What a Quant Actually Does Is Mostly Data Work

The Categories

Quant is not a single job. It covers several roles with different daily jobs and different skills

The distinction between researcher and developer is often blurred and companies increasingly expect both to be both. A researcher who can't implement his or her own ideas is slower than one who can

Where the Time Goes

The consistent report from the people doing the work is that a vast majority of the work is data

Acquire it clean it align timestamps handle corporate actions deal with reviews identify errors and build the infrastructure that makes it searchable. Only after all this does anything resembling research begin

The reason it is so important is that data errors produce false results that look exactly like the findings. A survival-biased universe a bad data field or an ill-adjusted stock split will produce a seemingly profitable strategy and nothing in the result distinguishes it from a real one

What the Data Work Actually Consists Of

The phrase data cleansing suggests removing obvious errors which is the smallest part. The substance is a series of specific problems each with its own way of silently ruining a result

Corporate actions. A stock that splits four for one shows an apparent 75 percent drop overnight in a series of unadjusted prices. Dividends spinoffs rights issues and mergers require history to be recast so that the performance series reflects what a holder actually earned. Getting the adjustment factors right and applying them consistently to price volume and shares outstanding is far from glamorous and is the difference between a signal and an artifact

Identifier mapping. Companies change names and tickers and tickers are reassigned to completely different companies after a delisting. Joining two data sets into one ticker without proper mapping via a permanent identifier will silently link one company's fundamentals to another company's prices and the join will be error-free

Timestamps. Each field needs to convey the moment it was met rather than the moment it describes and the two are almost never the same. Different providers register different conventions exchanges are located in different time zones and daylight saving time transitions change the relationship between them twice a year

Reviews. Economic statistics are published and then revised sometimes substantially. A database containing only the current value has erased the number anyone actually saw at the time so any strategy tested against it is operating on figures that didn't exist yet

The Security Master Is the Foundation

Beneath all that lies an unglamorous piece of infrastructure that determines whether the rest is trustworthy

A security master is the internal record of what each instrument is what identifiers it has maintained over time when it started and stopped trading and how it relates to the entity that issued it. Each data set is linked through it meaning that every error it contains is propagated to every piece of research the company produces

Building one properly requires including instruments that no longer exist with the dates and reasons why they ceased to exist because a universe assembled only from what is currently marketed has already had its flaws eliminated. That is the problem of survival that enters through the pipes and not through the research design and is invisible in the research code which seems totally correct

This is why companies treat their data infrastructure as a competitive asset and not a cost. Signals are difficult to maintain and the process that reliably produces a correct view of the story at any given time is long-lived expensive to replicate and improves with every problem it is already designed to solve

The Failure Rate

Most hypotheses don't work. A researcher can test many ideas for each one that survives into production and those that survive often yield less than the research suggests

This is a psychologically difficult work environment. The work is intellectually demanding the failure rate is high and the feedback loop on whether an implemented strategy actually works takes a long time due to the sample size issue

Companies achieve this by measuring processes rather than outcomes over short horizons which is easier said than implemented

The reason it is difficult to implement is that the quality of the process and the quality of the results appear identical over any period short enough to evaluate someone. A researcher who did a disciplined search and found nothing has done the job correctly and has nothing to show for it. A researcher who did a careless search and found something that worked for a year has a number attached to his name. Any organization that judges results over twelve months will systematically reward the latter and the cost of that comes later andIt is attributed to the market rather than to evaluation

What the Mathematics Actually Involves

The expected background is genuine: probability statistics linear algebra optimization and for derived functions stochastic calculus

What is used on a daily basis is more prosaic. Regression hypothesis testing and careful handling of sample size take up much more time than advanced techniques. Machine learning methods are widely used and the most difficult problem is almost always preventing them from adjusting to noise rather than applying it

Programming ability is not optional. Python dominates research and C++ is used when latency matters. Comfort with databases and version control is assumed

The Engineering Standards Are Not Optional Either

The programming requirement is usually described in terms of languages which understates it. What separates a productive research group from a slow one is primarily practice rather than syntax

The investigation has to be reproducible meaning that the same code with the same data returns the same answer months later. That requires versioning both the data and the code since a result that cannot be regenerated cannot be reviewed or debugged when it stops working in production

