique they know well and start grinding it even though some part of them already suspects it is wrong. They are trying to look competent. That instinct is the whole problem, and it is trained into people by every examination they have sat up to that point.

What the desk is actually screening for

Here is the thing that took me a while to understand, and which I only really understood after sitting on the other side of the table.

A trading desk is not hiring you to know things. It is hiring you to be reliable under uncertainty with money at stake. Nearly everything that happens in a research or trading seat is a version of that ninety seconds: incomplete information, a decision needed, no answer key, and a colleague who will ask you why. So that is what they test. Not because they are being clever about it, but because it is the closest thing to the job that fits into an hour.

Which means the interview is looking for a few specific behaviours:

Can you decompose something you have not seen? Faced with a novel problem, do you reach for structure — what is the simplest case, what happens at the boundary, what would I guess and why — or do you reach for a memorised method? Interviewers can tell the difference immediately, and the second one reads as brittle.

Can you be wrong out loud, cheerfully? This is the one nobody prepares for. You will be interrupted. You will be told your approach does not work. What the interviewer wants to see is whether you can say "ah — yes, that breaks because the variance isn't finite, let me think about it differently" and then actually think about it differently, without your confidence collapsing. What they are checking is whether they can tell you your model is wrong at 9am on a Tuesday and have you fix it rather than defend it.

Do you have calibrated uncertainty? When you say you are fairly sure, are you? People who round their confidence to zero or one hundred are expensive to work with. Anyone who has priced anything develops a feel for the middle.

Can you actually compute? Which brings me to the part everyone resents.

Mental arithmetic, and why it is not hazing

Every client I have worked with has resented the mental arithmetic drills for about the first month. I understand why. It feels like a trick, an artificial hurdle that has nothing to do with real work in an era of computers.

Two things are going on, and only one of them is what people assume.

The obvious one: market-making games and pricing questions are unplayable if arithmetic consumes your attention. If you are spending working memory on 17 × 23, you are not spending it on the actual question, which is what your spread should be given that the person quoting into you might know something.

The less obvious one, which matters more: fluent arithmetic gives you a sanity check that runs in the background. People who compute easily notice when an answer is off by an order of magnitude, because the wrongness is felt rather than derived. That instinct is worth a great deal on a desk, and interviewers are testing for it whether or not they would describe it that way.

It is also, and I say this without much sympathy, the single most trainable thing on this list. Twenty minutes a day for six weeks and you will not recognise yourself. Most people do not do it because it is boring, which is precisely why it remains a filter.

The stochastic calculus you need, and the large amount you do not

There is a widespread and expensive misconception that quant interviews are a viva on stochastic analysis. They are not, and preparing as though they are wastes months.

What actually comes up, repeatedly:

Itô's lemma, applied without hesitation — and, more to the point, understood well enough that you can say why the second-order term survives when it would vanish in ordinary calculus. If you can explain that in one sentence to someone who has not seen it, you understand it. If you can only apply the formula, you do not, and a good interviewer will find that out in two questions.

Change of measure, and specifically why anyone would want one. Girsanov as a piece of machinery is less interesting to an interviewer than whether you understand that pricing is an expectation under a measure chosen to make a particular process a martingale, and that this is a convenience rather than a claim about the world.

Martingale arguments, particularly optional stopping, because a startling number of interview problems are secretly a stopped martingale wearing a costume. Gambler's ruin, random walks on graphs, expected time to absorption — once you see the shape you stop solving these from scratch.

What almost never comes up: the fine print of stochastic integration, Lévy processes, rough paths, the measure-theoretic underpinnings in any depth. I say this as someone who genuinely enjoys that material and has done research in it. It is beautiful and it is largely irrelevant to getting hired. If you are reading Revuz and Yor to prepare for an interview, you are doing something you enjoy and calling it work. I have some sympathy, having done it myself.

The failure mode I see most

Strong mathematicians fail these interviews at a rate that surprises people, and it is almost always the same thing.

Their entire training rewards presenting finished work. A problem sheet, a paper, an exam script — you think in private, and you show the polished result. The whole apparatus of mathematical education is built to make you look like you knew it all along.

An interview inverts this completely. The interviewer cannot see whether you are good at mathematics from a correct answer, because plenty of people memorise correct answers. They can only see it by watching you work. So the candidate who thinks silently for two minutes and produces a correct answer has, perversely, given the interviewer almost nothing to go on, while the candidate who talks through three approaches, discards two with reasons, and gets most of the way there has given them everything.

This is trainable, but it is not trainable by reading. It takes someone sitting opposite you interrupting, pushing, and telling you that your approach fails — repeatedly, until being wrong in front of another person stops producing an adrenaline response. In my experience it takes about four sessions before someone stops flinching, and they are usually the four most useful sessions of the engagement.

What I would actually do with six weeks

If you have a live process and six weeks, this is roughly the shape I would want, and it is deliberately unbalanced away from what people expect.

Twenty minutes of arithmetic every single day, without exception, starting today. Not because it is the most important thing but because it is the thing that only works with time, so it has to start first.

Probability worked properly rather than broadly — conditional expectation until it is genuinely comfortable, then the standard family of problems until you recognise the shapes. Most interview probability is a small number of ideas in costume.

Itô and change of measure to real fluency, and then stop. Resist the urge to go deeper. The marginal hour is better spent elsewhere and you will feel the opposite.

Market-making games, played against an actual person who is trying to pick you off. These do not work alone; the entire content is adverse selection, and you cannot adversely select yourself.

And mock interviews run harder than the real thing, recorded, with someone watching how you behave rather than whether you are right. If the real interview feels slow by comparison, the preparation worked.

One last thing

Two of the five firms on your list are probably a poor fit for how you think, and you will not discover which two by reading their websites. Market-making, systematic research and low-latency development select for genuinely different people, and the failure to distinguish between them wastes an enormous amount of candidate effort every year.

Working out which is which, honestly, is worth more than another month of brainteasers.

Quant career programmes Start an enquiry