The path from research to production is where a lot of value is lost. A signal calculated one way on a laptop and another way on the live system will be different and the difference will be discovered as unexplained poor performance rather than a bug. Companies invest heavily to have research and production share the same code path for exactly this reason

Testing is more important here than in most software because the failure mode is silent. A defect in an ordinary application produces something visibly broken. A defect in the calculation of a signal produces a plausible number is traded and money is lost without ever generating an error

The Parts of the Week That Are Not Research

Even within a research role the proportion of time spent on original work is less than the title implies and the rest is not administrative filler

Production comes first when something breaks. A data source arrives late or is incorrectly formatted a signal is not calculated before the market opens or a supplier changes a field without warning and whoever owns that part of the pipeline handles it that morning regardless of what they were working on

Monitoring is continuous and not occasional. Live strategies produce daily numbers that must be analyzed and distinguishing ordinary bad days from a strategy that has stopped working is a decision made repeatedly without sufficient evidence

Then there's the explaining. In most companies a researcher has to convince a portfolio manager or a committee that a result is real which means presenting it to people whose job it is to find the flaw in it. That conversation is where a lot of research dies and properly preparing for it takes as much time as part of the analysis

Vendor management code review adding new data sets and rebuilding something that worked but can't be maintained fill out most of what's left. The mental image of a quant who thinks about math all day describes a fraction of the week and the people who thrive are usually the ones who find the work around them more interesting than tolerable

Getting In

The entry routes are narrower than in most finances and the filters are specific which is worth knowing early and not late

Backgrounds fall into mathematics physics statistics computer science and engineering. Advanced degrees are common and more necessary for derivatives pricing roles where the mathematics is genuinely deep than for systematic research roles where demonstrated ability with data and code outweighs credentials

The interview process is unusually concrete. Expect probability and statistics questions asked without warning mental arithmetic questions about how you would test a stated hypothesis and a coding assessment that is about clarity and correctness rather than exotic algorithms. Many companies also offer a take-home data set and what they read is whether you noticed problems in it

The trait that is analyzed at all times is the one described above. A candidate who presents a result with confidence is less interesting than another who presents the same result along with the reasons why it could be wrong and the checks he made to try to break it

For someone who is still studying the most useful preparation is not more finance. It is becoming truly fluent with data that is able to acquire a messy real data set clean it join it to another correctly and produce an honest analysis of it including the part where the interesting result turns out to be a data error. That skill transfers directly and is what the job mainly consists of

Where the Roles Sit

Hedge funds and proprietary trading firms hire researchers to find strategies and compensation is often tied to the performance of what they build

Banks hire desktop quants to set prices and hedge products and risk quants to model exposure and meet regulatory requirements. The work is more structured and the compensation less variable

Asset managers hire quants for portfolio construction factor research and risk systems falling somewhere in the middle in both style and compensation

The choice between them is mainly a choice about the form of feedback and the variation of remuneration. A researcher at a fund eventually finds out if they were right on a number attached to their name and receives corresponding compensation both ways. A desk quant at a bank is judged based on whether the price library is correct and whether the desk can cover which is a clearer standard with a narrower range of results

What Distinguishes People Who Are Good At It

Skepticism about their own results which is the most important trait given the ease with which false discoveries appear

Understand the markets and not just the mathematics since a statistically valid relationship without any economic mechanism is usually noise

Engineering discipline because research that cannot be reliably reproduced or implemented is not useful

And tolerance for the failure rate which is what most often causes people to leave

It's worth being specific about skepticism since it's a habit rather than an attitude. It seems like checking whether a surprising result survives missing the best month whether it holds up in the half of the sample you didn't look at first and whether the data supporting it is flawed before you show it to anyone. People who are good at this spend most of their effort trying to destroy their own findings and the ones who survive that treatment are the only ones worth defending

The Bottom Line

Quantitative roles are split between research development desk pricing and risk and they all involve a lot more data work than the title suggests. Hypothesis failure rates are high data errors reliably produce results that look like discoveries and the most valuable trait is skepticism about your own findings. The math is real and the daily bottleneck is often the data. Anyone deciding whether to pursue this should know that the lasting advantage in the field lies in process anddiscipline that surrounds it rather than in the cunning of any individual idea